94670 MIGRATION AND ECONOMIC MOBILITY IN TANZANIA: EVIDENCE FROM A TRACKING SURVEY Kathleen Beegle, Joachim De Weerdt, and Stefan Dercon* Abstract—This study explores to what extent migration has contributed to that migration patterns for marriage in rural India are con- improved living standards of individuals in Tanzania. Using a thirteen- year panel survey, we find that migration between 1991 and 2004 added sistent with the risk-sharing strategies of the initial house- 36 percentage points to consumption growth. Although moving out of hold. Recent evidence has highlighted not just the role of agriculture resulted in much higher growth than staying in agriculture, networks in facilitating migration from home areas, but also growth was always greater in any sector if the individual physically moved. As to why more people do not move given the high returns to how migration is closely linked to migrants’ access to social geographical mobility, analysis finds evidence consistent with models in networks in destination areas (Munshi, 2003) and to com- which exit barriers set by home communities prevent the migration of munity rates of out-migration (Kilic et al., 2009). some categories of people. Although the emphasis on the process of migration in most recent empirical work has provided many insights, few of these studies convincingly address the question of whether migration leads to improved living conditions. A I. Introduction major problem is having access to data that allow a careful and convincing assessment of the relative welfare of F INDING routes out of poverty remains a key issue for households and policymakers alike. A long-term vision of development suggests that poverty reduction is asso- migrants and nonmigrants, due to the standard evaluation problem: an individual cannot be observed to be both a ciated with intergenerational mobility out of rural areas and migrant and a nonmigrant. A few studies have access to agriculture and into urban nonagricultural settings. Physical experimental data, such as international migration lotteries and economic mobility seem to go hand-in-hand. Standard (McKenzie, Gibson, & Stillman, 2010), but most studies economic theory has multiple narratives of how physical have to work with nonexperimental data. Without experi- and economic mobility interact. The Lewis model offers a mental data, the key concern, unobserved heterogeneity stylized description of rural transformation, with sector affecting both outcomes and the process of migration, per- mobility of labor from agriculture into ‘‘modern’’ produc- sists. This leads to the quest for imaginative and convincing tion processes. At least in its original specification, the instruments for migration (see the review of the migration model suggests an initial gap in earnings between rural and and poverty literature by McKenzie & Sasin, 2007). An urban locations (Lewis, 1954).1 The Harris-Todaro model additional hurdle is the need for panel data to study migra- emphasizes the migration process and the fact that relative tion and economic mobility. The costs and difficulties of individual earnings incentives matter, so that both pull and resurveying means that attrition may be relatively high for push factors drive migration. A gap between rural and this group and may also result in the loss of some of the expected urban earnings drives migration. Unemployment most relevant households for the study of this process (or an informal sector offering low earnings) would never- (Beegle, 2000; Rosenzweig, 2003). theless allow an actual gap between urban and rural wages This paper uses unique data from a region in Tanzania to to persist, with the premium a function of the unemploy- address the question, What is the impact on poverty and ment rate (Harris & Todaro, 1970). Other work, such as on wealth of physical movement out of the original commu- the ‘‘new economics of migration’’ (Stark & Bloom, 1985), nity? Although we do not have experimental data, the nat- emphasizes that migration is part of a general livelihood ure of our data allows us to limit the potential sources of strategy for the initial household as a whole. Migration is unobserved heterogeneity. Building on a detailed panel sur- part of a welfare-maximizing strategy with a clear role for vey conducted in the early 1990s, we reinterviewed indivi- overall household income growth, but also a role for risk duals in 2004, making a notable effort to track individuals sharing. For example, Rosenzweig and Stark (1989) find who had moved. The tracking of individuals to new locations proves cru- cially important for assessing welfare changes among the Received for publication December 31, 2008. Revision accepted for baseline sample. The average consumption change of indi- publication March 16, 2010. viduals who migrated was more than four times greater than * Beegle: World Bank; De Weerdt: EDI, Tanzania; Dercon: Oxford University. that of individuals who did not move. Those who had We thank Karen Macours, David McKenzie, Peter Neary, two anon- moved out of Kagera by 2004 experienced consumption ymous referees, the editor of this journal, and seminar participants at the growth that was ten times greater compared with those who Massachusetts Avenue Development Seminar, Oxford University, and the World Bank for very useful comments. All views are our own and do not remained in their original community. These averages necessarily reflect the views of the World Bank or its member countries. translate into very different patterns of poverty dynamics 1 For example, Lewis (1954, p. 150) wrote: ‘‘Earnings in the subsistence for the physically mobile and immobile. For those who sector set a floor to wages in the capitalist sector, but in practice wages have to be higher than this, and there is usually a gap of 30 per cent or stayed in the community, the poverty rate decreased by more between capitalist wages and subsistence earnings.’’ about 4 percentage points over the thirteen years. For those The Review of Economics and Statistics, August 2011, 93(3): 1010–1033 Ó 2011 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1011 who moved elsewhere within the region, the poverty rate behind, implying that within-family migration may not decreased by about 12 percentage points, and for those who be random.2 We use two-stage least squares (2SLS) meth- moved out of the region, the poverty rate decreased by 23 ods to deal with this potential endogeneity. We assert that percentage points. Had we not tracked and interviewed peo- opportunities to migrate depend on the interaction of house- ple who moved out of the community, a practice that is not hold circumstances with the individual’s status and position carried out in many panel surveys, we would have seriously within the household at baseline. The 2SLS estimates show underestimated the extent to which poverty decreased dur- limited evidence of unobserved individual heterogeneity ing 1991 to 2004 in Kagera; we would have reported pov- affecting consumption growth. In short, unobservables at erty reduction at about half its true value. Clemens and the household level correlated with growth potential appear Pritchett (2008) raise similar concerns in the context of to matter, whereas individual heterogeneity does not. income growth and international migration. Furthermore, We explore two additional avenues of interest. First, does the tracking and reinterviewing enabled us to collect valu- migration to urban areas drive the results? Second, does able information about pathways out of poverty. migration capture changes in the sector of work that would Still, these statistics do not provide evidence that moving explain the consumption growth we observe? We find sug- out of the community leads to higher income growth. As gestive evidence that physical mobility has an independent noted above, we cannot observe the counterfactual: What effect beyond its association with moving out of agriculture would income growth have been for migrants had they not or moving to a more urban area. We use these results in migrated? We exploit some unique features of the data to conjunction with the literature on network externalities and address concerns about unobserved heterogeneity. First, poverty traps to explain why, if migration has such large individual fixed-effects regressions for movers and stayers payoffs, more people do not move. We conclude that the produce a difference-in-difference estimation of the impact findings are consistent with models in which exit barriers of physical movement, controlling for any fixed individual are set by home communities (through social and family factors that affect consumption. Second, we can control for norms), preventing migration of certain categories of people initial household fixed effects in the growth rate of con- when windows of opportunity arise. Being willing and able sumption because we observe baseline households in which to leave behind what you know appears to be a strong deter- some individuals migrate and others do not. This controls minant of economic mobility. There is no evidence of finan- for observable and unobservable factors fixed to the family cial constraints to migration. that can affect the growth rate of consumption. Thus, we In the next section, we provide the context of changes in identify the impact of migration on income using within- economic fortunes in Tanzania in the past decade. Section household variation in migration. Unlike most other studies III presents the data used in the analysis, and section IV of migration, our identification does not rely on household provides the basic indicators we use to assess economic and shocks, distances to possible destinations, or the existence welfare changes. Section V briefly describes the method we of family networks at the destination to identify the migra- use to assess the impact of migration, section VI presents tion decision. Such variables are likely to have an impact the results, and section VII carries out some robustness on the income of those migrating as well as those staying checks. Section VIII builds a narrative around the regres- behind, and so the exclusion restriction will not be satisfied. sions and aims to explain why more people do not migrate In our study, we are able to move beyond these approaches; when the benefits of doing so are so high. in addition to using panel data on migrants and nonmi- grants, we compare siblings and other relatives who were living together at the baseline. II. The Setting: Tanzania and Kagera, 1994–2004 These estimations address many possible sources of het- Between 1994 and 2004, Tanzania experienced a period erogeneity, such as (genetic) health and ability endow- of relatively rapid macroeconomic growth, attributed to lib- ments; risk aversion; wealth constraints; and market, risk, eralization, a renewed trade orientation, a stable political and environmental circumstances. We find that movement context, and a relatively positive business climate that out of the community results in 36% higher consumption helped to boost economic performance. Real GDP growth relative to staying. Comparison of the results with and with- was on the order of 4.2% per year between 1994 and 2004, out fixed effects suggests that migrants are more likely to and annual population growth was around 3.2% (United be from families with greater potential for growth in earn- Republic of Tanzania, 2004). There is also evidence that ings. growth accelerated in the last few years of the period com- A weakness of this approach, however, is the implicit pared with the 1990s. However, growth was not sufficiently assumption that within families, migration is random, broad-based to result in rapid poverty reduction. On the which is a strong assumption. For example, in view of the basis of the available evidence, poverty rates declined only standard Harris-Todaro model of individual migration, slightly, and most of the progress in poverty reduction was earnings differentials drive migration, so those who are observed to have migrated from within a household tend to 2 This is correct, even if, in equilibrium, when no further migration have had greater earning potential than those who stayed takes place, expected earnings are equal. 