, drillil)E',tlilli C X~~~~~~- 11 z X1 f~~ ,1 E 0 is the one-sided error compo- nent representing technical inefficiency.2 For the purpose of this analysis, we assume that the distribution of ui is derived from a N(O, y2 ) distribution trun- cated above at 0. Having estimated the model, we can obtain the residuals given IDy (3) £ = Yi -f (Xi, ) which can be regarded as estimates of the error terms, es. However, as shown by Jondrow and others (1982), e contains only imperfect information about u and makes it possible to obtain the mean technical efficiency over all observations. Jondrow and others (1982) show that a firm-specific measure of technical ineffi- ciency, that is, a point estimate of ui, can be obtained by calculating the mean of the conditional distribution of ui given Ei. That is, defining 62 = _72 + 02, J = cs2ue/o2, and 02 = 0202 /02 the conditional distribution of u given £ is that of a N - (w, 62) variable, truncated at 0. This distribution can be used to make inferences about u. The mean of the conditional distribution of u given e is shown by (4) E(ulc)= +6 [ f (-g/a.)] where f and F represent the standard normal density and cumulative density functions, respectively, and -I6j, = e'J where X=o 6a,. Equation 4 can thus be rewritten as (5) luE) = 6- f(X) (_6_)Y In equations 4 and 5, g,u and a. are unknown and are estimated by I. and A 0, > 0,, respectively. Equation 5 yields the point estimate of ui, which is then used to obtain firm-specific technical efficiency (TEJ) as given by (6) TEi = exp(-iLi). 2. u. measures technical inefficiency as a shortfall of output (yi) from its maximal possible value given by the stochastic frontier. Aw and Batra 67 IV. THE DATA AND THE EMPIRICAL MODEL SPECIFICATION In 1986 the Department of Statistics in Taiwan gathered information from more than 123,000 establishments for the 1986 census of manufacturing. In this article, we focus on 10 of the 20 two-digit Standard Industrial Classification (SIC) industries in the data base because data on technology variables such as R&D, training, and license expenditures, as well as foreign capital, are only avail- able for these industries. These 10 industries contribute 74 and 67 percent of total employment and total output, respectively, in the manufacturing sector and constitute 72 percent of total manufacturing establishments. Firm-level information includes data on the age or birth year of the firm; the rate of utilization of capacity; expenditures on raw materials, energy, and elec- tricity; the total volume and value of production; the value of net assets; and the workforce composition. For the first time, the 1986 census included informa- tion on the firm's investments in formal R&D and on-the-job training, its market orientation (with sales broken down by domestic and export sales), the value of its foreign capital, as well as its expenditures on foreign know-how. These fea- tures of the data set enable us to quantify the form of technology investments undertaken by firms formally and informally through their export sales. The production function we adopt is the translogarithmic production func- tion. The basic specification for the equation can be written as (7) In Q = ao + Y , ol n xi + Y2 Y, o ij In xi In x} + vi + ui . The dependent variable, value added, represented by Q, is measured in thou- sands of new Taiwan dollars and is calculated as the difference between the firm's value of output and the sum of its expenses on raw materials, energy, and electricity. The explanatory variables represented by x include labor and capital. Labor is classified into two groups: the number of nonproduction and produc- tion workers. Because nonproduction workers are generally more skilled than production workers, for simplicity, we refer to the former group as skilled labor and the latter as unskilled labor. The breakdown of total labor into the two groups enables us to control for the quality of labor in the production frontier estimates. This is common practice in the literature. Berman, Bound, and Griliches (1993), based on data assembled on the educational attainment by broad occu- pational groups in the United States for 1973, 1979, and 1987, argue that such occupational distinctions provide a reasonable separation between more skilled and less skilled labor. We measure capital as the value of the firm's net assets. Separate production frontiers are estimated for each of the two-digit indus- tries. For each of the 10 included industries, we could further subdivide firms into finer industry categories based on whether they perform R&D and training, have foreign capital or know-how, or export. The problem with all but the last subdivision is a practical one. The percentage of firms in a given two-digit indus- try (perhaps with the sole exception of the electrical and electronics industry) 68 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 with positive values of R&D or training expenditures or foreign capital is ex- tremely small (see table 1). Therefore, any finer disaggregation, especially by industry or by firms broken down by R&D or foreign capital status, is likely to result in some subdivisions with very small sample sizes or no observations. By contrast, the number of firms in any industry that export is substantial, so that sample size is not a problem in the breakdown of firms by export status. More important than the data issue, increasing evidence in the literature indi- cates that exporters are generally larger in size, are more capital-intensive, have better access to factor inputs and new technology, and are more productive than nonexporters. These results are documented by Aw and Hwang (1995) for Tai- wan and by Clerides, Lach, and Tybout (1996) for Colombia, Mexico, and Morocco. It follows that firms that sell in the export market are unlikely to operate with the same technological conditions as those that are oriented to- ward the domestic market. Thus, in the empirical model we allow the param- eters of the production function to differ for these two subgroups. The use of a single cross section has two limitations. The first is that our analysis does not permit us to conclude anything about the direction of causality between exports and efficiency or between investments in technology and effi- ciency. For instance, causality between exports and efficiency could run in either direction. A panel data set is necessary to determine the direction of causality. Our purpose is to use cross-sectional evidence to establish relationships in the data linking exports, formal technology investments, and technical efficiency. The second limitation, common in all estimations of production functions using cross-sectional data, is that the firm's choice of input may be correlated with inefficiency levels, measured by the error term, u. Again, access to panel data or to good instruments would solve the potential endogeneity problem. To minimize this problem in the absence of panel data, we take into account ob- served firm-specific characteristics that are likely to be correlated with efficiency (or inefficiency) and include them as regressors in our estimation of equation 7. These include the firm's age and capacity utilization rate. Age serves as a mea- sure of managerial experience. The absence of capacity utilization rate in most data bases is believed to be the main cause of measurement error in the capital input variables included in the production function (Mairesse 1990). Thus in- cluding the utilization rate variable may minimize the measurement error. To investigate the correlation between a firm's own investments in R&D and training as well as more direct access to foreign technology through li- censes and foreign capital, we also include in xi dummies for whether the firm has R&D or training expenditures and whether it has foreign investments or purchases of know-how. An important rationale for treating the variables, particularly those involving investments in technological capability or access to foreign technology, as binary variables is that our observations are flow measures of these activities. In the absence of stock measures, we use binary variables to proxy for the stock of knowledge accumulated through the firm's direct investments in or access to technology. For instance, it is very likely that Aw and Batra 69 once a firm begins to invest in R&D, it will continue to do so. Thus it is reason- able to assume that firms with positive expenditures on R&D have larger stocks of knowledge from past investments in R&D than firms with no current expen- ditures on R&D. The variables comprising xi in the final empirical model estimated sepa- rately for exporters and nonexporters are skilled labor, unskilled labor, capital, age, utilization rate, and indicator variables for three types of firms: those with positive investments in R&D and training (RT), those with foreign capi- tal or know-how expenditures (FK), and an interaction term between RT and FK. The coefficient of the interaction term reveals the additional techni- cal efficiency associated with investing in both R&D and training and FK, correlations that are above and beyond those arising from undertaking each activity separately. V. EMPIRICAL RESULTS We use the maximum-likelihood technique, using the Davidon-Fletcher-Powell algorithm to estimate the stochastic production frontier model specified in equa- tion 7. In order to allow production coefficients to differ between exporters and nonexporters, separate frontiers are estimated for the two groups. Tables 2 and 3 present the results of the estimates of the translogarithmic frontier regressions for exporting and nonexporting firms in each of the 10 two-digit manufacturing industries under study. Among both exporters and nonexporters, the capacity utilization variable is positive, implying that efficiency increases as full capacity is approached. This relationship is significantly different from 0 in half the industries among ex- porters and nine out of 10 industries among nonexporters. The magnitude of the efficiency effect of utilization is small, however, ranging from 0.2 to 0.4 percent. The age variable is not statistically significant, particularly among exporters. However, older firms operating in the domestic market are signifi- cantly more efficient in the electrical and electronics, plastics, paper and pub- lishing, and fabricated metals industries. The positive effect of age on effi- ciency is also documented for multinationals operating in the electronics industry in Taiwan (Chen and Tang 1987) and the weaving industry in Indone- sia (Pitt and Lee 1981). A striking pattern emerges from the parameter estimates for R&D and train- ing (RT). These estimates are positive in all 10 industries, and this result is inde- pendent of the export status of firms. They are statistically significant in nine of the 10 industries among nonexporters. Nonexporting firms that invest in R&D and training are between 13.5 percent (in textiles) and 31.2 percent (in iron and steel) more efficient than their counterparts that do not make such investments. For exporters, the RT variable is statistically significant in five of the 10 indus- tries. Firms that simultaneously export and invest in R&D and training are about 10-17 percent more efficient in the textile, clothing, iron and steel, machinery, Table 2. Stochastic Production Frontier Estimates for Exporters in Taiwan (China), 1986 Electrical Paper and Fabricated Iron and and Transport Variablea Textiles Clothing publishing Plastics metals Chemicals steel Machinery electronics equipment Constant 0.997 3.066- 2.642 4.695. 5.090* 5.392- 4.993w 3.649- 2.733- 5.132+* (0.831) (1.018) (1.722) (0.722) (0.836) (2.558) (1.517) (1.023) (0.725) (1.419) Log (skilled labor) -0.226 -0.043 -0.792 0.787- 0.180 -1.081 -0.229 0.282 -0.197 0.445 (0.239) (0.223) (0.242) (0.184) (0.207) (0.827) (0.399) (0.273) (0.192) (0.446) Log (unskilled labor) 0.365- 1.123"'* 0.242 0.366" 0.415- 1.133- 0.797*** 0.655- 0.4574'* 0.046 (0.187) (0.195) (0.349) (0.152) (0.183) (0.492) (0.292) (0.250) (0.138) (0.327) Log (capital) 1.138 * 0.283 0.906- 0.113 0.172 0.217 0.108 0.287 0.712' 0.206 (0.212) (0.248) (0.472) (0.182) (0.217) (0.609) (0.357) (0.281) (0.184) (0.377) Log (skilled labor2) -0.019 0.0007 -0.019 0.058*** 0.045** -0.025 -0.401 0.064- 0.014 0.084' (0.023) (0.019) (0.052) (0.017) (0.022) (0.088) (0.044) (0.027) (0.021) (0.045) Log (unskilled labor2) 0.055- 0.015 0.014 0.036- 0.024 0.074* 0.060- 0.012 0.004 0.033 (0.018) (0.016) (0.035) (0.010) (0.016) (0.042) (0.026) (0.023) (0.011) (0.029) Log (capital2) -0.043-' 0.013 -0.052 0.019 0.006 -0.007 0.015 0.008 -0.024 -0.002 (0.015) (0.017) (0.034) (0.012) (0.015) (0.037) (0.021) (0.019) (0.012) (0.026) Log (skilled labor x -0.068**' -0.066- -0.196*** -0.082- -0.124- -0.214'* -0.078' -0.058 -0.074"' -0.163... unskilled labor) (0.024) (0.021) (0.060) (0.017) (0.027) (0.097) (0.044) (0.038) (0.024) (0.056) Log (skilled labor x 0.087**' 0.055- 0.176- -0.056" 0.030 0.208- 0.078 -0.018 0.062+* -0.002 capital) (0.032) (0.029) (0.069) (0.023) (0.028) (0.102) (0.049) (0.037) (0.026) (0.058) Log (unskilled labor x -0.024 -0.069- 0.048 0.004 0.008 -0.064 -O.OS2 -0.014 0.006 0.054 capital) (0.027) (0.027) (0.053) (0.021) (0.026) (0.059) (0.038) (0.359) (0.019) (0.044) Utilization rate 0.001 0.003- 0.003 0.001- 0.004.' 0.001 0.002 0.001 0.002- 0.004 * (0.001) (0.001) (0.001) (0.0006) (0.001) (0.001) (0.002) (0.001) (0.0006) (0.001) Age -0.001 -0.0005 0.003 0.004* -0.003 -0.003 -0.001 -0.003 0.006 -0.005 (0.003) (0.003) (0.004) (0.002) (0.003) (0.005) (0.005) (0.002) (0.007) (0.003) R&D and training (RT) 0.153*** 0.172-* 0.091 0.051 0.053 0.127 0.139' 0.171*** 0.044 0.095' (0.045) (0.053) (0.098) (0.035) (0.043) (0.103) (0.073) (0.046) (0.033) (0.055) Foreign capital (FK)b 0.123 -0.191 -0.015 -0.023 0.073 0.155 0.252 0.159 -0.116 0.214 (0.125) (0.183) (0.275) (0.096) (0.096) (0.199) (0.090) (0.148) (0.071) (0.179) RTxFK -0.123 0.264 0.125 0.074 -0.001 -0.066 -0.304 -0.088 0.152' -0.103 (0.162) (0.225) (0.324) (0.128) (0.138) (0.22J) (0.236) (0.176) (0.084) (0.202) Note: Values are estimated using the maximum-likelihood technique for the model specified in equation 7 in the text. Standard errors are in parentheses. * Significant at 10 percent. ** Significant at 5 percent. * Significant at 1 percent. a. x indicates interaction between two variables. b. Positive foreign investment or purchases of foreign technology. Source: Authors' calculations. Table 3. Stochastic Production Frontier Estimates for Nonexporters in Taiwan (China), 1986 Electrical Paper and Fabricated Iron and and Transport Variablea Textiles Clothing publishing Plastics metals Chemicals steel Machinery electronics equipment Constant 4.364.. 3.659*** 6.026"' 5.880- 4.205-' 3.121 - 4.632-' 3.698"' 5.231- 2.731" (0.486) (0.922) (0.609) (0.527) (0.392) (1.702) (0.763) (0.615) (0.643) (0.727) Log (skilled labor) 0.051 -0.348 0.226 0.404-' 0.101 -0.371 -0.169 0.392"' 0.480- -0.016 (0.129) (0.208) (0.144) (0.129) (0.099) (0.404) (0.228) (0.148) (0.157) (0.206) Log (unskilled labor) 0.422"- 0.691"*' 0.851... 0.456... 0,445-' 0.898-' 0.819-' 0.0009 0.264- 0.014 (0.103) (0.169) (0.120) (0.097) (0.083) (0.354) (0.155) (0.126) (0.115) (0.151) Log (capital) 0.296... 0.451s -0.251 -0.115 0.328-' 0.451 0.158 0.507-' 0.031 0.770-' (0.127) (0.244) (0.165) (0.140) (0.108) (0.427) (0.192) (0.166) (0.170) (0.197) Log (skilled labor2) 0.009 0.003 0.032'** 0.055-' 0.035-' -0.010 0.012 -0.064- 0.058**' -0.005 (0.017) (0.023) (0.015) (0.016) (0.013) (0.045) (0.030) (0.019) (0.018) (0.026) Log (unskilled labor2) 0.031-' 0.017 0.065"** 0.026"' 0.038-' 0.047 0.037-' 0.003 0.020** 0.009 (0.010) (0.014) (0.011) (0.008) (0.008) (0.038) (0.016) (0.011) (0.009) (0.017) Log (capital2) -0.0006 -0.016 0.041**' 0.026"' -0.003 -0.0004 0.011 -0.019 0.017 -0.039"*' (0.009) (0.017) (0.011) (0.010) (0.008) (0.028) (0.012) (0.012) (0.012) (0.014) Log (skilled labor x -0.075S'* -0.095-' -0.089*** -0.083- -0.110"'* -0.098-' -0.125-' -0.090- -0.111-' -0.095-' unskilled labor) (0.018) (0.025) (0.016) (0.016) (0.014) (0.052) (0.029) (0.020) (0.018) (0.029) Log (skilled labor x 0.040-' 0.103-' 0.015 -0.015 0.032- 0.092-' 0.068-' -0.013 -0.016 0.061- capital) (0.017) (0.028) (0.020) (0.018) (0.014) (0.051) (0.029) (0.020) (0.022) (0.029) Log (unskilled labor x -0.0001 -0.020 -0.063-' 0.002 0.001 -0.066 -0.037.. 0.067**" 0.027* 0.063**' capital) (0.015) (0.023) (0.017) (0.013) (0.012) (0.047) (0.020) (0.018) (0.016) (0.022) Utilization rate 0.002- 0.004"' 0.002"' 0.002.. 0.002- 0.004- 0.003- 0.003"' 0.004- 0.004"' (0.0006) (0.001) (0.0005) (0.0006) (0.0004) (0.002) (0.0008) (0.0005) (0.0006) (0.0008) Age -0.0003 0.002 0.004'** 0.005-' 0.003-' 0.0005 0.003 -0.003 0.007- -0.0007 (0.002) (0.003) (0.001) (0.002) (0.001) (0.004) (0.003) (0.002) (0.002) (0.002) R&D and training (RT) 0.135- 0.271"' 0.135' 0.073 0.176"' 0.311- 0.312"' 0.180- 0.144- 0.218"' (0.061) (0.112) (0.055) (0.060) (0.042) (0.113) (0.092) (0.049) (0.047) (0.072) Foreign capital (FK)b 0.046 -0.245 0.065 0.036 0.072 0.032 0.465' 0.066 0.014 0.183 (0.219) (0.275) (0.194) (0.267) (0.120) (0.396) (0.260) (0.161) (0.152) (0.228) RTx FK 0.317 -0.268 0.014 -0.058 0.089 0.238 -0.934... 0.020 0.025 0.107 (0.294) (0.234) (0.306) (0.322) (0.178) (0.489) (0.363) (0.241) (0.194) (0.269) Note: Values are estimated using the maximum-likelihood technique for the model specified in equation 7 in the text. Standard errors are in parentheses. Significant at 10 percent. * Significant at 5 percent. " Significant at 1 percent. a. x indicates interaction between two variables. b. Positive foreign investment or purchases of foreign technology. Source: Authors' calculations. Table 4. Mean Efficiency Estimates and Relative Frontier Position of Exporting and Nonexporting Firms in Taiwan (China), 1986 Relative Mean Mean Level of homogeneity in frontier efficiency of efficiency of Mean efficiency levels of firms' Industry positiona exporters nonexporters difference testh Exporters Nonexporters Textiles 1.036 0.668 0.666 0.719 0.536*** 0.453- (0.043) (0.025) Clothing 0.992 0.996 0.700 73.054*"' 0.259*** 0.432-* (0.013) (0.036) Paper and publishing 1.044 0.627 0.688 10.042"' 0.577- 0.406- (0.077) (0.020) Plastics 1.040 0.694 0.690 1.19 0.421- 0.421- (0.029) (0.022) Fabricated metals 1.038 0.675 0.683 0.040 0.493-' 0.407°** (0.035) (0.015) Chemicals 1.014 0.688 0.651 3.266** 0.392 0.543- (0.199) (0.068) Iron and steel 0.962 0.693 0.666 3.058** 0.385- 0.615- (0.060) (0.041) Machinery 1.005 0.997 0.717 91.926- 0.253t** 0.390- (0.012) (0.021) Electrical and electronics 1.022 0.656 0.636 15.130*** 0.53S**' 0.SS7'** (0.031) (0.027) Transport equipment 1.030 0.758 0.677 17.765-' 0.305- 0.475-"* (0.046) (0.039) Note: Standard errors are in parentheses. * Significant at 5 percent. * * Significant at 1 percent. a. The ratio of the value added of exporters to the value added of nonexporters evaluated at the input levels of nonexporters. b. Reports the test statistics of the hypothesis that the mean efficiency level of exporters is not significantly different from the mean efficiency level of non- exporters. c. Measured by 2. Lower values of a2 indicate more homogeneous efficiency levels among firms in a given industry. Source: Authors' calculations. Aw and Batra 73 and transport industries. These estimates fall between the magnitudes of R&D elasticity, estimated at 9 percent for a sample of U.S. firms by Griliches (1987), and the marginal effect of R&D on technical efficiency, estimated at 26 percent by Huang and Liu (1994). A possible explanation for why the RT variable is not statistically significant in the other industries is that the relationship between RT and efficiency, par- ticularly in the highly export-oriented electrical and electronics and plastics in- dustries, may be complicated significantly by other characteristics of the firm. Exporting firms in these industries are likely to be more efficient because of their larger average size, greater skills, or better organizational and management ca- pabilities. Grouping firms by their export status is likely to take these factors into account, blurring the efficiency effects of the firm's own investments in technological capability. Another possibility is that in these industries the ex- port activity represents a good proxy for the degree of technological activity and sharing among firms and that this feature may have a more important effect on technical efficiency relative to the firm's own investments in RT. With the exception of one out of the 10 industries included this study, the coefficients on FK and its interaction with RT are not significantly different from 0 for nonexporting firms. Similarly, in the case of exporters, the FK coeffi- cients are not statistically significant in every industry. The only industry in which the interaction between FK and RT is statistically significant for export- ing firms is electrical and electronics. Exporting firms in this industry that invest simultaneously in both RT and FK are more efficient by 15.2 percent. This re- sult suggests that in industries like electronics, where technology is changing rapidly and likely to be more complex, firms that combine their own resources to adapt foreign knowledge obtained through informal (exports) or formal (for- eign investment) channels are more efficient than those that do not. From the results in tables 2 and 3, we use equation 6 to come up with a mea- sure of firm-specific technical efficiency for all firms in each of the 10 industries. Because we estimate separate production frontiers for exporters and nonexporters, it is important, in comparing the average efficiency of these two groups of firms, to take into account the position of the production frontier of exporting firms relative to their counterparts in the domestic market. To do this, we calculate the additional value added that can be generated by nonexporters if they com- bine the exporters' technology with their own inputs in production. We then compare this value added to that generated by the nonexporters using their own technology and inputs. To accomplish this, we first multiply the coefficient esti- mates of each group of firms separately by the input vector of the nonexporters.3 We then take the ratio of the estimated value added of exporters to that of nonexporters. A ratio greater than unity indicates that, using the same input vector (in this case, that of nonexporters), the estimated value added of using the 3. All inputs in the production frontier (equation 7), including the age, utilization rate, and technology variables, enter into the calculation. More specifically, each coefficient in the production function is multiplied by the mean value of the input corresponding to that coefficient. 74 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I technology of export firms exceeds that of using the technology of nonexporters. This ratio is used as the adjustment factor in comparisons of the average effi- ciency between exporters and nonexporters. The adjustment ratios are reported in column 1 in table 4. With the exception of the clothing and iron and steel industries, the adjust- ment ratio exceeds unity. For the remaining eight industries the adjustment ratio ranges from 1.005 to 1.044, suggesting that, holding inputs fixed, the frontier of the exporters as a group is generally above that of the group of nonexporting firms. (These results are generated with inputs of nonexporters. They do not change significantly when the input levels of the exporters are used instead.) The adjustment ratios for the more traditional industries average 1.043, while the corresponding average for the relatively more capital-intensive industries is only 1.018. (Textiles, clothing, paper and publishing, plastics, and fabricated metals are classified as traditional; chemicals, iron and steel, machinery, electrical and electronics, and transport equipment are classified as capital-intensive, modern industries.) Exporters and domestic market firms in the modern industries ap- pear to be more similar in production technology than are their respective coun- terparts in the traditional industries. Columns 2 and 3 of table 4 report the average within-group technical effi- ciency of exporters and nonexporters. The average efficiency of nonexporters generated by their production frontier estimates is divided by the adjustment factor to make them comparable to the mean efficiency of exporters. In eight of the 10 industries studied, the mean efficiency of exporters exceeds that of their domestic market counterparts. These differences range frora 2 to 28 per- cent in the more modern industries. Except for the clothing industry, the dif- ference in mean efficiency between exporters and nonexporters in the tradi- tional industries is either very small (textiles, plastics) or negati're (fabricated metals and paper and publishing). In fact, tests showing that the mean effi- ciency of exporters and nonexporters are equal cannot be rejected for textiles, plastics, and fabricated metals (column 4 in table 4). The mean difference be- tween the two subgroups is statistically significant in the other industries. Pa- per and publishing is the only industry where nonexporters are, on average, about 6 percent more efficient than exporters, with the difference being statis- tically significant. Mean efficiency of an industry is influenced by both the efficiency level of each firm in the industry and the distribution of efficiency levels among all the firms that comprise the industry. Higher firm-level efficiencies combined with a higher degree of homogeneity in efficiency (measured by ca) among firms in a given industry yield higher mean efficiencies than in an industry comprised of firms with very heterogeneous efficiency levels. The lower is a2, the more homo- geneous are the efficiency levels among firms in the industry. This could par- tially explain the larger and statistically significant difference in mean efficien- cies between exporters and nonexporters in some of the industries, such as machinery, transport equipment, and clothing. Exporters in these industries tend Aw and Batra 75 to have technical efficiencies that are more homogeneous (lower o2s) than the corresponding variance for nonexporters. In contrast, mean differences in the efficiency levels between the two subgroups are small in the textiles and fabri- cated metals industries due in part to their greater heterogeneity in technical efficiency, as indicated by the larger variances in technical efficiency among ex- porting compared with nonexporting firms. Overall, exporters generally have higher levels of technical efficiency than domestic market firms, although this difference is statistically significant mainly in the more modern industries. This finding is consistent with research examin- ing the productivity of exporting firms (Chen and Tang 1987; Aw and Hwang 1995; and Bernard and Jensen 1996). Given that technological change is likely to be more rapid in the modern industries, a firm's exposure to the international market may be a more important source of knowledge transmission here rela- tive to industries where technology changes less rapidly. Contact with foreign purchasers may explain only part of the higher efficiency observed for export- ers. The observed higher productivity of exporters relative to nonexporters may also reflect empirically unobserved firm-level characteristics that are positively associated with superior managerial or entrepreneurial skills or better access to and use of new or improved technology. Aw, Chen, and Roberts (1997) use this same data set to examine a related issue, namely, whether exporting firms have higher productivity before they enter the export market than they do after. We test for the sensitivity of the results to alternative specifications of the variables by treating R&D, training, foreign capital, and know-how purchases as separate variables and reestimate the production frontier equations including these variables separately as well as their interactions with each other. The sig- nificance and signs of the separate parameter estimates for R&D and training are very similar to those for the single dummy for both types of expenditures (RT), with significantly smaller magnitudes for the coefficients for training relative to R&D. Separating the effects of foreign capital and know-how generally leads to lower significance levels for know-how. As seen before, this is not surprising because the actual number of firms that have either know-how purchases or foreign capital is small. In general, the qualitative nature (signs and significance) of the results and conclusions remains essentially intact. The variances and mean efficiencies of all the industries remain the same. VI. SUMMARY AND CONCLUSIONS Taiwan's impressive growth in the last two decades has often been attributed to its emphasis on exports or, in the context of this article, the international marketplace. In addition to technology licensing and foreign capital, exports could potentially provide firms with a means of acquiring technology from abroad. We proxied firm-level efforts at modifying or adapting technology by their ex- penditures on R&D and on-the-job training. We distinguished the firm's use of its own resources to learn new technology from those technologies that are ac- 76 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 cessible to the firm through its contact with the foreign market, formally through licenses or DFI and informally through foreign purchasers. Using the stochastic frontier technique, we examined the correlation between technical efficiency and a firm's investments in R&D and training and its international linkages. Our findings confirm the positive correlations between exports and the level of productivity found in other developing countries using recently available mi- cro data (Tybout and Westbrook 1995 for Mexico; Roberts and Tybout 1997 for Colombia and Morocco). Contact with foreign purchasers by itself is associ- ated with higher levels of technical efficiency, particularly in the more modern industries. We found some evidence indicating that firm-level export activities have higher payoffs if they are accompanied by complementary investments in the development of in-house technological capabilities, although this relation- ship appears to be industry specific. More generally, our results suggest that efficiency and firm investments in R&D and training are positively correlated in all industries among both export- ers and nonexporters. This correlation is significant in nine out of the 10 indus- tries among nonexporters and in five industries among exporters. Thlus, although there appears to be an additional bang to exporters from simultaneous invest- ments in R&D and training in some industries, investing in R&D and training on its own appears to be significantly correlated with higher technical efficiency. In contrast the presence of foreign capital is generally not significantly correlated with technical efficiency. Taken together, our evidence for Taiwanese manufacturing firms in 1986 suggests that firm-level efficiency is clearly associated with informal contacts with foreign purchasers through the firm's export sales and its investments in R&D and training. The correlation between R&D and training and efficiency is also higher and more widespread across all manufacturing firms than that be- tween exports and efficiency. This study highlights the importance of the firm's own investments in techno- logical capability. In economies like Taiwan, these investments are likely to in- volve incremental modifications to adapt a given technology to fit the firm's specific situation. The significance of this incremental change in technology has been the focus of studies of firms in other developing countries such as the steel industry in Brazil (Dahlman and Fonseca 1987) and the petrochemical industry in Korea (Enos and Park 1988). As Taiwan enters into new and more technology- intensive and sophisticated production methods, although the sources of its new or improved technological information may change, the stimuLation of indig- enous technological effort in identifying, modifying, and assimilating foreign technology at the firm level has to assume an increasingly important role. A crucial caveat in the conclusion with respect to the importarce of develop- ing technological capability is that in Taiwan, as in the other East Asian tigers, the availability of a pool of skilled labor and a relatively competitive economic environment are key factors facilitating the efficient introduction of new tech- nology from abroad (Pack 1992 and Rodrik 1995). The presence of a substan- Aw and Batra 77 tial core of highly educated managers and technicians makes learning easier and facilitates the efficient introduction of the requisite technologies. At the same time, an open and competitive environment, among other things, reduces the cost of traded inputs and permits higher levels of productivity. In addition to maintaining an environment of economic stability and predictability, one of the key contributions of governments to technological capability at the micro level lies in their education policy and investment in training. Firms in developing countries need to be induced to invest more substantially in worker training through effective policy intervention because there are good reasons to expect firms to underinvest in this activity. Finally, the cross-sectional nature of our data set does not allow us to make conclusive statements on the direction of causality between exports or techno- logical investments and a firm's efficiency. The establishment of the issue of causality would provide a stronger basis for policy recommendations that di- rectly influence firm-level behavior. Panel data, which are currently being as- sembled for Taiwan, are needed to see if we can predict higher efficiency or productivity for firms that commit their own resources to enhance their techno- logical capability or that have direct access to foreign technology through ex- port contacts, or both. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Aigner, D. J., C. A. Knox Lovell, and Peter Schmidt. 1977. "Formulation and Estima- tion of Stochastic Frontier Production Function Models." Journal of Econometrics 6:21-37. Allen, T. J. 1977. Managing the Flow of Technology. Cambridge, Mass.: MIT Press. Aw, Bee Yan, Xiaomin Chen, and M. J. Roberts. 1997. "Firm-Level Evidence on Pro- ductivity Differentials, Turnover, and Exports." NBER Working Paper 6235. National Bureau of Economic Research. Cambridge, Mass. Processed. Aw, Bee Yan, and Amy Hwang. 1995. "Productivity and the Export Market: A Firm- Level Analysis." Journal of Development Economics 47(August):313-32. Bell, Martin, and Keith Pavitt. 1993. "Accumulating Technological Capability in Devel- oping Countries." In Proceedings of the World Bank Annual Conference on Devel- opment Economics 1992. Washington, D.C.: World Bank. Berman, E., John Bound, and Zvi Griliches. 1993. "Changes in the Demand for Skilled Labor within U.S. Manufacturing Industries: Evidence from the Annual Survey of Manufacturing." NBER Working Paper 4255. National Bureau of Economic Research, Cambridge, Mass. Processed. Bernard A. B., and J. B. Jensen. 1996. "Facts on the 'Export Boom' in U.S. Manufactur- ing: The Role of Entry Costs, Productivity, and Terms of Trade." MIT, Department of Economics, Cambridge, Mass. Processed. Chen, P. C., and H. C. Lee. 1988. The Development Strategy of the Manufacturing Industry. Taipei: Taiwan Economic Institute. 78 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Chen, T.-J., and D.-P. Tang. 1987. "Comparing Technical Efficiency between Import- Substituting and Export-Oriented Foreign Firms in a Developing Country." Journal of Development Economics 26(August):277-89. Clerides, Sofronis, Saul Lach, and James Tybout. 1996. "Is 'Learning-by-Exporting' Important? Micro-Dynamic Evidence from Colombia, Mexico, and Mo:rocco." NBER Working Paper 5715. National Bureau of Economic Research, Cambridge, Mass. Processed. Cohen, Wesley, and Daniel Levinthal. 1989. "Innovation and Learning: The Two Faces of R&D." Economic Journal 99(May):569-96. .1990. "Absorptive Capacity: A New Perspective on Learning and Innovation." Administrative Science Quarterly 35:128-52. Dahlman, Carl J. 1993. "Electronics Development Strategy: The Role of Government." In Bjorn Wellenius, Arnold Miller, and Carl J. Dahlman, eds., Developing the Elec- tronics Industry. Washington D.C.: World Bank. Dahlman, Carl J., and Francis Fonseca. 1987. "From Technological Dependence to Tech- nological Development: The Case of the USImINAS Steel Plant in Brazil." In Jorge Katz, ed., Technology Generation in Latin American Manufacturing Industries. London: Macmillan. Dahlman, Carl J., and Ousa Sananikone. 1990. "Technology Strategy in the Economy of Taiwan: Exploiting Foreign Linkages and Investing in Local Capability." Private Sector Development Department, World Bank, Washington, D.C. Processed. Dollar, David. 