77490 Has Rural Infrastructure Rehabilitation in Georgia Helped the Poor? Michael Lokshin and Ruslan Yemtsov This article proposes a research strategy to deal with the scarcity of data on benefici- aries for conducting impact assessments of community-level projects. Community-level panel data from a regular household survey augmented with a special community module are used to measure the impact of projects. Propensity score–matched difference- in-difference comparisons are used to control for time-invariant unobservable factors. This methodology takes into consideration the purposeful placement of projects and their interactions at the community level. This empirical approach is applied to infrastructure rehabilitation projects—for schools, roads, and water supply systems—in rural Georgia between 1998 and 2001. The analysis produces plausible results regarding the size of welfare gains from a particular project at the village level and allows for differentiation of benefits between the poor and the nonpoor. The findings of this study can contribute to evaluations of the impact of infrastructure interventions on poverty by bringing new empirical evidence to bear on the welfare and equity implications. A frequent problem in evaluating the impact of projects in developing econo- mies is the lack of data. A further complication is the increasing number of projects that target communities rather than individuals and rely on demand- driven placement, requiring special evaluation techniques and good-quality data to obtain robust results. Despite these difficulties, researchers are often called on to provide ex post assessments of a project’s impact. This article develops a strategy for meeting such requests with minimum data. Interest in evaluating the effectiveness of community-based infrastructure projects has grown in response to the increasing popularity of such programs. Michael Lokshin is senior economist in the Development Economics Research Group at the World Bank; his email address is mlokshin@worldbank.org. Ruslan Yemtsov is senior economist in the Europe and Central Asia Poverty Reduction and Economic Management unit at the World Bank; his email address is ryemtsov@worldbank.org. The research for this study was conducted as a part of the analytical work for the Georgia Poverty Update. Support from the Poverty Reduction and Economic Management Poverty Reduction Group trust fund is gratefully acknowledged. We thank Louise Cord, Martin Ravallion, and Dominique van de Walle for useful comments. We are also grateful for the support and contributions of Nodar Kapanadze, Alexander Kolev, and Zurab Sajaia. Special thanks to the Georgia Social Investment Fund team for making a wealth of information available. THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2, pp. 311–333 doi:10.1093/wber/lhi007 Advance Access publication August 31, 2005 Ó The Author 2005. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oupjournals.org. 311 312 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 Jalan and Ravallion (2003), Lee and others (1997), and Brockerhoff and Derose (1996) analyze improvements in water and sanitation facilities. Glewwe (1999), Hanushek (1995), and Kremer (1995) evaluate the impacts of school infrastruc- ture rehabilitation projects. Jacoby (2002) and van de Walle and Cratty (2002) analyze the effect of improvements in access to roads. For the most part researchers have had to apply matching techniques combining samples of ben- eficiaries with samples from regular household surveys, and they have used panel data or instrumental variables to deal with biases arising from the non- random placement of a project under evaluation. Following this practice, the empirical approach proposed here applies pro- pensity score–matched difference-in-difference comparison between project beneficiaries and a control group to purge biases arising from time-invariant unobservable community characteristics that might affect project outcomes. The approach also employs several innovations. First, a carefully designed commu- nity survey with sufficiently long recalls is used to compensate for the lack of baseline data. Second, repeated cross-sections from a household survey are aggregated to the community level to obtain longitudinal observations at the community level, with the community treated as a unit of observation. Third, the analysis explicitly considers all infrastructure micro-projects in every com- munity. Fourth, the methodology is applied to assessment of infrastructure rehabilitation projects rather than to the construction of new facilities, the focus of most studies so far. This approach is applied to data from Georgia, one of the poorest countries of the former Soviet Union. The analysis is conducted for all community-level infra- structure rehabilitation projects—school, roads, and water supply systems—in rural areas between 1998 and 2000. The proposed evaluation strategy produces defend- able results on the size of welfare gains from a particular project at the village level and can differentiate gains to the poor and gains to the nonpoor. The limitations of the approach are not trivial, however. It produces reliable results only for projects affecting a significant fraction of the population—projects large enough for a regular household survey to register their impacts. Consequently, the proposed strategy is probably best suited for large-scale projects deployed with little regard for the need for subsequent evaluation, such as disaster response operations or decentralized public investment programs in developing areas. I. COMMUNITY-LEVEL INVESTMENT PROJECTS IN GEORGIA Economic and political turmoil have led to a dramatic fall in living standards in Georgia, once one of the richest republics of the former Soviet Union. Indepen- dent Georgia inherited developed infrastructure facilities, but these have been deteriorating rapidly. Rural areas have been hit particularly hard, suffering from increasing economic marginalization and impoverishment (World Bank 2003b). The decay of infrastructure providing public services has resulted in deteriorat- ing nonincome indicators, particularly those related to child welfare—school Lokshin and Yemtsov 313 enrollment rates have fallen, maternal mortality ratios have risen, and infant mortality rates have remained high (World Bank 2003a). Georgia has few resources available for rehabilitating its severely decayed infrastructure. Donor funds have been required to finance basic maintenance of roads, repair of water and sanitation systems, and urgent rehabilitation of school facilities. In the early 2000s as many as 218 donor organizations were actively involved in such projects (UNOCHA 2003). Because of the backlog in deferred and neglected maintenance, the govern- ment and donors will continue to face difficult tradeoffs between capital invest- ments and spending on critical current needs. To make such choices, it is important to know which rehabilitation and maintenance programs address the most critical needs. Little is known, however, about the impact of these activities on households. Although every donor-financed operation includes an evaluation module, these evaluations often focus only on project-specific outputs. In the rare cases when beneficiaries have been the center of attention, lack of data