1012 THE REVIEW OF ECONOMICS AND STATISTICS in urban areas. According to the Household Budget Survey usually not as a unit, the 2004 round had more than 2,700 (HBS), between 1991 and 2000/01, poverty declined from household interviews (from the baseline sample of 912 39% to 36% in mainland Tanzania. The decline in poverty households). was steep in Dar es Salaam (from 28% to 18%) but minimal Although the KHDS is a panel of respondents and the con- in rural Tanzania (from 41% to 39%). cept of a ‘‘household’’ after ten to thirteen years is a vague For the purposes of this study, it is useful to consider the notion, it is common in panel surveys to consider recontact Kagera region specifically. The region is far from the capi- rates in terms of households. Excluding households in which tal and the coast, bordering Lake Victoria, Rwanda, Bur- all previous members were deceased (17 households with 27 undi, and Uganda. It is overwhelmingly rural and primarily people), the field team managed to recontact 93% of the engaged in producing bananas and coffee in the north and baseline households. This is an excellent rate of recontact rain-fed annual crops (maize, sorghum, and cotton) in the compared with panel surveys in low-income and high- south. Relatively low-quality coffee exports and agricul- income countries. The KHDS panel has an attrition rate that tural produce are the main sources of income. Mean per is much lower than that of other well-known panel surveys capita consumption was near the mean of mainland Tanza- summarized in Alderman et al. (2001), in which the rates nia in 2000. Likewise, the region appeared to mirror the rest ranged from 17.5% attrition per year to the lowest rate of of the country in terms of growth and poverty reduction: 1.5% per year, with most of these surveys covering consider- real GDP growth was just over 4% per year between 1994 ably shorter time periods (two to five years). and 2004, while poverty in Kagera is estimated to have Figure 1 charts the evolution of households from the base- fallen from 31% to 29% between 1991 and 2000/01 line to 2004. Half of all households interviewed were track- (Demombynes & Hoogeveen, 2007). ing cases, meaning they did not reside in the baseline com- The challenges of poverty reduction in Kagera seem to munities. Of those households tracked, only 38% were be representative for provincial Tanzania as a whole: some located nearby the baseline community. Overall, 32% of all pockets, such as Dar es Salaam, have had substantial households were neither located in nor relatively close to growth and poverty reduction, but this has not spread to the baseline communities. While tracking is costly, it is an other areas. This reflects the typical problem of landlocked, important exercise because migration and dissolution of agriculture-based economies: how to deliver poverty reduc- households are often hypothesized to be important responses tion if the main engine of growth appears to be elsewhere to hardship and a strategy for escaping poverty. Excluding (De Weerdt, 2010). these households in the sample raises obvious concerns regarding the selectivity of attrition. In particular, out-migra- III. The Data tion from the village, dissolution of households, and even marriage may be responses to changing economic or family The Kagera Health and Development Survey (KHDS) circumstances. Tracking surveys provide a unique opportu- was originally conducted by the World Bank and Muhim- nity to study these responses: who uses them, their effects, bili University College of Health Sciences (MUCHS) and and whether they get people out of poverty. consisted of about 915 households interviewed up to four Turning to the recontact rates of the sample of 6,352 times from fall 1991 to January 1994 at intervals of six to respondents, table 1 shows the status of the respondents by seven months (see World Bank, 2004, and http:// age group (based on their age at first interview in the 1991– www.worldbank.org/lsms/). The KHDS 1991–1994 serves 1994 rounds). The surviving older respondents were much as the baseline data for this paper. Initially designed to more likely to be located, which is consistent with higher assess the impact of the health crisis linked to the HIV- migration rates among the young adults in the sample. AIDS epidemic in the area, it used a stratified design to Among the youngest respondents, more than three-quarters ensure relatively appropriate sampling of households with were successfully reinterviewed. Excluding people who adult mortality. Comparisons with the 1991 HBS suggest died, 82% of all respondents were reinterviewed. Table 2 that in terms of basic welfare and other indicators, the shows the location of the respondents. Without tracking, KHDS can be used as a representative sample for this per- reinterview rates of surviving respondents would have iod for Kagera (although not necessarily for the rest of Tan- fallen from 82% to 52% (2,780 of 5,394 survivors). Non- zania; the results are available on request). local migration is important: restricting the tracking to The objective of the KHDS 2004 survey was to reinter- nearby villages would have resulted in 63% recontact of view all individuals who were household members in any survivors. Migration also proved to be an important factor round of the KHDS 1991–1994 and who were alive at the in determining whether someone was recontacted. Respon- last interview (Beegle, De Weerdt, & Dercon, 2006). This dents who were not traced were much more likely to reside effectively meant turning the original household survey into outside Kagera (43%) compared with their counterparts an individual longitudinal survey. Each household in which who were reinterviewed (8%). any of the panel individuals lived would be administered The consumption data come from an extensive consump- the full household questionnaire. Because the set of house- tion module administered in 1991 and again in 2004. The hold members at the baseline had subsequently moved, and consumption aggregate includes home-produced and pur- MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1013 FIGURE 1.—KHDS 2004: RECONTACTING RESPONDENTS AFTER TEN OR MORE YEARS TABLE 1.—KHDS INDIVIDUALS, BY AGE Age at Baseline, Reinterview Rate 1991–1994 Recontacted Deceased Untraced among Survivors <10 years 1,604 (77.1%) 160 (7.7%) 317 (15.2%) 83.5% 10–19 years 1,406 (73.2%) 104 (5.4%) 412 (21.4%) 77.3% 20–39 years 823 (63.3%) 285 (22.1%) 190 (14.6%) 81.2% 40–59 years 436 (70.6%) 147 (23.9%) 34 (5.5%) 92.8% 60þ years 163 (37.6%) 262 (60.4%) 9 (2.1%) 94.8% Overall 4,432 (69.7%) 958 (15.1%) 962 (15.1%) 82.2% Sample of individuals ever interviewed in KHDS 1991–1994 and alive at last interview. Age categories are based on age at first interview. TABLE 2.—KHDS REINTERVIEW RATES, BY LOCATION chased food and nonfood expenditure. The nonfood compo- Number Location % nent includes a range of nonfood purchases, as well as utili- ties, expenditure on clothing and other personal items, Baseline sample 6,352 Reinterviewed 4,432 transfers out, and health expenditures. Funeral expenses Same community 63.1 and health expenses prior to the death of an ill person were Nearby community 14.1 excluded. Monetary levels were adjusted to account for spa- Elsewhere in Kagera 14.4 Other region 7.1 tial and temporal price differences, using price data col- Other country 1.3 lected in the Kagera survey in 1991 and 2004, and, for Untraced 962 households outside Kagera, data from the National House- Kagera 56.6 Dar es Salaam 12.3 hold Budget Survey. Consumption is expressed in annual Mwanza 10.4 per capita terms. The poverty line is set at 109,663 Tanza- Other region 7.9 Other country 5.5 nian shillings (TSh), calibrated to yield for our sample of Don’t know 7.3 respondents who remained in Kagera the same poverty rate Deceased 958 as the 2000/01 National Household Budget Survey estimate Location for untraced respondents is reported by other household members from the baseline survey who were successfully located, interviewed, and able to provide location information on the respondent. In for Kagera (29%). At the time of the survey one U.S. dollar some cases, this information comes from other relatives or neighbors residing in the baseline communities. was worth around TSh 1,100. 1014 THE REVIEW OF ECONOMICS AND STATISTICS TABLE 3.—AVERAGE CONSUMPTION MOVEMENTS OF PANEL RESPONDENTS, BY 2004 LOCATION Mean 1991 Mean 2004 Difference in Means N Consumption poverty head count (%) Full sample 0.34 0.27 À0.07*** 4,116 Within community 0.35 0.31 À0.03*** 2,620 Nearby community 0.33 0.21 À0.11*** 577 Elsewhere in Kagera 0.36 0.24 À0.12*** 595 Out of Kagera 0.30 0.07 À0.23*** 324 Consumption per capita (TSh) Full sample 164,434 226,337 61,903*** 4,116 Within community 159,959 186,474 26,515*** 2,620 Nearby community 171,493 234,973 63,480*** 577 Elsewhere in Kagera 167,597 260,749 93,152*** 595 Out of Kagera 180,707 472,474 291,767*** 324 Food consumption per capita (TSh) Full sample 106,805 146,701 39,896*** 4,116 Within community 104,184 121,725 17,541*** 2,620 Nearby community 111,207 152,624 41,417*** 577 Elsewhere in Kagera 108,763 166,379 57,616*** 595 Out of Kagera 115,704 303,453 187,749*** 324 Nonfood consumption per capita (TSh) Full sample 57,629 79,636 22,007*** 4,116 Within community 55,775 64,748 8,973*** 2,620 Nearby community 60,286 82,348 22,062*** 577 Elsewhere in Kagera 58,834 94,369 35,535*** 595 Out of Kagera 65,003 169,021 107,018*** 324 Significance of the difference with the 1991 value using a paired t-test. *10%, **5%, ***1%. IV. Growth, Poverty, and Physical Mobility in Kagera TABLE 4.—DIFFERENCES IN CONSUMPTION AND POVERTY HEAD COUNT CHANGES, BY MOBILITY CATEGORIES In this section, we discuss changes in living standards t-Test for Equality overall and the changes for four mutually exclusive groups Average Change between based on residence in 2004: (a) still residing in the baseline N Change Both Subgroups community, (b) residing in a neighboring community, (c) Consumption per capita (TSh) residing elsewhere in Kagera, and (d) residing outside Stayed in community 2,620 25,940 t ¼ 13.93 Moved elsewhere 1,496 120,534 p ¼ 0.0000 Kagera. Stayed in same or 3,197 31,432 t ¼ 16.67 Table 3 shows that the basic needs poverty rate declined neighboring community 8 percentage points in the full sample. This figure masks Moved elsewhere 919 160,820 p ¼ 0.0000 significant differences in changes between subgroups based Stayed in Kagera 3,792 41,460 t ¼ 20.25 on migration. For those found residing in the baseline com- Moved elsewhere 324 281,064 p ¼ 0.000 munity, poverty rates dropped by 3 percentage points, but Poverty head count (%) rates dropped by 11, 12, and 23 percentage points for those Stayed in community 2,620 À0.034 t ¼ 5.41 Moved elsewhere 1,496 À0.140 p ¼ 0.000 who moved to neighboring communities, elsewhere in Kagera, and outside Kagera, respectively. A similar pattern Stayed in same or 3,197 À0.047 t ¼ 5.11 neighboring community is found for consumption per capita. Although mean con- Moved elsewhere 919 À0.162 p ¼ 0.000 sumption per capita grew by TSh 61,903 overall, or 38%, it Stayed in Kagera 3,792 À0.059 t ¼ 4.94 grew by only 17% for those found in the same community Moved elsewhere 324 À0.231 p ¼ 0.000 and by 37%, 56%, and 161% for those who moved to neigh- boring communities, elsewhere in Kagera, and outside Kagera, respectively. Dividing consumption into food and nonfood components gives the same result. The most basic mental samples (which are not mutually exclusive); it gives assessment of welfare changes would have been wrong if a more detailed picture of how inference on consumption we had focused only on individuals still residing in the growth and poverty reduction would have changed if we community, a practice found in many panel data surveys. had not tracked movers. It is apparent that inference from a We would have underestimated the growth in consumption ‘‘simple’’ panel survey of respondents continuing to reside by half of its true increase. within the original communities would have produced For the groups in table 3, the differences in consumption underestimates of actual consumption growth and poverty changes are statistically significant, as shown in table 4. reduction in this population. Excluding respondents who have relocated would omit These conclusions are robust across the distribution of those with greater rates of income growth and poverty consumption, as well as at the mean and the poverty line. reduction. Table 5 reports confidence intervals for the incre- Panel A in figure 2 depicts the cumulative density function MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1015 TABLE 5.