1992. "Exploring the Advantages of Backwardness: The Importance of Education and Outward Orientation." East Asia and Pacific Departmeni, World Bank, Washington, D.C. Processed. Dosi, Giovanni. 1988. "Sources, Procedures, and Microeconomic Effects cf Innovation." Journal of Economic Literature 36:1120-71. Enos, John, and W.-H. Park. 1988. The Adoption and Diffusion of Imported Technol- ogy: The Case of Korea. London: Croom Helm. Ferrantino, M. J. 1992. "Technology Expenditures, Factor Intensity, and Efficiency in Indian Manufacturing." Review of Economics and Statistics 74(November):689-700. Griliches, Zvi. 1987. "Productivity, R&D, and Basic Research at the Firm-Level in the 1970s." American Economic Review 76:141-54. Hewitt, T., and D. Wield. 1992. "Technology and Industrialization." In T. Hewitt, H. Johnson, and D. Wield, eds., Industrialization and Development. Clxford: Oxford University Press. Hobday, Michael. 1995. Innovation in East Asia: The Challenge to Japan. Cheltenham, U.K.: Edward Elgar Publishing Co. Hou, C. M., and Gee San. 1990. "National Systems Supporting Technical Advance in Industry: The Case of Taiwan." Taipei, Taiwan (China). Processed. Huang, C. J., and J.-T. Liu. 1994. "Estimation of a Non-Neutral Stochastic Frontier Production Function." Journal of Productivity Analysis 5:171-80. Jondrow, James, C. A. K. Lovell, I. S. Materov, and Peter Schmidt. 1982. "On the Esti- mation of Technical Inefficiency in the Stochastic Frontier Production Function Model." Journal of Econometrics 19:233-38. Kumbhakar, Subal, Soumendra Ghosh, and J. T. McGuckin. 1991. "A Generalized Pro- duction Frontier Approach for Estimating Determinants of Inefficiency in U.S. Dairy Farms." Journal of Business and Economic Statistics (July):279-86. Aw and Batra 79 Levin, Richard, Wesley Cohen, and David Mowery. 1987. "Appropriating the Returns from Industrial Research and Development." Brookings Papers on Economic Activ- ity 3:783-820. Levy, Brian. 1994. "Technical and Marketing Support Systems for Successful Small and Medium-Size Enterprises in Four Countries." Policy Research Working Paper 1400. Policy Research Department, World Bank, Washington, D.C. Processed. Lucas, Robert. 1993. "Making of a Miracle." Econometrica 61 (March):251-72. Mairesse, Jacques. 1990. "Time-Series and Cross-Sectional Estimates on Panel Data: Should They Be Equal?" In J. Hartog, G. Ridder, and J. Theeuwes, eds., Panel Data and Labor Market Studies. North-Holland: Elsevier Science Publishers. Meeusen, Wim, and Julien van den Broeck. 1977. "Efficiency Estimation from Cobb- Douglas Production Functions with Composed Error." International Economic Re- view 18(June):435-44. Mody, Ashoka. 1993. "Alternative Strategies for Developing Information Industries." In Bjorn Wellenius, Arnold Miller, and Carl J. Dahlman, eds., Developing the Elec- tronics Industry. Washington, D.C.: World Bank. Mowery, D. C. 1983. "The Relationship between Intra-Firm and Contractual Forms of Industrial Research in American Manufacturing, 1900-1940." Explorations in Eco- nomic History 20:351-74. Pack, Howard. 1992. "Learning and Productivity Change in Developing Countries." In G. K. Helleiner, ed., Trade Policy, Industrialization, and Development. Oxford: Clarendon Press. . 1993. "Technology Gaps between Industrial and Developing Countries: Are There Dividends for Latecomers?" In Proceedings of the World Bank Annual Con- ference on Development Economics 1992. Washington, D.C.: World Bank. Pack, Howard, and Larry Westphal. 1986. "Industrial Strategy and Technological Change: Theory vs. Reality." Journal of Development Economics 22(1):87-128. Pitt, Mark, and L. F. Lee. 1981. "The Measurement and Sources of Technical Ineffi- ciency in the Indonesian Weaving Industry." Journal of Development Economics 9(August):43-64. Roberts, Mark J., and James Tybout. 1997. What Makes Exports Boom? Washington, D.C.: World Bank. Rodrik, Dani. 1995. "Getting Interventions Right: How South Korea and Taiwan Grew Rich." Economic Policy 20(April):55-107. Taiwan, Department of Statistics. 1986. 1986 Census of Manufacturing. Taipei. Tybout, James, and M. D. Westbrook. 1995. "Trade Liberalization and Dimensions of Efficiency Change in Mexican Manufacturing Industries." Journal of International Economics 31(August):53-78. Westphal, Larry E. 1990. "Industrial Policy in an Export-Propelled Economy: Lessons from South Korea's Experience." Journal of Economic Perspectives 4(Summer):41-59. Westphal, L. E., Yung W. Rhee, and Gary Pursell. 1984. "Sources of Technological Capability in South Korea." In Martin Fransman and K. King, eds., Technological Capability in the Third World. London: Macmillan Press. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1: 81-104 Skills, Schooling, and Household Income in Ghana Dean Jolliffe This article examines the impact of cognitive skills on the income of households in Ghana. It uses scores on mathematics and English tests to measure cognitive skills and estimates the returns to these skills based on farm profit, off-farm income, and total income. The article uses Powell's censored least absolute deviations and symmetrically trimmed least squares estimators to estimate farm and off-farm income. In contrast to Heckman's two-step or the Tobit estimator, Powell's estimators are consistent in the presence of heteroscedasticity and are robust to other violations of normality. The results show that cognitive skills have a positive effect on total and off-farm income but do not have a statistically significant effect on farm income. This article estimates the effect of cognitive skills on the incomes of Ghanaian households. Scores on mathematics and English tests are used as measures of cognitive skills, and the returns to these skills are measured by estimating farm profit, off-farm income, and total income. Three features distinguish the analy- sis from that in much of the human capital literature. First, it measures human capital using performance on mathematics and English tests rather than years of schooling. Second, it measures the returns to human capital by estimating total household income and its components rather than just examining wage income, as is more typically done. Third, it incorporates information about the sample design into the estimation strategy. The choice of test scores, as opposed to years of schooling, for a measure of human capital is motivated by the argument that it is not school attendance that intrinsically increases a worker's productivity, but instead the skills obtained while in school. Following this argument, test scores serve to proxy for human capital better than years in school.' The use of test scores standardizes the mea- sure of human capital across numerous schools of varying quality. Just as col- leges in the United States use standardized tests to control for the variation in 1. One important caveat to this statement is that this article assumes that test scores are orthogonal to the estimated residuals. If endogeneity bias exists, there is no reason to believe that the bias is more or less of a problem for estimating returns to schooling or skills. Jolliffe (1996) tests for endogeneity bias of the estimated returns to schooling and finds that it is not significant for this sample of households. Dean Jolliffe is with the Food Consumption and Nutrition Division at the International Food Policy Research Institute. He began work on this article while employed as a consultant with the Policy Research Department at the World Bank. The author wishes to thank Paul Glewwe, Bo Honore, Hanan Jacoby, Chris Paxson, Cecilia Rouse, and four anonymous referees for comments. (D 1998 The International Bank for Reconstruction and Development/THE WORLD BANK 81 82 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I school quality of its applicants, so too does the use of tests control for the sig- nificant variation in school quality across Ghana. Several factors motivate the decision to measure the returns to human capital by estimating total household income, including farm and off-farm income. The majority of the human capital literature focuses on wage income, but the large majority of households in developing economies are self-employed workers, not wage earners. World Bank (1995) provides a detailed discussion of the typical composition of labor forces in developing countries. Grigg (1991) discusses the predominance of agricultural laborers in developing countries. Another reason to estimate total income rather than just farm profits or wage income is that a large portion of Ghanaian households earn income from nu- merous sources. For example, approximately 70 percent of Ghanaian house- holds engaged in farming have at least one household member who is engaged in some form of off-farm work. Estimating the returns to human capital in just one source of income presents a skewed picture of the returns to human capital for households engaged in numerous income-generating activities. This statement assumes that labor markets are inefficient and that it is not possible for house- holds to hire in and out labor of certain skill types. Two pieces of evidence support the assumption of incomplete labor markets. First, labor markets are not very active in Ghana. The average farming household only spends 5 percent of farm income on hiring outside labor. Second, the returns to skills vary dra- matically across economic activities. A strictly practical reason for using total household income to measure the returns to human capital stems from a feature of the data Income data from household surveys in developing countries are measured largely at the household level, and the data typically make it difficult to attribute portions of household income to individual household members. This is because the majority of laborers work either for household farms or for nonfarm, small household enterprises. This point is supported in more detail in Jolliffe (1996). The importance of incorporating sample design information into the estima- tion strategy results from the standard practice of using multistage sample de- signs for cross-sectional, nationally representative surveys. Multistage (or cluster- based) sample selection frequently results in the need to drop i:he assumption that residuals are homoscedastic. The presence of cluster-induced hetero- scedasticity has two important effects. First, standard errors that have not been corrected for the sample design dramatically underestimate the correct standard errors. Second, the standard estimators used for limited dependent variable analy- sis including sample selection estimators are biased in the presence of heteroscedasticity. Section I reviews the human capital literature pertaining to developing coun- tries and the use of test scores. It begins by discussing the literature on the re- turns to additional years of schooling in developing countries and then covers the literature on test scores and school quality. Section II discusses the survey Jolliffe 83 and sample design and reviews the estimation strategy and results. Section III presents some concluding comments. I. LITERATURE REVIEW The literature on returns to human capital in developing countries focuses predominantly on measuring the returns to additional years of schooling for wage earners. Psacharopoulos (1985 and 1994) summarizes the results from more than 55 such wage studies from Africa, Asia, and Latin America. These summaries present a consistent pattern of very large returns to primary educa- tion and somewhat smaller returns to secondary and postsecondary education. Psacharopoulos (1994) states that the average private rate of return to primary education in developing countries is 29 percent, while the returns to secondary and postsecondary education are 18 and 20 percent, respectively. The main prob- lem with the focus of these studies is that the majority of individuals in develop- ing countries are not wage earners. For example, about 20 percent of working individuals in Ghana are wage earners. Similarly, wage earners make up 15 per- cent of the workforce in India, 19 percent in Haiti, 20 percent in Nigeria, and 11 percent in Togo (World Bank 1995: table A-2). In order to draw conclusions about the returns to education for the majority of the population, it is necessary to assume that the benefits of education to farmers and other self-employed workers are the same as those accrued by wage earners. The largest component of the workforce in most developing countries is en- gaged in self-employed farm work. Examining the returns to education for these laborers results in a more representative assessment of the private value of edu- cation to the population. Jamison and Lau (1982) review the results of more than 35 studies that measure returns to the education of farmers in Africa, Asia, and Latin America. Most of these studies suggest that education has a positive effect on farm production, but the statistical significance of this result is often weak. In particular, Jamison and Lau's review finds no support for the hypoth- esis that there are any returns to education for farmers in Africa. The lack of a significant effect of schooling on farm profit has often been attributed to either the low technological level of production or the absence of technological change in Africa. Foster and Rosenzweig (1996) present evidence that technological change increases the returns to schooling. The contrast between the results of wage regressions and farm-profit regressions suggests that it is inappropriate to use results from one or the other to make inferences about the effects of educa- tion on the African workforce. The remaining component of household income (at least that resulting from labor) is self-employed, off-farm income. This component has been somewhat ignored in the human capital literature, in spite of the fact that self-employed, off-farm income is at least as significant as wage income in many developing countries. In the case of Ghana, 47 percent of households generate some self- employed, off-farm income, while 36 percent of households have at least one 84 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I wage-earning member. Blau (1986) finds that urban, self-employed men in Ma- laysia earn substantially more than wage earners with the same characteristics. Vijverberg (1993: 2) also notes that "family enterprises with one to four work- ers accounted for about 70 percent of manufacturing in India and Indonesia, 60 percent in the Philippines, and 40 percent in Korea and Columbia." Soon (1987) and Vijverberg (1993) have written two of the very few papers on returns to human capital that examine self-employed, off-farm work. They both find evi- dence that schooling asserts a positive influence on self-employment income. Vijverberg, though, finds little evidence of returns to cognitive skills. Soon states that the returns to schooling for the self-employed are significantly lower than for wage earners. Vijverberg uses a subsample of the data used here and selects households with self-employed, off-farm income from a nationally representa- tive sample but does not correct for the potential selection bias. By dropping more than 30 percent of the sample for various reasons, Soon also faces but does not deal with potential sample-selection bias. This problem is handled here by including all households and assigning zero self-employed, off-farm income to those not engaged in this activity. The strategy of focusing on only one source of income (such as wage or farm income) not only potentially suffers from sample-selection bias but also ignores the fact that many households generate income from several sources. For the individual who is engaged in both farm and wage work, the returns to human capital are the benefits derived in both sectors. For this reason, this study exam- ines total household income as well as farm and off-farm income. A motivation for using test scores as a measure of human capital instead of years of schooling is to control for variation in school quality. Another motiva- tion is the argument that it is not school attendance that increases a worker's productivity, but rather the skills learned. Test scores should serve as a better measure of these skills than years of schooling. This motivation is reasonable if there is evidence that school quality varies within countries and if school quality affects test score performance. Hanushek (1995) argues that school quality is an important determinant of the returns to education and that quality varies dra- matically within developing countries (and the United States). Hanushek further argues that the focus of education policy should switch from increasing access to schooling to improving the quality of schooling. There is also an extensive U.S. literature on whether school quality affects the rate of return to schooling. Rizzuto and Wachtel (1980) and Card and Krueger (1992a and 1992b) use U.S. census data on location and date of birth to match individuals with state-level data on school quality. All three studies find that school quality is an important determi- nant of the rate of return to education. Glewwe and Jacoby (1994) examine the test performance of middle school students in Ghana and find that the quantity and quality of school attendance have an impact on test scores. Behrman and Birdsall (1983) argue that ignoring differences in the quality of schools may bias the estimated returns to schooling. They use data from Brazil to show that the estimated bias can be large. Jolliffe 85 The literature on measuring the returns to cognitive skills is significantly more sparse than the returns-to-schooling literature. Boissiere, Knight, and Sabot (1985) and Knight and Sabot (1987) use household survey data on wage earners from Kenya and Tanzania to estimate the returns to cognitive skills, as well as other issues. Boissiere, Knight, and Sabot find positive and significant returns to cog- nitive skills and to schooling and suggest that these results refute the screening and credentialism theories of education. Knight and Sabot show that years of schooling are an important determinant of cognitive skills and that cognitive skills are important determinants of wages. Alderman and others (1996), using data from four districts in Pakistan, find that cognitive skills have a significant effect on rural wages. Glewwe (forthcom- ing), using a subset of the data used in this article, finds evidence that cognitive skills are important determinants of wages. He also shows that improvements in school quality have a higher return than additional years of schooling for wage earners. Jolliffe (forthcoming), using a different subset of the data used in this article, suggests that mathematics skills have a positive effect on the farm in- come of Ghanaian households, while English skills have no effect. In contrast to papers on the returns to cognitive skills, this article examines the effect of skills on total household income and not just wage or farm income. It is not enough to know if improved skills result in higher wage or farm income because increases in one type of income could be quite different from changes in the other type. For example, if it is found that improved skills do not improve farm income, households may still benefit from improved skills because their nonfarm income has increased. Moreover, this article recognizes the importance of sample design in the esti- mation strategy. The presence of design-induced heteroscedasticity requires large adjustments to standard errors and the use of estimators that are robust to the failure of the homoscedasticity assumption. Much of the literature discussed here focuses on certain types of occupations and therefore likely encounters some form of sample-selection bias. The sample selection results from either selecting farm households or wage earners from a nationally representative survey or simply from the fact that occupation is a choice variable. The papers that ad- dress the potential for sample-selection bias use the standard correction tech- niques, which are estimated by either a maximum-likelihood or two-step esti- mator. These correction techniques require homoscedastic residuals, and the existing literature relies heavily on this untenable assumption. This article handles sample selection by modeling farm and off-farm income as a type I Tobit model and uses estimators that are robust to heteroscedasticity. II. DATA AND ESTIMATION The analysis uses data from the Ghana Living Standards Survey (GLSS), a na- tionwide household survey carried out by the Ghana Statistical Service with technical assistance from the World Bank. As with most household surveys, the 86 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Table 1. Descriptive Statistics from the Ghana Living Standards Survey, 1988-89 Standard Variable Mean deviation Minimum Maximum Income Log total income 12.6 0.62 9.98 15.03 Log farm income 6.80 5.61 0.00 13.92 Log off-farm income 11.7 2.58 0.00 15.03 Household test scores and schooling Maximum English score 11.9 12.5 0.00 3.70 Maximum mathematics score 11.4 9.94 0.00 4.20 Average English score 7.60 9.48 0.00 35.00 Average mathematics score 7.56 7.58 0.00 40.00 Maximum years of schooling 8.12 4.70 0.00 23.00 Average years of schooling 4.16 3.45 0.00 20.00 Farm income Log (acres of land farmed) 2.11 1.90 0.00 6.55 Household average log of farm experience 1.71 1.16 0.00 4.11 Cluster/zone average day farm wage 342 114 0.00 600 Regional average In (price fertilizer) 7.57 0.24 7.05 7.93 In (price insecticide) 7.09 0.51 6.19 7.85 Price of maize 5.14 2.39 2.00 15.75 Price of okra 0.67 0.34 0.20 3.20 Price of cassava 1.13 0.84 0.01 4.00 Price of pepper 2.93 0.99 0.60 5.89 Off-farm income Log (value of business assets) 3.90 4.41 0.00 19.60 Household maximum log (off-farm experience) 1.62 1.15 0.00 4.51 Area average off-farm wage' Type 1 0.35 0.11 0.08 0.45 Type 2 0.47 0.15 0.20 0.65 Type 3 0.75 0.25 0.39 1.09 Type 4 0.82 0.32 0.38 1.53 Type 5 2.13 1.05 0.9'J 3.88 Household characteristics Log of household size 1.34 0.70 0.0( 3.64 Household gender composition' 1.49 0.28 1.0( 2.00 Number of males 15-24 years old 0.42 0.71 0.00 5.00 25-34 years old 0.25 0.45 0.00 3.00 35-44 years old 0.18 0.39 0.00 1.00 45-55 years old 0.14 0.34 0.00 1.00 Number of females 15-24 years old 0.42 0.66 0.00 5.00 25-34 years old 0.34 0.51 0.00 3.00 35-44 years old 0.19 0.41 0.00 2.00 45-55 years old 0.17 0.40 0.CO 3.00 Note: The sample consists of 1,388 households. a. The off-farm wages are regional averages of wages grouped into five occupational types. Type 1 occupations are the lowest-wage activities, and type 5 are the highest. The data on wvages come from the half of the sample that was not administered the supplemental education module. b. Household gender composition is the household average value of the binary variable that takes the value of 1 for men and 2 for women. Source: Author's calculations based on the Ghana Living Standards Survey for October 1988- September 1989. Jolliffe 87 sample is not a simple random draw but rather a two-stage sample. (For more details on the sample design, see Scott and Amenuvegbe 1989.) The survey, ad- ministered from October 1988 to September 1989, covers 3,200 households and contains detailed information on formal and informal labor activities, house- hold farm activities, expenditures, education status of household members, and many other determinants of household welfare. Table 1 provides basic descrip- tive statistics of the sample used here. See Glewwe and Twum-Baah (1991) for more information. Members from approximately half of the households, or 1,585 households, in the 1988-89 GLSS were randomly selected to be administered a series of cognitive skills tests. From this random sample, 197 households are dropped from the analysis, resulting in a sample of 1,388 households. The analysis uses data on members of these 1,388 households between the ages of 15 and 55. Adults over 55 are excluded because they were not administered the battery of tests, and children under 15 are excluded to ensure that children still in school are not treated as working adults. The large number of dropped households is due primarily to missing test score data from 163 of the households. In addi- tion to the households dropped because of missing test score data, 31 house- holds are dropped because there are no farm price data in their region of resi- dence. Two households are dropped because there are insufficient data to construct an estimate of their total consumption, and one household is dropped because there are no data on school levels within the household. Jolliffe (1996) presents a detailed comparison of the sample with and without the 197 dropped households and concludes that dropping the households does not significantly alter the composition of the sample. In particular, the rural/urban distribution of the data is the same, as is the geographic distribution. Similarly, the average values of education levels, per capita expenditure, and gender and age compo- sition of the households are essentially the same when the 197 households are dropped. Cognitive Skills and School Attainment In spite of ambitious literacy and education goals set during the 1960s (and largely pursued during the 1970s), school attainment for many Ghanaians is low. The GLSS data show that the average Ghanaian ages 15-55 has completed just six years of schooling, and less than 6 percent of this age category has any postsecondary education. The government of Ghana's goal of increasing the amount of schooling received by its citizens is nonetheless reflected in the data. Individuals ages 30-34 who were going to school during the late 1960s and early 1970s have on average seven years of school. This figure contrasts with an average level of four years of schooling for people ages 45-49, most of whom received their education prior to the country's independence. There is also a pronounced difference in school attainment between urban areas, where resi- dents have on average eight years of school, and rural areas, where they average 4.6 years. 88 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I The cognitive skills tests, which consisted of two mathematics and two En- glish comprehension tests, were administered to household members ages 9- 55. Both the mathematics and English tests consisted of a relatively easy test and a more difficult, advanced test. Those who correctly answered five or more (out of eight) questions on the easy test were given the advanced test. To con- struct one score for mathematics skills and one for English skills, the sum of the simple and advanced tests is taken. Zeros are assigned on the advanced test to those who did not receive a score of five or higher on the easy test. Those individuals who were unable to read the tests are also assigned a zero, as are most individuals with less than three years of schooling. The fieldwork was designed for tests to be administered only to individuals over nine years of age and with at least three years of schooling. This rule was not cornpletely fol- lowed, and about 10 percent (or 119) of the individuals with less than three years of schooling took the tests. The test scores follow the same patterns as years of schooling attained in that they are correlated with age and residence. For those individuals ages 30-34, the average mathematics and English scores are 9.8 and 11.1, respectively. Individu- als ages 45-49 scored significantly lower on both the mathematics (6.0 answers correct) and English tests (6.5 answers correct). Similarly, the average scores in urban areas were 10.8 (on the mathematics test) and 12.5 (on the English test), both of which are significantly higher than the average scores of 5.8 (mathemat- ics) and 5.7 (English) in the rural regions. Table 2 presents further evidence that schooling is an important determinant of the mathematics and English scores. Skills are modeled as functions of expe- rience, schooling, and ability. Dummy variables for relationship to head of house- hold are also included to control for any sort of gender or family-related bias resulting from position in the household. For example, dummy variables are used to designate whether the individual is a spouse, son or daughter, grandson or granddaughter, niece or nephew. These are tested for joint significance with the female dummy variable. A household-level, fixed-effects estimator is used to control partially for ability. This assumes that an important comnponent of an individual's ability is derived from the household, whether genetically or learned. The estimates in table 2 suggest that schooling is an important determinant of test scores, even controlling for experience and household fixed effects. One reason for using test scores rather than years of schooling is to control for variations in school quality. Table 3 presents evidence thait the quality of schooling is better in urban areas than in rural regions. The percentage of class- rooms in primary schools that cannot be used when it rains is 58 percent higher in rural regions than in the urban areas. Similarly, the percentage of rooms with- out blackboards in rural primary schools is twice as high as in urban ones. Ur- ban teachers have more experience and somewhat more schooling. The differences in school quality are associated with differences in student test performance, although the school quality variables were not measured when the individuals attended school. The implicit assumption is that the relative dif- Jolliffe 89 Table 2. Determinants of Cognitive Skills, Ghana, 1988-89 Variable English Mathematics Potential experience 0.221** 0.030 (0.1069) (0.0829) Potential experience squared -0.003' -0.001 (0.0018) (0.0014) Years of schooling 0.941-' 0.662-t (0.1887) (0.1463) Years of schooling squared 0.051>** 0.045**' (0.0093) (0.0072) Interaction: school and experience -0.005 0.003 (0.0051) (0.0040) Dummy: 1 = female, 0 = male -2.174* - -2.512- (0.5281) (0.4095) Intercept -0.011 3.198** (1.8265) (1.4163) Adjusted R2 0.71 0.74 Number of observations 2,295 2,295 Note: Values are fixed-effects estimates, controlling for household effects. The dependent variables are English and mathematics test scores. The 2,295 individuals come from 1,118 households that have more than one member between the ages of 15 and 55. Thirteen dummy variables for relationship to the head of household are included in the estimation, and, along with the gender dummy, they are jointly significant (p = 0.00) for both test scores. The household fixed effects are also jointly significant (p = 0.00) for both scores. Potential experience is defined as age minus years of schooling minus six. Standard errors are in parentheses. p-value is less than 0.1. p-value is less than 0.05. * p-value is less than 0.01. Source: Author's calculations. Table 3. Measures of Primary School Quality in Urban and Rural Areas of Ghana, 1988-89 Urban Rural Standard Standard Indicator Mean deviation Mean deviation Fraction of rooms unusable when wet 0.19 0.025 0.30 0.029 Fraction of rooms with a blackboard 0.96 0.013 0.90 0.015 Teacher's average experience (years) 10.97 0.383 8.26 0.312 Average teacher's schooling (years) 12.17 0.197 11.45 0.143 Source: Jolliffe (forthcoming: table 7.1). ference between urban and rural schools has not changed significantly over time. Urban individuals with six years of schooling scored 29 percent higher than rural individuals with the same years of schooling on the mathematics tests and 169 percent higher on the English tests. Similarly, individuals living in urban areas with eight years of schooling scored 19 and 80 percent higher than indi- viduals in rural areas on the mathematics and English tests, respectively. The differences in school quality highlight problems of using educational attainment 90 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I as a measure of human capital. In the event that an additional year of schooling produces different skill levels across groups of individuals, using levels of schooling will systematically mismeasure human capital and thus potentially bias estimates of the returns to human capital. Cognitive Skills and Household Income The returns to cognitive skills accrued by households are estimated using three separate household-level income functions: total income, farm income, and off- farm income. The mathematics and English test scores are included in all models as explanatory variables. Both the household average and maximum test scores of household members ages 15-55 are used. For more discussion on these two indicators, see Jolliffe (1996); the subsection on farm profit dravws heavily on Jolliffe (forthcoming). Farm revenues are measured as the sum of the value of all crops and animal products marketed in the last 12 months, the value of crops kept for seed and given away as gifts, and the value of home-consumed food and animal products. The value of farm output, Gf, is assumed to be a function of cognitive skills and the logs of land, labor, and crop inputs. The log of the farm production function Gf can be written as: (1) Gf = Gf (Af, Lf, Xf, Tj, u) where Af contains the log of acres of land cultivated and farming experience, both of which are treated as fixed inputs. Lf is the log of hours of household farm labor, Xf represents all other variable inputs (fertilizer and insecticide), iq measures cognitive skills (English and mathematics test scores), and v is an error term. Throughout the article, the subscript f denotes a farm variable, and the subscript o denotes an off-farm variable. As it is typically assumed that farmers maximize farm profits, not farm rev- enues, expenditures on insecticide, pesticide, hired labor, seed, rented land, stor- age, containers, and transportation are subtracted from farm revenues. Crops given as payments for other inputs are also deducted from farm revenue. The resulting measure of farm profit is conditional on the quantity of land and labor. The average value of this conditional, or restricted, farm profit is 144,604 cedis. During the year of the survey fieldwork, inflation in Ghana was 24 percent (Ghana Statistical Service 1991a). To correct for this all values in cedis are de- flated to October 1988. The average exchange rate during 1988 was 200 cedis to the U.S. dollar (Ghana Statistical Service 1991b). Solving for the input demand functions in terms of prices and substituting them back into equation 1 results in a restricted net income function Yf. The log of this is estimated as: (2) Yf= Yf(Af, Lf, pf , -,') where pf is a vector of input and output prices (fertilizer, insecticide, maize, Jolliffe 91 okra, cassava, and pepper prices and farm wages), and ,u is an error term. The measures of farm wages used are the wages for an adult male day laborer. Farm input and output prices are cluster averages. Estimating farm profit rather than farm output is both more consistent with economic theory and also likely to reduce the possibility of endogeneity bias. All of the chosen levels of the variable inputs, which are elements of the farm pro- duction function, are replaced with their prices. Even when markets are not functioning perfectly, it is still reasonable to assume that prices are exogenous to the farm household's decisions. The restricted farm profit function, equation 2, still contains the quantities of land and labor. Jolliffe (1996) shows that omit- ting land from farm profit regressions has no effect on the estimated return to human capital and argues that this implies that land either is fixed in the short run or is orthogonal to schooling and skill variables. The measure of off-farm income aggregates wage income and self-employment income. The decision to aggregate these two loses some information but helps to focus on the difference between farm and nonfarm income. The measure of wage income adjusts the wage rate by including all pecuniary remuneration for the labor supplied including commissions, bonuses, tips, allowances, and gratuities. The wage income is also adjusted to reflect the value of all nonpecuniary pay- ments including remuneration in the form of food, crops, animals, housing, cloth- ing, transportation, or any other form. The log of off-farm income, YO, is modeled as: (3) Y_= Y. (A.,L.,p, T,e) where A, contains the log of the value of business assets and off-farm work experience, L. is the log of hours of household off-farm labor, p0 is off-farm wages, ij measures cognitive skills (English and mathematics test scores), and e is an error term. Both the farm and off-farm income functions include hours of household la- bor, whose levels are chosen by the household. To correct for this potential endogeneity bias, farm and off-farm labor are modeled as functions of house- hold size, gender and age composition of the household, and farm and off-farm income. Household characteristics enter the labor functions because it is the total household level of labor supply, not individual-level labor supply, that is being modeled. Gender composition is included because of the cultural norms that dictate that men, women, and children will typically perform different work activities. Often the activities of women and children are not picked up in the measure of off-farm work. For example, the time spent collecting firewood or preparing food is not included in the measure of total hours worked. By using farm and off-farm income as determinants of labor supply, the model treats labor markets as imperfect. This model of labor supply allows equations 2 and 3 to be rewritten as: (2w) Yf = Yf[Af, Lt(Xh, Yf, Y), pf, r, A] = Yf (A, Xh X , X,m, A) 92 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I (3?) Y. = YJA., L.