and ad hoc choice of the control groups have limited the usefulness of evaluation. Evaluation attempts have focused exclusively on projects sponsored by the agency request- ing the evaluation, overlooking the many micro-projects implemented in paral- lel (and often with little coordination) by different donors in the same community. To help channel scarce public resources to their best uses, this article inves- tigates the welfare impact of various types of rural infrastructure rehabilitation projects and evaluates their targeting (or placement). It also provides evidence on whether such activities benefit the poor, a useful input to the implementation of a poverty reduction strategy in Georgia. II. DATA AND DEFINITIONS The data used for the analysis come from a household survey and a community survey. The household-level information is provided by the ongoing multitopic Survey of Georgian Households (SGHH). The survey, begun in July 1996 and conducted quarterly by the State Department of Statistics, collects information on the demographic characteristics of household members, their labor market activities, and their access to social services. One section of the questionnaire gathers information on income and consumption expenditures and ownership of assets. Modules collecting information on health and education outcomes were introduced in the first quarter of 2000. The household survey uses a two-stage stratified rotating sample of 2800 households, representative at the national, urban, rural, and regional levels. At the first stage 282 primary sampling units are randomly selected from the stratified list of 12,000 census units with probability proportional to size. In rural areas the primary sampling units roughly correspond to villages. At the second stage, 7–20 households are randomly selected from each primary sampling 314 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 unit. These households stay in the survey for four consecutive quarters and are then replaced by different households from the same sampling unit. This process continues until the list of households in the sampling unit is exhausted. At that point, another primary sampling unit is chosen from the same stratum. Each sampling unit tends to remain in the survey for years, making it possible to construct village-level panels spanning multiple periods. In total, the survey covers 174 rural population sites. Community-level data were collected through the Rural Community Infra- structure Survey (RCIS) conducted in May–June 2002 by the Georgian Opinion Research Business International with the support of the State Department of Statistics and the World Bank. The survey covers all 174 rural population sites from the household surveys. In addition, to expand the sample of beneficiaries, 75 villages not covered by the household surveys were selected from a list of 360 villages supplied by a major donor involved in the community projects, giving a total sample of 249 villages.1 A typical village in the sample benefited from multiple infrastructure projects—57 percent of survey sites reported having two or more projects carried out between 1998 and 2002 (the maximum was 15 projects). Forty-nine villages (20 percent of the sample) had no projects. One of the main purposes of the rural infrastructure survey is to collect retro- spective information on infrastructure projects. The survey questionnaire includes sections on the state of transport infrastructure, water supply systems, schools, kindergartens, and healthcare facilities. It also covers sources of livelihood for the local population based on detailed modules on agricultural and nonfarm activ- ities. One section of the questionnaire contains detailed questions on all infra- structure rehabilitation projects carried out since 1996: dates of initiation and completion of each project, source of funds, and sector. Information was collected from key informants in rural sites, such as local authorities, informal leaders, nongovernmental organizations, and social assistance workers. Because the survey covers all villages covered by the household surveys, there is complete overlap between these two sources of data. This permits the use of both community and household information for analysis.2 Evaluation Sample The community infrastructure survey collected information on 549 rehabilita- tion projects funded by local and international agencies covering schools 1. The data are from the Georgia Social Investment Fund. Detailed information on the Rural Community Infrastructure Survey sample frame, methodology, and questionnaire design can be found in GORBI and Georgia State Department for Statistics (2003). 2. Although this design feature of the community infrastructure survey compensates for the absence of proper baseline data for some indicators used in the analysis, the long recall period for the village-level outcome measures could introduce bias. The reliability of recall data is often questioned on the basis of framing bias (see, for example, Kahnemann 2003). To minimize such bias, group interviews were conducted to reduce individual heterogeneity in the responses of local informants. Lokshin and Yemtsov 315 (28 percent of projects), road infrastructure (27 percent), water supply systems (11 percent), medical facilities (6 percent), kindergartens (3 percent), and other infrastructure rehabilitation projects (25 percent). The three largest groups of interventions were evaluated—schools, road infrastructure, and water system rehabilitation projects. To fit the recall period of the rural infrastructure survey, the analysis uses projects that began on or after 1998 (the baseline) and that were completed by January 2001. That yielded a total of 144 projects in 106 villages. Impact Indicators Two sets of impact indicators were identified for each project, one drawn from the community infrastructure survey and one from the household survey. Community-level indicators based on the community infrastructure survey measure changes between 1998 and 2002.3 Village-level averages, which are based on data from the household survey, compare outcomes in 2000 and 2001. This arrangement is dictated by data availability and creates some disconnect between the two timeframes. It is not a big problem, however, because the majority of projects in the treatment group were completed in 2000.4 The indicators are listed in appendix table A1, and their values are shown in table 1, averaged across villages in the sample calculated at the beginning and at the end of the timeframe chosen for the analysis.5 To deal with the problem of several influential outliers, changes in some continuous variables were recoded into simple categorical variables, reporting the balance between positive and negative changes. Some outcome indicators reflect alarming trends in access to education, quality of road infrastructure, and availability of piped water. Only 68 percent of villages had all school-age children in school in 1998; by 2002 only 59 percent did. Household-level data suggest that close to 8 percent of children in an average village missed more than 30 days of classes in 2000. This indicator improved for all villages in the sample by the end of 2001. For 1998 as many as 91 percent of villages reported that the quality of their main roads was inadequate. This indicator improved considerably by 2002, but respondents in 71 percent of villages still complained about road quality. In more 3. The treatment group includes villages with projects completed before the beginning of 2001, allowing at least one year to pass before assessing project benefits. 