—SAMPLE SIZE, MEAN, STANDARD ERROR, AND 95% CONFIDENCE INTERVAL FOR INCREMENTAL SAMPLES N Mean SE 95% CI Change in consumption per capita (TSh) (1) ¼ Only those who remained in community 2,620 25,940 3,057 19,945 31,935 (2) ¼ (1) þ those who moved to neighboring communities 3,197 31,432 2,878 25,790 37,074 (3) ¼ (2) þ those who moved elsewhere within Kagera 3,792 41,460 2,985 35,609 47,312 (4) ¼ (3) þ those who moved outside Kagera Region (¼ full sample) 4,061 56,392 3,259 50,003 62,782 Change in poverty head count (%) (1) ¼ Only those who remained in community 2,620 À0.034 0.012 À0.058 À0.010 (2) ¼ (1) þ those who moved to neighboring communities 3,197 À0.047 0.011 À0.068 À0.025 (3) ¼ (2) þ those who moved elsewhere within Kagera 3,792 À0.059 0.010 À0.078 À0.039 (4) ¼ (3) þ those who moved outside Kagera (¼ full sample) 4,061 À0.068 0.009 À0.087 À0.049 FIGURE 2.—CUMULATIVE DENSITY FUNCTIONS OF CONSUMPTION PER CAPITA (TRUNCATED AT TSH 500,000) The vertical line is the basic needs poverty line (TSh 109,663). for consumption per capita for those who remained in the close to each other under the poverty line and diverge above same community. Panels B, C, and D show the cumulative it; for the other mobility categories, there is greater diver- density functions for respondents residing in neighboring gence. communities, elsewhere in Kagera, and outside Kagera. For Figures 3, 4, and 5 offer another cut of the data, compar- respondents who were located farther from their location in ing consumption of nonmovers to movers in 1991 when 1991, the differences between the functions for 1991 and both were living in the same community (panel A) and in 2004 are more pronounced. For people who remained in the 2004 (panel B). There is almost no difference between baseline community, the 1991 and 2004 distributions lie (future) nonmovers and movers in 1991, but by 2004, we 1016 THE REVIEW OF ECONOMICS AND STATISTICS FIGURE 3.—CUMULATIVE DENSITY FUNCTIONS OF CONSUMPTION PER CAPITA WITHIN COMMUNITY VERSUS NEARBY COMMUNITY (TRUNCATED AT TSH 500,000) The vertical line is the basic needs poverty line (TSh 109,663). FIGURE 4.—CUMULATIVE DENSITY FUNCTIONS OF CONSUMPTION PER CAPITA WITHIN COMMUNITY VERSUS ELSEWHERE IN KAGERA (TRUNCATED AT TSH 500,000) The vertical line is the basic needs poverty line (TSh 109,663). FIGURE 5.—CUMULATIVE DENSITY FUNCTIONS OF CONSUMPTION PER CAPITA WITHIN COMMUNITY VERSUS OUTSIDE KAGERA (TRUNCATED AT TSH 500,000) The vertical line is the basic needs poverty line (TSh 109,663). observe divergent income levels. The divergence is greater What drives the association between migration and between those who stayed and those who moved farther income growth? One plausible explanation is that migrants away (figures 4 and 5). are relocating to less remote, less poor areas. By 1991, 68% MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1017 TABLE 6.—MEAN AND MEDIAN CONSUMPTION GROWTH BY MOVE TO MORE OR TABLE 8.—MEAN CONSUMPTION GROWTH BY SECTOR ALLOCATION AND PHYSICAL LESS REMOTE AREA, 1991–2004 MOVEMENT, 1991–2004 Mean Median N Stayed in Moved out of Community Community All Did not move 0.13 0.16 2,147 Moved out of community 0.53 0.50 1,080 Stay in agriculture 0.18 0.29 0.22 Out of those that moved out of community: (1,248) (473) (1,721) Moved to more remote area 0.28 0.21 380 Move out of agriculture 0.42 1.04 0.67 Moved to similar area 0.46 0.45 378 into nonagriculture (201) (207) (408) Moved to less remote area 0.90 0.86 322 Stay in nonagriculture 0.11 0.88 0.44 Remoteness is based on the changes in classification among six possibilities: in order of remoteness, (88) (84) (172) island in Lake Victoria, remote village, connected village, urban center, district capital, and regional Move into agriculture À0.12 À0.00 À0.03 capital. from nonagriculture (157) (88) (245) Total 0.18 0.49 0.27 (1,694) (852) (2,546) TABLE 7.—MEAN AND MEDIAN CONSUMPTION GROWTH BY SECTOR ALLOCATION CHANGE, 1991–2004 Mean Median N ture (67%), and there was considerable growth for those Stay in agriculture 0.21 0.22 1,721 who started in nonagriculture. It is striking that the 10% Move out of agriculture into nonagriculture 0.69 0.67 408 who moved into agriculture from nonagriculture faced Stay in nonagriculture 0.43 0.43 172 Move into agriculture from nonagriculture À0.05 À0.03 245 declining consumption, suggesting that this is a sign of Total 0.28 0.27 2,546 hardship and possibly a means of coping with it. Table 8 reports consumption growth by both sector change and migration. A considerable number of people switched sec- tors without migrating but, within each category of sector of the sample was living in rural villages, of which a little status, migrants had much higher consumption growth than more than half were categorized by the survey team as nonmigrants. The main source of income matters for con- poorly connected in terms of infrastructure. The remainder sumption growth, but it is strongly related to migration as of the sample were living in (or close to) the regional capi- well. For example, those who moved out of agriculture tal, Bukoba (17%), or other small urban centers in Kagera while also moving out of their original community in this (14%). Table 6 investigates whether moving to a better- period more than doubled their consumption levels, while connected center (for example, from a poorly connected to those who switched into agriculture while staying within a better-connected village or from a rural area to an urban the community faced a 12% reduction in consumption. center) is correlated with higher consumption growth.3 This is indeed the case: about 10% of the sample moved to a V. Assessing the Impact of Migration on Consumption better-connected area, and they experienced 90% consump- Outcomes tion growth on average. For those who moved to a similar area, consumption increased by 46% on average, while those The correlations above do not resolve whether this con- who moved to a less urban or less-connected center experi- sumption growth is in fact directly related to migration or enced a lower increase at 28%. Clearly, it matters where whether it is spurious. To investigate this further, we ex- people move, but moving in itself seems to matter too.4 plore several empirical approaches. First, we employ a dif- Another plausible source of income growth for migrants ference-in-difference estimator, comparing the consump- is that they have moved to a different sector with respect to tion growth of those who moved with those who stayed in income. In table 7, we explore whether migration is corre- their baseline community. We define ln Cit as the natural lated with change in occupation or sector. Consumption logarithm of consumption per capita for individual i in per- growth was highest for those who moved into nonagricul- iod t, and Mi as a dummy that is 1 if the individual was found to have physically moved out of the original commu- nity between t and t þ 1, and 0 otherwise. The difference- 3 Tables 6 and onward are restricted to the sample in the main regres- in-difference specification is sions (N ¼ 3,227). From the full sample of 4,432, we exclude, in this order: 715 people who were not interviewed in wave 1 (they were inter- Dln Citþ1;t ¼ a þ bMi þ cXit þ dih þ eit; ð1Þ viewed in waves 2, 3, and/or 4), 15 people in one-person households, 267 people missing either wave 1 or wave 5 consumption expenditure, 120 people missing peer’s schooling, 2 people missing parental education, in which Dln Citþ1,t is (ln Citþ1 À ln Cit), the growth rate of and 86 people with incomplete data in wave 1. Tables 7, 8, and 12 have consumption per capita in the household in which i is resid- 2,546 observations because of missing occupational data for 2004. ing in the two periods. This specification controls for indivi- 4 In order to investigate the clustering of migration patterns, all house- holds were sorted into tracking zones, indicating the geographical area in dual fixed heterogeneity, which might have an impact on which they resided in 2004. Tabulating, for each tracking zone, the vil- the level of consumption in each period. This resolves a lage of origin of the households tracked in that zone did not reveal any large number of possible sources of endogeneity, such as discernable pattern of clustered migration. In each tracking zone, there was never any origin village that dominated, with the exception of vil- risk aversion or ability, which are likely to affect both lages that lie within or neighbor the tracking zone. migration and income outcomes. However, it does not 1018 THE REVIEW OF ECONOMICS AND STATISTICS address concerns about heterogeneity among families or tances), are likely to affect migration, but also may have individuals affecting growth in consumption and the migra- direct effects on consumption growth. tion decision. For example, current wealth may affect the Despite controlling for fixed individual heterogeneity ability to migrate as well as the potential to grow between t and both fixed and time-varying household-level heteroge- and t þ 1. McKenzie and Sasin (2007) discuss at length the neity (including initial growth paths) and the additional issue of endogeneity with respect to measuring the impact control variables, unobserved individual factors may still of migration on poverty, stating that work that does not affect migration as well as consumption growth. We extend identify causal relations provides ‘‘rather weak grounds for the analysis to 2SLS estimates, using three types of policy recommendations.’’ McKenzie et al. (2010) find that variables for instruments for the migration decision: pull ignoring selection led to overstating the gains from migra- factors, push factors, and variables reflecting social rela- tion from Tonga to New Zealand. tionships. Our data, while not experimental, still offer excellent The pull factors include age and baseline location. opportunities to control for a wide set of factors in this Migration opportunities and incentives are typically stron- respect. First, we have data on multiple individuals from ger for young male adults, as employment in low-skill and the original household, which allows us to control for any physically demanding activities is likely to be easier for initial household-level heterogeneity (dih) that may affect them. Similarly, if a family were to decide who should the growth of consumption by estimating equation (1) using migrate to capture opportunities, then allowing a young initial household fixed effects (IHHFE). The result is that male adult to go would seem sensible. Costs and informa- the impact of migration is identified using variation within tion needs for migration may well be affected by how far the initial household: differences between members of the the opportunities are located. We include an interaction same initial household, effectively controlling for initial term of the distance to the regional capital and whether the growth paths. Second, we can control for a set of individual- person is male and between 5 and 15 years old at the base- level factors that may affect consumption growth, and line (so between 18 and 28 in 2004) as a measure of the possibly migration as well, by including these as Xi in opportunities available.6 regression model (1). The variables used as individual con- Individuals may also be pushed into migration (or ditioning variables for the growth of consumption from families may decide to send someone) when shocks occur. baseline are individual variables (sex, age, education rela- We include a measure of economic shocks experienced by tive to age-specific peer groups,5 and marital status) and the household by including a measure of negative rainfall family background variables (number of biological children shock. Using data from 21 weather stations in Kagera from in the initial household at baseline interacted with the age- 1980 to 2004, each of the 51 baseline villages was mapped sex group of the children, the number of biological children to the nearest station, and 25-year average annual rainfall living elsewhere interacted with the distance to the regional was computed. The largest deviation of rainfall between capital, and the years of education of the biological mother 1992 and 2002 from the long-run average was identified. and father). We also include a variable indicating whether This rainfall shock variable was interacted with being in the the individual lost both parents between 1991 and 2004, 5-to-15 age group as a measure of this push factor (with allowing a separate effect if the individual was below age higher values defined as high-deviation rainfall). 