(Xh, Y,, Yf), p0, r, £] = Y0 (A, Xh, P, X, E) (4) Y = Yf + Yo= Y(A, Xh, p,n, v) where A is the farm and off-farm fixed inputs, Xh represents household charac- teristics, and p is a vector of farm and off-farm prices. Total household income, equation 4, is modeled simply as the sum of farm and off-farm income and is also estimated in log form. Because the total household variable is the sum of farm profit, enterprise profit, and wage earnings, it might be more accurately described as total household earnings. Heteroscedasticity The two-stage design of the GLSS sample typically results in the rejection of the assumption of homoscedasticity because observations drawn from within a cluster are likely to have characteristics that are more similar than observations drawn from different clusters. This difference between intra-cluster and inter-cluster correlations will most likely result in heteroscedasticity. Examining the Kish design effect is helpful to determine if the heteroscedasticity resulting from the sample design will introduce a large bias in the estimated standard errors. The square root of this measure of intra-cluster correlation gives an upper bound to the potential bias. (See Kish 1965 for more details.) The ordinary least squares (OLS) residuals from estimating the log of total income have a Kish clesign effect of 2.67. This means that the standard errors from OLS estimation rnay need to be corrected by as much as the square root of the design effect, or increased by 63 percent. Correcting the total income model for heteroscedasticity is fairly straightfor- ward. OLS estimation results in consistent parameter estimates for these models, but the residuals need to be corrected following Huber (1967). The Huber cor- rection is asymptotically equivalent to a jackknife estimate. The principle be- hind Huber's formula is that if the design effect is not significant, then the esti- mated standard error should not be affected by dropping any particular cluster. The correction employed by Huber is asymptotically equivalent to repeatedly estimating standard errors, dropping one cluster at a time until all clusters have been excluded once. The average value of these estimated standard errors is then the Huber-corrected standard error. Scott and Holt (1982) discuss the impact of sample design on the correction required for the OLS standard errors. Heteroscedasticity introduces more problems for estimating farm and off- farm income. Both of these variables are censored at zero, and the standard estimators used for the censored model, such as the Tobit or Heckman's two- step estimator, result in biased estimators. Arabmazar and Schmidt (1981) show that the bias resulting from the Tobit estimator can be large. Here farm and off-farm income are estimated using two estimators proposed by Powell, both of which are designed for the censored model and are consistent in the presence of heteroscedasticity. The first is Powell's (1986) symmetrically trimmed least squares (STLS) estimator. The second is Powell's (1984) censored Jolliffe 93 least absolute deviations (CLAD) estimator. CLAD and STLS estimators are both robust to heteroscedasticity of unknown form. The STLS estimator is more sensi- tive than the CLAD estimator to outliers but is more efficient. Both censored income functions are modeled as a type I Tobit model, and therefore the selec- tion process is modeled implicitly as being determined by the same variables that explain income. (See the appendix for more details on these estimators.) Reduced-Form Results Table 4 summarizes the results from estimating the three reduced-form in- come functions-total income, farm income, and off-farm income-with OLS, CLAD, STLS, and least absolute deviations (LAD) estimators. Table 4 also summa- rizes the F-statistics, which test the joint significance of the test scores.2 The primary result from estimating the impact of cognitive skills in the total income function is that the test scores are jointly significant whether the average value or maximum value of test scores is used and whether estimated by OLS or LAD. The weakest case is the LAD estimation using the maximum test scores. In this case, the F-statistic for the test of joint significance is 7.96, and the p-value is 0.0004.3 Each reduced-form function is estimated separately using the household aver- age and maximum test scores. For the sake of brevity, table 5 presents the full estimation results only for the model using the average test scores and the least squares estimators. The estimation results from using the maximum test scores, and the absolute-deviations estimators are qualitatively very similar and are pre- sented in full in Jolliffe (1996). To get a sense of the magnitude of the effect that test scores have on total income, consider the example of increasing both house- hold average scores by one standard deviation. An increase of this size means 7.6 more correct answers on the mathematics test and 9.5 more on the English test. From the results in table 5, this change in test scores results in an increase in total income of 9.6 percent. Repeating this exercise using maximum test scores and the LAD estimates, which are the lowest estimated returns to skills, results in an increase in total income of 6.2 percent. Increasing test scores by one standard deviation is a large change, yet the corresponding effect on total income is sig- nificantly smaller than would be expected from the wage studies reviewed in Psacharopoulos (1985 and 1994). In contrast to the total income estimates, the test scores do not appear to be important determinants of farm income. The hypothesis that jointly the test scores add nothing to the predictive power of the farm income model cannot be 2. The reported F-statistics result from Wald tests using robust estimates of the variance-covariance matrix. F-tests based on the sum of squares (or R-squared) are incorrect when homoscedasticity is violated. The Wald test is robust to heteroscedasticity. 3. The F-statistics for the joint significance of the (Huber-corrected) OLS estimates are 14.1 when using the maximum test scores and 11.6 when using the average test scores. Similarly, the F-statistic for the joint significance of the LAD estimates is 8.4 when using the average test scores. For all of these statistics, the probability of observing a statistic this large when the null hypothesis is true is essentially 0. 94 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Table 4. Summary of the Effect of Test Scores on Income, Ghana, 1988-89 Off-farm Estimator and variable Total income Farm income income Parameter estimates OLS and STLS Household maximum English score 0.001 0.012 0.015 (0.0018) (0.0081) (0.0173) Household maximum mathematics score 0.009- -0.002 0.049** (0.0027) (0.0104) (0.0214) Household average English score 0.003 0.007 0.000 (0.0028) (0.0112) (0.0242) Household average mathematics score 0.009" 0.000 0.086.. (0.0043) (0.0146) (0.0309) LAD and CLAD Household maximum English score 0.001 0.018 0.060-* (0.0028) (0.0174) (0.0230) Household maximum mathematics score 0.005 -0.010 0.008 (0.0034) (0.0223) (0.0269) Household average English score 0.003 0.014 0.055 (0.0041) (0.0255) (0.0344) Household average mathematics score 0.007 -0.005 0.032 (0.0049) (0.0313) (0.0405) F-statistics (joint significance of test scores) OLS and STLS Household maximum score 14.10 2.33 13.88 [0.00001 [0.0973] [0.0000] Household average score 11.60 0.57 13.43 [0.0000] [0.5656] [0.0000] LAD and CLAD Household maximum score 7.96 0.79 15.3 [0.00041 [0.4544] [0.0000] Household average score 8.36 0.29 11.35 [0.0002] [0.75051 [0.0002] Note: The total income functions are estimated by ordinary least squares (OLS) and least absolute deviations (LAD). All measures of income are in logs, and the test scores are in leve.s. The OLS standard errors are Huber-corrected for heteroscedasticity. The LAD standard errors are bootstrap estimates. The farm and off-farm income functions are estimated by symmetrically trimmed least squares (sTLS) and censored least absolute deviations (CLAD). The standard errors are bootstrap estimnates. All bootstrap estimates result from resampling 1,000 times. The F-statistics are from Wald tests for the joint significance of the mathematics and English test scores. The probability values of the F-statistics are reported in square brackets. The degrees of freedom for these statistics and tests of joint significance of other variables are reported in Jolliffe (1996). Standard errors are in parentheses. **p-value is less than 0.05. - p-value is less than 0.01. Souirce: Author's calculations. rejected for any of the specifications. This is the case whether the average values or maximum values of the test scores are used and whether estimated by STLS or CLAD. The test scores approach statistical significance only in. the case of STLS estimation using maximum test scores. In this case the F-statistic for the test of joint significance is 2.3, and the p-value is 0.097. Jolliffe 95 Table 5. Full Estimation Results of the Impact of Household Average Test Scores on Income, Ghana, 1988-89 STLS OLS, Off-farm Variable total income Farm income income Household average English score 0.003 0.007 0.000 (0.0028) (0.0112) (0.0242) Household average mathematics score 0.009- 0.000 0.086-> (0.0043) (0.0146) (0.0309) Log acres of land farmed 0.026' 2.614-' -0.015 (0.0151) (0.0417) (0.0904) Household average log of farm experience -0.026 0.624"' -0.925- (0.0197) (0.0639) (0.1414) Cluster/zone average day farm wage 0.004 0.139- -0.086 (0.0165) (0.0699) (0.1476) Regional average Log (price fertilizer) -0.707-' -1.086" -0.035 (0.1448) (0.5127) (1.0736) Log (price insecticide) -0.032 0.847-' 0.422 (0.1054) (0.2986) (0.6162) Price of maize 0.003- 0.002 0.009 (0.0009) (0.0027) (0.0058) Price of okra 0.006 -0.000 -0.031 (0.0045) (0.0180) (0.0404) Price of cassava -0.004 -0.005 0.022 (0.0031) (0.0084) (0.0178) Price of pepper 0.001 0.010* 0.004 (0.0016) (0.0056) (0.0129) Log business assets 0.013.. 0.000 0.169... (0,0042) (0.0134) (0.0272) Household maximum log (off-farm experience) 0.015 -0.165 1.993"' (0.0176) (0.0508) (0.1126) Area average off-farm wagea Type 1 3.526"* 5.582 3.877 (1.4029) (6.5144) (12.942) Type 2 -0.748" -0.435 0.177 (0.3704) (1.7914) (3.4869) Type 3 -1.386"- -1.388 -0.500 (0.5169) (2.4705) (4.8600) Type 4 -0.607'" -0.706 -0.125 (0.2079) (0.8647) (1.7211) Type 5 -0.011 -0.613... -0.426 (0.0529) (0.1336) (0.2756) Log of household size 0.323-' 0.009 -0.428' (0.0326) (0.1124) (0.2513) Household gender compositionb -0.044 0.259 0.155 (0.0707) (0.2463) (0.5068) Number of males 15-24 years old 0.010 -0.132 0.082 (0.0272) (0.0902) (0.1931) (Table continues on the following page.) 96 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Table 5. (continued) STLS OLS, Off-farm Variable total income Farm income income 25-34 years old 0.146-' -0.117 1.171- (0.0375) (0.1332) (0.2891) 35-44 years old 0.208't -0.203 0.310 (0.0379) (0.1547) (0.3374) 45-55 years old 0.146-' -0.275' 0.769-' (0.0402) (0.1658) (0.3635) Number of females 15-24 years old 0.102-" 0.099 0.518-' (0.0276) (0.0928) (0.1947) 25-34 years old 0.060 0.160 1.181- (0.0409) (0.1340) (0.2904) 35-44 years old 0.047 0.378'" 0.776' (0.0407) (0.1632) (0.3373) 45-55 years old 0.112-" 0.075 -0.572 (0.0357) (0.1490) (0.3341) Intercept 17.887'-' 2.542 0.559 (1.3403) (4.9031) (10.508) Adjusted R2 0.35 0.89 0.40 Number of observations 1,388 1,304 1,380 Note: The total income estimates are ordinary least squares (OLS), and the standard errors are Huber- corrected for heteroscedasticity. The farm and off-farm estimates are symmetricallly trimmed least squares (STLS), and the standard errors are bootstrapped. The bootstrapped estimates are generated by resampling the data 1,000 times. The bootstrapping procedure follows the classical assumptions that the independent variables are fixed, and it is the residuals that are resampled. The decision to include some of the vari- ables in log form and the others in levels was made on the basis of goodness of fit and variable significance. Standard errors are in parentheses. p-value is less than 0.1. p-value is less than 0.05. p-value is less than 0.01. a. The off-farm wages are regional averages of wages grouped into five occupational types. Type 1 occupations are the lowest-wage activities, and type 5 are the highest. The data on wages come from the half of the sample that was not administered the supplemental education module. b. Household gender composition is the household average value of the binary variable that takes the value of 1 for men and 2 for women. Source: Author's calculations. The results from estimating the impact of cognitive skills in the off-farm in- come functions are similar to the results from estimating total income. Namely, the test scores are jointly significant across all four specifications. Even the least- significant F-statistic strongly rejects the null hypothesis that test. scores have no effect. Two-Stage Results Up to this point no attempt has been made to determine whether the reduced- form estimates result from changes in farm and off-farm productivity or from changes in labor supply to these two activities. The positive effect of test scores Jolliffe 97 on off-farm income found in the reduced-form estimates could well be the result of increased productivity or increased effort in off-farm income activities (or both). Similarly the result that no returns are found in farm income could reflect off-setting productivity and labor supply effects. For example, increased test scores might improve farm productivity while decreasing labor supply to farm activities. To address this issue, farm and off-farm income are also estimated with labor supply explicitly included in the model and treated as an endogenous variable. For the two-stage model of farm profit, the sample is restricted to include only those households engaged in farming. Similarly, the off-farm income model with off-farm labor supply instrumented is also conditioned on the household engag- ing in off-farm activities. The farm labor supply function is instrumented using the determinants of off-farm income and the log of household size, while the off- farm labor supply function is instrumented by the determinants of farm income and the log of household size. The age and gender household composition vari- ables are used in both stages and can thereby affect income directly and indi- rectly through labor supply.4 It could be argued that the only valid variable to exclude from the second stage of estimation is household size. I do not follow this strategy both because the model implied by this argument is not identified and because there are other valid identifying instruments.5 For example, increases in off-farm wages may lead households to supply more labor to off-farm activities and thereby decrease farm income, but there is no compelling reason why off-farm wages directly affect farm income. Table 6 presents the results from these models. The estimates suggest that improved test scores have no joint effect on farm productivity and a positive effect on off-farm productivity. This provides evidence that the estimates from the unconditional, reduced-form off-farm model are due in part to increased productivity from improved cognitive skills. While the two-stage results for the off-farm function show that skills have a positive effect on off-farm income, the parameter estimates are about half the size of the unconditional estimates. This suggests that skills also increase income through either increased effort or an allocative effect. This result is similar to that of Yang (1997), who examines a sample of Chinese farm households and finds that schooling increases off-farm wages. He then argues that better-educated farm households allocate labor in response to increases in off-farm wages. 4. The farm and off-farm income models are also estimated with a set of the excluded variables included. The F-tests from these models fail to reject the null hypothesis that the excluded variables are jointly equal to 0. This implies that the over-identifying exclusion restrictions implied by the two-stage least squares model are valid. Similarly, F-tests of the joint significance of the excluded variables in the first-stage regression reject the null hypothesis that the excluded variables are jointly 0 in the determination of labor supply. This implies that the excluded variables are jointly significant determinants of farm and off-farm labor supply. 5. It is feasible to estimate this model for the farm income function, but not for the off-farm income function. Household size alone is not sufficient to identify off-farm labor supply. 98 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Table 6. Impact of Household Average Test Scores on Farm and Off-Farm Income Using Two-Stage Least Squares, Ghana, 1988-89 Variable Farm income Off-farm income Household average English score -0.011 0.013 (0.0075) (0.0106) Household average mathematics score 0.019" 0.032'* (0.0096) (C.0128) Log household farm labora 0.357'** (0.1357) Log acres of land farmed 0.526*' (0.0670) Household average log of farm experience 0.175* * (0.0612) Clusterlzone average day farm wage -0.022 (0.0442) Regional average Log (price fertilizer) -0.225 (0.3160) Log (price insecticide) -0.171 (0.1396) Price of maize 0.004 (0.0026) Price of okra 0.001 (0.0133) Price of cassava 0.008 (0.0088) Price of pepper 0.006 (0.0048) Log household off-farm laborb 0.916'** (0.1916) Log business assets 0.031" (0.0134) Household maximum log (off-farm experience) 0.042 (0.1063) Area average off-farm wagec Type 1 0.929 (2.2796) Type 2 0.155 (0.7379) Type 3 0.464 (0.9787) Type 4 0.422 (0.3825) Type 5 -0.041 (0.1031) Household gender compositiond 0.357'* -0.072 (0.1635) (0.2290) Number of males 15-24 years old -0.001 -0.076 (0.0590) (0.0922) Jolliffe 99 Table 6. (continued) Variable Farm income Off-farm income 25-34 years old 0.100 0.144 (0.0885) (0.1328) 35-44 years old 0.096 -0.088 (0.0908) (0.1552) 45-55 years old -0.079 0.122 Number of females (0.0965) (0.1754) 15-24 years old 0.102** 0.097 (0.0500) (0.0879) 25-34 years old 0.038 0.170 (0.0754) (0.1088) 35-44 years old 0.009 -0.009 (0.0957) (0.1475) 45-55 years old 0.000 -0.363'* (0.0863) (0.1733) Intercept 8.266'* 2.807'" (2.5S32) (0.8258) Adjusted R2 0.46 0.33 Number of observations 836 962 Note: The farm and off-farm parameter estimates are from two-stage least squares estimation with household labor instrumented. The standard errors are Huber-corrected, and household labor supply is treated as endogenous. The information on household age and gender composition is included in both stages because these characteristics may well affect productivity and thus income independently of their effect on labor supply. Only those households who generated positive farm income are included in the farm regression and similarly only those households with positive off-farm income are included in the off-farm regression. The decision to include some of the variables in log form and the others in levels was made on the basis of goodness of fit and variable significance. The standard errors are in parentheses. Standard errors are Huber-corrected for design effects. See Over, Jolliffe, and Foster (1996) for details. *p-value is less than 0.05. 4-p-value is less than 0.01. a. Household size and the off-farm variables are used to instrument household farm labor supply. b. Household size and the farm variables are used to instrument off-farm labor supply. c. The off-farm wages are regional averages of wages grouped into five occupational types. Type 1 occupations are the lowest-wage activities, and type 5 are the highest. The data on wages come from the half of the sample that was not administered the supplemental education module. d. Household gender composition is the household average value of the binary variable that takes the value of 1 for men and 2 for women. Source: Author's calculations. The reduced-form results presented in tables 4 and 5 differ from the two- stage results presented in table 6 for the model of farm profit. Although the two- stage least squares (2SLS) estimates of the farm profit model fail to reject the null hypothesis that the test scores jointly have no effect on farm profit, they do reject the null hypothesis that the mathematics score is not a significant determi- nant of farm profit. Conditional on selecting into farming activities, then, im- proved mathematics scores do seem to improve farm productivity. The section on reduced-form estimation emphasized the results from the joint tests of significance. This is partly because using semiparametric estimators and 100 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 robust estimates of the variance-covariance matrix results in inefficieni: estimates of the standard errors. Joint tests were necessary to draw any conclusions. The decision to present joint test results is also due to the high level of collinearity between the English and mathematics test scores and a suspicion that the data may not easily allow the English and mathematics effects to be disentangled. III. CONCLUSIONS The argument for using test scores in lieu of years of schooling to measure human capital is that years of schooling fail to capture any effect that school quality may have on the creation of human capital. This article shows that school quality does vary over Ghana and that the test scores reflect this variation. It also shows that the returns to skills, as measured in the total income models, are positive and significant. This result supplements the human capital literature because it makes clear that skills are rewarded and it provides evidence against the screening and credentialism theories of the returns to schooling. By estimating total income and its components, rather than focusing on just one source of income such as wage or farm income, the analysis makes state- ments both about the overall benefits of cognitive skills to households in Ghana and about the sources of income that are most affected by skills. Tlhis contrasts with the literature that focuses on specific types of workers in the economy (such as farmers or wage earners), so that the conclusions drawn are applicable only to workers in the same sector. The advantage of broader statements about the benefits accrued from skills is that they provide the policymaker with a clearer picture of the expected results from national education policies. The results presented here show that the returns to cognitive skills, as found in the total income estimates, are positive and statistically significant. Tests of the robustness of these results, using different household-level measures of skills and different estimation techniques, find that they are robust to the different measures of skills and over a wide class of non-normal error distributions. The decision to not select either farmers, wage earners, or self-employed fam- ily enterprises means that the analysis avoids explicitly modeling the sample selection process but must handle the resulting problem of numerous zero values when estimating farm and off-farm income. (All households that do not farm will have no farm income, and similarly all households that only farm will have no off-farm income.) The presence of sample-selection bias or censored depen- dent variables is a problem faced by most of the work discussed in the literature, but most of the papers address these problems by assuming that the underlying residuals are normally distributed. Yet, it is well known that the standard maximum-likelihood estimators for the censored model and for correcting sample- selection bias are inconsistent in the presence of heteroscedasticity as well as many other violations of normality (see Vijverberg 1987). It is now more widely acknowledged that most cross-sectional surveys are based on some sort of two- stage sample design, which typically means that the assumption of homo- Jolliffe 101 scedasticity is not tenable. For these reasons, this article uses Powell's CLAD and STLS estimators, which are robust to heteroscedasticity, for the estimation of farm and off-farm income. The estimates of farm and off-farm income suggest that none of the returns to skills is found in farm income and that only the off-farm income estimates pro- vide evidence of positive (and statistically significant) returns to skills. This re- sult is robust to the different measures of skills and over a wide class of non- normal error distributions. This result is somewhat tempered by noting that while the advantage of the STLS and CLAD estimators is their robustness to viola- tions of normality, their disadvantage is a loss of efficiency. Over most of the specifications for farm income, the returns to skills are positive but not signifi- cant. It is also tempered by noting the two-stage result that, conditional on se- lecting to engage in farm work, higher mathematics skills do seem to improve farm productivity. As a final comment, although the benefits of cognitive skills to Ghanaian households are not found in their effects on farm income, this does not mean that households engaged in farming do not benefit from improved cognitive skills. The GLSS data make clear that the typical household is engaged in numer- ous income-generating activities. Although a household's farm profitability might not improve, its total income will increase from the improved skills of house- hold members. APPENDIX. THE ESTIMATORS This article uses two estimators, both of which are consistent and asymptoti- cally normal for a wide class of symmetric error distributions with hetero- scedasticity of unknown form and a censored dependent variable. The first is Powell's (1986) symmetrically trimmed least squares (STLS) estimator. The sec- ond is Powell's (1984) censored least absolute deviations (CLAD) estimator. The STLS estimator is the f3 which minimizes: (A-1) I(xj'P > 0) [min( yi, 2xi3) - xip]2 where yi is the dependent variable, xi is the explanatory variable, and the indica- tor function, I, takes the value of 1 if the argument is true and 0 otherwise. This estimator is obtained by trimming the dependent variable and results in residu- als that are distributed over (-xif, xi,). The CLAD estimator is the j3 which minimizes: (A-2) I I yi - max(0, xi':) I. The consistency of this estimator rests on the fact that medians are preserved by monotone transformations of the data, and equation A-2 is a monotone trans- formation of the standard median regression that minimizes the absolute devia- tions. Koenker and Bassett (1978) discuss the properties of the median regres- sion or the least absolute deviations (LAD) estimator. 102 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. i The method of estimation used in this article for the CLAD estimator is Buchinsky's (1994) iterative linear programming algorithm (ILPA). The ILPA esti- mates a quantile regression for the full sample, then deletes the observations for which the predicted value of the dependent variable is less than 0. Another quantile regression is estimated on the new sample, and again negative predicted values are dropped. Buchinsky (1991) shows that if the process converges, thlen a local minimum is obtained. Convergence occurs when there are no negative predicted values in two consecutive iterations. All of the models estimated here converged, typically in fewer than 15 iterations. The STLS estimator used in this article is found by an iterative procedure simi- lar to the one used for the CLAD estimator. The first step to finding the STLS estimator is to use OLS estimates on the full sample. A new sample is created by dropping observations with negative predicted values and reassigning values to the dependent variable if it is larger than two times its predicted value. Another OLS regression is estimated on the new sample, and the process of trimming and reassigning values continues until convergence. The standard errors for the CLAD, STLS, and LAD estimators are bootstrapped by resampling the data 1,000 times. This bootstrap procedure results in stan- dard errors that are robust to violations of the assumption that the residuals are identically distributed. Breusch and Pagan's (1979) test statistics suggest that assuming identically distributed residuals is untenable and the standard errors need to be robust to violations of homoscedasticity. Similarly, the two- stage design of the sample is likely to result in rejection of the assumption that the residuals are independently distributed. An important caveat in in- terpreting the CLAD, STLS, and LAD estimates is that the bootstrap standard errors are not robust to violations of the independence assumption. This con- trasts with the OLS Huber-corrected standard errors, which are robust to vio- lations of both identically and independently distributed residuals. Jolliffe (1996) discusses a general bootstrap method for estimating standard errors that are robust to violations of the assumption of independently distributed residuals. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Alderman, Harold, Jere Behrman, David Ross, and Richard Sabot. 199 6. "The Returns to Endogenous Human Capital in Pakistan's Rural Wage Labour Market." Oxford Bulletin of Economics and Statistics 58(1):29-55. Arabmazar, Abbas, and Peter Schmidt. 1981. "Further Evidence o01 the Robustness of the Tobit Estimator to Heteroskedasticity." Journal of Econometrics 17(2):253- 58. Behrman, Jere, and Nancy Birdsall. 1983. "The Quality of Schooling: Quantity Alone Is Misleading." American Economic Review 73(5):928-46. Jolliffe 103 Blau, David. 1986. "Self-Employment, Earnings, and Mobility in Peninsular Malaysia." World Development 14(7):839-52. Boissiere, Maurice, John Knight, and Richard Sabot. 1985. "Earnings, Schooling, Abil- ity, and Cognitive Skills." American Economic Review 75(5):1017-30. Breusch, Trevor S., and Adrian R. Pagan. 1979. "A Simple Test for Heteroscedasticity and Random Coefficient Variation." Review of Economic Studies 47(5):1287-94. Buchinsky, Moshe. 1991. "The Theory and Practice of Quantile Regression." Ph.D. diss., Economics Department, Harvard University, Cambridge, Mass. Processed. .1994. "Changes in the U.S. Wage Structure 1963-1987: Application of Quantile Regression." Econometrica 62(2):405-59. Card, David, and Alan B. Krueger. 1992a. "School Quality and Black-White Relative Earnings: A Direct Assessment." Quarterly Journal of Economics 107(1):151-200. .1992b. "Does School Quality Matter? Returns to Education and the Characteris- tics of Public Schools in the United States." Journal of Political Economy 100(1):1-40. Foster, Andrew, and Mark Rosenzweig. 1996. "Technical Change and Human Capital Returns and Investments: Evidence from the Green Revolution." American Economic Review 86(4):931-53. Ghana Statistical Service. 1991a. Quarterly Digest of Statistics 9(4). Accra, Ghana. .1991b. "Statistical News Letter." A11/91. Accra, Ghana. Processed. Glewwe, Paul. Forthcoming. The Economics of School Quality Investments in Develop- ing Countries: An Empirical Study of Ghana. London: Macmillan Press. Glewwe, Paul, and Hanan Jacoby. 1994. "Student Achievement and Schooling Choice in Low-Income Countries." Journal of Human Resources 29(3):843-64. Glewwe, Paul, and Kwaku Twum-Baah. 1991. The Distribution of Welfare in Ghana in 1987-1988. Living Standards Measurement Study Working Paper 75. Washington, D.C.: World Bank. Grigg, David. 1991. The Transformation of Agriculture in the West. London: Basil Blackwell Press. Hanushek, Eric. 1995. "Interpreting Recent Research on Schooling in Developing Coun- tries." The World Bank Research Observer 10(2):227-46. Huber, Peter. 1967. "The Behavior of Maximum Likelihood Estimates under Non- standard Conditions." Proceedings of the Fifth Berkeley Symposium on Mathemati- cal Statistics and Probability 1:221-33. Jamison, Dean T., and Lawrence J. Lau. 1982. Farmer Education and Farm Efficiency. Washington, D.C.: World Bank. Jolliffe, Dean. 1996. "Schooling, Cognitive Skills, and Household Income: An Econo- metric Analysis using Data from Ghana." Ph.D. diss., Economics Department, Princeton University, Princeton, N.J. Processed. . Forthcoming. "The Impact of Cognitive Skills on Income from Farming." In Paul Glewwe, ed., The Economics of School Quality Investments in Developing Coun- tries: An Empirical Study of Ghana. London: Macmillan Press. Kish, Leslie. 1965. Survey Sampling. New York: John Wiley and Sons. Knight, John B., and Richard H. Sabot. 1987. "Educational Policy and Labour Produc- tivity: An Output Accounting Exercise." Economic Journal 97(385):199-214. Koenker, Roger, and Gilbert Bassett. 1978. "Regression Quantiles." Econometrica 46(1):33-50. 104 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Over, Mead, Dean Jolliffe, and Andrew Foster. 1996. "Huber Correction for Two-Stage Least Squares." Stata Technical Bulletin STB -29:SG46.29. Powell, James L. 1984. "Least Absolute Deviations Estimation for the Censored Regres- sion Model." Journal of Econometrics 25(3):303-25. . 1986. "Symmetrically Trimmed Least Squares Estimation for Tobit Models." Econometrica 54(6):1435-60. Psacharopoulos, George. 1985. "Returns to Education: A Further International Update and Implications." Journal of Human Resources 20(2):583-604. . 1994. "Returns to Investment in Education: A Global Update." World Devel- opment 22(9):1325-43. Rizzuto, Ronald, and Paul Wachtel. 1980. "Further Evidence on the Returns to School Quality." Journal of Human Resources 15(2):240-54. Scott, Chris, and Ben Amenuvegbe. 1989. Sample Designs for the Living Standards Sur- veys in Ghana and Mauritania. Living Standards Measurement Study Working Paper 49. Washington, D.C.: World Bank. Scott, Andrew J., and Tim Holt. 1982. "The Effect of Two-Stage Sampling on Ordinary Least Squares Methods." Journal of American Statistical Association 77(380):848- 54. Soon, Lee-Ying. 1987. "Self-Employment vs. Wage Employment: Estimation of Earn- ings Functions in LDCS." Economics of Education Review 6(2):81-89. Vijverberg, Wim. 1987. "Non-Normality as Distributional Misspecification in Single- Equation Limited Dependent Variable Models." Oxford Bulletin of Economics and Statistics 49(4):417-30. 1. 993. "Schooling, Skills, and Income from Non-Farm Self-Employment in Ghana." School of Social Sciences, University of Texas at Dallas. Processed. World Bank. 1995. World Development Report 1995: Workers in an Integrating World. New York: Oxford University Press. Yang, Dennis. 1997. "Education and Off-Farm Work." Economic Development and Cultural Change 45(3):613-32. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1: iOS-31 The Tragedy of the Commons in Cote d'Ivoire Agriculture: Empirical Evidence and Implications for Evaluating Trade Policies Ram6n L6pez Expansion of cultivated land diminishes the extent of forestlands or reduces the length of fallow periods and, hence, reduces the amount of natural vegetation. The increase in land under cultivation has a direct output-increasing effect at the cost of reducing natural capital and agricultural productivity. The evidence for western C6te d'lvoire is consistent with, and provides an explanation for, the declining agricultural productiv- ity observed in Sub-Saharan Africa during the past few decades. This article uses a theoretical model to determine the level of land cultivation that maximizes village income, using data from C6te d'Ivoire for 1985-87. An important part of the land is under common property, usually at the village level. The results show that farmers do not internalize even a small fraction of the external cost of bio- mass in their land allocation decisions. The lack of internalization of the social cost of the biomass resource leads to large income losses at the village level-as much as 14 percent of village income. These losses are many times larger than the usual estimates for conventional distortions. Natural vegetation represents an important factor of production in the context of traditional shifting cultivation. Overexploitation of this factor may cause sig- nificant loss of income among rural communities. Farmers in an area of western CUte d'Ivoire have overexploited the natural resources (forests and natural veg- etation in fallows) through excessive cultivation of communal lands and, hence, the reduction of fallows and forest areas (L6pez 1993). Individual cultivators considered at most 30 percent-and, more likely, a negligible fraction-of the social cost of natural vegetation or biomass when deciding how much land to clear for cultivation. These social costs include the negative effects that the clear- ing of land by one farmer has on other farmers, through soil degradation, flood- ing, and sliding, the reduction of fallow and forest areas as well as the decline in natural soil fertility caused by a reduction in the fallow part of the cultivation cycle. Ram6n L6pez is with the Department of Agricultural and Resource Economics at the University of Maryland. Financial support for this research was provided by the World Bank Research Committee, RPO 675-33. Research assistance was provided by Claudia Septilveda. The author would like to thank three anonymous referees for their helpful comments. C) 1998 The International Bank for Reconstruction and Development / THE WORLD BANK 105 106 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Rural communities have apparently failed to maintain a system of incentives and controls over individual cultivators that would induce a socially optimal allocation of land among forest, fallow, and cultivation and thus avoid the "trag- edy of the commons." These results give support to authors who have ques- tioned the effectiveness of indigenous forms of property in achieving a socially efficient allocation of natural resources (Perrings 1989; Sinn 1988; L6pez and Niklitschek 1991; Glantz 1977; and Allen 1985). In contrast, anthropologists, other social scientists, and several economists, most prominently Dasgupta and Maler (1990) and Larson and Bromley (1990), are of the opinion that commu- nities are able to develop controls on the use of common property resources to allow for their efficient exploitation. The evidence for western C6te d'Ivoire is consistent with and provides an explanation for the declining agricultural productivity observed in Sub-Saharan Africa over the past few decades (FAO 1986 and ADB, ECA, and OAU 1984). Sev- eral studies have illustrated the connection between declining agricultural pro- ductivity and agricultural intensification, which, in turn, is associated with popu- lation growth and the fall in biomass. A process that Geertz (1963) calls "agriculture involution" is documented, for example, by Jones and Egli (1984) in areas of fast population growth in Burundi, Rwanda, and Zaire; by Ludwig (1968) for the Ukare Island in Lake Victoria; and by Lagemann (I 977) for east- ern Nigeria. Niklitschek (1990) provides ample evidence of the close connection between rural population growth, expansion of area cultivated, reduction of fallows, and declining agricultural productivity in Sub-Saharan Africa. Do the results for western Cote d'Ivoire hold for other tropical areas where farmers practice traditional agriculture in the context of communal lands with restricted access? Using data for all regions of C6te d'Ivoire, this article provides further evidence on the efficiency of exploitation of common lands. Apart from providing results that are more representative for the country, this expanded data set also provides the degrees of freedom required for a substantially more disaggregated analysis. In particular, the analysis considers three agricultural outputs: tree crops (a capital-intensive output), cereals (land-intensive), and tu- bers and vegetables (labor-intensive). Using three outputs permits a much richer analysis of the effects of agricultural prices as well as certain economywide poli- cies than using an aggregate output. The article also evaluates agricultural price and trade policies in Cote d'Ivoire. In particular, it looks at how price and trade liberalization affect agriculture, taking into account their effects on both natural resources (fallow reduction and deforestation) and conventional factors of production (labor, purchased inputs). Agricultural price and trade policies do not represent first-best solutions to even- tual overexploitation of natural biomass. However, institutional and sociopolitical limitations make it unlikely that poor developing countries will be able to imple- ment such policies effectively in the near future. Hence, in devising further price and trade policy reforms, analysts should verify whether such policies impose additional tradeoffs by magnifying the overexploitation of natural resources. Lopez 107 Section I provides background information regarding the role of natural veg- etation or biomass in agricultural production in tropical areas. Section II pre- sents the theoretical framework used in the analysis, while section III presents details about the empirical method. Section IV discusses the data and estimation techniques, and section V presents the results. Section VI discusses the implica- tions of agricultural price and other economywide policies. Section VII concludes. I. NATURAL VEGETATION AS A FACTOR OF AGRICULTURAL PRODUCTION A system of relatively long rotation between crop cultivation and fallows (shift- ing cultivation) is a dominant practice in Cote d'Ivoire as well as in most other countries in Sub-Saharan Africa where population levels are not yet too large. The fallow period plays the important role of replenishing the fertility of the land by allowing natural vegetation to grow. The natural vegetation is incorpo- rated into the soil as natural fertility (usually as ashes after burning) at the time of cultivation. Fallow periods that are too short lead to insufficient growth of natural vegetation and consequently to low soil fertility, soil instability because the vegetation does not have a sufficiently strong root system to protect it at the time of cultivation, and cultivated areas that are not protected against flooding and sliding. Thus, the vegetation that is allowed to grow in the fallow period is a form of capital that accumulates and is eventually used at the time of cultiva- tion. The closed forest areas (that is, those areas not yet disturbed by cultiva- tion) also play important roles not only as a reserve of land for eventual use but also as protection against soil erosion, watershed destruction, and flooding. Expansion of the cultivated land diminishes the forestlands or reduces the length of the fallow periods and, hence, reduces the natural vegetation. The tradeoff between cultivated land and forest/fallow is clear: an increase in land under cultivation has a direct output-increasing effect at the cost of reducing the natural capital, thus reducing agricultural productivity. An optimal fraction of the land should be cultivated in order to maximize social income. If the level of land cultivated is above or below the optimum, income is reduced. The land is under common property (farmers have exclusive rights on the land usually for as long as they cultivate it), usually at the village level (that typically encom- passes hundreds of families). In the absence of communal controls, individual cultivators are likely to overexploit the natural resource by cultivating too much. In deciding how much land to cultivate, farmers in this case likely consider only the private costs, ignoring contemporary and intertemporal effects on other cul- tivators. If communal controls are adequate, individuals would behave as if they fully accounted for both the private and social costs of clearing land. The em- pirical model tries to elucidate whether this happens. There are of course substitutes for natural vegetation as a factor of produc- tion. The use of fertilizers could replenish the fertility lost as a result of shorter fallow periods. The construction of drainage and other infrastructure could re- 108 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 place part of the protective role of natural vegetation. These substitutes, how- ever, are likely to be imperfect particularly in humid tropical soils where the benefits of fertilizers and other chemicals are much reduced (Sanchez 1976). Thus, the availability of substitutes may allow for a reduction in natural vegeta- tion, but elimination of the natural vegetation is unlikely to be optimal. Further, in the absence of community controls on the use of natural resources, individual cultivators will not be prone to use usually expensive substitute inputs for natu- ral vegetation (and thus increase land productivity) as long as forest or fallow communal lands are available for clearing. Consistent with this is the fact that only a very small fraction of the farmers located in areas where fallow lands and forest still exist use chemical fertilizers. II. THE THEORETICAL MODEL Equation 1 defines the net revenue function for farmer j who is assumed to take the stock of natural vegetation as given, (1) Rk(w,x',O,p,k') = max [fiyi - -vLi':F(y',LP,O,k',xi) = O] Y' ,L1 where p is a vector of output prices; w is a vector of input prices other than biomass, land, and capital; xi is land cultivated by farmer j; 0 is the stock of village-level biomass per hectare; k' is capital; yi is a vector of output quantities produced by farmer j; Li is labor used by farmer j; and F(.) is the set of produc- tion possibilities. The function RI(-) satisfies all the conditions of a variable profit function in p and w (Diewert 1973) and is assumed to be increasing and concave in 0, x', and k'. The stock of biomass H per hectare can be defined as (2) H r- r(1- x/x) where r9 is the (average) density of biomass per acre in the land that is not culti- N vated, xV is the total land available, and x = E; xi is the total land under cultiva- i=h tion by the community or village (N is the number of cultivators in the village). L6pez (1993) has shown that the rate of extraction of biomass under shifting cultivation is proportional to the rate of land cultivation x/x-. Thus, if y is the natural increase of biomass in the areas not under cultivation, change in bio- mass density in the uncultivated land is (3) n = Y - YXI/ Y. The level of land cultivation that maximizes the income of the village (rather than the income of individual cultivators independently) is obta ined by Lopez 109 (4) max f , R'(iv;x' 7(1 - x x'x), p, ki) - cxi }erdt N subject to 7 - T I xil3- i=l 71(°) = 'lo, Ixi < j. where c is the private cost of land clearing per unit of land, r is the discount rate, t is time, and a dot over the variable name reflects a change over time. The optimization (expression 4) is a benchmark for comparing actual alloca- tions. The solution of expression 4 defines the allocation that is socially optimal from the community or village point of view. It maximizes the wealth of the community. The model does not assume this optimization; expression 4 is pro- vided only for the purpose of comparing its solution with the observed alloca- tions. The actual decentralized allocations by individual cultivators may not be consistent with expression 4, depending on the institutional conditions prevail- ing, the level of monitoring, and transaction costs. See, for example, Baland and Platteau (1996) for the nature of the games that could lead to efficient allocation of common resources through collective action and other forms of cooperation. In the steady state, the first-order conditions of expression 4 imply that -. ~~~~~~~iN _i (S) R'(.)= c+ R3(.) +,ulY, j=..N X i=l where Rk2 -- and u = (1- x / ) 3 _ is the shadow value of biomass Ax] i=lr+xIx desity il. Equation 5 is the key benchmark for deriving the empirical methodol- ogy. It says that if cultivators choose socially optimal levels of land cultiva- tion, the marginal revenue of the last unit of land cleared for cultivation should equal the private cost of land clearing (c) plus the marginal revenues forgone by all cultivators that one additional unit of land cleared causes now Q1 K, 3 and in the future (rl (uIx)) due to the fall in rl that increasing xi causes. The second right-hand term in equation 5 measures the forgone marginal income for all cultivators caused by reducing the area covered with natural vegetation given a level of biomass density (ril) per acre of noncultivated land. The third right- hand term in equation S captures the present value of all future forgone incomes due to the fall in rq that will occur in the future as the rotation period is reduced as a consequence of increasing xi. 110 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Using the definition of u, the last two right-hand terms in equation 5 can be combined, and thus equation 5 can be rewritten as (6) R2(') = Cz1,...,N r + z+ N where z = x'/x is the (endogenous) rate of depreciation of the stock of bio- mass density r. If individual cultivators make decisions completely independently and face no constraint from their communities, they would only consider I/Nth of the external effects (of the second right-hand term in equation 6) as part of the cost of increasing xi. In this case, a Nash equilibrium would arise as the solution of the N equations: ( 7) R2' ( = C + _- E- R'3 ( ,i .. Nx r+zi=r Depending on the effectiveness of the village controls on individual cultivators' decisions, the equilibrium vector of x = (x', x2,.. , xN) will be given by equa- tion 6, or by equation 7, or an intermediate solution if controls exist but are imperfect. To model these intermediate outcomes empirically, I postulate a more general specification for equation 6 or equation 7: (8) R2 (.) = c + --I RI(.),..,N x r+zi~=i where X is an efficiency parameter that could fluctuate between Il/N and 1, de- pending on the efficiency of the village controls. The closer to 1 is parameter X, the more efficient are the land allocation decisions. The specification of the model assumes a fixed stock of physical capital, k'. It may be argued that this is inconsistent with the dynamic nature of the model and, therefore, that a mechanism of capital accumulation needs to be incorpo- rated explicitly into the theoretical model. There are several ways of doing this. One possibility is to assume the extreme opposite of the assumption made so far, that capital is instantaneously adjusted to optimal levels instead of being fixed. In this case, it should be clear that the land allocation equation (equation 7 or 8) remains intact, except that the revenue function Ri(.) would include the rental price of capital instead of the stock of capital. A less extreme possibility is that capital is neither fixed nor fully variable, but quasi-fixed. That is, a gradual process of capital accumulation takes place through time. To model this, there are several avenues depending on what source pre- vents an instantaneous adjustment of the capital stock. Appendix A presents a Lopez 111 growth model that explicitly incorporates the two state variables, the stock of biomass and the stock of capital. Except for the addition of the biomass state variable, the model is a conventional neoclassical growth model. That is, it as- sumes that the main reason for slow capital adjustment is the existence of a utility function that is strictly concave in consumption and that capital accumu- lation is financed entirely through the savings of the individuals. Appendix A shows that in the steady state the land allocation decision is still ruled by a specification analogous to equation 7. That is, at least in the steady state, the specification used is in a sense quite robust to the assumption used regarding fixity, quasi-fixity, or full variability of the capital variable. The empirical model is based on the assumption of a steady state. Dropping this assumption would change the specification of the land equation. The right- hand side of the land equation would be the same except that a new term involv- ing the change in biomass, 6, would have to be added (L6pez 1993). If 6 < 0, the true shadow value of biomass is underestimated in a steady-state specification; if 6 > 0, the model would overestimate the shadow value of biomass (L6pez 1993). Because biomass decreases through time, the steady-state specification under- estimates the social value of biomass. That is, the social cost of expanding x exceeds the value assumed in equation 6. This implies that the steady-state assumption will provide a value for X that is larger than the true value of A. So the estimated X will be an upper bound of the true X if biomass declines through time. A potential limitation of the model is that it assumes that producers are risk neutral. Poor farmers tend to be risk averse and, hence, if cultivated land is risk- decreasing, farmers may behave as if the true marginal cost of cultivated land were less than the right-hand side of equation 8. It may be argued, then, that the model might estimate a X value below unity because a misspecification is associ- ated with ignoring risk rather than because institutions fail to assign land effi- ciently. However, two aspects considerably weaken this objection. First, the spe- cific variable that dictates the value of X. is the shadow value of biomass for the community. If the omitted variable, risk level, causes a downward bias in X, the level of risk and the shadow value of biomass would be positively correlated. It is hardly plausible, however, to argue that the shadow value of biomass for a community is correlated with the level of risk or uncertainty prevailing in the community. Second, the empirical model in section III does not use actual values for the private costs of cultivating land, c. Instead it uses an estimate of the private cost coefficient that can change across villages (these are the village fixed effects). So, if the level of risk or uncertainty faced by the farmers varies more across villages (due to climatic and agroecological differences or to proximity to markets) than through time, as might be expected, its effect would be captured by the village fixed effects without biasing the estimates of B. Another potential problem of the model is that it ignores the existence of credit market imperfections. Some analysts have argued that imperfect capital markets and low incomes drive farmers to behave myopically. One way of con- sidering capital market imperfections is through the interest rate. Basu (1989), 112 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 for example, argues that most farmers in developing countries have access to the informal credit market, where the main mechanism of allocation is (usually high) interest rates rather than credit rationing. If farmers face credit constraints in the formal capital market, they have to rely on informal credit sources, where inter- est rates are much higher than interest rates in the formal market. Although the model obviously ignores credit market imperfections, it does not rely on the existence of perfect credit markets. In fact, it explicitly allows the discount rate to assume various values in the empirical estimation, within a wide range. The analysis does not use market interest rates as a proxy for r, but rather experi- ments with a broad range of values for r and considers the sensitivity of the estimates of X to changes in r. There is still a potential problem if the levels of r vary across farmers rather than being constant, as implicitly assumed in equation 8. The problem in this case could be that the omitted variable rj is correlated with the level of the shadow value of biomass, which is the variable that determines X. However, the shadow value of biomass is village-specific, not farmer-specific. Hence, village fixed ef- fects will control for the effect of the possible intervillage variations in the inter- est rate. The only remaining problem would be if the interest rate varies system- atically through time. {II. THE EMPIRICAL MODEL The empirical model requires the specification of the normalized net revenue function, which is assumed to be identical for all farmers. The revenue function is normalized by the farm price of tree crops. Equation 9 is a normalized qua- dratic functional form for the revenue function: (9) Ri =Ao + Aw + A2xi + A3H + A4ki + Asp, + A6Po + 2 A11w2 + 2 A22X'2 + X A3 02 + X A44kj2 + X2 A5p2 + 2 A66p2 + A12wx' + A13wO + A14wk1 + A15wp, + A16WpO + A23x'0 + A24k1X + A25x'p, + A26X'PO + A340kl + A35ep, + A36P0O + A4.5k'pc + A46k'po + IBkHk + Y-YVh+E k k where RK = R'/q is the net revenue normalized by the price of tree crops (q); w - iivlq; PC and po are the prices of cereals and other crops (tubers, vegetables), re- spectively, also normalized by q; Hk are household-specific characteristics; Vb are village characteristics; Ai, Aqj, Bk, and yh are fixed parameters; and es is an error term assumed to be normally distributed with zero mean and finite vari- ance. The village characteristics also interact with the prices and wage rate. These interactive terms are omitted in equation 9. The well-known symmetry or reci- Lopez 113 procity conditions that arise from the implicit optimization assumption in the context of well-behaved production technologies are imposed on equation 9. The symmetry conditions imply that Aji = A1i for all i j. This considerably reduces the number of parameters to be estimated. Using equation 8, an explicit functional representation for farmer j's demand for land to cultivate is (10) Xi = c-A2 _A12 W_-_A23 0 __A24k _ A25 p _ A26 p A22 A22 A22 A22 A22 A22 + ; 9 (1 +r) N [A3 + A13w+ A230+ A34 ki + A35ps + A36po] +s A22 r +z where E0 is an additive disturbance assumed to satisfy the same properties as £R, and x - xi is the total area cultivated in the village. The function R in equation 9 is assumed to satisfy the usual regularity assumptions: it is convex in p and w; it is increasing and strictly concave in xi, 0, and ki; and the marginal revenues of 0, k', and x are increasing in the level of the other factors. These conditions imply that A22, A33, and A44 are all negative and that A23, A24, and A34 are all positive. These conditions assure that cultivated land is decreasing in c and in- creasing in 0 and ki. Using Hotelling's lemma, the labor demand and output supply equations im- plicit in equation 9 are -Li = Al + A1lw + A12x' + A130 +A14k' + A1lsp + A16Po + £j , (11) Q'i = A5 + Alsw + A25xi + A350+ A45k' + A55pc + A56Po + j,c 00 Qb = A6 + A16w + A26X' + A360 + A46k' + A56P, + A66Po + j where Li is labor demand by farmer j, and Qi and Q are the supply of cereals and other crops, respectively. Equation 11 omits the household-specific charac- teristic effects and the village characteristic effects. The supply of perennial crops (or tree crops) is obtained by noting that Qi = Ri - wLi - PCQi - poQo. Thus, ( 12) QP=A0+ A2x' +A30+ A4ki _ 2 Al w2- 2 A5sp2 - Y2A66p + 2A22X'2 + 34 A3302 +YXA + E B,-IB + X y TV6 + k k 6 The system of equations 10 and 11 and either equation 9 or 12 is jointly estimated using a maximum likelihood procedure. The joint estimation of the system permits the identification of all the relevant coefficients. The X coefficient can be identified conditional on a fixed level of the discount rate r. The estima- tion strategy consists of estimating alternative values of X for a plausible range of discount rates. Given the symmetry conditions (that is, Aii = A1i), the coeffi- 114 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I cients inside the square brackets in equation 10 (A3 and A,3, i = 1,.. ., 6) are all identified by the rest of the system or by the linear part of equation 10 itself. Hence, as q and z are observed variables, X can be exactly identified if r is also known. But since r is unknown, alternative fixed values are postulated for r. The value of k will, of course, vary directly with the postulated level of r. This re- flects the fact that any land allocation can be considered optimal if the discount rate is large enough. Thus values for X are estimated within a plausible range of r. The key question is how high the discount rate r would have to be to make the observed resource allocation socially efficient. The normalized quadratic is a flexible functional form, in the sense that it provides a second-order approximation to any functional form. This is one of the functional forms that Diewert (1973) proposes for revenue flnctions and has been frequently used in the literature (Binswanger and Evenson 1984 and Diewert and Morrison 1988). This form imposes little a priori restrictions on the matrix of elasticity of substitution and on the implicit production technolo- gies. Unlike other commonly used functional forms (such as Cobb-Douglas and Constant Elasticity of Substitution), the normalized quadratic allows for differ- ent elasticities of substitution, it does not impose separability, and it allows for nonhomotheticity.' The function does impose the price homogeneity conditions that production theory predicts for input demand and output supply equations.2 IV. DATA AND ESTIMATION The data come from three sources-the Living Standards Survey (LSS) con- ducted annually in C6te d'Ivoire between 1985 and 1988, remote sensing data based on satellite images of 20 villages scattered throughout the country for the years 1985 to 1988, and field visits to the villages. The total area covered by the remote sensing analysis is about 450,000 hectares. The LSS and remote sensing data sources were carefully matched for each of the villages and years consid- ered. The LSS information allows the creation of a panel data set for a sample of 16 farm households in each of the 20 villages for the 1985-87 period. The LSS data used in this article concern information on farm production, labor hired, household member's work on the farm, land cultivated, other inputs used by each of the farm households, as well as demographic characteristics. The analy- sis also uses village-level information such as the number of households per vil- lage. (See appendix B for data definitions.) The remote sensing data provided information about the total land under cultivation in the village, land under fallow per village, land under closed forests 1. Some of the flexible functional forms do impose quasi-homotheticity and certain separability conditions (see L6pez 1985). 2. Some of the implicit input demand and output supply equations are linear in the normalized prices, and one of the output supply functions is nonlinear in prices. As Diewer: (1973) shows, this feature imposes no special peculiarities on the underlying production technology. Lopez 115 per village, and average biomass density of the fallow land in each of the three years considered. Village biomass is estimated by multiplying the total land area under fallow by the biomass density index obtained from the remote sensing analysis. Also included is an estimate of biomass calculated in the same way for forest areas that have not been cultivated recently. Some of the regressions only use the fallow biomass, others use the sum of fallow and forest biomass, and still others use the two measures as separate variables. Preliminary regressions showed that the fallow biomass was the key to agricultural productivity and that includ- ing forest biomass added very little to the regressions. Therefore, the regressions reported in this article use only fallow biomass per hectare. The matching of the LSS household-level data and remote sensing village-level data provided a panel data set that combines individual household information with information on natural resource endowments at the village level. Field visits to the villages provided data concerning the agricultural area that is under the sphere of influence of each village, that is, the area that is considered "property" of the village group. Moreover, the field visits provided certain es- sential qualitative information on issues such as land allocation decisions, the state of the communal system, and the functioning of labor markets within the communities. Three main conclusions that emerged from the field visits are im- portant for the empirical analysis. First, shifting cultivation is the dominant prac- tice in all the villages considered. Second, the system of land allocation for culti- vation of the fallow areas is not transparent. In some villages, a village council is still quite important in determining the extent of land given to cultivation. In most villages, however, the decisions appear to be much more decentralized, with little input from the village chiefs. Additionally, in several areas an incipi- ent and spontaneous process of privatization is taking place, and one of the mechanisms for obtaining land rights is to cultivate the fallow lands continu- ously. In general, the assessment of people in the field is that the traditional system of land allocation is changing mainly because of the reduced power of the village leaders and the process of gradual privatization. Third, in most cases, an active and seemingly efficient labor market operates within the villages. Table 1 gives an overview of the land allocation information provided by the remote sensing analysis for C6te d'Ivoire as a whole. Table 1 shows, even at the aggregate country level, a significant degree of change in land allocation over the three-year period from January 1985 to January 1988. In particular, a decrease in the forest area of almost 9 percent and an increase in the area cultivated of about 2.5 percent a year are quite remarkable. At the same time the fallow area shows a less conspicuous but significant downward trend. The village-level data that are actually used in the regression analysis exhibit an even greater variability through time than the aggregate data for the country. The remote sensing data also include a measure of the average biomass density per hectare in the fallow land. In order to make intervillage comparisons of fallow meaningful (a village may have a large fallow area but very little vegeta- tion density, and another may have a smaller area but much greater vegetation 116 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Table 1. Changes in Land Allocation in C6te d'Ivoire, 1985 and 1988 (hectares) Percentage Land area 1985 1988 change Total area surveyed 447,132 447,132 0.0 Under closed forest 125,410 114,217 -8.9 Under fallow 172,391 169,585 -1.6 Cultivated 132,222 142,073 7.5 Source: EARTHSAT (1991). cover), the fallow area of each village is weighted by its corresponding average biomass density. Two alternative definitions of biomass are used in the empirical model. A narrow definition only includes average biomass per hectare in fallow lands. A broad definition includes all biomass in fallow lands as well as in forestlands (biomass is, in this case, normalized by the total fallow plus forest areas). The two measures turn out to be highly correlated, and the econometric results are only marginally different when either measure is used. Equations 9, 10, and 11 are estimated jointly as a system restricting the Ai and A i coefficient across equations using FIML (full information maximum like- lihood) estimators. Assuming that the variables N, w, 0, and £xi are not corre- lated with the error terms, the system is recursive (rather than siraultaneous), a feature that greatly facilitates the estimation. Furthermore, a random effects model is used, allowing for a component of each error term that is comrnon to a given year's observations corresponding to households located in the same village. That is, if Es5Z is the total residual in equation s for household i located in village j at time t, assume that es, = Pst + AS-, where ps is the common-to-the-village component and A' is the idiosyncratic component. E(p10A#,i) = 0 is assumed for all k. The variance of the across-village estimated component (ps) is signifi- cantly larger than the variance of the idiosyncratic component in all three equa- tions. This means that imposing a single residual would give inconsistent esti- mates of the standard errors of the regression. Another important issue is the validity of the assumption that the residuals are uncorrelated with the exogenous variables. In particular, of most concern is the possibility that the biomass variable could be correlated with the error terms. This would, of course, lead to inconsistent estimates of the parameters. The endogenous variables (household revenue, cultivated area, labor demand, and the supply of the various crops) are explained by village-level variables as well as by household characteristics. A problem of simultaneity may arise if the village explanatory variables and the household dependent variables are correlated with important but unobserved variables. The key question concerns the source of variability of the village-level variables. A positive correlation between farm revenues or output and biomass may arise if some villages have greater biomass than others because they have better Lopez 117 soil and climate that allow the natural vegetation to grow faster. These better climatic and soil characteristics would lead to higher agricultural productivity in the villages that have more biomass. Thus any positive correlation between out- put and biomass could be spurious. In order to deal with this issue, village- specific characteristics (including soil and climate) are controlled for using vil- lage dummy variables. Using village effects should control for unobserved variables such as climate and soil quality as well as for other village characteristics such as infrastructure and distance to markets. Thus, the use of village dummies is likely to eliminate the influence of such variables from the coefficients associated with the other explanatory variables. The use of village-specific dummies (rather than house- hold dummies) is an adequate procedure because the biomass variable used is also defined at the village level. This procedure is possible because the analysis uses panel data for three years for all the relevant variables. A remaining problem, however, is that the variation of biomass through time in a village may be associated with changes in weather. Agricultural productiv- ity is also likely to be associated with fluctuations in weather, and, hence, a spurious correlation between biomass and agricultural productivity may arise. The extent of the biases induced in the coefficient of biomass will depend on the importance of the fluctuations in weather. Other estimates (not reported here) are obtained using instrumental variables for the various terms involving the biomass variable. These include average family size as well as two regional dummy variables. The results, particularly the sign and statistical significance of the co- efficients, do not change with the instrumental variable procedures. These vari- ables are not likely to be correlated with the weather variable but are good instruments, in the sense that they explain a high proportion of the changes in biomass. There are large differences in agroecological conditions across the main re- gions in the country (West Forest, East Forest, and Savanna). In particular, the inland Savanna has a different pattern of production than the tropical forest belt in the west and in the east. Thus, using different village intercepts in the estimation of the revenue function may not be enough to capture all the pro- duction and ecological differences across the three regions. This would call for allowing the slope coefficients to vary across regions as well. Unfortunately, given the large number of coefficients of the revenue function, this is not fea- sible. However, using a more restricted approach with one aggregate output variable, L6pez (1993) estimates a similar revenue function for the western region of C6te d'Ivoire, which has enough village observations to estimate the revenue function. The effect of biomass on gross revenues is indeed higher in the west than in the country as a whole, but the differences are not great. For the west, the estimates for the contribution of biomass are 0.18-0.20, only slightly larger than those for the country as a whole. The contribution of land to gross revenues is slightly larger in the forest region (with an elasticity of about 0.41 compared with 0.38 in the analysis for the whole country), but the 118 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I contribution of labor is significantly smaller in the west (about 0.10 compared with 0.18 for the whole country). V. THE RESULTS Table 2 shows the FIML estimates of the system of equations 9, 10, and 11. The estimates were obtained assuming a discount rate of 0.2. Several other esti- mates were obtained by specifying different discount rates. In general, the re- sults are robust to the changes in the assumed discount rate, and only the k coefficient changes numerical values, but without altering its sign and degree of statistical significance. In table 2 the estimates in the first and second columns, respectively, were obtained with and without using village dummies. The estimates are mostly (but not completely) consistent with the properties of the revenue function postu- lated by the optimization model. In particular, the revenue function estimated is monotonically increasing in land cultivated, biomass, and capital as well as in the output prices and is monotonically decreasing in the wage rate. The normal- ized revenue function estimated is not, however, convex in prices. Although the All and A66 coefficients have the expected signs, the fact that A155 is negative implies a violation of the convexity property (although this coefficient is not statistically significantly different from 0). The revenue function estimated is also strictly concave in land cultivated, biomass, and capital, implying that the marginal revenue functions are downward-sloping as expected. The most important finding concerns the resource efficiency parameter X. As can be seen in table 2, the parameter X is positive but not statistically different from 0 in either of the estimates presented. The model was also estimated as- suming even higher discount rates (up to r = 0.6), and the X parameter is never statistically significant.3 This result implies that farmers do not internalize even a small fraction of the external cost of biomass in their land allocation decisions.4 Thus, farmers are cultivating too much land to maximize the village income. Or, equivalently, given the tradeoffs between land cultivated and the stock of biomass, the village wealth can be increased by reducing the amount of land cultivated and, conse- quently, allowing for more natural biomass. Moreover, the income loss associ- ated with this inefficiency is quite large. Several reasons might explain why such a clear inefficiency occurs in a situation where village farmers exploit the biomass resource in an essentially 3. In the L6pez ( 1993i study, a statistically significant K was found for the west of C6te d'lvoire at least under certain specifications; the highest X found was about 0.3. 