4. Omitting these cases would reduce the number of usable observations, which is a critical constraint in the study. We chose to retain all the villages but to exercise care in interpreting results. 5. Note the difference in the definition of ‘‘before’’ and ‘‘after’’ in community infrastructure survey and household survey indicators. Also, the number of observations reported for the household survey in table 1 differs across indicators. For example, data on school enrollment rates are available for all 102 villages with household survey data, but information on ambulance arrival time is available for only 68 villages. This attrition is clearly related to the low frequency with which rural residents use emergency care. 316 T A B L E 1 . Summary Statistics for Main Outcome Indicators Beforea Aftera Change Outcome indicator Source Number Mean SD Mean SD Mean SD All children are enrolled in school RCIS 249 0.683 0.466 0.590 0.493 À0.092 0.353 Number of pupils RCIS 246 293.781 278.019 283.129 292.574 0.244b 0.431 Number of graduates RCIS 249 23.450 21.955 21.635 22.520 0.339b 0.474 Access to educationc RCIS 0.289 0.454 School enrolment rate SGHH 102 0.973 0.079 0.978 0.044 0.006 0.076 Share of pupils missing more than 30 days SGHH 102 0.081 0.194 0.063 0.159 À0.009b 0.111 Unsatisfactory schooling conditions SGHH 102 0.051 0.142 0.054 0.111 À0.015b 0.153 Expenditures on schooling SGHH 100 11.755 10.456 48.470 45.612 1.119d 1.114 Incidence of respiratory diseases (child) SGHH 102 0.126 0.182 0.079 0.137 À0.065 0.163 Time to district capitalc RCIS À30.721 46.553 THE WORLD BANK ECONOMIC REVIEW, VOL. Subjective assessment of road (bad) RCIS 224 0.911 0.286 0.711 0.454 À0.223 0.428 Barter trade RCIS 249 0.494 0.501 0.498 0.501 0.004 0.168 19, Small enterprises RCIS 249 0.462 0.500 0.486 0.501 0.024b 0.268 Time for ambulance to arrive SGHH 68 0.668 0.388 0.528 0.402 À0.113b 0.334 NO. Sales of agricultural products SGHH 104 158.122 244.790 114.760 150.807 À0.322d 0.959 2 Female off-farm employment SGHH 103 0.141 0.141 0.129 0.121 0.001b 0.081 Nonagricultural employment SGHH 103 0.161 0.114 0.154 0.105 0.001b 0.065 Household transport expenditures per capita SGHH 104 2.196 2.851 1.642 1.814 À0.133d 1.208 Incidence of trauma SGHH 102 0.001 0.005 0.001 0.005 À0.001 0.004 (Continued) TABLE 1. Continued Beforea Aftera Change Outcome indicator Source Number Mean SD Mean SD Mean SD c New water sources RCIS 249 0.092 0.290 Number of livestockc RCIS 249 0.663 0.474 Piped water in the household SGHH 103 0.561 0.444 0.565 0.429 0.006 0.227 Hours of piped water supply SGHH 103 8.380 9.456 8.788 9.533 0.767b 4.852 Incidence of water-borne diseases (total) SGHH 102 0.004 0.008 0.008 0.028 0.000b 0.008 Incidence of water-borne diseases (child) SGHH 91 0.008 0.028 0.022 0.097 0.007b 0.032 Expenditures on bottled water SGHH 89 1.608 0.872 1.583 0.895 À0.019d 0.421 Note: Values are averaged across villages in the sample calculated at the beginning and the end of the timeframe chosen for analysis. a For RCIS, ‘‘before’’ is 1998 and ‘‘after’’ is 2002; for the SGHH, ‘‘before’’ is 2000 and ‘‘after’’ is 2001. b Recoded change indicator: the share of changes in a positive direction minus the share of changes in a negative direction. Some villages have missing values for an indicator before or after project completion; thus the indicator of change may differ from a simple ratio of before and after project indices. c Based on a direct change question in the survey. d Change in the log of per capita values. Lokshin and Yemtsov 317 318 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 than half the villages, it took more than four hours for an ambulance to respond to a call. In 1998 and 2002 only about 56 percent of rural households were connected to a piped water supply. Piped water was available an average of eight hours a day. The high and increasing incidence of waterborne diseases among children is of particular concern. By 2001 as many as 2 percent of children below age seven reported illnesses related to poor water quality in the month preceding the survey. III. METHODOLOGY The criteria for project placement vary among agencies operating in Georgia, but in most cases placement criteria take into account the extent of poverty or its correlates, the state of infrastructure in a village, or regional characteristics. Many projects rely on demand-driven targeting mechanisms. Whether a parti- cular village gets a project can depend on the village’s ability to seek support from implementing agencies. Villages are chosen by project managers based on characteristics that could be correlated with the expected outcomes of a project. Because of such nonrandom placement, a simple comparison of outcomes between villages with projects and villages without projects would be invalid. If selection of a village for a project is based purely on observable character- istics, a propensity-score matching method can be used to correct for selection bias (Rosenbaum and Rubin 1983; Rubin 1973). The propensity score measures the probability that a project is implemented in a village as a function of that village’s observed preintervention characteristics. Villages with projects (the treatment group) are matched with villages without projects (the control group) on the basis of the propensity score. Following Chen and Ravallion (2003), outcome measure Iit for a project in ith village at date t is defined as: ð1Þ Ã Iit ¼ Iit þ Git I Di ; à where Iit is the outcome for a village if the project is not implemented, and GitI is the gain to village i from an outcome attributable to a project. Then the estimate of the average impact of the project on a treatment village (dummy variable Di ¼ 1) can be decomposed as: à à ð2Þ EðIit j Di ¼ 1ÞÀ EðIit j Di ¼ 0Þ EðIit j Di ¼ 1Þþ EðGit I j Di ¼ 1Þ ÀEðIit j D ¼ 0Þ: From equation 2 the estimation bias amounts to à à ð3Þ EðIit j Di ¼ 1Þ À EðIit j Di ¼ 0Þ: There is no bias in a simple comparison of the means between treatment and control villages if the terms of equation 3 are equal. The cross-sectional Lokshin and Yemtsov 319 propensity score method assumes that conditional on a set of observed char- acteristics X, à à ð4Þ EðIit j Di ¼ 1; XÞ ¼ EðIit j Di ¼ 0; XÞ: Thus, the cross-sectional propensity score method produces an unbiased esti- mate of the project effect if project placement is based purely on a village’s observed characteristics. However, some unobserved characteristics of the village that are correlated with project outcomes might also be correlated with project placement. This correlation can introduce bias in the estimation of project impact. For example, an active parent group might lobby the village authorities to pursue a school rehabilitation project. This same group of active parents might then become involved in the education process and positively affect school outcomes for their children. If the evaluation does not take into account the differences in parental activity between treatment and control villages, the effectiveness of the school project will be overestimated. If the