15 at baseline. Quite a few of these variables, such as edu- Finally, norms and social circumstances are likely to cational level, marital status, parental death, or having chil- affect migration. In particular, within a household, who is dren living elsewhere (offering opportunities for remit- able or expected to migrate is likely to be determined by the individual’s position in the household. We include indi- cators for being the head or spouse of the household head at 5 We used the variable ‘‘years of schooling completed relative to peers’’ the baseline. We expect these two positions in the house- rather than a straight ‘‘years of schooling completed’’ for two reasons. First, a substantial number of individuals in the sample were younger than hold would make it less likely that the person would leave 18 at the baseline and therefore had not necessarily completed their edu- relative to others in the household. Age rank among those cation. As such, years of schooling at the baseline might be less correlated between 5 and 15 (with the oldest receiving the highest with a move by 2004 than, say, eventual completed years of schooling. Second, akin to this concern, years of schooling is highly correlated with value) is also included. These indicators are unlikely to age for individuals of school-going age. The regressions also include a set determine the consumption growth of the household, but of age variables, defined in broad age groups (for ease of interpretation may well affect whether a person is allowed, chosen, or and discussion of results). One consequence could be that years of school- ing at the baseline would pick up at least some age effect. To address this chooses to migrate. Finally, close family members, the clo- concern, rather than use education in years, we constructed a variable of sest relatives of the household head, sons and daughters, education relative to peers: the absolute deviation of education levels may have different probabilities of leaving the household’s compared with mean education of age-specific peers at the baseline for those younger than 18 and relative to other adults for the rest of the sam- community, compared with other residents, such as cousins ple. This purges the education variable of an age effect it would otherwise or nephews. Local norms on marriage are patrilocal: girls pick up. All the regressions below were repeated using a straightforward ‘‘years of schooling’’ variable rather than our ‘‘years of schooling relative 6 to peers’’ variable. Neither the results nor their interpretation were The noninteracted variables are all included as determinants of con- affected. sumption growth via Xi and dih. MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1019 are expected to move to the community of their husbands the next section on alternative definitions of the consump- after marriage, and husbands are expected to stay where tion aggregate, excluding transfers out). It also seems coun- their father was based. We include an indicator for being ter to the theory that the migration decision is part of a the son of the household head at baseline. Although both household-level maximization strategy (although it cannot sons and daughters of the head may be expected to be more preclude that this is partly true). likely to stay in the community than other initial household For the first-stage results in table 10, in terms of basic members, patrilocality would make this probability higher diagnostics, our set of excluded instruments appears strong for boys than for girls. and valid: the Cragg-Donald (F) test shows a value of 11.70 In sum, this means we are using a set of six instruments. for the movement dummy and 9.07 for the distance regres- Although we will show in the appendix that statistically sion. Especially in the former case, it is comfortably above convincing and close to identical results are obtained by the level of 10 often recommended for rejecting weak using subsets of these instruments, we focus on the full set instruments (and in the latter case, still with relative limited of instruments in the discussion of the results. bias in the tables in Stock & Yogo, 2002). The results Although our main measure of migration (Mi) is an indi- are also robust to exclusion of any of the instruments (see cator for having moved, we also substitute this for the log table A2). of the distance moved (kilometers from the original com- Some interesting patterns explaining migration emerge munity of the location in which the individual was found in from table 10. First, education offers strong and convex 2004, ‘‘as the crow flies,’’ set to 0 for nonmovers). We also effects in leaving one’s community. Being unmarried, espe- extend the multivariate analysis to explore whether moving cially being female and unmarried, is correlated with a to a more urbanized area or changing employment sectors higher probability of migration (consistent with patrilocal- plays a role in increasing consumption growth. ity, whereby females move out of the paternal location at the time of marriage). When looking more specifically at VI. Regression Results the identifying instruments, we find significant effects, con- sistent with expectations: positional variables in the house- Table 9 presents the basic results for the initial household hold matter, with the head and spouse less likely to leave, fixed effects (IHHFE) and 2SLS estimates. (Table A1 pre- as are children of the head (relative to others belonging to sents the means and standard deviations for the covariates.) the household). The effect is, however, considerably larger We estimate the regressions using an indicator for having (more negative) for male children of the head—again con- moved and a measure of the distance of the move. The sistent with patrilocality, as marriage norms make sons 2SLS estimates in columns 3 and 4 use the six instruments more likely to be expected to stay in the community than defined above. In table 10, we present the first-stage results daughters. Older members among the children in the house- of regressions explaining migration or the distance traveled hold are more likely to migrate, possibly reflecting some in migration. kind of pecking order, given the opportunities available. Before turning to the variables of interest, we briefly dis- Rainfall shocks increase the probability of leaving. Finally, cuss the coefficients on the control variables. Recall that all pull factors, like the interaction of being young, male, and effects are identified using variation within initial house- residing close to the regional capital, increase the probabil- hold. Those who are relatively better educated at baseline, ity of leaving. The results are also consistent for the regres- relative to their peers and within the household, experi- sions with the dummy variable for migration and with the enced much higher consumption growth, and the effect is distance-migrated variable. In short, although not aiming to strongly convex. Having an educated father has an addi- obtain a structural model, we find suggestive correlates for tional effect on growth. The younger cohort did consider- the process of migration from within households. These ably better, as did males still unmarried at baseline. include better income opportunities (education and distance Turning to the migration variables, we observe in the to the regional capital), norms of settlement and marriage, IHHFE regression a larger and statistically significant and other social factors. impact of migration on consumption growth. Moving out of The 2SLS results (IV with fixed effects) in columns 3 the community resulted in a 36 percentage point increase and 4 of table 9 are almost identical to the IHHFE results. in consumption growth over the thirteen-year period. As They are slightly less statistically significant (as can be migrants move farther from their baseline community, the expected from IV regressions given their lower efficiency) impact is greater. These effects are large, with migration but still significant at 5%. Thus, there is no evidence that resulting in a large divergence in income between people unobserved individual time-varying heterogeneity affects who initially lived together—usually parents, siblings, and the noninstrumented results. For the distance variables, the other close relatives. Because this is the impact comparing results are marginally smaller (the coefficient is 0.10 com- within families, it nets out any transfers from migrants to pared with 0.12), suggesting limited evidence of a positive nonmovers. That is, if migrants sent remittances back to bias in the earlier results (migrants traveling longer dis- their origin households, then the estimates in table 9 are a tances are those with somewhat higher unobserved con- lower bound of the impact of moving (see also the results in sumption growth potential, consistent with expectations). 1020 THE REVIEW OF ECONOMICS AND STATISTICS TABLE 9.—EXPLAINING CONSUMPTION CHANGE: IHHFE AND 2SLS WITH IHHFE (1) (2) (3) (4) IHHFE IHHFE 2SLS with IHHFE 2SLS with IHHFE Moved outside community 0.363*** 0.378** (0.025) (0.150) Kilometers moved (log of distance) 0.120*** 0.104** (0.006) (0.043) Individual characteristics at baseline Deviation of years schooling from peers 0.013** 0.009 0.013** 0.010 (0.006) (0.006) (0.006) (0.006) Squared deviation of years schooling from peers 0.004*** 0.004*** 0.004*** 0.004*** (0.001) (0.001) (0.001) (0.001) Male À0.004 À0.009 À0.003 À0.010 (0.038) (0.037) (0.038) (0.037) Unmarried À0.023 À0.020 À0.027 À0.011 (0.056) (0.054) (0.064) (0.060) Unmarried male 0.141*** 0.131*** 0.144*** 0.123** (0.045) (0.044) (0.053) (0.049) Both parents died À0.006 0.013 À0.006 0.010 (0.084) (0.081) (0.083) (0.082) Above 15 and both parents died 0.050 0.024 0.048 0.033 (0.100) (0.098) (0.101) (0.100) Years of education mother À0.003 À0.004 À0.003 À0.003 (0.006) (0.005) (0.006) (0.006) Years of education father 0.008* 0.007 0.008* 0.007 (0.005) (0.005) (0.005) (0.005) Biological children residing in household at baseline Male children 0–5 À0.028 À0.029 À0.028 À0.028 (0.031) (0.030) (0.030) (0.030) Female children 0–5 À0.027 À0.024 À0.027 À0.025 (0.030) (0.029) (0.030) (0.029) Male children 6–10 0.009 0.014 0.009 0.014 (0.035) (0.034) (0.035) (0.034) Female children 6–10 À0.045 À0.056 À0.046 À0.055 (0.038) (0.037) (0.037) (0.037) Male children 11–15 0.012 0.017 0.012 0.016 (0.036) (0.035) (0.036) (0.035) Female children 11–15 À0.000 À0.006 À0.000 À0.007 (0.035) (0.034) (0.035) (0.034) Male children 16–20 0.010 0.001 0.010 0.001 (0.041) (0.040) (0.041) (0.040) Female children 16–20 À0.085* À0.093** À0.085* À0.094** (0.044) (0.043) (0.044) (0.043) Male children 21þ 0.033 0.026 0.033 0.028 (0.045) (0.044) (0.045) (0.044) Female children 21þ À0.073 À0.094* À0.072 À0.094* (0.055) (0.054) (0.055) (0.054) Number of children residing outside household À0.000 0.002 À0.000 0.001 (0.011) (0.011) (0.011) (0.011) Kilometers from regional capital  number outside children 0.000 0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.000) Age at baseline (1991–1994) 5–15 years 0.143*** 0.139*** 0.140*** 0.149*** (0.030) (0.029) (0.043) (0.040) 16–25 years 0.059 0.059 0.056 0.069 (0.039) (0.038) (0.049) (0.045) 26–35 years 0.108* 0.105* 0.107* 0.108* (0.065) (0.063) (0.065) (0.063) 36–45 years 0.132* 0.130* 0.130 0.135* (0.080) (0.078) (0.081) (0.079) 46–55 years 0.149 0.163* 0.148 0.164* (0.091) (0.088) (0.090) (0.088) 56–65 years 0.118 0.123 0.118 0.124 (0.098) (0.096) (0.098) (0.095) 66þ years 0.180 0.168 0.179 0.172 (0.121) (0.118) (0.120) (0.118) Constant À0.023 À0.013 (0.064) (0.063) Cragg-Donald 11.86 9.33 Sargan statistic 6.26 7.28 Sargan p-value 0.28 0.20 Number of observations 3,227 3,227 3,227 3,227 Standard errors are in parentheses. Significance at ***1%, **5%, *10%. TABLE 10.