4. According to the theoretical specification, the smallest possible value of X is IIN rather than 0. Because the number of households in most villages is very large, fluctuating beeween 742 and 1,320 with a mean of 937, the lower bound of X is very close to 0, at about 0.001. L6pez 119 Table 2. Estimation Results for the Revenue Function, Cote d'Ivoire, 1985-87 Using village Without using village Variable dummy variables dummy variables Constant -2.52 -2.97 (-2.85) (-3.44) Agricultural wage, w -1.09 -1.08 (-6.09) (-6.02) Land cultivated by farmer j, xi -1.39 -1.31 (-1.71) (-1.60) Stock of village biomass, 0 1.40 1.36 (7.75) (8.24) Capital owned by farmer j, ki 2.34 2.31 (4.54) (4.49) Price of cereals, p, 0.95 0.82 (2.18) (1.88) Price of other crops, p, 0.02 -0.05 (0.05) (-0.11) Agricultural wage squared, w2 0.11 0.11 (2.05) (1.95) Land cultivated by farmer j squared, (x')2 -1.22 -1,26 (-15.82) (-16.48) Stock of village biomass squared, 02 -0.20 -0.19 (-7.52) (-7.97) Capital owned by farmer j squared, (kl)2 -0.33 -0.33 (-1.38) (-1.37) Price of cereals squared, (p )2 -0.38 -0.47 (-1.24) (-1.53) Price of other crops squared, (p0)2 0.38 0.38 (0.97) (0.95) Interaction with agricultural wage Land cultivated by farmer j, wx' 0.00 -0.008 (0.00) (-0.08) Stock of village biomass, wO -0.003 -0.0022 (-0.33) (-0.28) Capital owned by farmer j, wki -0.10 -0.10 (-2.90) (-2.77) Price of cereals, wp, -0.30 -0.30 (-3.30) (-3.30) Price of other crops, wp0 -0.18 -0.20 (-1.83) (-1.87) Interaction with land cultivated by farmer j Stock of village biomass, xiO 0.04 0.05 (1.23) (1.37) Capital owned by farmer j, kix' -0.66 -0.67 (-10.44) (-10.54) Price of cereals, xip, 0.81 0.84 (3.55) (3.61) Price of other crops, x'p0 0.69 0.72 (2.56) (2.64) (Table continues on the following page.) 120 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Table 2. (continued) Using village Without using village Variable dummy variables dummy variables Interaction with stock of village biomass Capital owned by farmer j, Ok' 0.03 0.04 (1.06) (1.08) Price of cereals, Op, -0.15 -0.14 (-7.23) (-6.S 1) Price of other crops, Op. -0.06 -0.05 (-3.72) (-3.33) Interaction with capital owned by farmer j Price of cereals, kip, 0.36 0.:36 (4.12) (4.11) Price of other crops, kip, 0.15 0.15 (1.78) (1.69) Interaction between price of cereals and 0.67 0.65 price of other crops (2.81) (2.71) Biomass efficiency parameter, k 0.03 0.008 (0.30) (0.07) Private cost of land clearing, c 1.39 1.55 (2.13) (2.46) Dummy year 1986 -1.92 -2.57 (-3.19) (-5.47) Dummy year 1987 -1.79 -2.33 (-2.00) (-405) Number of observations 458 458 Note: Full information maximum likelihood was used to estimate the model (equations 9, 10, and 1 1 in the text), assuming a discount rate of 0.2. Five household demographic characteristics were included in the regressions: age of household head, education of household head, number of children, number of female adult members, and distance to health care facilities. t-statistics are in parentheses. Source: Author's calculations. closed-access form. The monitoring and other transaction costs involved in the implementation and design of institutional mechanisms to exploit the resource efficiently might be very large, particularly given the high popula- tion density prevailing in most villages. This may prevent the development of such mechanisms. Another explanation could be that most farm communi- ties may operate at subsistence levels and, thus, cannot afford investing in the resource by reducing the cultivated area. Equivalently, the true discount rate that communities use is even higher than the 60 percent annual rate used as an upper bound in the estimation. It could be that their true discount rate is almost infinity. The analysis shows, however, that the way in which farm- ers respond to price incentives does not seem to be consistent with the sub- sistence story. The significance of the coefficients associated with biomass (the A3j coeffi- cients) and the net positive effect of biomass on total revenues when evaluated at the data points show the importance of biomass as a factor of production. Using L6pez 121 equation 9, the effect of biomass on total revenues is alnR e (13) a = [A3 + A330 + A13w + A23X + A34ki + A35P, + A36P. I -. al1nO R The effects of land cultivated and capital can be derived in similar fashion. Table 3 presents the values of these elasticities evaluated at the mean sample levels. The values of these elasticities are not very different from the coefficients ob- tained when a Cobb-Douglas aggregate production function is estimated (see L6pez 1993). The contribution of biomass to gross agricultural revenues is comparable to the contribution of labor and less than that of capital and cultivated land. In any case, the estimates in table 3 show that biomass makes an important contribu- tion to agricultural revenues and that a further loss of biomass would likely have a serious impact on agricultural production. The estimated contribution of biomass to gross agricultural revenues for C6te d'Ivoire is very similar to the values estimated for Ghana (L6pez 1997). In fact, the Ghana estimates range between 0.15 and 0.20, compared with 0.17 for C6te d'Ivoire. Moreover, several agronomic studies in tropical countries have shown that the fallow period makes a large contribution to agricultural productivity. Ellis and Mellor (1995), for example, have found that by reducing fallow peri- ods from five to two years in Zaire, the nutrients left in the soils decline signifi- cantly. Nitrogen falls by more than 50 percent, calciumlmagnesium and potas- sium fall by about 45 percent. This dramatic fall in nutrients is likely to have serious detrimental effects on yields, particularly if few fertilizers are used, as happens in most of Sub-Saharan Africa. Thus, the agronomic evidence is consis- tent with the estimates here for the effect of biomass on farm revenues. Table 4 shows the partial elasticities of supply for the three outputs, land cultivated, and labor demand with respect to the output prices and the wage rate. The most striking finding shown in table 4 is the fact that, although higher prices of cereals and other annual crops cause a very large expansion in area cultivated, an increase in the price of tree crops induces a reduction in the area Table 3. Implicit Factor Shares in the Total Value of Production, Cote d'Ivoire, 1985-88 Factor Sharea Biomass 0.17 Land cultivated 0.38 Capital 0.23 Labor 0.18 Note: Shares are evaluated at the means of the variables. See section V in the text. a. All of the factor share values are significant at the 10 percent level. Source: Author's calculations. 122 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Table 4. Elasticities of Supply for Outputs, Land Cultivated, and Labor in C6te d'Ivoire, 1985-88 Output price Output supply or Other annual Tree Wage factor demand Cereals crops crops rate Cereals -0.39 0.70* -0.02* -0.29* Other annual crops 0.66* 0.19 -0.74* -0.11 Tree crops -0.12* -0.68 0.58* 0.22 Land cultivated 0.60* 0.49* -0.95* -0.14* Total labor 0.29* 0.11 -0.24 -0.16* * Significant at the 10 percent level. Note: Values are partial elasticities with respect to output price or the wage rate, evaluated at the mean values of the variables. Source: Author's calculations. cultivated. One possible explanation for this result is that annual~ crops are much more land-intensive than tree crops, which are likely to be capital-intensive. This may give rise to a Rybcynski-type of effect. An increase in the price of tree crops induces two conflicting effects. First, an increase in the relative profitability of tree crops causes an expansion of the area planted with tree crops. Second, a fall in the relative profitability of cereals and other annual crops causes a reduction in the area planted with annual crops. The empirical results suggest that the second effect dominates; that is, the reduc- tion in the land devoted to annual crops is greater than the increase in the area devoted to tree crops. This implies that an increase in the price of tree crops leads to less land cultivated, longer fallow periods, and more natural biomass. Why does the area under annual crops fall more than the increase in the area under tree crops? Tree crops and nontree crops compete for other resources that are more or less in fixed supply (capital, credit, managerial skills). Therefore, the increase in the price of tree crops might absorb a sufficiently large volume of these resources from the annual crop sectors to force the area cultivated under those crops to fall more than the increase in the demand for land for tree crops. Tree crops and nontree crops compete not only for land but also for other re- sources in more or less inelastic supply. The extent of cultivated land, however, significantly increases if the prices of the three agricultural outputs rise proportionally, as shown by the negative and significant wage elasticity in table 4. Given the homogeneity conditions, the wage elasticity equals minus the sum of the three output price elasticities. That is, a 10 percent rise in all agricultural prices would lead to an increase in the total area cultivated by about 1.4 percent. Or, equivalently, a 10 percent fall in wages main- taining output prices constant would cause a similar increase in area cultivated. The long-run elasticity of biomass with respect to cultivatecl land is about -1.3 when evaluated at the average levels of the variable. This means that an across-the-board increase in farm output prices of 10 percent would cause the stock of biomass to fall about 1.3 percent in the long run. This, in turn, would Lopez 123 have negative consequences for agricultural productivity and income. The fact that producers do not consider the total social cost of biomass in their land allocation decisions (as shown by the fact that X is not significantly different from 0) implies that biomass is overexploited and that agricultural income is below its optimal. A further increase in agricultural prices aggravates the distor- tion by causing further losses in biomass and long-run agricultural productivity that outweigh the short-run benefits of expanding the land under cultivation. According to the econometric estimates, the agricultural income of an aver- age village could be increased by 14 percent in the long run if the total cost of biomass were internalized by individual cultivators. This represents a large loss, many times larger than the losses usually estimated for price or trade distor- tions. The main source of this income loss is the fact that land is overcultivated by about 23 percent. That is, a land allocation that would maximize the income of the average village requires a 23 percent decrease in the amount of cultivated land. An increase in all three agricultural prices or a fall in wages would sub- stantially increase the magnitude of these losses by increasing the incentives to cultivate land. VI. AGRICULTURAL PRICE POLICIES AND TRADE REFORMS From the previous discussion it is clear that policies that improve all or most agricultural commodity prices are likely to reduce agricultural productivity and to cause more deforestation and a reduction of fallow periods, further deterio- rating the natural resources. The net effect on agricultural income of raising agricultural prices across-the-board is still positive, however. The reason for this is that the fall in agricultural productivity and the increased direct cost of clear- ing more land do not offset the increase in farm revenues due to the higher farm commodity prices. That is, farmers have a net gain, but only because of the redistributional effect of improved relative prices. Similarly, policies that have a negative impact on wages are also likely to be deleterious for natural resources and agricultural productivity. By contrast, price policies that improve the terms of trade of agriculture by reducing the taxation of the tree crop sector alone are likely to be of the win-win type. They may induce social gains through conventional price efficiency (by reducing the distortionary tax on tree crops) and by reducing total land culti- vated and, hence, increasing biomass and agricultural productivity. Thus, the effect of trade liberalization on the environment and on national income is likely to vary according to the initial structure of protection. If trade reform improves the prices of all agricultural subsectors or if it causes a rise in the relative price of land-intensive (nontree crops) compared with capital-intensive agricultural commodities (tree crops), it may have deleterious effects for the environment and may even cause a net fall in national income. That is, complete trade liberalization without, for example, stronger communal 124 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 control institutions could be counterproductive not only for the environment but also for national income. Thus, trade liberalization makes even more urgent the need for strengthening village institutions. Fine-tuning of trade policies is not, however, feasible in many developing countries. Although this might be easier than applying first-best policies, some of the same institutional weakness that makes first-best policies difficult to apply could affect the implementation of trade policies as well. If, however, the initial distortions tax tree crops more than nontree crops, trade liberalization would improve the prices of tree crops relative to annual crops as well as relative to the rest of the economy. In this case, trade liberaliza- tion or, more generally, the removal of price distortions is likely to be a win-win policy causing an even greater expansion of income and an improvement in the level of natural resources. Economywide reforms that induce a fall in wages are likely to impose greater pressures on the natural resource base of agriculture. Macroeconomic policies such as real devaluation and measures to reduce fiscal disequilibria typically cause a fall in real wages. Thus, if those policies are not accompanied by comple- mentary measures to internalize the costs of natural resources, the loss of bio- mass, with the consequent fall in agricultural productivity over the medium term, is likely to be significant. According to my estimates, a macroeconomic adjust- ment that causes real wages to fall 10 percent, for example, would cause an increase in area cultivated of about 1.4 percent. This, in turn, would reduce biomass over the long run about 1.8 percent. This fall in biomass reduces the supply responsiveness of agriculture in the long run by a wide margin. VII. CONCLUSIONS Biomass is an important factor of production, accounting for roughly 17 per- cent of the agricultural gross domestic product in C6te d'Ivoire. The large losses of forest and the considerable reduction in fallow periods observed In Cote d'Ivoire over the past few decades have implied a considerable loss of productive natural capital that is likely to have reduced the productivity of labor and other resources. Individual cultivators act as if the biomass resource has no social costs be- yond the purely private costs of clearing the land. The hypothesis that communi- ties develop adequate controls on the use of communal resources to maximize their collective income does not appear to be valid for villages in C6te d'Ivoire. There are several possible reasons for this. Extremely fast pop-ulation growth may have caused a significant increase in monitoring and transaction costs in the villages. The greater density of the village population is likely to have sub- stantially increased the pressure on village resources and made it more difficult to control for violations of the implicit or explicit village regulations on the use of communal resources. At the same time, the significant immig.ration into cer- tain villages from both inside and outside C6te d'Ivoire and the rapid western- Lopez 125 ization of traditional values may have deteriorated the village hierarchies, mak- ing social controls more difficult to implement. The lack of internalization of the social cost of the biomass resource leads to large income losses at the village level. This article estimated losses on the order of 14 percent of village income. These losses are many times larger than the usual estimates for conventional distortions. The main response of annual crops to price incentives is to increase the area cultivated. The output response of tree crops, however, relies much less on ex- panding the area cultivated. An improvement in tree crop prices, keeping annual crop prices constant, is likely to cause a net reduction in cultivated area and an improvement in the natural resource. Thus, the potential for win-win policies exists in cases where, initially, tree crops (typically export crops) are taxed while annual crops (commonly import substitutes) are protected. In this case, the re- moval of trade distortions is likely to induce both the gains in price efficiency and a reduction in environmental losses. By contrast, reforms that cause an in- crease in the prices of all or most agricultural commodities (or reforms that reduce wages) are likely to deepen the environmental distortion and, thus, to reduce agricultural productivity in the long run. The analysis for C6te d'Ivoire refutes the hypothesis that common property resources are efficiently allocated; in other places common resources might be more efficiently used. Similar tests for villages in the eastern part of Ghana pro- duced results similar to those for C6te d'Ivoire (L6pez 1997). More case studies like these are needed to analyze whether common resources in developing coun- tries are allocated efficiently and to elucidate the conditions that determine the degree of efficiency in the allocation of common resources. It is possible that an efficient allocation of common resources requires certain specific conditions to facilitate collective action and that such conditions are not always present in most poor countries. In general, it appears that efficiency of the commons tends to be present in communities with low population density where the transaction and monitoring costs are low (Baland and Platteau 1996). The paradox is that it is precisely in cases of high and rapidly increasing population density that collec- tive action is most needed to achieve an efficient use of common resources. APPENDIX A. A NEOCLASSICAL GROWTH MODEL OF CAPITAL AND BIOMASs ACCUMULATION As a benchmark model, assume that the village maximizes the sum of the discounted present values of the utilities of the village households, subject to the budget constraint and the biomass constraint. (A-1) max fXU(yi)e-rtdt o i 126 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 subject to (i) k, k' ={R'[w, xi, (l - Ex'Il)ki] - kj- -x _ y} (ii) n = Y -(nE Xi/X) (iii) ki(O) = k/, ... N;i(O)=ino; Yxi<- where U(-) is an increasing and strictly concave utility function of the village households, / is consumption of household j, r is the discount rate, t is time, k' is capital, R' is revenue of farmer j, w is input prices other than biomass, land, and capital, xi is land cultivated by farmer j, 11 is the average density of biomass per acre in the land that is not cultivated, x is total land available, 8 is the (con- stant) rate of depreciation of the capital stock, c is the private cost of clearing the land, fis the change in ', k is the change in k, Y is the natural increase of biomass in the areas not under cultivation, xi is the total land under cultivation by farmer i, and kTI is the initial level of capital of farmerj. The first-order conditions of this problem are (A-2) (i) U'(yi) = j =l, ............,N (ii) R2 =c + rlYRi3(-) + i=,, (iii) r = ( Y + C - R4)E- j=1X.., (iV) (r ( x/)g - 31(l x/3) (v) A-1, parts i and ii where e is the shadow value of the stock of capital, R is the shadow value of biomass, a dot over a variable indicates rate of change, and subscripts indicate partial derivatives. It is important to note that for E to be identical for all house- holds, it is necessary to assume that R' is identical for all households. That is, assume that in equation A-2 the production technology is identical for all farm- ers and that they face the same prices w (conditions that are sufficient to assure that xi is equal for all farmers) and start with identical capital stocks. This is equivalent to postulating the model using a representative consumer/producer fiction, as is usually done. Equation A-2, part ii, dictates the land allocation decisions in the short run. It is analogous to the short-run land allocation derived from equation 4 in the text, with the only exception being that now the shadow value of capital (or marginal utility of consumption), £, appears. Equation A-2, part ii, indicates that the so- cial cost of expanding the cultivated area should be weighted by the marginal utility of consumption. That is, if e is large-that is, householcls are poor in the sense that their consumption is low, in which case U'(yi) is high-the social cost of expanding the area cultivated (and, hence, of depleting biomass) is less than if Lopez 127 households have a higher consumption level and consequently a lower U'() and E. This implies that, ceteris paribus, as a way of financing their capital buildup, capital-poor societies tend to deplete biomass more intensively than richer soci- eties. Thus, this model, unlike the one in the text, allows for a mechanism by which poverty may lead to greater resource degradation beyond the discount rate and the direct production substitution effects associated with capital or other factors. Compared with the model in the text, the steady state model has the same land allocation decision, but in addition it has an equilibrium equation for the optimal capital stock. It thus makes it possible to determine simultaneously the long-run equilibrium values of x and k, (ii) R4()=r+11i=Lr. ,) N A3)(iii,1 c - - Y, R=(-) i=i where the functions R2(.), Ri(-), and R4j(.) are evaluated at the steady-state equi- librium values of x*, k*, and il derived from the solution of equation A-3. The land allocation equation (A-3, part i) is analogous to equation 7 in the text, ex- cept that it needs to be evaluated at k* instead of at k. For the empirical model, however, apart from being more difficult to imple- ment, it is not clear that using this specification (equation A-3), which imposes a certain structure on the level of k, is necessarily superior to a specification based on equation 7 in the text, which does not impose any specific mechanism for the derivation of k. A more eclectic empirical approach is simply to instrumentalize variable k. APPENDIX B. DATA Part of the data used in this article comes from the Living Standards Survey (LSS) conducted in C6te d'Ivoire. About 50 percent of the sample is reinterviewed every year. These data provide information on demographic, labor, and produc- tion characteristics as well as the time allocation of households. The data used correspond to the period 1985-87. Statistics on agricultural land and biomass density at the village level were provided by a special study done for this project by EARTHSAT (1991) based on satellite images of 20 villages in the western region of Cote d'Ivoire for the years 1985, 1986, and 1988. These statistics were incorporated into the LSS panel data. Matching the LSS and EARTHSAT data gives a panel data set of household 128 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 observations that combine the usual individual characteristics of the households with information on biomass density and area at the village leveiL. Detection of Biomass Change Based on Remote Sensing Data Land was first classified into five categories: forestlands (areas with large and highly dense trees that do not appear to have been cultivated for at least 50 years), low-impact agriculture or bush fallow, high-impact agriculture (includ- ing sites of active agriculture or recently fallowed fields with negligible regenera- tion), human settlement areas (including villages and areas of construction), and other lands (including unproductive lands with little or no soil). The 1985 and 1986 imagery was geometrically corrected to fit the image space of the 1988 imagery, allowing comparative evaluations of the same areas over time. A principal components analysis was carried out on all imagery to normalize the data and to isolate radiometric anomalies. A statistical evaluation was carried out to determine the mean digital number value for each land-cover category, as represented by the principal component vectors from the various dates of imagery. Pixel counts were therL carried out to determine the quantity of change in each of the five categories of land. The pixels were then converted to hectares (the pixel size was 28.5 square meters). From this information, it was possible to determine the change in area for each of the five categories of land over time for each of the village areas. The digital numbers of band four were divided by those of band three to pro- duce a vegetation density index for the forest and low-impact agriculture areas for each of the study units. The total biomass value in fallow was defined as the vegetation density index times the areas under low-impact agriculture and that of the forest area as the forest vegetation index times the areas under forest. Thus, an index of total biomass in each land category was obtained for each village area. Variable Definitions and Summary Statistics Table B-1 provides the means and standard deviations for some of the most important variables. All the figures are in 1985 CFA francs. The deflator used was the African Food Price Index with base 1985 = 100. Table B-1. Means and Standard Deviations of Selected Variables Used in the Regressions for C6te d'Ivoire, 1985-88 Variable Mean Standard deviation Net revenue per household (1985 CFA francs) 668,990 647,200 Land cultivated per household (hectares) 7.2 4.23 Biomass per hectare of fallow (index number) 6.50 3.56 Total days of work per household per year 1,428 867 Real wage rate (1985 CFA francs per hour) 450 216 Number of agricultural tools or implements per household 12.3 11.6 Source: Author's calculations. Lopez 129 The variables used as well as their definitions are presented below, where i denotes the index for household i = 1,.. , n, and j denotes the village. (i) Rk, - Si, + PHLI, + SKi,t + HCi + HPzt - wi + LH - wiA LHMA,t - WHPjLHMPE' (ii) WH=A -w +bmWM + CwC where bM LHMAm,FC jb=1 HA F F m M cC F,C LHMA (iii)WHP -CFWF + CMWM + CCwC, where CM,F,C -LHMP 1 LHMP LH LHMP (iv) w' -aMWF + aHAwHA + aHMpwHP, where aM = LLaHMp = LL aHA LHMA LL ,aM,aHMp,aHA=1 (v) LLI, LH,i + LHMA' + LHMPl (vi) oi - i(-z') Xi (vii) z _ - Real revenue of farmer i in village j i, = Real total sales to the market of farmer i in village j PHLI = Real value of payments in kind to hired labor by farmer i in village i SKi Real value of seeds kept by farmer i in village j HG' = Real value of home consumption of farmer i in village j HP' = Real value of home production or transformation of farmer i in village j L Hi Number of days hired by farmer i in village j during the last 12 months LHMA, = Number of days worked in the field by members of the household i in village j during the last 12 months U/p = Agricultural real wage per day earned by a female in village j tt& = Agricultural real wage earned per day by a male in village j ul'c = Agricultural real wage earned per day by a child in village j u4HA = Average agricultural real wage earned per day by a member of a household working in his or her own field in village j U/HP = Average agricultural real wage earned per day by a member of a household working in the transformation of agricultural products in village j w' = Average real wage earned by a member of the household working in his or her own field or in the transformation of his or her agricultural products or by a worker hired by farmer i in village j 130 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 LU4 = Total number of days worked in the field or in home production by members of household i plus days hired by farmer i iri village j C = Total biomass in village j, normalized by its total larid area I' = Average biomass density per acre of land under fallow in village j (from EARTHSAT) Xi = Total land cultivated in village j (from EARTHSAT) X = Total agricultural land (cultivated plus fallow) in village j in hectares (from EARTHSAT) Z, = Fraction of cultivated land to agricultural land in village j y = Marginal biomass density xi, = Land cultivated by farmer i in village j in hectares n i=l ' = Average land cultivated in village j in hectares n Ni = Number of farms in village j Ei' = Years of education of the head of household i in village j L, = Number of family members in household i of village j 7il = Number of tree crops in farm i of village j E = Minimum value of E' in the sample = 0 L = Minimum value of LX in the sample = The sum of the agricultural tools and implements used by household i in village j T = Minimum value of Ti in the sample = 0 r = Time discount rate REFERENCES The word "processed" describes informally reproduced works that nmay not be com- monly available through library systems. African Development Bank (ADB), Economic Commission for Africa (ECA), and Organi- zation of African Unity (OAU). 1984. "The African Food Crisis and the Role of the African Development Bank in Tackling the Problem." Paper presented jointly by the staff of the ADB, ECA, and the OAU to the symposium on the future of ADB and the food crisis of Africa, Tunis, Tunisia, May. Processed. Allen, Julia. 1985. "Wood Energy and Preservation of Woodlands in Semi- and Devel- oping Countries, The Case of Dodoma Region, Tanzania." Journai of Development Economics 19(1):59-84. Baland, J. M., and J. P. Platteau. 1996. Halting Degradation of Natural Resources. Oxford: FAo and Clarendon Press. Basu, K. J. 1989. "Rural Credit Markets: The Structure of Interest Rates, Exploitation, and Efficiency." In Pranab Bardhan, ed., The Economic Theory of Agrarian Institu- tions. Oxford: Clarendon Press. Binswanger, Hans, and Robert Evenson. 1984. "Labor Demand in North lndia Agriculture." In Hans Binswanger and Mark Rosenzweig, eds., Rural Labor Markets in Asia: Contrac- tual Arrangements, Employment, and Wages. New Haven, Conn.: Yale University Press. L6pez 131 Dasgupta, Partha, and Karl-Goran Maler. 1991. "The Environment and Emerging De- velopment Issues." In Proceedings of the World Bank Annual Conference on Devel- opment Economics. Washington, D.C.:World Bank. Diewert, W. E. 1973. "Functional Forms for Profit and Transformation Functions." Journal of Economic Theory 6(3):284-316. Diewert, W. E., and Catherine Morrison. 1988. "Export Supply and Import Demand Functions: A Production Theory Approach." In Robert Feenstra, ed., Empirical Meth- ods for International Trade. Cambridge, Mass.: MIT Press. EARTHSAT (Earth Satellite Corporation). 1991. "Ivory Coast Change Detection: Quanti- tative Assessment for Basic Land-Cover Categories for the Years 1985, 1986, and 1988." Bethesda, Md. Processed. Ellis, Stephen, and Albert Mellor. 1995. Soils and Environment. London: Routledge. FAO (Food and Agricultural Organization of the United Nations). 1986. "Natural Re- sources and Human Environment for Food and Agriculture in Africa." Environment and Energy Paper 6. Rome. Processed. Geertz, Clifford. 1963. Agriculture Involution: The Process of Ecological Change in Indonesia. Los Angeles: University of California Press. Glantz, Michael. 1977. Desertification: Environmental Degradation in and around Arid Lands. Boulder, Colo.: Westview Press. Jones, William I., and Roberto Egli. 1984. Farming Systems in Africa: The Great Lakes Highlands of Zaire, Rwanda, and Burundi. World Bank Technical Paper 22. Wash- ington, D.C.: World Bank. Lagemann, Johannes. 1977. Traditional African Farming Systems in Eastern Nigeria. Munchen: Weltforum Verlag. Larson, B. A., and D. W. Bromley. 1990. "Property Rights, Externalities, and Resource Degradation." Journal of Development Economics 33:235-62. L6pez, Ram6n. 1985. "Structural Implications of a Class of Flexible Functional Forms for Profit Functions." International Economic Review 26(2):593-601. . 1993. "Resource Degradation, Community Controls, and Agricultural Produc- tivity in Tropical Areas." Department of Agricultural and Resource Economics, Uni- versity of Maryland, College Park. Processed. . 1997. "Environmental Externalities in Traditional Agriculture and the Impact of Trade Liberalization: The Case of Ghana." Journal of Development Economics 78(1):508-16. L6pez, Ram6n, and Mario Niklitschek. 1991. "Dual Economic Growth in Poor Tropi- cal Areas." Journal of Development Economics 36(2):189-211. Ludwig, H. D. 1968. Ukare: Ein Sonderfall Tropischer Bodennutzung in Raum des Victoria-Sees. Munchen: Weltforum Verlag. Niklitschek, Mario. 1990. "Economic Growth in Sub-Saharan Africa: The Implication of the Environmental Degradation." Ph.D. diss., University of Maryland, College Park. Processed. Perrings, Charles. 1989. "An Optimal Path to Extinction? Poverty and Resource Degra- dation in the Open Agrarian Economy." Journal of Development Economics 30(1):1- 24. Sanchez, Pedro. 1976. Properties and Management of Soils in the Tropics. New York: John Wiley and Sons. Sinn, Hans-Werner. 1988. "The Sahel Problem." KIKLOS 41(2):187-221. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1: 133-53 Economic Parameters of Deforestation Joachim von Amsberg In theory, economic instruments should overcome the market failures that lead to excessive deforestation. Secure property rights could he established and enforced to eliminate the open access problem. In practice, the size of the welfare loss that arises from market failures in the forest sector in the absence of such first-best policies is determined by the incentives, prices, and policies faced by those who make decisions about land use. In many cases, the effects of policies on deforestation are not straight- forward. For example, there are conflicting views on whether an increase in the price of logs leads to an increase or a decrease in deforestation. The effect of a change in the price of logs has particular relevance for the controversial debate about the effect on deforestation of a ban on log exports or other trade restrictions that lower the domes- tic price of logs. This article provides an analytical framework for determining the effects of changes in economic policies and parameters on deforestation. It models dynamic, profit- maximizing land-use choices and obtains unambiguous comparative static results by distinguishing between unmanaged and managed forests. The results suggest that mea- sures to reduce the producer price of logs could be a second-best policy to reduce the pressures on the frontiers of unmanaged forests and to protect biodiversity. Property rights to forests in frontier areas are rarely established or enforced. As a result of open access, deforestation (the conversion of forested lands to other uses) can be excessive. Even when property rights are established, forested lands provide external benefits that do not accrue to the owner, government forester, or other decisionmaker. These external benefits include stabilization of the re- gional and global climate, conservation of the soil, prevention of floods, preser- vation of biodiversity, and gathering of nontimber products by individuals who do not own the forest. These externalities can be another reason for excessive deforestation. In theory, economic instruments would overcome the market failures that lead to excessive deforestation. Secure property rights could be established and enforced to eliminate the open access problem. External benefits of forests could be internalized by taxes on deforestation or subsidies for the maintenance of forestlands equal in amount to the external benefits. Such first-best policies would Joachim von Amsberg is with the Brazil Country Management Unit, Latin America and the Caribbean Region, at the World Bank. The author would like to thank Ken Chomitz, Jeffrey Hammer, Muthukumara Mani, David Wheeler, and two anonymous referees for their helpful comments. This article is a summarized version of the World Bank Policy Research Working Paper 1350, "Economic Parameters of Deforestation." © 1998 The International Bank for Reconstruction and Development/THE WORLD BANK 133 134 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I lead individuals to make efficient land-use decisions through the operation of market forces. In practice, however, governments rarely use first-best polices such as Pigouvian taxes. Some externalities are international in nature (carbon sequestration and biodiversity conservation), and individual countries have no incentive to implement globally efficient policies. Other reasons for the absence of efficient policies are political (for example, the owners of forests have better representation than the beneficiaries of positive forest externalities). In addition, the establishment and enforcement of secure property rights are costly. In the absence of first-best policies, the size of the welfare loss that arises from market failures in the forest sector is determined by the incentives, prices, and policies faced by those who make decisions about land use. Economic param- eters, such as transportation costs, royalty structure, trade policy, foreign ex- change policy, and productivity changes in the forest sector as well as in agricul- ture, influence the patterns of deforestation through their effects on the incentives of those individuals making choices about land use. Therefore, two questions arise in the absence of first-best policies for forest management and land use. First, which policies should be avoided because they would increase the welfare loss arising from excessive deforestation? Second, which second-best policies can be implemented to reduce the welfare loss arising from excessive deforestation? In many cases, the effects of policies on deforestation are not straightforward. For example, there are conflicting views on whether an increase in log prices leads to an increase or a decrease in deforestation. In one view, lower log prices reduce logging profits and the incentives for logging and hence reduce deforesta- tion. In the opposing view, lower log prices reduce the profitability of forestry and hence encourage the conversion of forestlands to other uses such as agricul- ture (see Vincent 1990; Brandon and Ramankutty 1993; and Sharma and others 1994). The effect of changes in log prices has particular relevance for the contro- versial debate about the effect on deforestation of a ban on log exports or other trade restrictions that lower the domestic price of logs. This article is related to three strands of the theoretical literature. First, an extensive forestry literature builds on Faustmann (1968, originally published in 1849) and examines the effects of changes in various economic parameters on the optimal management of a forest (see Jackson 1980; Chang 1983; Nautiyal and Williams 1990; Hyde and Newman 1991; Thiele 1995; and Thiele and Wiebelt 1994). Most of these papers use comparative statics analysis to deter- mine the effect of changes in production costs, discount rate, and various taxes on the optimal rotation age and the optimal management intensity for a given forest. These models rarely consider possible changes in land use. Second, static land-use models have been used to analyze the optimal use of land at a given point in time. This work was pioneered by von Thiinen (1826), applied to for- estry by Ledyard and Moses (1974), and recently used by Chomitz and Gray (1996). Third, the effect on land use of changes in the price of logs has been explored in recent work that applies land-use models to deforestation problems (Deacon 1994; Deininger and Minten 1996; Southgate 1990; Kishor and von Amsberg 135 Constantino 1993; Hyde, Amacher, and Magrath 1993; and Barbier and Rauscher 1993). In addition, several authors have analyzed the empirical relationship be- tween economic parameters and deforestation (see Barbier and others 1995 and Cropper and Griffiths 1994). None of these works, however, has produced un- ambiguous results with regard to the directional impact on deforestation of ap- parently simple changes, such as a drop in the price of logs. This article provides an analytical framework for determining the effects of changes in economic policies and parameters on deforestation. The framework allows the systematic analysis and reconciliation of opposing views on the effect on deforestation of changes in the price of logs. A simple theoretical land-use model also provides results on the effects on deforestation of specific policy changes, such as the imposition of a ban on log exports. Section I outlines the modeling approach. Section II presents a formal model of the comparative statics of land use. Section III discusses tentative policy implications. Section IV concludes with a discussion of extensions and further research. 1. MODELING APPROACH This article analyzes the links between economic parameters and deforesta- tion through a theoretical model of profit-maximizing choice of land use. Fol- lowing von Thiinen's (1826) approach, it assumes that land is put to the use that maximizes the present value of profits to the decisionmaker. The analysis of land-use dynamics is based on a formal comparative statics model, similar to those in the traditional forestry literature. It differs from previous work by si- multaneously incorporating two elements that are critical for understanding deforestation processes: a distinction between different types of forests and the dynamic nature of land-use decisions involving forests. First, the analysis clearly distinguishes between managed and unmanaged for- ests. In unmanaged forests, net timber growth is zero because decaying timber offsets biological growth. Logging of such a mature forest can be modeled like the mining of a nonrenewable resource (see Lyon 1981). Unmanaged forests would include primary forests and mature second-growth forests. Managed for- ests, such as plantation forests, by contrast, are planted in order to be harvested at regular intervals. Even though the dichotomy of managed and unmanaged forests is somewhat extreme, the distinction not only simplifies the analysis but also clarifies the often opposite impact of a policy on managed and unmanaged forests. Moreover, the distinction is relevant from a policy perspective because unmanaged and managed forests provide different types of environmental exter- nalities. Unmanaged forests typically have higher value for the conservation of biodiversity, while managed forests (depending on the subsequent use of forest products) can provide greater benefit in terms of carbon sequestration. The second difference from previous work is the analysis of land-use changes in a dynamic context. A static analysis based on a comparison of returns to different land uses at one point in time can be misleading. The relevant question 136 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 is not only whether deforestation would occur on a given piece of land but also when it would occur. For example, the introduction of forest plantations could increase logging of unmanaged forests in the short run but slow down deforesta- tion in the long run when the plantation output reaches the market. The timing of excessive deforestation is important from a policy point of view because it determines the effectiveness of corrective policies that are taken at a specific point in time. Therefore, the analysis of policy impacts has to be based on the comparison between different land-use patterns through time. In a dynamic context, land-use decisions depend on not only current log prices but also the expectation about future prices. The analysis assumes certainty and rational expectations. Therefore, agents determine their profit-maximizing be- havior in the first period for all times in the future based on the expected path of log prices. In the absence of unanticipated shocks, there is no difference between the expected and realized price path and the expected and realized behavior. As a result of geographic conditions and anticipated changes in log prices, deforestation rates can increase or decrease over time without a change in policy. Because deforestation rates can change without policy changes, the relevant ques- tion for analyzing the effect of policy changes is not whether deforestation rates fall or rise after a change in policy occurs, but whether deforestation rates differ from what they would have been if the change in policy had not occurred. This comparison of the actual with the counterfactual scenario is the natural realm of theoretical modeling. The analysis models a policy change as an unanticipated shock that changes price expectations and, therefore, profit-maximizing behav- ior. The analysis focuses on the change in behavior that results from such unan- ticipated policy changes. II. A FORMAL MODEL OF DYNAMIC LAND USE This section contains a partial equilibrium model of profit-maximizing land use to determine dynamic land use as a function of an exogenous path of log prices over time. The model analyzes the timing of land-use changes for each specific parcel of land. It derives results for spatial land-use changes by combin- ing the changes in the timing of land-use changes for each class of land (land with the same locational characteristics). Initially, all land is covered with unmanaged forest. There is no profit to the owner of an unmanaged forest until it is converted. After converting the unmanaged forest, the owner puts the land to its profit-maximizing use, either as managed forest or as farm land. The decision to convert an unmanaged forest depends on the profit or loss at the time of conversion (value of logs-if sold- minus clearing or logging costs) and the profits from alternative land use after logging (farming or managed forest). In this model, the value of logs represents all forest products, including latex, fruits, nuts, and fuelwood. The model is equally applicable in cases where (a) the unmanaged forest is logged, the logs are sold, and the land is subsequently cultivated, (b) the unmanaged forest is logged, von Amsberg 137 but the land is left idle after logging, or (c) the removal of logs is not profitable, and the forest is simply cleared for subsequent farming or managed forestry. After the initial conversion of the unmanaged forest, the owner may switch between different alternative land uses. Of course, on some lands logging might not occur in finite time. The following diagram shows the sequence of possible land uses: Managed forest Unmanaged forest tj Ito Farming where tr is the time of converting unmanaged forest to managed forest or farm- ing, and ta is the time of switching from managed forest to farming or from farming to managed forest. Equation 1 defines the present value of profit from profit-maximizing land use, fia, in land class i from the time of conversion of unmanaged forest to infinity. Superscript a refers to the profit-maximizing land use-either managed forest or farming-after logging the unmanaged forest. The land class is for a particular parcel of land and represents a generalization of von Thiinen-type distance from market, including other location-specific factors such as slope and fertility. (1) fl"i (t'W, k) e rsia(s, k)ds :ti. where k is a parameter that represents the effect of exogenous policy changes on the log price, 7r is profit in each period, and s is the time passed after conversion of the unmanaged forest at tl. Superscript u refers to unmanaged forest, and r is the discount rate of the decisionmaker. Equation 2 defines the land expectation value, LEV, the present value of the sum of conversion profit and subsequent land-use profits. (2) LEV' = e-rt 7iu (tiu k) + flia (tiu, k). The model is based on an exogenous log price path. The model, thus, applies to a situation of perfectly elastic demand, that is, for log exports of a small country. For the log price path, the analysis assumes that Pt > 0 and that (ptt/ pt) < r. (Here and throughout subscripts denote partial derivatives.) This assump- tion appears eminently reasonable given that (p,, / pt) < r is satisfied for any constant rate of price increase less than the discount rate, r. The assumption is also consistent with empirical observation and with the results from theoretical models of nonrenewable resource extraction (logs from unmanaged forests) with 138 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 increasing extraction costs and a renewable back-stop technology (logs from managed forests) that would make log prices rise at a declining rate. Market simulations support the price path assumptions (see section IV). The effect of policy interventions that would depress log prices is expressed in the form of a parameter k that enters the log price with the following character- istics: Pk < 0 and (Ptk / Pk) < r. An increase in k either reduces the level of the log price path or reduces the price at any time in some other form; however, an increase in k does not reduce the slope of the price path more than permitted by (Ptk / Pk) < r. With profit functions increasing in log prices, the profit from both conversion and subsequent cultivation will decline with a drop in log prices, resulting in the following properties: iTc < 0, rl 'a < 0. The Remaining Unmanaged Forests How do changes in the log price path affect the area of unm,anaged forests that will ultimately remain? The unmanaged forest will never be converted in any class of land in which the LEV is less than zero for any finite time of conver- sion. By contrast, all land will ultimately be converted in classes of land in which the LEV is greater than zero at least at some time. With these properties, LEV decreases, for all tm, with a drop in the log price path, k. With a lower log price path, there is no land class in which unmanaged forest that is ultimately con- verted would not also have been converted with higher prices. 1-lowever, some classes of land that would ultimately be converted under a higher log price path may not be converted at all under lower prices. Hence, the unmanaged forest area that will ultimately be converted is equal or less under a lower log price path. Up to this point, the model is general enough for the results to hold indepen- dent of the property rights regime and the specific production functions dis- cussed in the following sections. It also applies for a Faustman-type rotation model for managed forests in which forest intensity and rotation period are chosen to maximize the LEV. Conversion of Unmanaged Forests with Secure Property Rights The next step is to analyze how economic parameters affect land use during the time of transition from all-unmanaged forest to the final land use. To pro- vide stylized answers to this question, this section assumes very simple produc- tion functions. The analysis is carried out first for the case of secure property rights and then for the case of open access. Under secure property rights, conversion profits and profits from subsequent cultivation accrue to the owner or decisionmaker. Even under secure property rights, externalities occur in the form of nontimber benefits of the forests that do not accrue to the landowner, such as climatic and soil stabilization, biodiversity conservation, and nontimber forest products. The case of secure property rights is also applicable where logging decisions are made by a government that cares about logging revenues but ignores nontimber benefits of the forest. Such gov- von Amsberg 139 ernment behavior appears reasonable for a variety of reasons. In contrast to nontimber benefits, logging often generates government revenues from stump- age fees. Some nontimber benefits such as climate and soil stabilization will accrue in the future, possibly after the tenure of the current government. A con- centrated logging industry can generate lobbying pressure on the government more easily than the less-organized recipients of nontimber benefits can. Finally, some nontimber benefits may accrue as international externalities. Under secure property rights, the owner of each piece of unmanaged forest- land maximizes the LEV by choosing the optimal time for converting the unmanaged forest, the optimal subsequent land use, and possibly the optimal time for switching later from managed forest to farming or from farming to managed forest. For simplicity, the analysis uses Leontief (constant coefficient) technology for the production of logs. In real-life forestry, there is clearly some substitutability between timber land and effort. Different logging intensities and technologies can be observed in logging operations throughout the world. De- tailed analysis shows that the main result of this section-conversion of unmanaged forests proceeds less rapidly with lower log prices-continues to hold under very reasonable conditions even with variable logging effort (von Amsberg 1994). The two inputs to production are unmanaged forestland and logging effort (with effort representing all inputs other than land, for example labor and capital such as chainsaws). The profit from converting unmanaged forest is 7tiu(t, k) = liup(t, k) - ci', where l.8 is the quantity of logs that can be sold once at the time of converting (logging) one unit of land of unmanaged forests in land class i, and cia is the cost of converting one unit of land of unmanaged forest in land class i and transporting logs to the market. For simplicity it is assumed that a managed forest produces a constant timber crop. The model abstracts from the question of optimal effort and optimal rota- tion periods in the managed forest and focuses squarely on the question of land conversion. The profit from a managed forest is nim(t, k) = erdlimp(t, k) - cim, where pim is the quantity of logs that can be produced each period by cultivating one unit of land in land class i with managed forest, ci- is the cost each period of cultivating one unit of land in land class i with managed forest and transporting logs to the market, and d is the fixed rotation period until the harvest of a managed forest. Land that is left idle after logging yields zero profits. Profits from farming are assumed to be independent of p, cu, cm, and r. The profits from profit- maximizing land use after converting unmanaged forest are assumed to be nondeclining.' With these assumptions, the following properties result: 7tt > 0, 1. The assumption about nondeclining profits refers to the time of logging, t", and not to the time passed since land conversion, s. For simplicity, the model does not allow for profits to depend on s. Even though not shown formally, the basic intuition of this model would not change if profits were declining in s as long as the present value of future alternative land uses at the time of conversion would be nondeclining in t". With this extension, the basic results would carry through also in the cases of shifting or nutrient-mining agriculture where agricultural profits would typically decline in s (but not in t). 140 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I 7u < ° nu < 0, (Tt, / nu,) < r, 7rc = 0, (nuk / 7iu) < r, itf> 0, it < 0, 7nt'< 0, Ir < 0, and ia,Ž 0. (To simplify notation, superscript i is omitted from expressions that hold for all i.) Land use after the logging of unmanaged forest is determined by maximiz- ing profits by choice of land use over time (managed forest or farming). The optimal time of logging the unmanaged forest is determined by maximizing the present value of returns from logging and subsequent profit-maximizing cultivation (s is the integration variable, running from the time of logging, tu, to infinity): (3) max LEV = ert. tu(tuk) + Jersnla(s,k)ds t. f tU subject to the condition that max LEV > 0 because logging would not take place within finite time if max LEV < 0. The first-order condition is: (4) LEVt =- =ertr [ (tu) -rtu(tu) -_ia(tu ) 0 atu I where an asterisk denotes the LEV-maximizing conversion time. The intuition of this first-order condition is that at the optimal time of conversion, the rate of appreciation of logs in the unmanaged forest, due to the increasing log price, must equal the forgone returns from logging as well as alternative cultivation of the land. The effect of changes in the parameters on the optimal time of conversion, t8<, is determined by solving the total derivatives of equation 4 withl respect to k, cu, and r for (dt" / dk), (dtu / dcu), and (dt"* / dr), respectively: dtu' LEVtk + nk -tk- > 0 dk -LEVtt 7tUt-r74t-1rt (5) dtU= LEV k k 4tk > 0 dc -LEV,,t iru- rcu-ICa dtu' LEVt, - - dr -LEVt, nu -r7cut - > +r Hence, on any piece of land, a reduction in the log price path (an increase in k) delays the profit-maximizing logging time. An increase in the cost of logging also delays logging. If profits from logging are positive and greater than the reduction in profits from land cultivation with an increase in the discount rate, then an increase in the discount rate advances deforestation. von Amsberg 141 Conversion of Unmanaged Forests with Open Access Unmanaged forests typically involve frontier situations with poorly defined property rights and some form of open access. Angelsen (1996), Schneider (1995), Mendelsohn (1994), Mahar (1989), Anderson and Hill (1990), and Binswanger (1989) have modeled such situations of frontier land use and land races. In these models, property rights are granted only for colonists who invest resources (which typically means that they clear the land). Such a policy regime has been analyzed in the case of Brazil but is common in other countries as well. Mendelsohn (1994) shows that development (or conversion of unmanaged forest) will occur wher- ever the value of land is positive; however the rents of land with values above zero will be at least partially dissipated through the investments necessary to establish property rights. The following modification to the basic model ana- lyzes how changes in log prices affect unmanaged forest conversion under this particular policy regime. If access to the unmanaged forest is open and property rights are acquired by clearing and cultivating land, conversion does not take place at the profit- maximizing time but as soon as the sum of profit or loss from conversion and the present value of profits from subsequent cultivation rises to zero. All lands with positive conversion profits would already have been converted in the past. The condition that determines the time of conversion is thus: (6) LEV = nu (tu', k) + fe-r(s-t" ) a (s, k)ds = 0. The comparative statics results can be formally derived, similar to the case of secure property rights. For the open access case, the algebra is tedious, but the results are rather obvious. Therefore, the formal derivation of the following comparative statics result is not shown here: (7) d > 0. dk Under open access, a drop in the log price path delays the logging time, as in the case of secure property rights. This is intuitively obvious by observing that LEV in equation 6 is increasing in tu and decreasing in k (ntt > 0 and lta > 0 by assumption, and itk < 0 and ik < 0 because profit functions are increasing in output price). Therefore, any increase in k has to be offset by an increase in t2, or vice versa, in order for equation 6 to hold. The Switch between Managed Forests and Agriculture After unmanaged forest has been converted, the land will be put to the profit- maximizing use, which may be farming or managed forest. If there is a change in the relative profitability of these activities, there may be a later switch from one 142 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I to the other. If there is a switch, the optimal time of the switch from managed forest to agriculture (tat), or from agriculture to managed forest, is determined by the condition of equal profits in both land uses: (8) q:f (ta') = m (ta v k) where superscript f refers to farming. Solving the total derivative of equation 8 with respect to k, cm, and r for (dta*1dk), (dta*ldcm), and (dta'ldr), respectively, gives the comparative statics ef- fects. If profits from managed forests are expected to rise faster than profits from agriculture (nt > irc), then ta' marks the time of optimal conversion from agriculture to managed forestry. With the assumptions on profit functions made above: dta* - _ __m > d_ > dk 7tf I-Tt ta m (9) d> m ___ > dc- t-1t- dtu - 7 :_ o dr nf _ nm If, however, Tr, < R4 (in this case, ta* marks the time of optimal conversion from managed forestry to agriculture), then all signs are reversed: (dta*ldk) < 0, (dt"*/dcm) < 0, and (dta*ldr) < 0. These results simply show that the switch from farming to managed forests, if it occurs, is delayed by factors that reduce the profits from managed forests (a drop in the log price path, an increase in the cost of managed forests, or an increase in the discount rate). A switch from managed forests to farming, if it occurs, is advanced by the same factors. Land-Use Changes The analysis has produced unambiguous results on the timing of land conver- sion for any land class i. Because this analysis is valid for an,y land class, it implies results for aggregate land-use changes over time. A drop in the log price path delays the possible conversion of unmanaged forest to other uses (including managed forests), delays the possible switch from farming to managed forest, and advances the possible switch from managed forest to farming, all for any land class i. Therefore, at any time after the drop in the log price path, there will be more or equal land under unmanaged forests and less or equal land under managed forest than if the price drop had not occurred. The effect on the aggre- gate area of agriculture is ambiguous. A drop in the log price path reduces the conversion of unmanaged forests and, thus, retains a larger area of unmanaged forests. At the same time, how- ever, a lower log price path reduces the area under managed forests. Keeping in von Amsberg 143 mind the distinction between unmanaged and managed forests, the intuition of the main result is easily explained. The conflicting views about the effects of log price changes on deforestation arise from the dual nature of forestland as stor- age for logs and as an input to the production of logs. This article reconciles the two opposing views by analyzing the distinct impacts of changes in the price of logs on different types of forests, which are characterized by the difference in the importance of land as storage for logs or as an input to log production. A higher log price path increases the logging of unmanaged forests that are used to store logs but that are no longer productive. With a higher log price path, the logging of more remote, unmanaged forests with higher site-specific extraction costs becomes profitable, and the logging of unmanaged forests increases. By con- trast, managed forestlands are productive. A higher log price path increases the profitability of log production and results in more land being devoted to log production. Therefore a higher log price path leads to a smaller area of unmanaged forest and a larger area of managed forest. Figures 1 and 2 illustrate the translation of results for the timing of conver- sion of a specific land class to results for aggregate land use over time. The land- use graph in figure 1 shows different land classes on the vertical axis, with higher land classes representing increasingly unfavorable conditions for cultivation, for example increasing transport costs in a von Thunen-type model. The horizontal axis represents time beginning with a situation in which all land is covered with unmanaged forest. In good locations (near the horizontal axis), agriculture is relatively more profitable than forestry. Conversion of unmanaged forest would begin at these most favorable locations and, as the price rises along the log price path, proceed to less favorable locations. At sufficiently high log prices, man- aged forestry becomes profitable as shown in the example in figure 1. Once the log price stabilizes, no further conversion of unmanaged forests occurs. Figure 2 illustrates the effect of a drop in the log price path (an increase in k). For each land class, the conversion of unmanaged forests-if it occurs-is de- layed (compared with the dark line that represents the base case), the switch from agriculture to managed forests-if it occurs-is delayed, and the switch from managed forestry to agriculture-if it occurs-is advanced. As a result of these changes in the timing of conversion of specific land classes, there are changes in aggregate land use at any specific time. An unanticipated drop in the log price path leads to an increase in the area under unmanaged forests and a reduction in the area under managed forest at any time after the shock. Table 1 summarizes the results of the analysis. It compares aggregate land use at any time after a hypothetical shock with a situation in which the shock would not have occurred. Additional results, not all analytically derived here, include the following (see von Amsberg 1994). * An increase in conversion (logging) costs for unmanaged forests (or a logging fee per unit of unmanaged forest) produces an increase in unmanaged forests and a decrease in managed forests and farming area. 144 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Figure 1. Log Price and Land Use in the Base Case Log price path Price Time Land use Land class Time rI I Agriculture N anaged forest IUnmanaged forest von Amsberg 145 Figure 2. Log Prices and Land Use in the Log-Price Drop Case Log price path Price Time Base case Log-price drop case Land use Land class Time El Agriculture 0Managed forest UUnmanaged forest Note-. Black lines depict the base case land use for comparison. 146 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 * An increase in the decisionmakers' discount rate produces a reduction in unmanaged forest if logging is relatively profitable (see equation 5) and an increase in the area of unmanaged forest otherwise. * An increase in farming profits produces a reduction in the area of both unmanaged and managed forests and an increase in the area of farming. * A reforestation subsidy per unit of land produces a reduction in the areas of unmanaged forests and farming and an increase in the area of managed forests. III. POLICY IMPLICATIONS In a very simple land-use model, a drop in the log price path leads to a delay in the conversion of unmanaged forests in all land classes. The quantity of unmanaged forests that is ultimately preserved is the same or larger under a lower log price path. The area of managed forests is reduced under a lower log price path. Great care needs to be taken in applying the results of a simple theo- retical model directly to complex real-life policy situations. However, the main result of the basic model and its underlying basic intuition appear to be robust enough to suggest some implications for the policy debate on timnber trade re- strictions, agricultural intensification, and changes in the cost of capital. Table 1. Policy Interventions and Changes in Land Area Used for Unmanaged Forest, Managed Forest, and Farming Policy intervention Unmanaged forest Managed forest Farming Drop in log price Increase Decrease Decrease at the unman- caused by log unit aged forest margin; tax or log export increase at the ban managed forest ma rgin Increase in conver- Increase Decrease at the Decrease at the sion costs (logging unmanaged forest unnanaged forest tax per land unit) margin; no effect margin; no effect at at the agriculture the managed forest margin margin Increase in the Decrease if logging Uncertain effect if Increase if logging discount rate unmanaged forest logging unman- unmanaged forest is is relatively aged forest is relatively profitable; profitable; relatively profit- uncertain otherwise increase otherwise able; decrease otherwise Increase in farming Decrease Decrease Increase profitability Reforestation sub- Decrease Increase Decrease sidy (per area unit) von Amsberg 147 Timber Trade Restrictions Many timber-exporting countries have imposed log export bans (LEBS) or high log export taxes (see Crossley 1993). LEBS were imposed primarily with the ob- jective of promoting domestic processing and the export of higher-valued sawnwood or manufactured goods. Even though LEBs were conceived as instru- ments for the protection of infant industry, they have implications for logging rates, and a lively debate centers on the environmental effects of LEBS (see Goodland and Daly 1994). LEBS, most other timber trade restrictions, as well as consumer boycotts in importing countries lower the price of logs in the export- ing country. Following a log export ban in Costa Rica, for example, domestic log prices have fallen to 20-60 percent of international price levels (Kishor and Constantino 1993). This article suggests a differentiated approach to analyzing whether lower log prices increase or decrease deforestation. A lower log price path would tend to reduce the logging of unmanaged forests but, at the same time, would also tend to reduce the area of managed forests. At any time, there would be more unmanaged forest and less managed forest than otherwise.2 This result is consis- tent with earlier findings that lower domestic log prices encourage wasteful log- ging and processing techniques (Repetto and Gillis 1988). In contrast to Repetto and Gillis (1988), however, this article suggests that the reduced logging inten- sity resulting from a lower log price path would go along with reduced logging (and larger remaining areas) of unmanaged forests and reduced areas of man- aged forests. Policies other than LEBS that reduce log prices include import restrictions by log-importing countries and consumer boycotts of tropical timber. Such policies would tend to reduce the pressure for logging unmanaged forests and therefore assist the conservation of biodiversity and other external benefits associated with unmanaged forests. The same measures would lead to reduced incentives for managed forestry and a decline in the area devoted to managed forests. Thus these policies have positive effects on the external benefits associated with unmanaged (old-growth) forests and negative effects on the external benefits associated with managed (plantation) forests. Because the effect on the com- bined area of managed and unmanaged forests is ambiguous, no statement can be made about the effect on external benefits that are associated with both types of forests. However, FAO (1992) estimated that 82 percent of the tropical forest area logged between 1981 and 1990 was in previously unlogged (unmanaged) forests. This figure would suggest the relative importance of the positive effect of lower log prices on unmanaged forest conservation compared with the nega- tive effect of reduced managed forests. 