preintervention differences between treatment and control villages are assumed to be the result of time-invariant unobserved factors, the difference-in-difference method can be used to correct for possible bias. The preproject difference in outcomes may be subtracted from the postpro- ject differences for the same villages. The underlying assumption in this method is that the time trend in the control group is an adequate proxy for the time trend that would have occurred in the treatment group in the absence of an intervention, or ð5Þ EðIi1 à À Ii0 à j Di ¼ 1; XÞ ¼ EðIi1 à À Ii0 à j Di ¼ 0; XÞ: The mean difference in difference for the outcome is estimated by taking the expectation of equation 1 over all N sample villages using equation 5: ð6Þ E½ðIi1 À Ii1 Ã Þ À ðIi0 À Ii0 Ã Þ j Di ¼ 1Š ¼ E½Gi1 I j Di ¼ 1Š: If the outcomes in period 0 are not correlated with project placement, equation 6 estimates the mean changes in outcomes for the treatment villages. This study uses the matched difference-in-difference method, which com- bines propensity score–matching and difference-in-difference methods. Recent studies by Heckman and others (1997, 1998) have argued that combining these methods can substantially reduce the bias found in other nonexperimen- tal evaluations. First, villages from the control and treatment groups are matched using propensity score matching. This matching removes the selection bias due to the observed differences between treatment and control villages. 320 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 Then the difference-in-difference method is applied to correct for possible bias due to the differences in time-invariant unobserved characteristics between the two groups. To evaluate the impact of the project, the changes in outcome measures are compared between matched villages from the treatment and control groups. There is another form of bias that these methods cannot remove, which arises from time-variant unobservable characteristics correlated with both project placement and the outcomes of the intervention.6 In particular, project place- ment could be based on unobserved community characteristics that are corre- lated with changes in the expected project outcomes. However, there are reasons to believe that this bias may not arise in the context of micro-projects in rural Georgia. The project placement procedures used by the implementing agencies are based on formal criteria that capture exclusively the current state of affairs. Thus, placement can reduce (but not completely eliminate) possible bias from time-variant unobservables. IV. RESULTS The project placement mechanism for each type of intervention is modeled first as a function of a large set of variables from the community infrastructure survey that include village-level aggregates on geographic, demographic, and socioeconomic conditions (table 2). The model also controls for the presence of other projects in the same village. For example, in the specification that models the probability of a village participating in a school rehabilitation project, two dummy variables are included to reflect the presence of road and water projects. The probit estimates for three types of interventions are shown in table 2. The adjusted pseudo-R2 of these estimations ranges from 0.156 for the school projects to 0.393 for the water projects. These are acceptable levels of explana- tory power. A high R2 could indicate the existence of fundamental differences between the characteristics of project and nonproject villages, which would make the formation of a proper control group very problematic. Only a few coefficients in the table are significant—the indicator of natural disasters, for example. This should not be taken as a sign of problems in forming a control group because the empirical specifications include many correlated variables and the purpose of the estimation is to calculate the propensity score and not model an underlying selection mechanism. 6. This problem is thought to be severe for infrastructure programs in poor areas if the deficient state of infrastructure in the initial period not only attracts the rehabilitation project, but also reduces future growth (Jalan and Ravallion 1998). T A B L E 2 . Probit Estimates of the Probability of a Village Participating in a School, Road, or Water Project School Project Road Project Water Project Summary Statistics Coefficient SE Coefficient SE Coefficient SE Mean SE School project 0.202 (0.25) À0.727 (0.54) 0.245 (binary) Road project 0.129 (0.294) À0.359 (0.581) 0.185 (binary) Water project À0.494 (0.498) 0.008 (0.462) 0.068 (binary) Total population 0.04 (0.147) 0.039 (0.145) À0.008 (0.262) 7.046 (1.082) If internally displaced person in the village À0.027 (0.254) À0.146 (0.246) À0.466 (0.421) 0.482 (binary) Agriculture only À0.157 (0.258) À0.09 (0.247) À0.716 (0.457) 0.394 (binary) Experienced disaster 0.172 (0.057) À0.032 (0.06) 0.053 (0.08) 1.956 (2.087) Experienced flood À0.479 (0.352) 0.308 (0.325) 0.403 (0.621) 0.129 (binary) Mountain area 0.054 (0.252) 0.197 (0.253) 0.46 (0.477) 0.442 (binary) Alpine area 0.165 (0.346) 0.077 (0.334) 0.795 (0.625) 0.241 (binary) Distance to district center À0.129 (0.129) À0.034 (0.143) 0.143 (0.259) 2.452 (0.808) Distance to market 0 (0.002) À0.004 (0.005) 0.003 (0.003) 24.388 (49.308) Rail road 0.475 (0.315) À0.075 (0.385) À7.569 0.104 (binary) Interstate highway À0.017 (0.267) À0.592 (0.274) 0.257 (0.432) 0.390 (binary) Asphalt road À0.011 (0.257) 0.702 (0.247) À0.505 (0.499) 0.297 (binary) Number of schools À0.027 (0.138) 0.24 (0.136) 0.476 (0.252) 1.197 (0.802) Number of large enterprises À0.009 (0.034) À0.053 (0.036) 0.088 (0.089) 14.940 (3.097) Small enterprise 0.204 (0.253) 0.084 (0.249) À0.183 (0.442) 0.462 (binary) Police station 0.474 (0.319) À0.247 (0.334) 0.179 (0.546) 0.169 (binary) Post office À0.209 (0.244) 0.293 (0.24) 0.282 (0.402) 0.558 (binary) Restaurant À0.1 (0.32) 0.118 (0.31) À0.931 (0.735) 0.161 (binary) Proportion of households with a phone À0.199 (0.442) 0.501 (0.405) 1.669 (0.548) 0.088 (0.284) Proportion of households with a toilet À0.589 (0.507) À0.156 (0.566) À2.468 (0.945) 0.956 (0.206) Unreliable electric power supply 0.128 (0.213) À0.026 (0.226) 1.179 (0.416) 0.462 (binary) Proportion of households with piped water 0.413 (0.234) 0.014 (0.276) 0.176 (0.457) 0.229 (0.421) Proportion of buildings with wooden walls 0.107 (0.346) 0.891 (0.267) À1.087 (0.628) 0.333 (0.472) Lokshin and Yemtsov Proportion of buildings with dirt floors 0.147 (0.449) À0.034 (0.519) À7.348 0.056 (0.231) Trade by the roadside À0.367 (0.312) 0.692 (0.294) 0.633 (0.533) 0.181 (binary) 321 (Continued) 322 TABLE 2. Continued School Project Road Project Water Project Summary Statistics Coefficient SE Coefficient SE Coefficient SE Mean SE Regional dummy variables Kaheti (omitted category) 0.137 (binary) Inner (Shida) Qartli 1.133 (0.448) 0.153 (binary) Lower (Kvemo) Qartli 0.489 (0.51) 0.141 (binary) Samskhe-Djavakheti 1.102 (0.49) 0.141 (binary) Achara 0.929 (0.726) 0.068 (binary) THE WORLD BANK ECONOMIC REVIEW, VOL. Guria 0.626 (0.594) 0.092 (binary) Samegrelo 0.963 (0.666) 0.076 (binary) 19, Imereti 1.045 (0.496) 0.193 (binary) Constant À1.327 (1.332) À1.155 (1.302) À2.311 (2.526) NO. Sample size 249 249 249 249 2 Adjusted pseudo-R2 0.156 0.181 0.393 Source: Authors’ computations