—FIRST-STAGE REGRESSIONS OF TABLE 9 (1) (2) Moved Distance moved Baseline covariates: Excluded instruments Head or spouse À0.218*** À0.634*** (0.038) (0.147) Child of head À0.097*** À0.423*** (0.032) (0.123) Male child of head À0.114*** À0.334** (0.037) (0.144) Age rank in household  age 5–15 14.390* 65.346* (8.003) (30.884) Kilometers from regional capital  male  age 5–15 À0.001*** À0.002** (0.000) (0.001) Rainfall shock  age 5–15 0.002** 0.007** (0.001) (0.003) Individual characteristics at baseline Deviation of years schooling from peers 0.012** 0.071*** (0.005) (0.018) Squared deviation of years schooling from peers 0.003** 0.014*** (0.001) (0.004) Male À0.017 À0.010 (0.030) (0.116) Unmarried 0.137*** 0.464** (0.048) (0.187) Unmarried male À0.105** À0.244 (0.042) (0.164) Both parents died À0.029 À0.261 (0.066) (0.253) Above 15 and both parents died 0.113 0.562* (0.079) (0.304) Years of education mother 0.012*** 0.040** (0.004) (0.017) Years of education father À0.002 À0.000 (0.004) (0.015) Biological children residing in houshold at baseline Male children 0–5 À0.001 0.008 (0.024) (0.093) Female children 0–5 À0.001 À0.010 (0.024) (0.092) Male children 6–10 À0.001 À0.059 (0.028) (0.107) Female children 6–10 À0.006 0.038 (0.030) (0.116) Male children 11–15 À0.011 À0.083 (0.028) (0.110) Female children 11–15 À0.035 À0.077 (0.027) (0.105) Male children 16–20 À0.022 À0.006 (0.032) (0.125) Female children 16–20 À0.031 À0.036 (0.035) (0.134) Male children 21þ 0.020 0.127 (0.036) (0.137) Female children 21þ À0.016 0.127 (0.044) (0.169) Number of children residing outside household À0.008 À0.043 (0.009) (0.033) Kilometers from regional capital  number outside children 0.000** 0.001** (0.000) (0.000) Age at baseline (1991–1994) 5–15 years 0.284*** 0.886*** (0.054) (0.210) 16–25 years 0.206*** 0.603*** (0.031) (0.118) 26–35 years 0.079 0.246 (0.051) (0.198) 36–45 years 0.135** 0.403* (0.063) (0.243) 46–55 years 0.079 0.095 (0.071) (0.276) 56–65 years 0.046 0.068 (0.078) (0.300) 66þ years 0.056 0.246 (0.095) (0.366) Number of observations 3,227 3,227 Standard errors are in parentheses. Significance at ***1%, **5%, *10%. Linear probability model (column 1) and OLS (column 2) with household fixed effects. 1022 THE REVIEW OF ECONOMICS AND STATISTICS Still, the difference is remarkably small. The conclusion is directly by the instruments, does yield jointly significant strong: being able to move out of the village or community instruments (at 10%). appears to be an important factor for consumption growth. If those who moved had stayed behind, our evidence sug- VII. Robustness gests that they would not have done as well. The fact that there is little change going from the IHHFE We perform a variety of checks to verify the robustness to the 2SLS results does not suggest that there is no selec- of the findings. First, we use alternative definitions of the tion process in the migration decision. For example, it could consumption aggregate, in particular excluding transfers be expected that more able people migrate. There is some out, which could be an important driver of our results if evidence that this occurs, yet this heterogeneity is not at the remittances to one’s origin village are large. We have data individual level but at the household level. Estimating the on transfers sent between the 2004 households of the same 2SLS without IHHFE increases the coefficient on migrant origin. The size of these remittances is on average only a status by almost a third (from 0.37 to 0.57). This is consis- small percentage of total consumption. Our findings are tent with the proposition of positive selection among house- robust to excluding this component. holds: individuals from households with high earning Second, we check the role of the configuration of the potential migrate. Within the household, there seems to be data. Our outcomes are household-level measures of con- no unobserved heterogeneity in terms of earning potential sumption per capita in levels and growth, assigned to indi- among those who do or do not migrate. viduals. We re-structure the data to the 2004 household These results are not driven by the lack of a parsimonious level in tables A3 and A4 (using average characteristics as set of instruments or relatively weak instruments. The controls and appropriately defined household-level aggre- results are similar when restricting the instrument set. When gated instruments). The results are similar and consistent focusing only on the relational variables (head, spouse, son, regardless, of analyzing the data at the individual or house- daughter), the Cragg-Donald (F) statistics become 14.5 and hold level. 11.3, the Sargan is not rejected, the coefficient on physical Third, concerns may be raised that changes in household movement stays at 0.36, and the distance variable becomes size and composition in new households in 2004 are driving 0.097, virtually identical to the results in table 9. the results. Table A5 shows that migrant households are The validity of our interacted instruments assumes that smaller in terms of members or adult-equivalent members. they do not capture different growth rates (for example, Table A6 repeats the analysis using adult-equivalent units because of different labor markets) across these groups rather than household size as the denominator and finds within households. Although growth rates might be influ- essentially similar results enced by the distance to the regional capital, rainfall Fourth, we investigate whether lack of common support shocks, gender, age, and other characteristics, additively, drives the results. The coefficients in the IHHFE regressions there is no evidence to suggest that the interaction of these are identified from the sample households that had split up would capture different growth rates outside the migration from the baseline. Restricting the sample to the 2,940 indi- effect. To explore this point, we exploit the fact that the viduals from at least two split-offs in 2004 yields identical 1991–1994 baseline data consist of four waves. The wave 1 results in both the IHHFE and 2SLS estimations. We further data were used as the baseline for this paper because the refine this by examining the sample of individuals from ori- consumption recall period was identical to the follow-up gin households that split off into at least one household that survey (wave 5). We use the three interim waves (2–4), moved by 2004 (N ¼ 2,520) and the sample of individuals which have similar recall periods, to check the validity of from origin households that had at least one split-off that our interacted instruments. Using a measure of annual con- remained in the village (N ¼ 2,777). These samples yield sumption per capita growth for 1992–1993, we can check identical results for both IHHFE and 2SLS. Restricting the whether our instruments, appropriately defined for this per- sample further to baseline households that had at least iod, jointly or individually explain the baseline consump- one split-off that moved and one that remained in the vil- tion changes. We find that they do not, giving further confi- lage (N ¼ 2,357) yields identical IHHFE results, but has dence that the exclusion restriction is valid for our 2SLS estimates of 0.23 and 0.68 for the migration indicator instruments: the instruments do not influence growth except variable and distance variables, significant at 10% and 7%, through migration. Of course, this regression of baseline respectively, and with IV diagnostics that remain sound. growth rates on our instruments can be valid only if migra- Taken together, these sample restrictions do not cast doubt tion can be plausibly omitted from it. We do find that 1992 on the validity of the results, although they suggest that the was the year with the lowest and 1993 the third-lowest size of the effects may be slightly lower than indicated in migration rates of all the years between 1992 and 2004, table 9. suggesting that the omission of the migration variable from As an alternative to the fixed-effects model and the two- the regression should not to lead to specification errors. As stage estimation, we investigate a number of matching can be expected, the same exercise for the regressions in models. Of course, the advantage of matching techniques is table 9, with the endogenous moved variable replaced that they ensure comparison of like-with-like, with less MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1023 TABLE 11.—EXPLAINING CONSUMPTION CHANGE: IHHFE, WITH CHANGE IN SCHOOLING AND INTERACTIONS (1) (2) (3) (4) IHHFE IHHFE IHHFE IHHFE Coefficient/se Coefficient/se Coefficient/se Coefficient/se Moved outside community 0.364*** 0.262*** (0.025) (0.033) Kilometers moved 0.120*** 0.099*** (0.006) (0.009) Gains in years of education 0.018*** 0.018*** 0.005 0.010* (0.005) (0.005) (0.006) (0.005) Gains in education  moved dummy 0.033*** (0.007) Gains in education  kilometers moved 0.006*** (0.002) Deviation of years schooling from peers 0.017*** 0.013** 0.018*** 0.014** (0.006) (0.006) (0.006) (0.006) Squared deviation of years schooling from peers 0.005*** 0.004*** 0.005*** 0.004*** (0.001) (0.001) (0.001) (0.001) Number of observations 3,226 3,226 3,226 3,226 ***p < 0.01, **p < 0.05, *p < 0.1. Number of observations is 1 less than table 9 due to 1 excluded observation missing 2004 education. Other controls reported in table 9 are included but not reported. restrictive functional form assumptions and omission of and 2 indicate that even controlling for educational gains, noncomparable observations. The disadvantage is that they the premium remains high and virtually identical to the ear- ignore potential unobservables (such as ability) that drive lier results. Including the interaction terms in columns 3 and selection into migration. Across an array of different match- 4 shows nevertheless that some of the gains may well work ing techniques (Gaussian, nearest neighbor, Epanechnikov), through education: the returns to moves are higher for those the main findings on the impact of migration on consump- who added more years of schooling, although for those with- tion gains are remarkably robust to the results in table 10 out additional schooling since the baseline, the returns to (results not presented). movement are still considerable. Finally, we examine the role of time-varying factors, spe- cifically education. If migration itself is the result of the indi- VIII. Migration Incentives, Social Constraints, and vidual’s efforts to increase his or her level of education, we Windows of Opportunity might be capturing the gains to a migration-education bun- dle rather than to migration per se. For secondary schooling, The regressions in tables 9 and 10 are suggestive of how often in the form of boarding schools in Tanzania, this is a the relatively traditional and tightly knit society of Kagera plausible concern. At the primary level, few people migrate reacted to growing economic opportunities in the past dec- for schooling opportunities, and tertiary education is limited ade. There is substantial movement out of these commu- for this sample. Descriptively, we find that the likelihood of nities, and those moving capture a substantial premium a move is not correlated with additional grades of schooling when measured in consumption terms. At the same time, conditional on age, suggesting that the moves we observe the high premium may indicate opportunities that have not are not specifically driven by demand for education. Explor- been taken. ing this further, we repeated the regressions in table 9 but In this section, we build a narrative around these results this time included the years of education gained between in four steps. First, we argue that there are windows of rounds, as well as the interaction of schooling gains with the opportunity that arise over time and space in the region, and migration variables. We have to be cautious in interpreting people need to move in order to take advantage of these these results, as surely migrating between rounds and years opportunities. Second, we complete the discussion of the of education gained between rounds are bound to be joint drivers of migration; the regressions in table 10 used house- decisions. Nevertheless, the results can at least explore hold fixed effects and therefore do not reveal correlates of whether the observed premium is just driven by education the constraints to migration at the household and commu- gains. Table 11 shows the results (only reporting the migra- nity levels. Third, we discuss how social norms can prevent tion and education variables, but the independent variables some people from moving. Finally, we argue why such wel- are otherwise identical to those in table 9).7 Our findings are fare-reducing constraints may be imposed by society on its robust to including education gains. The results in columns 1 members, thus providing a potential answer to the question of why more people do not move if the payoffs are so high. 7 At the end of this section, we discuss what our results imply Only the IHHFE and not the 2SLS results are reported here. When estimating the first column using 2SLS with the same instruments as in terms of standard models of migration and qualify this before, but also including the education gained in the first and second discussion by offering a few alternatives that cannot be stages, gave close to identical results as the IHHFE regressions (not rejected given the data available. shown). In the first-stage regression, the variable additional years of edu- cation gained is not correlated with the likelihood of moving, conditional The economic landscape in the Kagera region, as in other on age. regions in Africa, has been changing in the past two decades. 