2. If protection is declining over time or the domestic processing industry gains some efficiency over time, the price depressing effect of an LEB would decline over time. The effect is thus well represented by the model, with an increase in k with Pk < 0 and (Pk I Pk) < r. 148 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I The positive effect of LEBS on unmanaged forests should not be misinterpreted as an endorsement or a recommendation for LEBS. First, the effects of real-life LEBS include political economy effects that are not captured by the simple model presented here. Second, due to reduced logging and processing efficiency and increased logging wastes, LEBs and other trade restrictions are clearly inferior to first-best policies (for example, a charge for the conversion of forestland equal to the external benefits of unmanaged forests). Even in the context of the simple model presented in this article, LEBS can be justified as second-best policy instru- ments only if first-best instruments are impossible to implement and if the ben- efits of reduced logging outweigh the efficiency costs imposed on the economy as a result of the price distortions from trade restrictions. In policy terms, re- moving LEBS in the absence of efficient first-best policies for protecting forests will increase the pressure on unmanaged forests. Other Policies Policymakers sometimes claim that agricultural intensification programs as well as forest plantation projects reduce the pressures to convert unmanaged forests. Within the conceptual framework presented here, agricultural improve- ments, such as increased yields from improved seed varieties or improved farm- ing practices, would reduce pressures on forests only if they reduce the potential profitability of agriculture on currently forested lands. This would occur only if the demand for the agricultural product is very inelastic (for example, in the case of subsistence agriculture). Agricultural improvements would then reduce the prices of agricultural outputs and, thus, the profitability of agriculture. In this case, the same quantity of agricultural output would be produced on a smaller area of land, and pressures for deforestation would be reduced (for the subsis- tence case, see Angelsen 1996). By contrast, if demand for the agricultural product is elastic (for example, in the case of an export crop), agricultural improvements would increase the po- tential profitability of agriculture on currently forested lands. The area of agri- culture would expand at the expense of managed and unmanaged forests, and agricultural progress would unambiguously increase deforestation. If agricul- tural intensification does not change the potential profitability of agriculture on currently forested lands (for example, because irrigation systems are installed in currently cultivated areas only), there would be no effect on forestry. Several other policies increase producer prices and, thus, lead to increased productivity of land use in either agriculture or managed forestry. In the case of export goods, devaluation of the national currency increases the profitability of agriculture, managed forestry, and logging of unmanaged forests. Devaluation therefore contributes to increased conversion of unmanaged forests. Road building increases the producer prices paid to farmers and foresters, particularly in more remote and, therefore, often unmanaged forest areas. Road building is particu- larly harmful to the conservation of unmanaged forests, increasing the profit- ability not only of alternative cultivation but also of logging itself (see also Chomitz von Amsberg 149 and Gray 1996). Although higher producer prices reduce logging waste, they also go along with more logging of unmanaged forests. Measures that reduce decisionmakers' discount rates include improved access to credit and more secure tenure. Lower discount rates unambiguously increase the area of managed forests because they reduce the cost of waiting for trees to mature. The effect on unmanaged forests at the agricultural margin depends on the profitability of logging. If logging is profitable by itself (logs are typically sold in the market), a lower discount rate slows the logging of unmanaged for- ests because it reduces the opportunity cost of leaving the timber standing in the forest. If logging is not profitable by itself (logs typically are not sold but are burnt), land clearing is an investment that has costs (labor, equipment) and is made for obtaining the benefits of alternative land use. A lower discount rate stimulates this investment and advances the logging of unmanaged forests. The latter situation is reported for parts of the Brazilian Amazon (see Schneider 1993). Empirical evidence that the availability of credit advances deforestation is also provided by Oz6rio de Alameida and Campari (1995), Barbier and Burgess (1996), Pfaff (1997), and Andersen (1997). At the frontier between unmanaged and managed forests, a lower discount rate can also lead to increased conversion if the higher returns to plantation forestry outweigh the reduction in opportu- nity costs of the standing unmanaged forests. Kishor and Constantino (1993) makes this point in a static context. IV. EXTENSIONS AND FURTHER RESEARCH This article suggests a new conceptual approach to the analysis of economic determinants of forestland use. However, the model has limitations that reduce its direct applicability to policymaking situations. One limitation of the model is the assumption of an exogenous log price path and, hence, the assumption that log output does not influence log prices. von Amsberg (1994) contains a simulation model with the same structure underly- ing the model, but with an endogenous log price path. The simulation model determines the profit-maximizing land use and profit-maximizing forestry effort for a finite number of land classes (representing lands with differing transporta- tion costs) for a finite number of time periods assuming that all land is covered with unmanaged forests in the first period. The analysis compares the supply of logs that results from these land-use choices with the demand for logs for the same price at different levels of demand elasticity. The model is rerun until a log price path is found at which the log market clears in all periods. This simulation model allows analysis of dynamic land use in a situation in which log prices respond to supply. Such a situation would be expected, for example, for a local fuelwood market or for a large log-exporting country. The simulations also illustrate the theoretical results of the basic model and could be used together with location-specific data to estimate deforestation effects em- pirically in specific real-life policy situations. In all simulations, the resulting iSO THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 equilibrium price path for logs shows the characteristic of declining rates of increase, consistent with the assumptions of the basic model with exogenous log prices. The simulations produce seven major results that are consistent with the theoretical results derived here. First, a tax on log sales (simulated by having the market clear for consumer prices that are equal to producer prices plus tax) leads to a reduction in the producer price compared to the base case. Consistent with the results of the analysis with the basic model, the reduced producer price path leads to a reduc- tion in logging of unmanaged forests and a decrease in the area with managed forests. The area with unmanaged forests increases, while the area with man- aged forests decreases. Agriculture contracts at the margin with unmanaged for- ests and expands at the margin with managed forests. Figure 2 shows the results of this simulation and the comparison with the base case. Second, a charge levied per area of unmanaged forests logged (like a Pigouvian tax for the reduction of external benefits from the standing natural forest) leads to a reduction in logging of unmanaged forests. This reduction in logging leads to a reduction in managed forests at the extensive margin. Log prices are some- what higher than in the base case, and the margin between agriculture and man- aged forests shifts in favor of managed forests. Third, a reduction in transportation costs (for example, as the result of road improvements) leads to increased pressure on the frontier and an expansion of agriculture and managed forest at the expense of unmanaged forests. The effect of road building on the log price path and logging intensity is ambiguous be- cause the reduction in transportation unit costs and the increase in distance due to increased logging operate in opposite directions. Fourth, an increase in agricultural productivity for a product with infinitely elastic demand (for example, exports of a cash crop from a small country) lead to an increase in agricultural area. The resulting increase in the log price path shifts the area of managed forests into the area of unmanagecd forests, which decline. At the other extreme, an increase in agricultural productivity for a prod- uct with inelastic demand (for example, a pure subsistence crop) leads to a de- cline in the agricultural area, a fall in log prices, and a reduction in the logging of unmanaged forests. Fifth, if demand for logs is highly elastic (the case of small timber-exporting countries), an increase in the productivity of managed forestry creates additional pressures to convert unmanaged forests. However, if demand for logs is inelastic (for example, where timber supplies fuelwood for the local market), increased supply of logs from plantations reduces the price of logs and, thus, reduces the pressure to convert unmanaged forests. As in the case of agriculture, demand for logs in a real-life situation is neither fully elastic nor fully inelastic. The resulting net effect from the introduction of plantations is ambiguous and depends on case-specific demand elasticities. In certain cases, the increase in productivity of managed forests increases the logging of unmanaged forests in the short run because of the additional demand for managed forestland. In the long run, von Amsberg 151 however, as production from managed forests enters the market, logging of unmanaged forests is reduced. In the theoretical case of total absence of man- aged forestry, logs are a nonrenewable resource with increasing extraction costs. In this case, the price path shows an increasing rate of price increase. Sixth, an increase in the decisionmaker's discount rate (for example, as a result of a reduced time horizon or more uncertain tenure) leads to an expansion of agriculture into managed forest areas because log prices are lower and the returns to forestry are better than the returns to agriculture due to the longer growth period for trees. However, logging of unmanaged forests increases only slightly if timber rents at the margin of unmanaged forests are relatively low or even negative. In these cases, clearing land is an investment that is less profitable with a higher discount rate. Increasing security of tenure alone does not drasti- cally reduce deforestation. Seventh, open access to the unmanaged forests drastically advances logging. In the long run, however, the remaining unmanaged forest area is the same with open access and secure property rights because in both cases all lands with posi- tive conversion profits are ultimately logged. Under open access, the log price is initially lower (because of excessive supply from still abundant forests), later higher (because excessive logging leads to higher transportation costs), and fi- nally equal to the case of secure property rights. Important additional research in three areas would strengthen the analysis. First, many of the parameter values that are used in the simulations could be estimated empirically for specific locations, as is done for a related model for the case of Belize by Chomitz and Gray (1996). This would allow quantitative pre- dictions to be derived for specific policy interventions. Second, a model of a forest as a stock of homogenous timber is clearly unrealistic. In particular, unmanaged forests consist of a variety of tree species with highly different eco- nomic values. Even though some of the qualitative effects of this heterogeneity of timber are captured in the production function for logs employed in this ar- ticle, a modeling approach closer to the physical realities of a natural forest would be desirable but would require additional empirical work. Finally, impor- tant economies of scale in land use, both internal (for example, lumpy invest- ments necessary for forestry and agriculture) and external (for example, mini- mum area for biodiversity conservation), are not addressed in the current model. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Andersen, Lykke. 1997. "Modeling of the Relationship between Government Policy, Economic Growth, and Deforestation in the Brazilian Amazon." Working Paper 2. Department of Economics, University of Aarhus, Denmark. Processed. Anderson, Terry L., and Peter Hill. 1990. "The Race for Property Rights." Journal of Law and Economics 33:177-97. 152 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Angelsen, Arild. 1996. "Deforestation: Population or Market Driven? Different Ap- proaches in Modeling of Agricultural Expansion." Working Paper 9. Chr. Michelsen Institute, Bergen, Norway. Processed. Barbier, Edward B., N. Bockstael, J. C. Burgess, and I. Strand. 1995. "The Linkages between Timber Trade and Tropical Deforestation-Indonesia." World Economy 18(3):411-42. Barbier, Edward B., and Joanne C. Burgess. 1996. "Economic Analysis of Deforestation in Mexico." Environment and Development Economics 1(2):203-39. Barbier, Edward B., and Michael Rauscher. 1993. "Trade, Tropical Deforestation, and Policy Interventions." Environmental and Resource Economics 4:75-90. Binswanger, Hans. 1989. "Brazilian Policies That Encourage Deforestation in the Ama- zon." Environment Department Working Paper No. 16. World Bank, Washington, D.C. Processed. Brandon, Carter, and Ramesh Ramankutty. 1993. Toward an Environmental Strategy for Asia. World Bank Discussion Paper 224. Washington, D.C.: World Bank. Chang, Sun Joseph. 1983. "Rotation Age, Management Intensity, and the Economic Factors of Timber Production: Do Changes in Stumpage Price, Interest Rate, Regen- eration Cost, and Forest Taxation Matter?" Forest Science 26(2):267-77. Chomitz, Kenneth M., and David A. Gray. 1996. "Roads, Land Use, arLd Deforestation: A Spatial Model Applied to Belize." The World Bank Economic Review 10(Septem- ber):487-512. Cropper, Maureen, and Charles Griffiths. 1994. "The Interaction of Population Growth and Environmental Quality." American Economic Review 85:250-54. Crossley, Rachel. 1993. "Log Export Bans: Are They Economically and Environmen- tally Beneficial? " Agriculture Department, World Bank, Washington, D.C. Processed. Deacon, Robert T. 1994. "Deforestation and the Rule of the Law in a Cross-Section of Countries." Land Economics 70(4):414-30. Deininger, Klaus, and Bart Minten. 1996. "Poverty, Policies, and Deforestation: The Case of Mexico." Policy Research Department, World Bank, Washington, D.C. Processed. FAO (Food and Agriculture Organization of the United Nations). 1992. Forest Resources Assessment 1990: Tropical Countries. Rome. Faustmann, Martin. 1968. "Calculation of the Value Which Forestland and Immature Stands Posses for Forestry." In M. Gane, ed., Martin Faustmann and the Evolution of Dis- counted Cash Flow, pp. 27-55. Commonwealth Forestry Institute Paper 42. Reprinted from an article originally published in 1849. Oxford: Commonwealth Forestry Institute. Goodland, Robert, and Herman Daly. 1994. "If Tropical Timber Export Bans Are So Perverse, Why Are There So Many?" Environment Department, World Bank, Wash- ington, D.C. Processed. Hyde, William F., Gregory S. Amacher, and William Magrath. 1993. "Deforestation, Scarce Forest Resources, and Forestland Use: Theory, Empirical Evidence, and Policy Implications." Rural Development Department, World Bank, Washington, D.C. Pro- cessed. Hyde, William F., and David H. Newman. 1991. Forest Economics and Policy Analysis: An Overview. World Bank Discussion Paper 134. Washington, D.C.: World Bank. Jackson, David H. 1980. The Microeconomics of the Timber Industry. Boulder, Colo.: Westview Press. von Amsberg 153 Kishor, Nalin M., and Luis F. Constantino. 1993. "Forest Management and Competing Land Uses: An Economic Analysis for Costa Rica." LATEN Dissemination Note 7. Latin America and the Caribbean Technical Department, World Bank, Washington, D.C. Processed. Ledyard, John, and Leon N. Moses. 1974. "Dynamics and Land Use: The Case of For- estry." In R. E. Grieson, ed., Public and Utility Economics. Lexington: Heath- Lexington. Lyon, Kenneth S. 1981. "Mining of the Forest and the Time Path of the Price of Tim- ber." Journal of Environmental Economics and Management 89(4):330-44. Mahar, Dennis J. 1989. Government Policies and Deforestation in Brazil's Amazon Region. Washington, D.C.: World Bank. Mendelsohn, Robert. 1994. "Property Rights and Tropical Deforestation." Oxford Eco- nomic Paper 46:750-56. Nautiyal, J. C., and Jeremy S. Williams. 1990. "Response of Optimal Stand Rotation and Management Intensity to One-Time Changes in Stumpage Price, Management Cost, and Discount Rate." Forest Science 36(2):212-23. Oz6rio de Alameida, Anna Luiza, and Joao S. Campari. 1995. Sustainable Settlement in the Brazilian Amazon. New York: Oxford University Press. Pfaff, Alexander S. 1997. "What Drives Deforestation in the Brazilian Amazon? Evi- dence from Satellite and Socioeconomic Data." Policy Research Working Paper 1772. Policy Research Department, World Bank, Washington, D.C. Processed. Repetto, Robert, and Malcolm Gillis, eds. 1988. Public Policies and the Misuse of Forest Resources. Cambridge, U.K.: Cambridge University Press. Schneider, Robert R. 1993. "Land Abandonment, Property Rights, and Agricultural Sustainability in the Amazon." LATEN Dissemination Note 3. Latin America and the Caribbean Technical Department, World Bank, Washington, D.C. Processed. . 1995. Government and the Economy on the Amazon Frontier. World Bank Environment Paper 11. Washington, D.C.: World Bank. Sharma, Narendra P., Simon Rietbergen, Claude R. Heimo, and Jyoti Patel. 1994. A Strategy for the Forest Sector in Sub-Saharan Africa. World Bank Technical Paper 251. Washington, D.C.: World Bank. Southgate, Douglas. 1990. "The Causes of Land Degradation along Spontaneously Ex- panding Agricultural Frontiers in the Third World." Land Economics 66(1):93-101. Thiele, Rainer. 1995. "Conserving Tropical Rain Forests in Indonesia: A Quantitative Assessment of Alternative Policies." Journal of Agricultural Economics 46(2):187- 200. Thiele, Rainer, and Manfred Wiebelt. 1994. "Policies to Reduce Tropical Deforestation and Degradation: A Computable General Equilibrium Analysis for Cameroon." Quar- terly Journal of International Agriculture 33(2):162-78. Vincent, Jeffrey R. 1990. "Don't Boycott Tropical Timber." Journal of Forestry 88 (4):56. von Amsberg, Joachim. 1994. "Economic Parameters of Deforestation." Policy Research Working Paper 1350. Policy Research Department, World Bank, Washington, D.C. Processed. von Thunen, Johann Heinrich. 1826. The Isolated State. New York: Pergamon Press. THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1: 155-74 Why Has Poland Avoided the Price Liberalization Trap? The Case of the Hog-Pork Sector Anning Wei, Waldemar Guba, and Richard Burcroff II Price liberalization in the agrifood economy in the transition economies is likely to slip into a trap: food prices rocket up, consumption declines, but food supply does not catch up and even contracts. However, during the transition period following the 1989 price liberalization, the Polish hog-pork sector succeeded in avoiding this trap. By conducting market structure and econometric analysis, this article looks for the rea- sons for this success. In the Polish hog-pork sector the restructuring of state-owned enterprises and the emergence of private firms introduced an effective price transmission mechanism be- tween the processing-retailing and farm levels. This mechanism allowed farm supply to respond to changed demand and to take advantage of increased retail prices. Such a relatively efficient marketing system was made possible by a relatively stable macro- economic environment and limited government intervention. In August 1989, the Polish government removed most price controls and subsi- dies in the agrifood sector. Immediately after price liberalization, food prices soared, and price margins between processed and raw agricultural products in- creased considerably. This was predictable because retail food prices had been suppressed under the planning regime to favor urban consumers. Polish policymakers and many international observers worried that escalating retail prices for food would become a driving force of hyperinflation and that state- owned or former state-owned processing enterprises might maneuver to enlarge price margins in their favor. In this event, farms would benefit little from price liberalization, and agricultural production would not improve despite high re- tail prices for food, which would dampen consumer demand. Concerns that Poland might fall into such a "price liberalization trap" were not without foun- dation. Reform of the agrifood sector in many transition economies did fall into this trap, and some such as Russia and the Ukraine struggle with this trap even today. This article is part of the project Determinants of Price Efficiency in Agrifood Markets of the Transition Economies, financed by the World Bank's Research Support Budget (RPo 680-14) and the Ford Foundation. Anning Wei and Richard Burcroff II are with the Rural Development Department at the World Bank. Waldemar Guba is with the Agricultural Policy Analysis Unit at the Ministry of Agriculture and Food Industry in Poland (MAFI). The authors appreciate the comments and suggestions from many colleagues and friends, particularly those of Derek Baker, Csaba Csaki, Michel Debatisse, Andrzej Kwiecinski, Richard Lacroix, Wladyslaw Piskorz, William Tomek, and three anonymous reviewers. © 1998 The International Bank for Reconstruction and Development/THE WORLD BANK 155 156 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Seven years after price liberalization, it is of interest to assess the extent to which the Polish agrifood sector has transcended the price liberalization trap. Poland is leading the transition process occurring in former socialist countries, and the experiences of Poland can shed light on reforms in other transition econo- mies still struggling to overcome the price liberalization trap. The key problem with price liberalization is unresponsive supply, particularly the supply of agricultural products from farms. Two factors may cause a slug- gish supply response. First, state-owned or collectively owned farms may not be very sensitive to market signals. Second, the food processing and marketing sys- tem and farm input supply industries may be dominated by monopoly forces- squeezing farms between high prices for their inputs but low prices for their outputs. Unlike most other transition economies, agriculture in Poland was domi- nated by private farms even prior to reform. In 1989 on the eve of reform, private farms worked 75 percent of all agricultural land (World Bank 1994). Therefore, to avoid the price liberalization trap, the primary challenge for Po- land lay in the marketing system. In a marketing system where the transactions across vertical marketing chains- from farmers to processors, processors to wholesalers, and wholesalers to retail- ers-are not regulated by competitive market forces, changes in consumer de- mand might not lead to significant changes in farmgate prices. Similarly, changes in farm production and farmgate prices might not be passed on to consumers through changes in retail prices. In the absence of competitive forces in process- ing and marketing, margins between farmgate and retail price: would be ex- plained not only by processing and marketing costs but also by monopoly inter- ventions. This was the situation in Poland before reform. Then, procurement and retail prices for most important foodstuffs were fixed, substantial subsidies were granted to processors and farmers, and segmentation across vertical mar- keting levels was high (Kwiecinski and Quaisser 1993). If a marketing system is segmented and the transmission of prices between farmgate and retail is ob- structed, price liberalization will not lead to an effective supply response. This article studies the behavior of hog and pork prices in post->reform Poland in order to determine what factors influenced the margin between the two prices. In particular, we are concerned with evidence of whether or not the Polish mar- keting system gave rise to an effective mechanism for transmitting prices among hog farmers, processors, and retailers. Although we must be cautious in general- izing to other agricultural industries, we consider the hog-pork industry more typical than atypical. In 1994 the agriculture and food processing industries together accounted for almost 14 percent of Poland's gross domestic product (GDP) and 29 percent of national employment (Central Statistical Office, Statis- tical Yearbook, 1995). Within agriculture, hog production is the single most important commodity, accounting for 20 percent of total agricultural produc- tion and 30 percent of marketed production. Within agricultural processing in- dustries, meat processing is the single most important activity., and pork is the dominant meat commodity, accounting for almost 23 percent of the value of Wei, Guba, and Burcroff 157 total production (OECD 1995).1 In addition to its economic significance, the hog- pork sector is among the most competitive agrifood markets in Poland and, in this sense, may be somewhat atypical, yet prototypical. Composed of a large number of private marketing and processing firms, it is subject to only limited government intervention. Examining this sector focuses on the most dynamic aspects of devel- opment of the agrifood market in Poland since price liberalization. Section I examines recent developments in the hog-pork marketing system relevant to the degree of competitiveness. Section II reports results of an econo- metric analysis designed to test whether the new marketing system has an effec- tive mechanism for transmitting prices among different market players. One reliable indicator of the price transmission mechanism lies in the process of de- termining the price margin. If there is an effective mechanism for transmitting prices, the movement of the margin between hog and pork prices will be ex- plained by various processing and marketing costs. Section III summarizes the analysis and offers policy recommendations indicated by our results. I. DEVELOPMENTS IN THE MARKETING SYSTEM AND GOVERNMENT INTERVENTION Livestock and meat production as a whole declined after 1989 as feed prices increased faster than output prices. The elimination of subsidies increased pro- duction costs, and the liquidity crisis of state farms, which occupy 20 percent of Poland's arable land, reduced supply from the state sector (World Bank 1994). On the demand side, decline in real per capita income limited the consumption of most meat products other than pork. In contrast to the general trend of re- duced production following price liberalization, Polish hog production increased from 18.8 million head in the first quarter of 1990 to 22.8 million head by the third quarter of 1993. This growth was spurred by higher demand for pork than for other meat and dairy products. Also, hog production traditionally relied heavily on homegrown feeds. Thus it was less influenced by increases in the price for manufactured feeds and by the phasing out of subsidies for feed con- centrate. Following drought and very low potato and grain harvests in 1992, pork production collapsed late in 1993. Production on private farms began to recover after 1993, and in 1995 it approached the peak level of 1992, only to decline again in 1996 (Central Statistical Office, Statistical Yearbook, 1996). The fluctuation of hog-pork supply indicated that farmers and processors were responsive to market signals at least to some extent. This was largely due to developments in the marketing system. The Marketing System The greatest changes in the marketing system were in the processing and re- tail sector. The privatization of state-owned retail stores and processing firms 1. In 1992 other important agricultural industries were potatoes (15 percent of total production), milk (14 percent), fruit and vegetables (11 percent), wheat (7 percent), poultry and eggs (7 percent), cattle (S percent), and sugar beets (3 percent). 158 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Table 1. Ownership Structure of the Agricultural Processing Industry in Poland, 1989-93 (number of firms of different kinds) Type of firm 1989 1990 1991 1992 1993 Enterprise 141 279 434 544 671 State-owned 37 59 67 76 71 Private 104 220 367 468 600 Sole proprietorship 18 359 1,467 3,385 7,934 Source: Poland, Central Statistical Office (1991-96a). and the large number of private processing and marketing firms that entered the agrifood industry brought about changes in the behavior of individual partici- pants and the relationships among them. These changes were at the center of the development of a new market system in Poland. As the era of central planning drew to a close in Poland, the meat processing subsector was dominated by large state-owned companies. Twenty-five plants were spread evenly throughout the country, and each was a monopoly force in its regional market. Since 1990, the sector has attracted large numbers of smaller, privately owned companies. State-owned firms have been required to undergo privatization or restructuring into separate functional components that are rented out to groups of private individuals. Table 1 describes this transformation up to 1993. Of the revenue generated by firms that employ more than 50 people, the share of state-owned firms declined from nearly 100 percent in 1989 to 60 per- cent in 1993. Had small firms with fewer than 50 employees also been included, the share of the state sector would have been less than 50 percent: because most small farms are privately owned (Grudzinska 1994). This trend continued after 1993 according to a field investigation we conducted in 1995 (Guba and others 1995) 2 The degree of industrial concentration in the meat industry is not high. Of the revenue generated by firms with 50 or more employees, the largest nine enter- prises accounted for only 20 percent. The 55 largest enterprises accounted for only 65 percent of revenues, leaving a large share for small private firms (Grud- zinska 1994). Processing firms primarily serve local demand. Nearby markets (within a ra- dius of 100 kilometers) account for 70 percent of sales of the average firm. Thus the domestic market remains largely regionalized. Field visits conducted in 1995 indicate, however, that interregional markets are becoming increasingly impor- tant for some processing firms and that private dealers (intermediaries) are be- coming a more effective source of raw materials for processors (Guba and oth- ers 1995). 2. We interviewed 40 farmers, traders, processors, and retailers on the subject of agricultural marketing in September 1995. Wei, Guba, and Burcroff 159 There is one organized livestock auction, which is run by PEKPOL in Radomsko. It has not been successful, and many participants now bypass it. The auction's poor success is partially due to the size of Polish farms. Many farms are so small that owners who visit the auction have few animals to sell. Competition in hog procurement is growing, and procurement channels for agricultural products have become more diversified since 1989. Direct contact between farmers and processors is the most common channel. A large process- ing firm typically has its own procurement stations, through which it buys some 60 percent of its requirements. A small, local meat plant may obtain up to 90 percent of its hog supply directly from farmers (Office of International Policy Service and Mitama Ltd. 1995). Whether the procurement agents are indepen- dent or associated with a processor, they must be willing to extend the geo- graphical scope of operations in order to obtain their hog supply. Meat processing plants are often vertically integrated enterprises. The plant will own a slaughterhouse to produce carcasses for further processing. The car- cass market itself is very thin and exists only when there are surplus carcasses that a processing plant cannot handle. Processing workshops produce different cuts of meat and other high value added products such as ham. They may also process beef as well as pork. Many meat processing companies have organized their own wholesale distri- bution systems in which they sell directly to retailers and sometimes consumers. It is estimated that processing firms control 30 percent of retail sales (World Bank 1994). As plants seek more cost-effective means of distributing their prod- ucts, direct suppliers have stepped up their activity, operating through small, private outlets. Most retail outlets in Poland were rapidly privatized at the beginning of the economic transition. In addition, new retail outlets have sprung up. Although retailers can now freely choose their suppliers, they tend to remain loyal to a particular plant once a link has been established. Compared with the processing and retail sectors, changes in the farming sec- tor took place in the marketing behavior of farmers rather than in the structure of farm ownership. With 75 percent of agricultural land in private hands even before the reform, privatization of state and collective farms was not as impor- tant in Poland as in other transition economies and proceeded rather slowly. The share of land owned by private farms increased from 76 to 80 percent from 1990 to 1994. Private farms are very small in scale, occupying on average about 6.5 hectares of agricultural land (OECD 1996). Hog production mirrors the struc- ture of landholdings. Small farms hold most of the hog population in small herds, and hog production is just one of many kinds of economic activities in which they engage. Large, specialized hog farms are rare. The size of private farms, in terms of both landholding and hog production, changed very little during 1990-95, as shown in table 2. About half of private farms operate with less than 5 hectares of land; about half of private hog farms raise fewer than six hogs. 160 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 Table 2. Size of Private Farms in Poland, 1990-95 (percent) Indicator 1990 1991 199 5 Area in hectares 1.01-1.99 17.7 - 20.9 2.00-4.99 35.1 - 33.7 5.00-6.99 14.9 - 13.4 7.00-9.99 14.9 - 13.4 10.00-14.99 11.3 - 10.7, 15.00 or more 6.1 - 8.0 Number of hogs held 2 or less - 30.3 30.4 3-5 - 21.1 20.3 6-10 - 16.7 16.0 11-20 - 18.2 16.9 21-50 - 11.3 12.2 50 or more - 2.4 4.5 -Not available. Source: Poland, Ministry of Agriculture and Food Industry (1996). It is not surprising that small private farms in Poland do not specialize. A 1993 survey of farms in Poland found that a typical private farm has a small or medium-size tractor and employs three workers with 10 years of education.3 Little hired labor, rented land, or debt financing are used. Farms tend to plant half their land with cereal and use the remaining space for pasture, feed, and cash crops. Farm herds typically contain less than 10 cattle and roughly 10 hogs. The ratio of marketed to total production is greater than 50 percent for crops, 76 percent for hogs, and almost 100 percent for dairy and cattle (Euroconsult and the Centre for World Food Studies 1994). These characteristics have changed very little during the seven years after price liberalization. However, the dramatic changes that occurred in processing and retail sectors placed private farms in a new market environment and changed their market behaviors. Farmers can now sell their animals to processors in the region, to intermediaries, or at farm markets, depending on which purchlaser offers the highest price. There are now four channels for sales: direct sales to slaughter- houses or sales to purchase units, dealers, or wholesale markets. Farmers sell 55 percent of their marketed livestock directly to slaughterhouses, 30 percent through collection stations owned by slaughterhouses or cooperatives, and 15 percent through intermediaries who visit local markets and farms. Less than 1 percent is sold through wholesale animal markets (Office of International Policy Service and Mitama Ltd. 1995). 3. The survey contained 116 households from the Plockie region and 100 from the Pilskie region. Together these two regions are representative of the national agricultural economy. Plockie is above the national average in terms of agriclimate, agricultural production, price of arable land, proportion of population engaged in agriculture, and share of private agricultural land. Pilskie is below the national average in these categories (Euroconsult and Centre for World Food Studies 1994: annex 1.4). Wei, Guba, and Burcroff 161 Only a few big farms use formal contractual agreements when selling di- rectly to processing firms. The simplest form of contractual agreement is one in which a farm sells its hogs to the processor at market prices prevailing at the time of sale. More complicated contracts specify the time, price, and quality of what will be delivered, as well as preferential credits provided by the processor such as piglets, concentrated feed, technical services, or transportation. Al- though formal contracts offer an additional degree of certainty that is attrac- tive to farmers, contracts with supplemental clauses are too costly for many processors to provide. Hog farmers are price-takers in the market. They can choose whom to sell to but cannot determine the price. Even large hog farms are not able to nego- tiate prices with buyers. The usual mode of operation is for big processing firms in a region to set procurement prices and for small processors to follow suit. Overall, Poland's new marketing system for hogs and pig products is diversi- fied and competitive. At each marketing level, there are many choices among channels. Farmers can sell to many different buyers, and retailers can purchase from many different processors. A relatively competitive market in the hog-pork sector permits prices at both retail and farm levels to fluctuate, leading various market players to adjust their level of activity. At the same time, market fluctuations and diversified marketing channels increase the risk for various market players, especially farmers, leading to discontent among farmers and concern by the government. The Agricultural Market Agency (AMA) has responded by intervening in the hog-pork market with procurement activities aimed at stabilizing farmers' income and preventing further declines in hog procurement prices. Government Intervention AMA initiates procurement when it considers market prices too low. Each of eight regional offices can contract with meat processing plants to pur- chase carcasses, and these plants then purchase live hogs from farmers in an assigned region at designated prices. AMA does not buy directly from farmers. The carcasses it buys are resold when the agency considers prices of retail pork too high. AMA also manages strategic reserves. Every year it sells a portion of what it has bought and stored and then purchases new car- casses to maintain its reserves. Because pig prices are seasonal, dropping to their lowest point in June and peaking from September to December, AMA makes its purchases during the second quarter of the year, when prices are depressed. AMA'S intervention is significant in some regions, in some years, and in some seasons, but not everywhere or all the time. The market mechanism is basically operational because price transmission is effective between the retail and farm level. This judgment is subject to more rigorous econometric verification in the next section. 