based on RCIS 2002. Lokshin and Yemtsov 323 School Rehabilitation Projects Typically, school projects in Georgia focus on improving school buildings: repair- ing roofs, windows, and floors; replacing pipes; installing sanitary and heating equipment; and repainting walls. These projects may yield several types of ben- efits to the community. School rehabilitation may improve both enrollment and attendance rates. Better heating and repaired windows could be particularly important in Georgia, where some rural schools close for several weeks in winter because of frigid classrooms (Orivel 1998). Changes in household expenditures on schooling can be used as an indicator of the private response to investment in school rehabilitation. The subjective assessments of schooling conditions provide a useful check on results based on objective measures. School rehabilitation projects were completed in 61 villages (about a quarter of all villages) in the community infrastructure survey sample by 2001. Thirty- seven of these villages were also covered by the household survey. The initial (unmatched) control group was constructed from villages without school pro- jects and villages with incomplete school projects at the end of 2001. For the community infrastructure survey data three outcome indicators are reported at the community level based on difference-in-difference estimation of the impact of school rehabilitation projects for the unmatched control group and propensity score match–constructed control group (table 3). No significant differences are detected between the treatment group and the two control groups for the share of villages reporting that all children are enrolled in school, T A B L E 3 . Difference-in-Difference Estimates of the Average Impact of School Rehabilitation Projects Unmatched Sample Matched Sample Treatment Control Control Outcome Indicators Group Group p-Value Group p-Value RCIS All children are À0.066 À0.101 0.251 À0.066 0.500 enrolled in school If number of pupils increased 0.328 0.216 0.056 0.190 0.051 If number of graduates increased 0.373 0.327 0.268 0.237 0.059 Access to education has improved 0.361 0.266 0.089 0.213 0.036 SGHH School enrollment 0.059 À0.004 0.102 0.000 0.117 Share of pupils missing more À0.057 À0.001 0.063 0.020 0.019 than 30 days Unsatisfactory schooling conditions À0.020 À0.014 0.584 À0.013 0.611 Expenditures on schooling 1.249 1.094 0.365 1.544 0.772 Incidence of respiratory diseases À0.120 À0.056 0.083 À0.056 0.160 in children Source: Authors’ computations based on the RCIS and the SGHH. 324 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 which declined for all groups between 1998 and 2002. In approximately a third of project villages the number of pupils and the number of graduates increased while decreasing in the control group (this difference is statistically significant). The village-level subjective assessment indicator shows a significant change in the perception that access to education improved between 1998 and 2002 for the treatment villages but not for the control villages. A more detailed set of outcome indicators is estimated for the household survey data. In treatment villages, primary and secondary school enrollment rates increased by 6 percentage points between 2000 and 2001; there was no change in matched control villages. However, this difference is only marginally significant. A more responsive indicator of school attendance shows clearer benefits from school improvements. The share of pupils missing classes dropped by more than 5 percentage points in treatment villages, and it increased by 2 percentage points in the matched control group. The health impact of school rehabilitation is substantial. The incidence of respiratory diseases among school-age children declined by 12 percentage points in villages with a project compared with a decline of slightly more than 5 percentage points in villages without a project. No significant changes in parents’ assessments of schooling conditions were detected. Overall, the estimation results fit the prior expectations. In Georgia, where primary education is compulsory, the most sensitive gauge of project impacts is changes in attendance, the outcome indicator for which the results are the most significant. It can also be speculated that if school rehabilitation projects induce a positive response in one indicator in treatment villages, this could lead to improvements in other indicators that intuitively are less sensitive to this type of intervention. Improvement in the health status of school-age children is one example. Improvements in Road Infrastructure Road and bridge rehabilitation often means repaving existing roads, restoring road structures damaged or destroyed by flooding and earthquakes, and widen- ing road intersections and bridges. Such rehabilitation can reduce commuting time and improve access to markets. Investments in roads and bridges are likely to generate new income opportunities for agricultural households, with impacts far beyond the project site.7 Several labor market studies have identified off- farm employment, an activity highly dependent on transportation, as the driving force behind welfare change in Georgia (Bernabe ` 2002; Yemtsov 2001). Poor access to product markets appears to constrain growth and to perpetuate barter trade (Cord and others 2003). 7. According to studies of other countries in Eastern and Central Europe, poor road quality can add 28–44 percent to transportation costs for local producers and to commuting costs for rural dwellers. Lowering transportation costs will have a dramatic effect on the poor, because poor households generally tend to be located in very remote areas (World Bank 2003c). Lokshin and Yemtsov 325 By the end of 2000 road improvement projects were completed in 41 villages, or 19 percent of the community infrastructure survey sample; 36 of these villages were also covered by the household survey. The initial control group was constructed from all villages without road or bridge projects completed between 1998 and 2001. The most immediate outcome indicator of a road rehabilitation project—time spent commuting to the district center—shows a 36-minute reduction in project villages, but these gains are not statistically different from changes for the control group (table 4). Indicators linked to the economic impact of projects show more pronounced trends. The share of villages with active nonagricultural small and medium-size enterprises increased in project villages, a statistically significant change com- pared with the propensity score–matched control group. The share of villages reporting barter exchange among the main channels for marketing agricultural products dropped significantly as a result of the road projects while increasing in control villages. Subjective assessments reflect no reaction to road rehabilita- tion interventions. Off-farm employment and female wage employment rates increased in vil- lages affected by road rehabilitation but declined in the control villages. Indi- cators reflecting changes in the per capita market sales of agricultural products, however, showed no improvement in the treatment villages. Time for an ambu- lance to arrive improved in 24 percent of the treatment villages. This compares favorably with the worsening of this indicator in the propensity