1024 THE REVIEW OF ECONOMICS AND STATISTICS TABLE 12.—(CONTINUED) TABLE 12.—EXPLAINING CONSUMPTION CHANGE: IHHFE, CHARACTERISTICS OF THE MOVE (1) (2) IHHFE IHHFE (1) (2) IHHFE IHHFE 56–65 years 0.135 0.127 Characteristics of the move (0.096) (0.095) Move to more remote area 0.176*** 66þ years 0.190 0.169 (0.036) (0.118) (0.117) Move to similar area 0.274*** Constant À0.015 À0.007 (0.034) (0.063) (0.062) Move to more connected area 0.661*** Number of observations 3,227 3,227 (0.037) Standard errors are in parentheses. Significance at ***1%, **5%, *10%. Kilometers moved 0.073*** (0.011) Distance moved if to similar area 0.032** Growth opportunities are continually being introduced and (0.015) Distance moved if to more connected area 0.070*** eliminated across time and space, as the refugee crisis (0.013) abates, links with war-ridden bordering countries change, Individual characteristics at baseline and more localized negative and positive shocks manifest Deviation of years schooling from peers 0.010* 0.008 (0.006) (0.006) themselves with various degrees of severity. People need to Squared deviation of years 0.004*** 0.004*** be physically (geographically) mobile in order to respond schooling from peers (0.001) (0.001) to the opportunities. To elaborate on this point, we decom- Male À0.004 À0.008 (0.037) (0.037) pose the results of table 9 further to examine the role of the Unmarried À0.008 À0.006 location of a move (more or less remote areas) and moves (0.054) (0.054) associated with sector changes (out of agriculture into non- Unmarried male 0.127*** 0.121*** (0.044) (0.043) agricultural activities). Table 12 disaggregates the migra- Both parents died 0.005 0.025 tion variable into three categories of migration (moving to a (0.082) (0.081) more or less connected or urbanized area). Even moving to Above 15 and both parents died 0.053 0.020 (0.098) (0.097) a less connected area is still correlated with higher growth Years of education mother À0.004 À0.004 compared with not moving, but moving to a more con- (0.005) (0.005) nected area results in consumption growth that is 66 percen- Years of education father 0.006 0.006 (0.005) (0.005) tage points higher than not moving. The same result is Biological children residing in household at baseline found using the distance of the move in the second column. Male children 0–5 À0.021 À0.023 Although where individuals move matters for the magni- (0.030) (0.030) Female children 0–5 À0.026 À0.025 tude of the effect, any movement has the potential to be (0.029) (0.029) welfare improving. Male children 6–10 0.008 0.015 In table 13, we interact migration with change in sector (0.034) (0.034) Female children 6–10 À0.048 À0.056 (out of agriculture). We pool people who moved out of agri- (0.037) (0.036) culture and those who remained in nonagriculture. Both Male children 11–15 0.023 0.022 groups had statistically indistinguishable findings in all (0.035) (0.035) Female children 11–15 À0.010 À0.011 regressions. The first column shows that moving out of agri- (0.034) (0.033) culture is strongly linked to higher consumption growth (as Male children 16–20 0.012 0.002 noted above in the descriptive statistics). The next two col- (0.040) (0.040) Female children 16–20 À0.085* À0.095** umns show a large and positive impact of moving, even (0.043) (0.043) after controlling sector shifts; there is also a strong interac- Male children 21þ 0.023 0.020 tive effect of this sector shift with physical movement out (0.044) (0.044) Female children 21þ À0.090* À0.099* of the village. In other words, it is not just the move out of (0.054) (0.054) agriculture that accounts for the large growth differential; Number of children residing À0.001 0.003 outside household (0.011) (0.011) migration as physical movement out of the village has Kilometers from regional capital  0.000 0.000 strong additional and complementary effects. number outside children (0.000) (0.000) Tables 12 and 13 thus show that movement in itself is Age at baseline (1991–1994) 5–15 years 0.141*** 0.143*** important. A logical—for economists, perhaps even tautolo- (0.029) (0.028) gical—consequence of this is that constraints to movement 16–25 years 0.063* 0.066* are impediments to growth for whomever they happen to (0.038) (0.038) 26–35 years 0.107* 0.102 constrain. To estimate the migration premium, the regres- (0.063) (0.063) sions in tables 9 and 10 use initial household fixed effects. 36–45 years 0.130* 0.131* Any initial household and community characteristics are (0.078) (0.077) 46–55 years 0.164* 0.166* therefore a black box. Although this improves inference (0.088) (0.088) regarding the migration premium, it also offers an incom- plete narrative of why certain people migrate and therefore TABLE 13.—EXPLAINING CONSUMPTION CHANGE: IHHFE, MOVING OUT OF AGRICULTURE (1) (2) (3) IHHFE IHHFE IHHFE Characteristics of the move Moved outside community 0.195*** (0.034) Kilometers moved (log of distance) 0.073*** (0.011) Moved out of agriculture 0.407*** 0.126*** 0.175*** (0.034) (0.044) (0.040) Moved outside community and out of agriculture 0.449*** (0.059) Distance moved  moved out of agriculture 0.075*** (0.015) Individual characteristics at baseline Deviation of years schooling from peers 0.013* 0.011* 0.010 (0.007) (0.006) (0.006) Squared deviation of years schooling from peers 0.004** 0.003** 0.003** (0.002) (0.001) (0.001) Male À0.059 À0.020 À0.030 (0.042) (0.040) (0.040) Unmarried À0.005 À0.059 À0.049 (0.063) (0.061) (0.060) Unmarried male 0.079 0.128** 0.132*** (0.051) (0.050) (0.049) Both parents died À0.066 À0.045 À0.022 (0.113) (0.108) (0.108) Above 15 and both parents died 0.110 0.076 0.045 (0.126) (0.121) (0.120) Years of education mother 0.005 À0.002 À0.005 (0.007) (0.007) (0.007) Years of education father À0.003 0.001 0.000 (0.006) (0.006) (0.006) Biological children residing in household at baseline Male children 0–5 À0.048 À0.047 À0.044 (0.035) (0.034) (0.033) Female children 0–5 À0.029 À0.018 À0.020 (0.034) (0.033) (0.033) Male children 6–10 0.023 0.014 0.024 (0.039) (0.038) (0.037) Female children 6–10 À0.057 À0.056 À0.067 (0.043) (0.041) (0.041) Male children 11–15 0.018 0.010 0.023 (0.041) (0.039) (0.039) Female children 11–15 0.004 0.010 0.008 (0.039) (0.037) (0.037) Male children 16–20 À0.024 À0.009 À0.011 (0.046) (0.044) (0.044) Female children 16–20 À0.100** À0.103** À0.111** (0.049) (0.047) (0.047) Male children 21þ 0.027 0.013 0.005 (0.053) (0.051) (0.050) Female children 21þ À0.141** À0.103 À0.119* (0.067) (0.064) (0.064) Number of children residing outside household 0.006 0.004 0.004 (0.013) (0.012) (0.012) Kilometers from regional capital  number outside children 0.000 0.000 0.000 (0.000) (0.000) (0.000) Age at baseline (1991–1994) 5–15 years 0.177*** 0.132*** 0.135*** (0.049) (0.047) (0.047) 16–25 years 0.058 0.029 0.029 (0.057) (0.055) (0.054) 26–35 years 0.063 0.082 0.078 (0.083) (0.079) (0.079) 36–45 years 0.077 0.085 0.077 (0.098) (0.094) (0.093) 46–55 years 0.103 0.128 0.133 (0.110) (0.106) (0.105) 56–65 years 0.091 0.105 0.113 (0.119) (0.114) (0.114) 66þ years 0.195 0.246* 0.233 (0.156) (0.149) (0.148) Constant 0.084 0.054 0.057 (0.082) (0.079) (0.078) Number of observations 2,546 2,546 2,546 Standard errors are in parentheses. Significance at ***1%, **5%, *10%. 1026 THE REVIEW OF ECONOMICS AND STATISTICS why the migration premium remains so high. To explore TABLE 14.—EXPLAINING MIGRATION, HOUSEHOLD, AND COMMUNITY CORRELATES this, we replaced the initial household fixed effect in table (1) (2) 10 by a set of household and community variables as mea- Probit: OLS: Kilometers Moved out of Moved (Log sured at baseline. We include characteristics of the head of Community of Distance) the household (age, sex, and years of education), household Sex of household head 0.012 À0.105 size, a dummy for whether the household is a farm house- (0.026) (0.093) hold, and wealth characteristics (land for cultivation, con- Age of household head 0.001 0.009*** sumption per capita, the value of the physical capital stock, (0.001) (0.002) Education of household head À0.005 0.000 and the flooring quality in the dwelling). The community (0.004) (0.015) variables included are the cluster (village) means of all the Household size À0.000 À0.014 above characteristics, the distance to Bukoba (the regional (0.003) (0.012) Primary occupation is farming À0.064** À0.355*** capital), and whether the community is remote (defined as (0.031) (0.099) an area not closely connected to an urban center). Acres of land cultivated À0.003 À0.014* The results are shown in table 14 (only the additional (0.002) (0.008) Consumption per capita À0.021 À0.674 variables are reported as inference on the role of individual (in millions of TSh) (0.124) (0.419) characteristics was superior in table 10). The table offers Value of physical assets À0.008* À0.024* the marginal effects from a probit regression on whether an (in millions of TSh) (0.005) (0.014) Good flooring in dwelling À0.003 0.097 individual moved and the coefficients from an OLS regres- (0.031) (0.105) sion explaining the logarithm of the distance migrated. Kilometers from cluster to Bukoba À0.000 À0.000 We find suggestive evidence of some factors that matter (0.000) (0.001) Remote community À0.047** À0.078 for the migration decision. First, we find some weak evi- (0.020) (0.068) dence that migration is higher in communities where farm- Cluster mean of household characteristics ing was still the most important activity for a higher number Sex of household head À0.207* 0.068 (0.120) (0.411) of families, at least with respect to the distance of moves Age of household head 0.002 0.017** (column 2). The median community in the sample has 83% (0.002) (0.008) of households mainly involved in agriculture. However, Education of household head 0.005 0.017 (0.013) (0.043) controlling for this, being a farmer decreases the probability Household size 0.007 0.049* of any move and the distance migrated. Those with more (0.008) (0.028) land do not migrate as far as those with less land. Taken Primary occupation is farming 0.110 0.500* (0.087) (0.289) together, this suggests that migration is to some extent dri- Acres of land cultivated À0.001 À0.012 ven by a move out of agricultural settings, but those not (0.007) (0.023) Consumption per capita 0.256 1.545 involved in agriculture and faced with land pressure are (in millions of TSh) (0.362) (1.224) more likely to move away and move farther. Value of physical assets 0.007 0.013 Second, the regression results reported in table 10 show (in millions of TSh) (0.015) (0.053) Good flooring in dwelling À0.028 À0.152 that education matters in explaining the migration of the (0.084) (0.286) individual, but table 14 shows that the educational level of Constant À0.825 the family or community does not matter: there is no more (0.755) Number of observations 3,119 3,119 or further migration from households or settings with more Standard errors are in parentheses. Significance at ***1%, **5%, *10%. Column 1 presents the mar- education. Similarly, although living in better-connected ginal effects from a probit estimation. Individual characteristics included in table 10 specifications are also included here, but are not presented in the table. The sample is slightly reduced (from 3,227) due to areas is positively associated with the probability of mov- missing information on baseline value of physical stock for 108 households. ing, it is not associated with distance, and being closer to the regional capital is not associated with either migration measure. Overall, there is little evidence of credit or wealth con- Finally, are there any wealth or credit constraint effects? straints; if anything, there is a tendency for more migration Migration may be a costly activity, requiring a serious from poorer households. investment. It is also an indivisible investment. Thus, some This still leaves open the question as to why more people households may