162 THE WORLD RANK ECONOMIC REVIEW, VOL. 12, NO. I II. MOVEMENTS OF THE PRICE MARGIN AND ITS DETERMINANTS The first indication of a competitive market is that prices fluctuate in response to changes in supply and demand. From January 1990 to March 1996, monthly prices for both hogs and pork were driven by inflation and rose steadily,4 and the price margin in nominal terms between hog and pork increased substan- tially. Many complained that retail prices rose too fast, the price margin was too large, and both consumers and farmers were suffering as a result of price liberal- ization. However, if properly deflated, both retail and farm prices declined steadily in real terms. Moreover, retail prices fell even faster than farm prices. Contrary to the common impression derived from nominal price movements, the market margin actually declined, not increased, after January 1990, as figure 1 shows. Another important fact embedded in figure 1 is that farm prices moved in the same pattern as retail prices. Almost always, both prices moved in the same direction, although the magnitude of changes might be different. In a separate study, we conducted a cointegration test on the two series using fohansen's full information maximum-likelihood model and found that retail and farm prices were cointegrated. However, because retail and farmgate prices can be cointegrated even if monopoly forces dominate the marketing chain, we turn to an indirect, but more conclusive, test of competitiveness. If there is an effective process for transmitting prices between farmgate and retail, and no monopoly forces manipulate the prices, the margin between pork and hog prices will be determined by the costs of processing and marketing. If such costs cannot ex- plain observed margins, it is likely that monopoly forces are operative, which would impede the price transmission mechanism operating between farmgate and retail and prevent price liberalization from yielding an effective supply re- sponse. Moreover, examining the determinants of price margins, besides indi- cating if there is an effective price transmission mechanism, can shed light on other important considerations. Our study allows us to assess the impact of labor costs, material costs, government intervention, and, most important, mar- ket risk on price margins, which can indicate how market efficiency might be further enhanced. Determination of Price Margins Starting from the theory that price margins are the sum of marketing services (Tomek and Robinson 1990), we model the hog-pork price differential as a function of processing and marketing costs. In particular we follow the model and empirical work of Brorsen and others (1985).5 Assuming that firms exhibit decreasing absolute risk aversion (the more wealthy the firm, the less risk-averse it will be), Brorsen and others (1995) show that the solution of the firm's utility 4. Price data are derived from Central Statistical Office monthly price series. See the appendix for a full explanation of data sources. 5. Some other relevant studies are Buccola (1989), Gardner (1975), Holloway t1991), Kinnucan and Forker (1987), Shroeter and Azzam (1991), and Wohlgenant and Mullen (1987). Wei, Guba, and Burcroff 163 Figure 1. Movements of Real Prices of Hog and Pork in Poland, 1990-96 Price (zloty) 16,.D000 14.i000 V 12,000 - 10O D . .. .. - - - - - - - - [- - 4,000 - - - - - - - - -s - #--Hog_ 4.000 0 II +-4t++u ; D 1 1 & 1 i 2 :1 i : 1 1 1 iIII:1 i I I: II I: Ic 1 4 HP1.1i l1f l ii i990 1991 1992 1993 1994 1995 1996 Source: Poland, Central Statistical Office (1991-96a). maximization problem under competitive conditions and plausible assumptions regarding its production function implies that the expected margin, M, is a func- tion of w, the firm's initial wealth, q, a vector of input prices, 6, a measure of the uncertainty of the price the firm will receive for its output, and Y, the firm's total output. (1) M = M(w,q,a,Y). Moreover, their comparative static analysis shows: (2) am > (3) a>0 (4) am> or <0 aq Equation 2 implies that output should be positively related to price margin. Equation 3 implies that an increase in output price uncertainty should unam- biguously increase the price margin. And equation 4 implies that a change in input price can have either a positive or a negative impact on price margin. 164 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Based on this theoretical model, Brorsen and others (1995) estimate regressions to determine if increases in the uncertainty of free on board (FOB) mill prices and retail prices of wheat flour increase farm-to-mill and mill-to-retail wheat price margins, respectively. This article adopts Brorsen's framework with certain modifications. Process- ing and marketing costs are divided into two categories, labor and material cost. This refinement enables us to determine if changes in wages are the driving force behind enlarged price margins. The monthly wage rate published by the Central Statistical Office for the agricultural processing industry is used for labor costs. Because no wage rates for the agricultural processing industry are available for 1990, the wage rate for the whole processing industry is used as an approxima- tion for that year. Exact data are not available for material costs. The price index of fuel and lubricant oil is used as a proxy for material costs of processing and marketing. To make the measurement of labor cost comparable to material cost, the monthly wage rate is converted into an index using January 1990 as the base. Both the wage and material cost indexes are then deflated by dividing them by the consumer price index (cpi). L denotes the ratio of the wage index to the cPI, and C denotes the ratio of the fuel index to the cpi. In transition economies, it is likely that macroeconomic risk strongly affects all markets, so that using price uncertainty in a given market may underestimate the real risk that agents face. Macroeconomic risk can reflect uncertainty in input and financial markets as well as in output markets. Also, the correlation between individual output price and macroeconomic risks is likely to be high. For these reasons, macroeconomic uncertainty, instead of pork price uncertainty, is used. The inflation rate is the most readily available indicator of macroeconomic uncertainty. At time t (a given month) macroeconomic risk is measured as the standard deviation of (cPIt-2,cPit-1,cPI,cPII+1,cPI1+2). In other words, it is assumed that the pork industry's perception of macroeconomic risk is based on the past two months' experiences, current observation, and expectation for the next two months. The magnitude of the standard deviation is positively related to the absolute level of cPi; that is, the standard deviation tends to be large when the cpI is high. But this may not be true for risk measurements, because when the level of cpi is high for the period but the variation is small, the resulting high standard deviation exaggerates the risk the industry actually Eaces. Following common statistical practices, the standard deviation is divided, by the mean of the observations, which gives a coefficient of variation free of the biases caused by variation in the level of the cPi. This measurement is denoted as R. No data are available for monthly pork production, but monthly hog pro- curement data are available. Because there is generally a shortage of hogs for pork processors and there is practically no independent market for carcasses in Poland, monthly hog procurement reasonably approximates monthly pork pro- duction; therefore, we denote volume of output by Y. Wei, Guba, and Burcroff 16S Finally, it is important to include the impact of AMA interventions in hog pro- curement. A dummy variable is used to represent government intervention. In months when AMA intervened in procurement, the dummy variable is 1; other- wise it is 0. In sum, the price margin, M, is presumed to be a function of the marketing and processing labor cost, L, material cost, C, macroeconomic risk, R, pork output, Y, and government intervention, I. (5) Mt = f(L, C, R,Y,I,) Table 3 provides basic information on the variables in equation 5. Comparing the minimum with the maximum and the standard deviation with the mean, M, C, L, Y, and R all displayed significant variation between 1990 and 1996. Varia- tion in our measure of macroeconomic risk, R, was particularly great where the maximum was 10 times the minimum. The mean of the ratio of the material cost index to the cpi is less than 1, which means that our proxy for input prices in- creased slower than the general rate of inflation. Similarly, since the mean of the ratio of the wage index to the cPi is greater than 1, wage rates in agricultural processing industries increased faster than the general rate of inflation. AMA in- tervened in 21 out of 74 months, consistent with their policy of buying only at certain times, usually during the first half of the year. Because nonstationarity of variables can lead to spurious results, augmented Dickey-Fuller stationarity tests are conducted on each variable. All are weak-form stationary, meaning that they have finite and constant means and variances. Assuming that there is a linear relationship between the price margin and the five independent variables, equation 5 can be further specified as: (6) M,= a+bL,+b2C,+b3R,+b4Y +b5I,+e, Table 3. Descriptive Statistics for Poland, 1990-96 Standard Variable Mean deviation Minimum Maximum Price margin in real terms,a M 5,461.4 1,177.6 4,245.1 8,784.9 Ratio of fuel index to consumer price index, C 0.698 0.085 0.540 0.920 Ratio of wage rate index to consumer price index, L 1.169 0.092 0.977 1.356 Hog procurement,b Y 70.2 15.9 43.4 105.8 Coefficient of variation of consumer price index, R 0.048 0.023 0.010 0.137 Note: There were 76 monthly observations for January 1990 to April 1996. However, construction of the coefficient of variation of consumer price index (R) limited the study to only 72. a. The unit of price margin is zloty per kilogram. The nominal prices of hog procurement and pork retail are deflated by the consumer price index. The price margins are the difference between the two prices. b. The unit of hog procurement is 1,000 tons live weight. As discussed in the text, hog procurement is used as a proxy for total pork production. Source: Authors' calculations. 166 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I where a and b are coefficients, and e is an error term. Equation 6 is estimated on the monthly data. Although one month might be long enough for farms or firms at different marketing levels to finish adjusting to market signals, the possibility cannot be excluded that the impacts on price margins of changes in some pro- cessing or marketing costs may last more than a month. In this case, certain lagged independent variables should be included. However, there is no indica- tion which, if any, independent variables have effects lasting longer than one month, making dynamic specifications difficult and ad hoc. We follow the sug- gestion of Pindyck and Rubinfeld's (1991) to use a combined regression-time- series model under such circumstances. Equation 6 is a structural model, and the error term e in equation 6 may contain some variance in M that is not explained by the five structural vari- ables. An autoregressive and moving average (ARMA) process can be performed on the series of et and be incorporated into equation 6. This ARM.A process can help explain any variance in M that cannot be explained structurally. The analy- sis of autocorrelation and partial autocorrelation of the series of et suggests that a first-order autoregressive process is sufficient to capture effects not ex- plained by the structural model. The short autoregressive result rends to imply that most of the price transmission between farm and retail is accomplished within a month and is captured by the structural model. In any case, equation 6 is appended as: (7) Mt = a + b1L, + b2Ct + b3R, + b4Yt + b5lt + T + t-l where -t + Oil,, is the autoregressive process of et in equation 6. Equation 7 itself is neither a moving average nor an autoregressive process, it is a combination of a structural model and a time-series structure. An auto- regressive process is applied to the residuals of equation 6 and is introduced to capture changes in price margins that cannot be explained by the structural vari- ables (L, C, R, Y, and I). Estimation Results The estimation of equation 7 takes two steps. First, equation 6 is estimated to obtain the series et. Second, equation 7 is estimated using ordinary least squares to estimate simultaneously the structural and time-series part of the model as suggested by Pindyck and Rubinfeld (1991). This two-step estimation is done first for the whole time period, January 1990 through April 1996. However, price liberalization introduced a big shock, and all variables were highly volatile during 1990-91 but calmed clown starting in 1992. Therefore, the time period is split in two-January 1990 through Decem- ber 1993 and January 1992 through April 1996 (1992-93 is common to both periods). Significant differences between the initial and long-term reactions to price liberalization, if any, should be revealed by differences in the equations estimated for the two time periods. Wei, Guba, and Burcroff 167 Diagnoses of residuals of the three estimations show that there is a multicorrelation problem. Ordinary least squares estimators in the presence of multicollinearity remain unbiased and efficient, and the major undesirable con- sequence is that the variance of coefficients is exaggerated. However, in all three equations, the R2 for the regression exceeds the R2 of any independent variable regressed on the other independent variables. For this reason, no further correc- tive steps are deemed necessary (for a justification, see Kennedy 1993). The Durbin-Watson statistic cannot be used to test the existence of autocorrelation with the lagged error term included in equation 7. Instead, the Breusch-Godfrey Lagrange Multiplier test is conducted (for the testing proce- dure, see Maddala 1992). The results appear in the last row of table 3. In all cases, the null hypothesis of no further autocorrelation is accepted at a 1 percent significance level. All the estimation results are consistent with the theoretical model developed by Brorsen and others (1985); that is, all the coefficients have signs predicted in equa- tions 2 through 4. The coefficient of the lagged error term, 0, is statistically signifi- cant in all three cases, which means that some dynamic effects are left out of the structural model. Including the lagged error term improves the explanatory power of the model. The adjusted R2 of the model without the lagged error term is 0.73 for the whole period, 0.69 for the first period, and 0.59 for the second period. Results for the whole period, the first period, and the second period are shown in table 4. For the whole period, the coefficients of all independent variables, except material costs, are highly statistically significant, and R2 for the model is 0.8. The structural variables explain most of the variation in price margins, which implies that the market mechanism is functional. Most notable, macroeconomic risk, measured by the variation coefficient of the cpi, has a large, highly significant, positive coefficient. The large magnitude of the coefficient suggests that the price margin is very sensitive to macroeco- nomic risk. If any single factor should be blamed for enlarging price margins after price liberalization, it would appear to be macroeconomic risk. Pork output, Y, has a positive and statistically significant coefficient, which is also consistent with Brorsen's theoretical model. An increase in quantity mar- keted is correlated with an increase in the market margin. However, the output coefficient is not significant for the second time period. The coefficient of labor cost, L, is statistically significant and negative except in the second period. As mentioned before, the wage rate has consistently in- creased faster than the cpi, driving up processing and marketing costs. Signifi- cant negative coefficients of the AMA dummy variable, I, indicate that AMA inter- vention did reduce the price margins of pork products. AMA procurement interventions raised hog procurement prices at critical junctures, which proces- sors apparently were not able to pass entirely on to consumers, meaning that some degree of competitiveness is evident in pork processing and retailing mar- kets. The coefficient of material costs, C, is statistically significant. 168 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Table 4. Market Margin Determination of the Hog-Pork Market in Poland, January 1990-April 1996 January 1990 to January 1990 to Jajuary 1992 to Dependeitt variable April 1996 February 1993 April 1996 Material cost 500.30 -1,506.5 1,586.3*** (0.45) (-0.67) (2.92) Labor cost -3,305.1`** -3,951.6.*. -112.18 (-4.34) (-3.64) (-0.22) Agricultural Marketing -459.5*** -487.1* -164.4* Agency intervention (-2.93) (-1.75) (-1.96) Pork output 28.0*** 30.9*** 3.45 (5.51) (4.60) (1.04) Variation coefficient of 18,328.0*-$* 24,436*** 8,535** consumer price index (4.75) (4.04) (2.76) Lagged error term, 6 0.52*** 0.57*** 0.48*** (4.56) (3.75) (4.58) Constant 6,257.1 ** 7,952.2*** 3,378.7*** (4.81) (4.19) (4.48) Observation 72 45 49 Adjusted R2 0.80 0.77 0.58 Breusch-Godfrey lagrange multiplier test 0.020*-* 0.335*** 0.275*** Note: t-ratios are in parentheses. Significant at 10 percent. Significant at 5 percent. * Significant at 1 percent. Sozurce: Authors' calculations. The estimation for the first period produces similar results as that for the whole time period. The sign of all statistically significant coefficients is the same, although the magnitude is somewhat different, as would be expected. However, the estimation for the second period provides rather different results. The coef- ficient of material cost is positive and statistically significant, while the coeffi- cient of labor costs is no longer significant. The positive coefficient of material cost means that an increase in material costs enlarged market margins in the second period. It might be the case that since consumers' income recovered from the deterioration at the beginning of transition, and the demand for meat prod- ucts became less price-elastic, processors and retailers were able to pass on the cost increases to retail prices. It is noteworthy that the magnitude of the coeffi- cient of macroeconomic risk is reduced substantially in the sec:ond period. It is only one-third of that in the first period. Of course, as economic stability was achieved, macroeconomic risk was reduced. But it also appears that the impact of macroeconomic risk on market margins declined considerably as well. This may be because market players became more able to bear and absorb risk so that market prices and margins were less likely to overreact to risk than before. Or it is possible that the dominance of macroeconomic risk over sectoral mar- kets decreased, so that the measurement of macroeconomic risk covered a smaller proportion of the actual risk facing a hog-pork marketing firrn than before. Wei, Guba, and Burcroff 169 III. CONCLUSIONS AND DISCUSSION Price liberalization in the agrifood sector of transition economies is likely to slip into a trap: food prices rocket up, but if higher retail prices are not transmit- ted to higher farmgate prices, food supply does not respond, retail prices remain high, consumer demand declines, and production stagnates or even contracts. Fortunately, the hog-pork market in Poland appears to have avoided this trap. After the 1989 price liberalization, real prices for hog and pork did not increase; they declined. The supply of live hogs and pork kept rising until 1992. After 1993, the hog-pork market adjusted to the impacts of the 1992 drought and the shift of consumers' preference from meat products to vegetables and fruits.6 Thus the hog-pork sector safely crossed over the price liberalization trap. The most important factor contributing to this success would seem to be the restructuring of the pork processing and retailing industries. Private firms blos- somed, and state-owned firms were subjected to competitive pressures. Corre- spondingly, the market behavior of processors and retailers became active, ag- gressive, and profit-oriented. Second, although most private farms are operated on a small scale, and their degree of specialization and commercialization re- mains low, once situated in a competitive market they explored diversified mar- keting channels and reacted actively to market signals. Emergence of a competitive system allowed price transmission across market- ing levels to become a feature of the hog-pork industry in Poland, as confirmed by the results of this article that the price spread between raw and processed pork products was determined mostly by market forces rather than being ma- nipulated by monopoly forces. The smooth price transmission mechanism pre- vented the farming sector from being isolated from the retail food market. Farms benefited from increased competition in the processing and retail industries and received increases in farmgate prices commensurate with increases in retail prices of pork. Price liberalization succeeded in eliciting increases in farm supply. Except for AMA procurements, there is little significant government interven- tion in the hog-pork sector. Our results indicate that AMA intervention does ben- efit farmers by increasing farmgate prices of live hogs when the supply of hogs is strong, without dominating, much less superseding, the marketing system. Most important, this study of margin determination showed that the price spread between raw and processed pork products is very sensitive to macroeco- nomic risk. An increase in macroeconomic risk substantially increases the price spread. Apparently macroeconomic uncertainty has profound cost implications for processing and marketing firms. High risk leads to high costs, and high costs prevent the price transmission mechanism from working efficiently among farm- ers, processors, and retailers. The Polish government has an effective stabiliza- tion program, and the inflation rate declined from 586 percent in 1990 to 70 percent in 1991 and to 22 percent in 1995 (OECD 1996). There is good reason to 6. The share of vegetable fats in total fat consumption increased from 33 percent in 1990 to about 60 percent in 1995 (OECD 1996). 170 THE WORLD BANK ECONOMIC REVIEW, VOL. 12. NO. 1 believe that the stable macroeconomic environment made it possible for the marketing system to operate effectively and helped the hog-pork sector to avoid the price liberalization trap. Gains from an even more stable macroeconomic environment should not be underestimated. The case of hog-pork markets in Poland is a concrete example of the benefits of global stabilization policies. This article focused on vertical marketing chains in the hog--pork sector. An- other study analyzes spatial market integration in the hog-pork sector and finds that regional hog-pork markets are well linked. More important, these linkages are well explained by economic variables such as transportation costs and dif- ferences in demand and supply among regions, leading to the conclusion that hog-pork markets in Poland are spatially integrated. The hog-pork sector in Poland is not alone in avoiding the price liberalization trap. Our other studies (Wei and others 1996; Wei, Guba, ancL Krzyzanowska 1997) show that the marketing system functions fairly well in the wheat, flour, and bread sector and functions very well in the dairy sector-for similar reasons as in the hog-pork sector. Competitive processing and retail industries, market-responsive private farms, limited and indirect government intervention, and a stable macroeconomy ap- pear to be the major factors that allowed the Polish hog-pork sector to avoid the price liberalization trap. This contrasts sharply with conditions in other transi- tion economies. In the Ukraine, state-owned processing enterprises were priva- tized very slowly, and little private capital entered the industry. Moreover, al- though state and collective farms were restructured, the new ownership and management conditions do not produce effective incentives, and farms are still not sensitive to market signals. Finally, the macroeconomic environment in the Ukraine is still very volatile, creating high risks for agrifood enterprises, which this article implied can be a major impediment to effective price transmission. It is hardly surprising that the Ukrainian agrifood sector is still locked in the price liberalization trap. Although reform in the hog-pork sector in Poland succeeded in avoiding the price liberalization trap, its successes should not be overstated. The hog-pork marketing system continues to be characterized by a large number of very small private farms, underdeveloped intermediary markets, and large, vertically inte- grated processing enterprises capable of engaging in predatory pricing. Also, the state has demonstrated a continued willingness to intervene. Its favored parastatal (the AMA) continues to enjoy privileged access to finance and an ability to thwart the dynamism of the private sector. And the domestic hog-pork market is still protected by high tariffs. To achieve a higher level of marketing efficiency, the first step is to overcome the size limitations of small private farms. Recent interviews found that private farmers and, more surprisingly, processors considered that only modest increase in farm size was possible over the next 15 years (Guba and others 1995). Nor can the imbalance in market power between processors and farmers be resolved by decreasing the scale of processing plants. In fact, economies of scale, if any- Wei, Guba, and Burcroff 171 thing, warrant further consolidation in the processing industry. One way to re- solve the imbalance between size and power would be to develop new farmer cooperatives (usually termed "marketing groups" to avoid unwanted associa- tions with the former system of cooperatives). A few hog farmers are in the process of organizing new cooperatives in order to (a) purchase production in- puts collectively, especially feed, (b) sell pigs to processors as a group, (c) share technology among members, and (d) construct feed production facilities for the group. Organized into marketing and service cooperatives, farmers would be better able to bargain with processors over price and would have better access to mar- ket and technical information. Organization could help compensate for the se- vere size disadvantage presently facing small, individual farms. Meat processing plants complain about the uncertain quality of animals, the timing of supply delivery, and the high unit costs of dealing with so many small farmers, which explains why processors concur with farmers over the desirability of farmer organization. Cooperatives may indeed provide a solution beneficial to both parties. Several processors have even expressed willingness to assist farmers in forming marketing groups. Primary wholesale markets (from farms to processors) and secondary whole- sale markets (from processors to retailers) are absent in the Polish hog-pork marketing system. However, development of wholesale markets is more difficult due to the large number of small production units in both farming and process- ing. Wholesale markets could put useful pressure on processing firms. The qual- ity and health requirements for products could improve if trade were conducted by multiple participants in primary wholesale markets. Also changes in con- sumer demand might be communicated to processors more readily by secondary wholesale markets. The quality of market information should also improve. If new farm cooperatives were established and wholesale markets developed, AMA would have less reason to intervene to stabilize the market. Unlike most devel- oped countries where wholesale markets have been created by local governments and operated as nonprofit organizations, there has been little action of this sort in Poland. Farmers, processors, and traders alike have generally approved of AMA'S price stabilization interventions but have been somewhat critical of the price levels chosen and the late timing of interventions. AMA decisionmaking procedures should be improved. Reserve management and intervention activities often run counter to each other and need to be separated. Although the total amount of reserves need not be disclosed to the public, purchase and resale should be more transparent to market participants. In this way, market uncertainty would not be exaggerated by AMA's reserve management interventions. AMA uses preferen- tial credits for its intervention activities and enjoys a de facto monopoly on most kinds of bulk and cold storage. These practices should be phased out. In the beginning of the transition, many analysts and industrialists in Poland argued that imports should be restricted in order to protect a weak domestic 172 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. 1 market. More than seven years into the transition, the hog-pork sector has ma- tured and should now be capable of meeting international competition. Poland will join the European Union at some point in the future. Bridging the gap be- tween domestic and international markets, and not delaying until the last mo- ment to do so, is a high-priority task. APPENDIX. DATA SOURCES All data used in the analysis come from the Central Statistical Office in War- saw and cover national and regional procurement prices, retail prices, procure- ment, production, and consumption volumes, population, and income. Procurement Data Monthly procurement data are published regularly by the Cen-tral Statistical Office. The first level of aggregation is performed by procurement agents who cover monthly activity. Those agents are defined as "economic units making procurement a part of their commercial activity." As such, they a re obligated to report to the local voivodship statistical office. Once aggregated, the data are passed to the Central Statistical Office in Warsaw. Procurement units report value and volume of procurement, and this information makes it possible to calculate the weighted average at each aggregation level. This procedure reflects procurement characteristics (value, volume, price) made by local agents, includ- ing transactions outside a particular voivodship. Twice a year, recalculation is done to present the volume of procurement from local producers. Retail Data The Central Statistical Office collects prices on 1,007 consumner goods and services to monitor the inflation rate and calculate the consumer price index. The data collection procedure includes 307 representative regions (cities and gminas); 28,000 retail outlets including local markets are selected and moni- tored once a year. For the purpose of this study, retail data have been updated. Original data used to construct variables came from Central Statistical Office statistical reports. REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Brorsen, B., W. J. Chavas, W. R. Grant, and L. D. Schnake. 1985. "Marketing Margins and Price Uncertainty: The Case of the U.S. Wheat Market." American Journal of Agricultural Economics 67:521-28. Buccola, S. T. 1989. "Pricing Efficiency in Agricultural Markets: Issues, Methods, and Results." Western Journal of Agricultural Economics 14:111-21. Wei, Guba, and Burcroff 173 Euroconsult and Centre for World Food Studies. 1994. "Farm Restructuring and Land Tenure in Reforming Socialist Economies: A Comparative Analysis of Eastern and Central Europe." Agriculture and Rural Development Department, World Bank, Wash- ington, D.C. Processed. Gardner, B. L. 1975. "The Farm-Retail Spread in the Competitive Food Industry." American Journal of Agricultural Economics 57:399-409. Grudzinska, Agnieszka. 1994. "Transition of Socialist Agriculture and Agri- Processing: The Case of Polish Meat Processing Industry." Discussion draft. Agricul- ture and Natural Resources Department, World Bank, Washington, D.C. Pro- cessed. Guba, Waldemar, Anning Wei, Jerzy Dabrowski, and Wanda Chmielewska. 1995. "Emerging Trends in Polish Agri-Food Markets: A Report of Field Investigation." Agriculture and Natural Resources Department, World Bank, Washington, D.C. Pro- cessed. Holloway, G. J. 1991. " The Farm-Retail Price Spread in the Imperfectly Competitive Food Industry." American Journal of Agricultural Economics 73:979-89. Kennedy, P. 1993. A Guide to Econometrics, 3d ed. Cambridge, Mass.: MIT Press. Kinnucan, H. W., and 0. D. Forker. 1987. "Asymmetry in Farm-Retail Price Transmis- sion for Major Dairy Products." American Journal of Agricultural Economics 69:285- 92. Kwiecinski, Andrej, and Wolfgang Quaisser. 1993. "Agricultural Prices and Subsidies in the Transformation Process of the Polish Economy." Economic Systems 17(2):125-54. Maddala, G. S. 1992. Introduction to Econometrics, 2d ed. New York: Macmillan Press. Office of International Policy Service and Mitama Ltd. 1995. Masterplan for Wholesale Marketing. FAPA project P9205-05-20. Warsaw. OECD (Organisation for Economic Co-operation and Development). 1995. Review of Agricultural Policies: Poland. Paris. .1996. Agricultural Policies, Markets, and Trade in the Central and Eastern Eu- ropean Countries (CEEus), the New Independent States (Nis), Mongolia, and China- Monitoring and Outlook 1996. Paris. Pindyck, R. S., and D. L. Rubinfeld. 1991. Econometric Models and Economic Fore- casts, 3d ed. New York: McGraw-Hill Press. Poland, Central Statistical Office. 1991-96a. Statistical Bulletin. Various issues. War- saw. . 1991-96b. Statistical Yearbook. Various issues. Warsaw. Poland, Ministry of Agriculture and Food Industry. 1996. Agriculture and Economics Policy Unit (SAEPR). Warsaw. Shroeter, J., and A. Azzam. 1991. "Marketing Margins, Market Power, and Price Un- certainty." American Journal of Agricultural Economics 73:990-99. Tomek, W. G., and K. L. Robinson. 1990. Agricultural Product Prices, 3d ed. Ithaca, N.Y.: Cornell University Press. Wei, Anning, Waldemar Guba, Richard Burcroff II, Jerzy Dabrowski, and Wanda Chmielewska. 1996. "Market Efficiency and Government Policy in the Polish Wheat Market." Paper presented to the European Association of Agricultural Economist conference, Edinburgh, U.K. Processed. 174 THE WORLD BANK ECONOMIC REVIEW, VOL. 12, NO. I Wei, Anning, Waldemar Guba, and Z. Krzyzanowska. 1997. "Market Development and Government Policy in Polish Dairy Sector." Discussion draft. Agriculture and Natural Resources Department, World Bank, Washington, D.C. Processed. Wohlgenant, Michael K., and John D. Mullen. 1987. "Modeling the Farm-Retail Price Spread for Beef." Western Journal of Agricultural Economics 12:119-25. World Bank. 1994. "Poland Livestock Sector Review: Constraints and Opportunities." Central Europe Department, Washington, D.C. Processed. Coming in the next issue of TlHE WORLD BANK ECONOMIC REVIEW May 1998 Volume 12, Number 2 * Half a Century of Development Economics: A Review Based on the Handbook of Development Economics by Jean Waelbroeck A symposium on regionalism and development, including... * An Introduction by Maurice Schiff and L. Alan Winters * An Analysis of Nontraditional Gains from Regional Trade Agreements by Raquel Ferndndez * Endogenous Tariff Formation: The Case of Mercosur by Marcelo Olarreaga and Isidro Soloaga * Trading Arrangements and Industrial Development by Diego Puga and Anthony J. Venables * Regional Arrangements and Industrial Development by Maurice Schiff and L. Alan Winters * Regional Integration and Economic Growth by Athanasios Vamvakidis /j/3/2 /J7