score–matched T A B L E 4 . Difference-in-Difference Estimates of the Average Impact of Road and Bridge Rehabilitation Projects Unmatched Sample Matched Sample Treatment Control Control Outcome Indicators Group Group p-Value Group p-Value RCIS Travel time to district center À35.577 À29.235 0.295 À27.692 0.287 Subjective assessment of road (bad) À0.261 À0.214 0.259 À0.326 0.731 Barter trade À0.044 0.015 0.038 0.044 0.022 Small enterprises 0.044 0.020 0.308 À0.044 0.052 SGHH Time for ambulance to arrive À0.238 À0.070 0.088 0.035 0.092 Sales of agricultural products À0.324 À0.322 0.496 À0.184 0.287 Female off-farm employment 0.006 À0.001 0.397 À0.011 0.274 Nonagricultural employment 0.013 À0.004 0.199 À0.003 0.241 Household transport À0.165 À0.121 0.442 À0.407 0.758 expenditures per capita Incidence of trauma À0.001 À0.001 0.396 À0.003 0.764 Source: Authors’ computations based on the RCIS and the SGHH. 326 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 control group. The difference between the control and treatment groups in the rate of road accidents is not statistically significant. Some of the effects from road rehabilitation projects could be difficult to capture because of their data requirements and long-run nature. For example, indicators of improved road safety impose high demands on data coverage because accidents occur rarely. Water System Rehabilitation Projects Water projects include a wide range of works—installing new or repairing existing communal water tanks, installing water treatment equipment, fitting new pumps, repairing or installing pipes, and rehabilitating wastewater manage- ment networks. Benefits could include a reduction in the incidence of water- borne disease (Jalan and Ravallion 2003), less reliance on more expensive alternatives to piped water, and more time for child schooling and for produc- tive activities among adults, particularly women. Coverage was less extensive for water rehabilitation projects than for school or road projects. In the community infrastructure survey, 17 villages (7 percent of the sample) had a water system rehabilitation project completed by the end of 2001. Only nine villages in this group were also covered by the household survey. The small number of cases make this an important test of the limits of the proposed evaluation strategy. The impact evaluation estimations show that the range of drinking water supply options expanded in 24 percent of project villages (table 5). In the control group, only 8 percent of villages in the unmatched and 6 percent in the matched sample reported a new water supply option available between 1998 T A B L E 5 . Difference-in-Difference Estimates of the Average Impact of Water System Rehabilitation Projects Unmatched Sample Matched Sample Treatment Control Control Outcome Indicators Group Group p-Value Group p-Value RCIS New channels of water supply 0.235 0.082 0.086 0.059 0.041 Increase in livestock 0.647 0.664 0.447 0.529 0.272 SGHH Piped water in the household 0.110 0.002 0.216 À0.063 0.243 Hours of piped water supply 0.980 0.779 0.565 À0.785 0.665 Female wage employment À0.055 0.004 0.181 À0.020 0.382 Incidence of waterborne À0.006 0.001 0.196 À0.001 0.123 diseases (total) Incidence of waterborne diseases 0.000 0.007 0.037 0.000 in children Expenditure on bottled water À0.018 À0.018 0.500 À0.027 0.514 Source: Authors’ computations based on the RCIS and the SGHH. Lokshin and Yemtsov 327 and 2002. Coverage of piped water supply increased 11 percent in the treatment villages compared with no change or even slight deterioration of coverage in the control groups. The number of hours that piped water is available increased sub- stantially in the project villages while declining considerably in the matched control group. Comparison of changes in the incidence of waterborne diseases shows a marginally significant effect. Other impact indicators show changes in the expected direction (with the exception of changes in the female employment rate), but the differences between the treatment and control group averages are insignificant. Difficulties in observing significant effects of water rehabilitation projects could be linked to three factors. First, water projects were the least ‘‘popular’’ in rural Georgia according to the community infrastructure survey, resulting in too small a sample to capture the effect. Second, it is difficult to extract specific indicators reflecting improved access to water from a regular multitopic survey. Third, a distinct feature of water projects is partial coverage of the population. In many villages only certain clusters of houses are connected to pipes and therefore are direct beneficiaries of this kind of intervention. As a result, the effect observed at the village level may not fully reflect the heterogeneity in impact among project beneficiaries. This issue is addressed in the next section. Distributional Impact of Infrastructure Rehabilitation Projects Households within the same village may benefit differently from a particular project. Jalan and Ravallion (2003) find that piped water projects, for example, have different impacts for poor and nonpoor households in India. To assess whether infrastructure rehabilitation projects had different impacts on the living standards of poor and nonpoor households in Georgia the main outcome indicators were reconstructed using subsamples of poor and nonpoor households from each village covered by the household survey. Community- level impact indicators from the community infrastructure survey were omitted because these cannot be differentiated for the poor and the nonpoor. Poor households seem to have benefited more than nonpoor households from school rehabilitation projects (table 6). The most sensitive indicator—improvements in school attendance—shows that school rehabilitation has a significant effect on the poor. The share of children from poor households missing classes declined by 11 percentage points as compared with about 2 percentage points for nonpoor households. Similarly, health outcomes improved more among children from poor households than from nonpoor households. Changes in school enrollment rates, however, demonstrate a better response for children from nonpoor households, whereas differences in changes in private educational expenditures are ambiguous. The distributional impact of road rehabilitation projects varies for different outcome indicators. The nonpoor clearly benefited more in improved access to emergency medical assistance and in opportunities for nonagricultural employ- ment. Female off-farm employment rates, on the other hand, show greater positive change among the poor. Interpreting results for the agricultural product sales indicator is more complex (World Bank 2003c). In recent years the sales of T A B L E 6 . Poor versus Nonpoor: Difference-in-Difference Estimates of the Average Impact 328 of the Project for Three Types of Interventions Poor Nonpoor Differencea Treatment Group Control Group p-Value Treatment Group Control Group p-Value p-Value School rehabilitation School enrollment 0.012 0.000 0.403 0.041 À0.008 0.168 0.345 Share of pupils missing more À0.110 0.023 0.034 À0.035 0.012 0.050 0.099 than 30 days Unsatisfactory schooling conditions 0.000 0.015 0.085 À0.023 À0.020 0.545 0.200 Expenditures on schooling 1.586 1.201 0.294 1.267 1.614 0.201 0.290 Incidence of respiratory À0.130 À0.054 0.244 À0.081 À0.069 0.436 0.313 diseases in children Road and bridge rehabilitation Time for ambulance to arrive À0.188 À0.250 0.545 À0.237 0.039 0.105 0.572 Sales of agricultural products À0.251 0.231 0.150 À0.226 À0.094 0.290 0.480 Female off-farm employment 0.053 0.013 0.162 0.002 À0.012 0.317 0.100 Nonagricultural employment À0.002 0.018 0.330 0.021 À0.001 0.150 0.270 THE WORLD BANK ECONOMIC REVIEW, VOL. Household transport expenditures 0.077 0.636 0.262 À0.118 À0.476 0.155 0.203 per capita 19, Incidence of trauma 0.000 À0.005 0.082 À0.002 À0.002 0.465 0.340 Water system rehabilitation NO. Piped water in the household 0.121 0.062 0.368 0.105 À0.099 0.225 0.536 2 Hours of piped water supply 0.747 0.654 0.527 1.058 À1.434 0.310 0.411 Female wage employment 0.002 0.063 0.145 À0.083 À0.008 0.317 0.138 Incidence of waterborne diseases, total À0.007 À0.004 0.361 À0.005 0.007 0.142 0.407 Incidence of waterborne diseases 0.000 0.000 0.000 0.000 in children Expenditure on bottled water 0.482 0.168 À0.054 À0.051 0.494 Source: Authors’ computations based on the RCIS and the SGHH. a Difference in means between poor and nonpoor households from treatment group villages. Lokshin and Yemtsov 329 agricultural products plummeted for the whole country, and the decline was particularly strong for rich households, which had been better integrated into markets. This is what the impact analysis results show here, suggesting that road quality is not the main driver in this process. The key benefits from water projects are related to improvements in health status, which were found mainly among nonpoor households. Changes in the incidence of waterborne diseases among poor households were not statistically different between treatment and control groups. V. CONCLUSION This analysis of the impact of community-level investments in infrastructure rehabilitation in rural Georgia on household well-being combined household- and community-level survey data, controlling for time-invariant unobservable characteristics at the community level by applying propensity score–matching difference-in-difference comparisons. The results indicate that improvements in school infrastructure produced nontrivial gains in school enrollment rates, raised school attendance, and reduced health risks for school-age children. Road and bridge rehabilitation projects generated clear economic benefits at the community level. The number of small and medium-size enterprises increased, and the importance of barter trade fell. Access to emergency medical assistance improved unambiguously. For water system rehabilitation interventions, the most unambiguous effect is the reduction of the incidence of waterborne diseases. The impact of water projects measured by other indicators is less clear-cut. To a large degree, the ambiguity is related to the small number of project villages in the sample. Community-level interventions had different distributional impacts. School reha- bilitation improved school attendance and children’s health status among the poor more than it did among the better off. Road projects benefited the poor and nonpoor in different ways. The nonpoor gained more from improved accessibility to emergency medical assistance. Expansion of nonagricultural job opportunities favored women from poor households. That better-off households fully accounted for the observed decline in the incidence of waterborne diseases suggests that the benefits of water rehabilitation projects accrue mostly to the nonpoor. It is encouraging to see such richness in the results considering that the analysis relied on modest additional data collection. The methodology demon- strates that evaluation of project impact is possible even in the absence of proper baseline survey data. Carefully designed community surveys (collecting retro- spective information) in combination with ongoing nationally representative household surveys could provide a feasible and low-cost alternative to standard before-and-after techniques, perhaps stimulating wider use of robust impact assessment methodologies for community-level projects in developing countries. Nevertheless, it is important to emphasize that proper baseline data are crucial for a credible evaluation. Using retrospective data to substitute for baseline data, as 330 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 APPENDIX T A B L E A . 1 . Definition of Indicators Indicator Definition All children are enrolle d in school Dummy, whether all children in the village 7–15 years old are enrolled in school Number of pupils Total number of pupils enrolled in all primary and secondary schools located in the village Number of grads Total number of pupils graduating in a given year from all schools located in the village Access to education Assessment by key informant of the change (1998–2002) in the access to education by village population School enrollment rate Share of village school-age children (7–15 years old) currently enrolled in school Share of pupils missing more than Share of village children enrolled in school who 30 days missed more than 30 days in the last school year Unsatisfactory schooling conditions Share of parents assessing conditions as unsatisfactory in the school in which their child is currently enrolled Expenditures on schooling Average lari per child attending school in past 12 months for transport, textbooks, fees, and other school expenses Incidence of respiratory diseases (child) Share of children enrolled in schools who suffered from respiratory disease over the past 30 days Time to district capital Reduction in time between 1998 and 2002, in minutes, to reach the district capital by the most usual means of transportation Subjective assessment of road (bad) Share of villages where key informants assess the main road quality as ‘‘bad’’ or ‘‘very bad’’ Barter trade Share of villages where barter is listed among the three most important ways to sell agriculture products Small enterprises Share of villages which have operating small or medium-size manufacturing or construction firms (fewer than 10 employees) Time for ambulance to arrive Average time for emergency assistance to arrive at households that called for an ambulance in the past 30 days Sales of agricultural products Average sales of crops and animal products over the last quarter, in lari per household Female off-farm employment Share of women in the working age population who were employed for wages for at least one hour during the last week Nonagricultural employment Share of working age population employed for pay or self-employed outside agriculture in the last three months Household transport expenditures Spending on transportation services and means of per capita transport over the past quarter in lari per capita Incidence of trauma Share of population reporting suffering from trauma and burns in the past 30 days (Continued) Lokshin and Yemtsov 331 T A B L E A.1. Continued Indicator Definition New water sources Share of villages where new water supply options become available between 1998 and 2002 Number of livestock Share of villages where key informants report considerable or some increase in livestock owned between 1998 and 2002 Piped water in the household Share of household that have a piped water supply in or near their dwelling Hours of piped water supply Hours water was available per day on average over the last three months for households with piped water Incidence of waterborne diseases (total) Share of population suffering diarrhea, gastrointestinal, or parasitic infections in the past 30 days Incidence of waterborne diseases Share of children under age five suffering from (child) diarrhea, gastrointestinal, or parasitic infection in the past 30 days Expenditures on bottled water Purchases of bottled water over the last three months, in average lari per household If internally displaced person Dummy variable that takes a value of 1 if the village in the village has at least one internally displaced person from Abkhazia in residence Agriculture only Dummy variable that takes a value of 1 if the key informants report that there is no economic activity in the village except agriculture Experienced disaster Total number of floods, earthquakes, droughts, hails, fires, landslides, and livestock epidemics between 1997 and 1999 Experienced flood Village experienced at least one flood between 1997 and 1999 Railroad Whether the village had an operating railroad station in 1998 Interstate highway Whether a highway of regional importance passed through the village in 1998 Asphalt road Whether the main road in the village was paved in 1998 Number of schools Number of primary and secondary schools operating in the village in 1998 Number of large enterprises Number of operating enterprises within 20 km of the village with more than 10 employees in 1998 Small enterprise Whether the village had a small or medium-size enterprise operating in 1998 Police station (post, restaurant) Whether the village had a police station (post office, restaurant, or roadside cafe ´) in 1998 Proportion of household with a phone Share of households in the village that were connected to a fixed telephone line in 1998 Proportion of household with a toilet Share of households in the village that were using latrines in 1998 Unreliable electric power supply Key informant in the village reported electric power supply to the village of less than 24 hours a day in 1998 (Continued) 332 THE WORLD BANK ECONOMIC REVIEW, VOL. 19, NO. 2 T A B L E A.1. Continued Indicator Definition Proportion of households with Share of households in the village that have a piped piped water water supply near their dwelling in 1998 Built with wooden (dirt) walls (floors) Key informant estimate of share of buildings in the village with wooden walls (dirt floors) in 1998 Trade by the roadside Dummy variable taking a value of 1 if one of three main trade channels for the village was selling products by roadsides in 1998 Source: SGHH and RCIS. Note: Currency values are in nominal terms, with 1 lari approximately equal to $0.50. The inflation rate in Georgia in 1998–2002 was 4 percent a year, and food prices were essentially unchanged. is done here, risks introducing recall bias, which can influence the precision of the results. Retrospective data cannot fully substitute for baseline data, but they can serve as a defensible fallback when first-best options are not feasible. Also important, the proposed strategy fails to produce robust results for projects affecting a small fraction of the population. For example, most of the results for water system rehabilitation projects are only marginally significant or insignificant. Data limitations required dropping health clinic and kindergarten rehabilitation projects from the analysis. Thus, the proposed methodology is probably best suited for large-scale com- munity-driven micro-projects deployed with little regard for subsequent evalua- tion, such as emergency or disaster response operations. Government-run decentralized public investment programs in developing economies are another good candidate. Projects of this type play an important role in developing economies, and assessing their effectiveness can help in making informed choices about their focus, scope, and delivery mechanisms. REFERENCES ` , S. 2002. ‘‘A Profile of the Labour Market in Georgia.’’ International Labour Organization and Bernabe UN Development Programme, Tbilisi. Brockerhoff, M., and L. Derose. 1996. ‘‘Child Survival in East Africa: The Impact of Preventive Health Care.’’ World Developmetnt 24(12):1841–57. Chen, S., and M. Ravallion. 2003. ‘‘Hidden Impact? Ex-Post Evaluation of an Anti-Poverty Program.’’ Journal of Public Economics 86(1):123–53. Cord, L., R. Lopez, M. Huppi, and O. Melo. 2003. ‘‘Growth and Rural Poverty in the CIS7 Case Studies of Georgia, the Kyrgyz Republic, and Moldova.’’ Paper prepared for the Lucerne Conference of the CIS-7 Initiative, January 20–22, Lucerne, Switzerland. Glewwe, P. 1999. The Economics of School Quality Investments in Developing Countries: An Empirical Study of Ghana. London: Macmillan. GORBI(Georgian Opinion Research Business International) and Georgia State Department for Statistics. 2003. ‘‘Fiscal Community Survey Report.’’ Tbilisi. Lokshin and Yemtsov 333 Hanushek, E. 1995. ‘‘Interpreting Recent Research on Schooling in Developing Countries.’’ World Bank Research Observer 10(2):227–46. Heckman, J., H. Ichimura, and P. Todd. 1997. ‘‘Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Program.’’ Review of Economic Studies 64(4):605–54. Heckman, J., H. Ichimura, J. Smith, and P. Todd. 1998. ‘‘Characterizing Selection Bias Using Experi- mental Data.’’ Econometrica 66(5):1017–99. Jacoby, H. 2002. ‘‘Access to Markets and the Benefits of Rural Roads.’’ Economic Journal 110(465): 713–37. Jalan, J., and M. Ravallion. 1998. ‘‘Are There Dynamic Gains from a Poor-Area Development Program?’’ Journal of Public Economics 67(1):65–85. ———. 2003. ‘‘Does Piped Water Reduce Diarrhea for Children in Rural India?’’ Journal of Econometrics 112(1):153–73. Kahnemann, D. 2003. ‘‘Maps of Bounded Rationality: Psychology for Behavioral Economics.’’ American Economic Review 93(5):1449–75. Kremer, M. 1995. ‘‘Research on Schooling: What We Know and What We Don’t: A Comment.’’ World Bank Research Observer 10(2):247–54. Lee, L., M. Rosenzweig, and M. Pitt. 1997. ‘‘The Effects of Improved Nutrition, Sanitation, and Water Quality on Child Health in High-Mortality Populations.’’ Journal of Econometrics 77(1):209–35. Orivel, F. 1998. ‘‘Cost and Finance of Education in Georgia.’’ World Bank, Washington, D.C. Rawlings, L., and N. Schady. 2002. ‘‘Impact Evaluation of Social Funds.’’ World Bank Economic Review 16(2):213–17. Rosenbaum, P., and D. Rubin. 1983. ‘‘The Central Role of the Propensity Score in Observational Studies for Causal Effects.’’ Biometrika 70(1):41–55. Rubin, D. 1973. ‘‘The Use of Matched Sampling and Regression Adjustment to Remove Bias in Observa- tional Studies.’’ Biometrics 29(1):185–203. UNOCHA(United Nations Office for the Coordination of Humanitarian Affairs). 2003. ‘‘Directory of Humanitarian and Developments Sectoral Programs.’’ Tbilisi. van de Walle, D., and D. Cratty. 2002. ‘‘Impact Evaluation of a Rural Road Rehabilitation Project.’’ World Bank, Washington, D.C. World Bank. 2003a. ‘‘Achieving the Human Development MDGs in ECA.’’ Washington, D.C. ———. 2003b. ‘‘Georgia Country Assistance Strategy.’’ Washington, D.C. ———. 2003c. ‘‘Georgia: Trade Diagnostic Study.’’ Washington, D.C. Yemtsov, R. 2001. ‘‘Labor Markets, Inequality and Poverty in Georgia.’’ IZA Discussion Paper 251. Institute for the Study of Labor, Bonn.