not be able to afford any migration to take do not migrate, given the high returns. The individual-level advantage of the high return or can afford to have only variables in the regressions in table 10 suggest that particu- some members migrate. From figures 3, 4, and 5, the sam- lar types of people within families can go and not others. ples of future movers and nonmovers started off with rela- Factors include individual education (with a convex effect), tively similar wealth distributions. The evidence from table being unmarried and female (consistent with considerable 14 is mixed. Among the household indicators, only the migration for marriage by girls), being of a particular age value of physical assets is associated negatively with mov- group when rainfall shocks occur, and a series of positional ing. Neither consumption per capita nor good flooring is variables in the household (including an age pecking order associated with migration. None of the community-mean and gender effects). Distance to the regional capital, a migra- wealth variables are statistically associated with migration. tion pull factor, matters specifically for young males. MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1027 There is scope for further interpretation of these findings that equilibrium has not yet been attained, with continuing within the local social context. The regressions suggest that streams of migrants, so our results are inconsistent with the an individual needs to be in a position to move in order to migration equilibrium, provided that expectations reflect take advantage of geographic and time-specific economic true conditions.8 We cannot discount that we may be obser- opportunities, while at the same time a number of crucial ving the migration process in a state of disequilibrium as social constraints in place may prevent an individual from part of a dynamic adjustment process to a long-run equili- doing so. Social and family norms interacting with pull brium with equal returns in expectation. Nevertheless, the (nearby towns) and push (shocks) factors are determinants scale of the disequilibrium is not easily explained given the of who may be allowed (or chosen) to move. available data and the results in tables 10 and 13. There appear to be windows of opportunity—being in the Other interpretations can nevertheless be offered as to right place at the right time—that certain categories of peo- why migration may be limited despite high returns in con- ple can take advantage of: not having social and family con- sumption terms. Consumption may be a poor measure of straints in a window of time when physical mobility has the overall net welfare benefit of migration. People may large payoffs. Missing these windows implies remaining find the alienation from their original home environment trapped in a low-return environment. But this still begs the costly in subjective terms. For example, a recent resurvey question of why we do not see more migration given these of the ICRISAT households in six villages in Maharashtra high returns and why barriers remain in place if they are and Andhra Pradesh in India used a similar tracking metho- welfare reducing. dology to the current paper (Dercon, Krishnan, & Krutikov, Our results are consistent with the literature that links 2009). Tracking all the people living in the original 240 network externalities to poverty traps and so endogenizes ICRISAT households from 1975 to 1984, the study found exit barriers in the village. In Hoff and Sen (2006), the kin- that by 2004, those who had migrated had a premium of ship group decides how high to set the exit barrier for its about 20% in consumption, controlling for initial household members. They start from the observation that kin who fixed effects. However, again controlling for initial house- have moved and remain loyal to their kinship group at hold fixed effects, there was a negative premium on being a home will sometimes need to undertake actions with nega- migrant in regressions with subjective well-being or subjec- tive consequences for their employers (securing jobs for tive assessment of overall wealth as the dependent variable. kin) or landlords (sharing housing). This creates an entry In short, migrants had higher consumption in real terms but barrier for anyone with obvious, strong kinship ties to their lower subjective well-being compared with those from the home village. In order to overcome such entry barriers, an same original households who did not migrate, possibly as individual may have to sever ties with his or her kinship if a premium in terms of the former is required to compen- group, implying the loss of a productive element (from the sate for the latter. In such circumstances, there is no reason kinship group’s point of view). To avoid this ex ante, the that the consumption of migrants would ever equate to the kinship group may decide to manipulate exit barriers, rais- consumption of nonmigrants; a gap would remain. ing them through social norms about migration in order not to lose productive members. Hoff and Sen’s model finds IX. Conclusion that it may be in the interest of the kinship group to prevent some of its members from taking advantage of economic This paper explores the impact of migration on poverty opportunities. and living standards in Tanzania. We use a unique thirteen- Our results offer an empirical qualification of this basic year panel data set, offering information on split-off house- result and suggest that exit barriers are not equal over time holds and migrants. Assessing the impact of migration on because they depend on interactions among gender, age, living standards is particularly difficult because we cannot age rank, and the degree of connectedness to the household observe someone to be a migrant and remain in the original head. Furthermore, our results suggest that exit barriers are community at the same time. A relatively simple differ- binding constraints only when geographic and time-specific 8 push or pull factors offer a window for economic advance- In principle, deviation of the estimated average difference between urban and rural living standards expost could also be possible in equili- ment through migration. brium in a Harris-Todaro-style model. For this to explain the apparent Are our results consistent with standard models of migra- less-than-optimal migration levels, it would need to be the case that the tion? As in Harris and Todaro (1970), higher benefits gap in living standards between urban and rural areas was expected to be much lower than is now apparent. This could come about if expectations appear to drive migration. In their model, the assumption of were not rational; if there was an unexpected higher level of urban wages unemployment in the urban sector allows the persistence of by 2004; or, for example, if there was much higher employment in well- a wage premium between urban and rural areas, linked to paid jobs. The premium seems to be too high for this to be a sufficient explanation, most likely because although urban wages may be high, urban unemployment in a context of imperfect labor mar- unemployment levels for particular groups are also high. More than 20% kets. In equilibrium, expected urban earnings (across work- of the urban population between ages 18 and 34 is unemployed, in the ers and the unemployed) are equal to rural earnings (Harris sense that they are looking for work and available for work. Unemploy- ment is double this figure using the national definition, which includes & Todaro, 1970). As a result, observing living standards in those with marginal or precarious job situations (United Republic of Tan- urban areas to be above those in rural areas would suggest zania, 2007). 1028 THE REVIEW OF ECONOMICS AND STATISTICS ence-in-difference model is used to assess the impact of There appear to be barriers to physical movement, so that migration on consumption levels, thereby controlling for potential returns are unexploited. Our evidence suggests fixed individual heterogeneity in determining the level of that just as in Hoff and Sen (2006), some of these barriers consumption. Furthermore, we can identify the impact of may be exit barriers that result in less-than-efficient levels migration on the growth rate of consumption using within- of migration. Alternatively, other welfare costs related to household variation in the subsequent migration of indivi- migration may also limit migration. dual members. This initial household fixed-effects estimator controls for the unobserved heterogeneity in the growth rate of consumption that is common among baseline household REFERENCES members. Finally, a number of specific individual factors Alderman, Harold, Jere R. Berman, Hans-Peter Kohler, John Maluccio, are added as controls, and IV estimates are also presented. and Susan Cotts Watkins, ‘‘Attrition in Longitudinal Household We avoid identifying the migration decision based on the Survey Data: Some Tests for Three Developing-Country Sam- household circumstances (such as shocks, distance to poten- ples,’’ Demographic Research 5:4 (2001), 79–124. Beegle, Kathleen, ‘‘Economic Mobility in Indonesia and Vietnam: What tial areas of destination, and existence of family networks) Missing Data Can’t Tell Us,’’ mimeograph (2000). used in most studies of migration impacts, but that are Beegle, Kathleen, Joachim De Weerdt, and Stefan Dercon, ‘‘Kagera unlikely to satisfy the exclusion restriction. Health and Development Survey (KHDS), 2004, Basic Informa- tion Document,’’ mimeograph (2006). The identified effects are remarkably large and robust: Clemens, Michael, and Lant Pritchett, ‘‘Income per Natural: Measuring migrants experienced 36 percentage points higher consump- Development for People Rather Than Places,’’ Population and tion growth compared with those who stayed behind. We Development Review 34 (2008), 395–434. De Weerdt, Joachim, ‘‘Moving out of Poverty in Tanzania: Evidence from also find that transfers from migrants to nonmigrating Kagera,’’ Journal of Development Studies 46 (2010), 331–349. household members are relatively limited. We find sugges- Demombynes, Gabriel, and Johannes Hoogeveen, ‘‘Growth, Inequality tive evidence that it matters where one moves. Moving to and Simulated Poverty Paths for Tanzania, 1992–2002,’’ Journal of African Economies 16 (2007), 596–628. more connected areas has substantially higher returns, but Dercon, Stefan, Pramila Krishnan, and Sonya Krutikov, ‘‘Migration, even moving to a less connected area is correlated with Risk-Sharing and Subjective Well-Being,’’ Oxford University, higher growth. Moving into nonagricultural activities con- mimeograph (2009). Fields, Gary S, ‘‘Dualism in the Labor Market: A Perspective on the tributes to consumption growth, but this growth is always Lewis Model after Half a Century,’’ Manchester School 72 (2004), greater in any sector or activity if the individual physically 724–735. moved. Harris, John, and Michael Todaro, ‘‘Migration, Unemployment and Development: A Two-Sector Analysis,’’ American Economic Relative to the theory, this paper offers evidence consis- Review 60 (1970), 126–142. tent with features in the original Lewis model, with persis- Hoff, Karla, and Arjit Sen, ‘‘The Kin System as a Poverty Trap?’’ (pp. tent wage differentials, at least at this stage of the rural 91–115), in Samuel Bowles, Steven Durlauf, and Karla Hoff (Eds.), Poverty Traps (Princeton, NJ: Princeton University Press, transformation.9 Relative to the new economics of migra- 2006). tion, the evidence is less conclusive. The robust difference Kilic, Talip, Gero Carletto, Benjamin Davis, and Alberto Zezza, in welfare levels between those who migrated and those ‘‘Investing Back Home: Return Migration and Business Ownership in Albania,’’ Economics of Transition 17 (2009), 587–623. who did not is not consistent with either complete altruism Lewis, W. Arthur, ‘‘Economic Development with Unlimited Supplies of or risk sharing, although it is possible that the results reflect Labor,’’ Manchester School 22 (1954), 139–191. McKenzie, David, and Marcin J. Sasin, ‘‘Migration, Remittances and partial risk sharing and some transfers. We offer suggestive Human Capital: Conceptual and Empirical Challenges,’’ World evidence that some transfers indeed occur, but they are rela- Bank Policy Research working paper no. 4272 (2007). tively limited from those who migrated long distances. McKenzie, David, John Gibson, and Steven Stillman, ‘‘How Important Is Selection? Experimental v. Non-Experimental Measures of the Because we have not constructed a structural model of Income Gains from Migration,’’ Journal of the European Eco- migration (only a first stage in a 2SLS procedure), our evi- nomic Association 8 (2010), 913–945. dence does not shed full light on the migration process. Munshi, Kaivan, ‘‘Networks in the Modern Economy: Mexican Migrants in the United States Labor Market,’’ Quarterly Journal of Econom- However, we provide suggestive evidence that within- ics 118 (2003), 549–597. family social structures matter for who gets the opportunity Rosenzweig, Mark, ‘‘Payoffs from Panels in Low-Income Countries: Eco- to migrate, how far they go, and who, therefore, can move nomic Development and Economic Mobility,’’ American Eco- nomic Review 93 (2003), 112–117. up economically in Tanzania. This evidence also helps us Rosenzweig, Mark, and Oded Stark, ‘‘Consumption Smoothing, Migra- to understand better how our results relate to the predictions tion, and Marriage: Evidence from Rural India,’’ Journal of Politi- of the Harris-Todaro model. Unlike the equilibrium condi- cal Economy 97 (1989), 905–926. Stark, Oded, and David Bloom, ‘‘The New Economics of Labor Migra- tions in Harris-Todaro, on average the premium on migra- tion,’’ American Economic Review 75 (1985), 173–178. tion is positive and rather large: in expectation, earnings do Stock, James, and Motohiro Yogo, ‘‘Testing for Weak Instruments in Linear not appear to be equalized, suggesting that equilibrium has IV Regression,’’ NBER technical working paper no. 0284 (2002). United Republic of Tanzania, ‘‘The Economic Report, 2004,’’ President’s not been obtained, and returns to migration remain high. Office, Planning and Privitisations (2004). ——— ‘‘Analytical Report for the Integrated Labor Force Survey’’ (National Bureau of Statistics, Ministry of Planning, Economy and 9 This interpretation of Lewis (1954) is still debated, and not necessarily Empowerment, 2007). a feature of subsequent dual economy models. For a discussion, see Fields World Bank, ‘‘User’s Guide to the Kagera Health and Development Sur- (2004). vey Datasets,’’ mimeograph (2004). MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1029 TABLE APPENDIX TABLE A1.—SAMPLE MEANS AND STANDARD DEVIATIONS Mean s.d. Change in (logged) consumption per capita 0.26 (0.77) Moved 0.33 (0.47) Distance moved (kms) 35.02 (145.01) Distance moved variable: log(kms þ 1) 1.06 (1.78) Baseline covariates: Excluded instruments Head or spouse 0.26 (0.44) Child of head 0.49 (0.50) Male child of head 0.25 (0.43) Age rank in household  age 5–15 0.00 (0.00) Kilometers from regional capital  12.04 (38.43) male  age 5—15 Rainfall shock (annual centimeters À18.51 (28.71) deviation)  age 5–15 Individual characteristics at baseline Deviation of years schooling from peers À0.25 (2.24) Squared deviation of years schooling 5.08 (9.50) from peers Male 0.47 (0.50) Unmarried 0.69 (0.46) Unmarried male 0.36 (0.48) Both parents died 0.05 (0.22) Above 15 and both parents died 0.02 (0.15) Years of education mother 2.72 (3.02) Years of education father 4.23 (3.32) Biological children residing in household at baseline Male children 0–5 0.15 (0.45) Female children 0–5 0.14 (0.45) Male children 6–10 0.10 (0.36) Female children 6–10 0.09 (0.34) Male children 11–15 0.10 (0.36) Female children 11–15 0.11 (0.38) Male children 16–20 0.06 (0.29) Female children 16–20 0.06 (0.28) Male children 21þ 0.05 (0.26) Female children 21þ 0.04 (0.21) Number of children residing outside 0.64 (1.83) household Kilometers from regional capital  number 44.62 (184.60) outside children Age at baseline (1991–1994) 5–15 years 0.35 (0.48) 16–25 years 0.20 (0.40) 26–35 years 0.08 (0.27) 36–45 years 0.07 (0.26) 46–55 years 0.06 (0.23) 56–65 years 0.04 (0.20) 66þ years 0.02 (0.12) Number of observations 3,227 TABLE A2.—ALTERNATIVE SETS OF INSTRUMENTAL VARIABLES Excluded Instrumental Variable: Age Rank in Kilometers from Avgerage Head or Child of Male Child Household  Regional Capital  Rainfall Shock  All Included Spouse Head of Head Age 5–15 Male  Age 5–15 Age 5–15 Moved outside community Coefficient 0.38 0.40 0.49 0.30 0.32 0.38 0.36 Standard error 0.15 0.20 0.16 0.16 0.15 0.16 0.16 Cragg-Donald 11.86 7.62 12.36 12.32 13.58 12.54 12.89 Kilometers moved (log of distance) Coefficient 0.10 0.10 0.16 0.08 0.08 0.10 0.10 Standard error 0.04 0.05 0.05 0.05 0.04 0.04 0.05 Cragg-Donald 9.33 7.45 8.81 10.11 10.29 10.23 9.87 Each coefficient is generated from a separate regression based on the 2SLS specification in table 10. 1030 THE REVIEW OF ECONOMICS AND STATISTICS TABLE A3.—EXPLAINING CONSUMPTION CHANGE: IHHFE AND 2SLS, HOUSEHOLD-LEVEL RESULTS (1) (2) (3) (4) IHHFE IHHFE 2SLS 2SLS Moved outside community 0.321*** 0.520*** (0.038) (0.154) Kilometers moved (log of distance) 0.112*** 0.146*** (0.009) (0.045) Individual characteristics at baseline Deviation of years schooling from peers 0.013 0.007 0.011 0.005 (0.010) (0.010) (0.010) (0.010) Squared deviation of years schooling from peers 0.006*** 0.006*** 0.006*** 0.005** (0.002) (0.002) (0.002) (0.002) Male 0.085 0.067 0.115 0.076 (0.096) (0.093) (0.098) (0.093) Unmarried À0.114 À0.117 À0.144 À0.132 (0.101) (0.098) (0.104) (0.100) Unmarried male 0.179* 0.182* 0.208** 0.197** (0.102) (0.099) (0.104) (0.101) Both parents died À0.079 À0.042 À0.080 À0.032 (0.136) (0.133) (0.136) (0.133) Above 15 and both parents died 0.124 0.066 0.092 0.034 (0.175) (0.171) (0.177) (0.175) Years of education mother À0.004 À0.006 À0.008 À0.009 (0.010) (0.009) (0.010) (0.010) Years of education father 0.016* 0.015* 0.017* 0.015* (0.009) (0.009) (0.009) (0.009) Biological children residing in household at baseline Male children 0–5 À0.049 À0.042 À0.049 À0.040 (0.074) (0.072) (0.074) (0.072) Female children 0–5 0.014 0.019 0.040 0.034 (0.078) (0.076) (0.081) (0.078) Male children 6–10 À0.112 À0.097 À0.101 À0.087 (0.090) (0.087) (0.090) (0.088) Female children 6–10 À0.181* À0.209** À0.179* À0.216** (0.090) (0.088) (0.090) (0.088) Male children 11–15 0.046 0.051 0.042 0.051 (0.076) (0.074) (0.076) (0.073) Female children 11–15 À0.046 À0.045 À0.018 À0.031 (0.083) (0.081) (0.086) (0.082) Male children 16–20 0.040 0.024 0.054 0.026 (0.088) (0.086) (0.089) (0.085) Female children 16–20 À0.200** À0.214** À0.174* À0.206** (0.105) (0.102) (0.107) (0.102) Male children 21þ 0.046 0.041 0.035 0.035 (0.105) (0.102) (0.105) (0.102) Female children 21þ À0.204 À0.237* À0.166 À0.229* (0.126) (0.122) (0.129) (0.122) Number of children residing outside household À0.001 0.001 À0.004 0.000 (0.022) (0.021) (0.022) (0.021) Kilometers from regional capital  number outside children 0.000 0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.000) Age at baseline (1991–1994) 5–15 years 0.245*** 0.246*** 0.215*** 0.229*** (0.062) (0.060) (0.066) (0.063) 16–25 years 0.060 0.066 0.047 0.058 (0.071) (0.069) (0.072) (0.069) 26–35 years 0.152 0.160 0.196 0.183 (0.131) (0.127) (0.135) (0.130) 36–45 years 0.141 0.167 0.155 0.161 (0.168) (0.163) (0.168) (0.163) 46–55 years 0.206 0.264 0.278 0.312* (0.187) (0.182) (0.194) (0.193) 56–65 years 0.144 0.180 0.243 0.231 (0.199) (0.193) (0.212) (0.206) 66þ years 0.336 0.344 0.384* 0.357 (0.236) (0.229) (0.238) (0.230) Constant 0.025 0.019 (0.129) (0.125) Cragg-Donald 13.68 9.10 Sargan statistic 5.67 7.76 Sargan p-value 0.34 0.17 Number of observations 1,909 1,909 1,909 1,909 Standard errors are in parentheses. Significance at ***1%, **5%, *10%. TABLE A4.—FIRST-STAGE REGRESSIONS OF TABLE A3 (1) (2) Moved Distance Moved Baseline covariates: excluded instruments Head or spouse À0.179*** À0.479* (0.063) (0.254) Child of head À0.058 À0.343* (0.048) (0.193) Male child of head À0.157*** À0.455** (0.056) (0.227) Age rank in household  age 5–15 À7.379 À1.269 (12.036) (48.763) Kilometers from regional capital  male  age 5–15 À0.001* À0.001 (0.000) (0.001) Rainfall shock  age 5–15 0.001*** 0.002*** (0.000) (0.000) Individual characteristics at baseline Deviation of years schooling from peers 0.007 0.070** (0.007) (0.030) Squared deviation of years schooling from peers 0.003** 0.018*** (0.002) (0.006) Male À0.126* À0.233 (0.070) (0.282) Unmarried 0.095 0.364 (0.075) (0.306) Unmarried male À0.022 À0.073 (0.079) (0.322) Both parents died 0.019 À0.276 (0.098) (0.396) Above 15 and both parents died 0.160 0.974* (0.125) (0.507) Years of education mother 0.015** 0.060** (0.007) (0.028) Years of education father À0.010 À0.019 (0.007) (0.027) Biological children residing in household at baseline Male children 0–5 0.016 À0.011 (0.053) (0.216) Female children 0–5 À0.082 À0.246 (0.056) (0.229) Male children 6–10 À0.007 À0.162 (0.065) (0.263) Female children 6–10 0.005 0.221 (0.066) (0.267) Male children 11–15 0.037 0.045 (0.055) (0.224) Female children 11–15 À0.087 À0.274 (0.061) (0.246) Male children 16–20 À0.061 À0.040 (0.063) (0.257) Female children 16–20 À0.105 À0.169 (0.075) (0.304) Male children 21þ 0.063 0.239 (0.075) (0.305) Female children 21þ À0.156* À0.168 (0.091) (0.369) Number of children residing outside household 0.006 0.001 (0.016) (0.064) Kilometers from regional capital  number outside children 0.000** 0.001** (0.000) (0.000) Age at baseline (1991–1994) 5–15 years 0.454*** 1.351*** (0.072) (0.292) 16–25 years 0.032 0.046 (0.052) (0.209) 26–35 years À0.227** À0.745* (0.094) (0.383) 36–45 years À0.089 À0.389 (0.121) (0.491) 46–55 years À0.340** À1.521*** (0.135) (0.548) 56–65 years À0.433*** À1.549*** (0.144) (0.582) 66þ years À0.239 À0.727 (0.170) (0.688) Number of observations 1,909 1,909 Standard errors are in parentheses. Significance at ***1%, **5%, *10%. 1032 THE REVIEW OF ECONOMICS AND STATISTICS TABLE A5.—HOUSEHOLD SIZE AT BASELINE AND FOLLOW-UP, BY MOBILITY CATEGORIES, MEAN (MEDIAN) Household Size Household Size: Adult Equivalent 1991 2004 1991 2004 N Same village 7.71 5.98 6.15 4.94 2,150 (7.0) (6.0) (5.7) (4.6) Neighboring community 8.20 4.93 6.59 3.87 400 (7.0) (5.0) (5.9) (3.4) Elsewhere in Kagera 7.65 4.47 6.17 3.55 437 (7.0) (4.0) (6.0) (3.2) Outside Kagera 8.45 4.45 6.74 3.69 251 (7.0) (4.0) (6.1) (3.1) Adult equivalence is defined following the National Bureau of Statistics with varying weights by age and sex. TABLE A6.—EXPLAINING CONSUMPTION CHANGE: IHHFE AND 2SLS ADULT EQUIVALENT CONSUMPTION (RATHER THAN PER CAPITA) (1) (2) (3) (4) IHHFE IHHFE 2SLS with IHHFE 2SLS with IHHFE Moved outside community 0.363*** 0.426*** (0.024) (0.143) Kilometers moved (log of distance) 0.117*** 0.123*** (0.006) (0.041) Individual characteristics at baseline Deviation of years schooling from peers 0.014** 0.010* 0.013** 0.010 (0.006) (0.006) (0.006) (0.006) Squared deviation of years schooling from peers 0.004*** 0.003** 0.004*** 0.003** (0.001) (0.001) (0.001) (0.001) Male À0.010 À0.016 À0.008 À0.016 (0.036) (0.035) (0.036) (0.035) Unmarried 0.043 0.048 0.030 0.045 (0.053) (0.051) (0.060) (0.057) Unmarried male 0.087** 0.076* 0.099* 0.079* (0.043) (0.041) (0.051) (0.046) Both parents died 0.007 0.026 0.009 0.027 (0.079) (0.077) (0.079) (0.077) Above 15 and both parents died 0.032 0.008 0.026 0.005 (0.095) (0.093) (0.096) (0.095) Years of education mother À0.002 À0.003 À0.003 À0.003 (0.005) (0.005) (0.006) (0.006) Years of education father 0.008* 0.007 0.008* 0.007 (0.005) (0.005) (0.005) (0.005) Biological children residing in household at baseline Male children 0–5 À0.041 À0.041 À0.041 À0.041 (0.029) (0.028) (0.029) (0.028) Female children 0–5 À0.027 À0.025 À0.027 À0.024 (0.028) (0.028) (0.028) (0.028) Male children 6–10 0.001 0.005 0.000 0.006 (0.033) (0.032) (0.033) (0.032) Female children 6–10 À0.042 À0.052 À0.042 À0.053 (0.036) (0.035) (0.035) (0.035) Male children 11–15 0.001 0.005 0.001 0.005 (0.034) (0.033) (0.034) (0.033) Female children 11–15 À0.014 À0.019 À0.011 À0.018 (0.033) (0.032) (0.033) (0.032) Male children 16–20 0.005 À0.003 0.006 À0.003 (0.039) (0.038) (0.039) (0.038) Female children 16–20 À0.056 À0.064 À0.054 À0.064 (0.042) (0.041) (0.042) (0.041) Male children 21þ 0.013 0.006 0.011 0.005 (0.043) (0.042) (0.043) (0.042) Female children 21þ À0.066 À0.087* À0.063 À0.087* (0.052) (0.051) (0.053) (0.051) Number of children residing outside household 0.002 0.004 0.003 0.004 (0.010) (0.010) (0.010) (0.010) Kilometers from regional capital  number outside children 0.000 0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.000) MIGRATION AND ECONOMIC MOBILITY IN TANZANIA 1033 TABLE A6.—(CONTINUED) (1) (2) (3) (4) IHHFE IHHFE 2SLS with IHHFE 2SLS with IHHFE Age at baseline (1991–1994) 5–15 years 0.200*** 0.199*** 0.187*** 0.199*** (0.028) (0.027) (0.041) (0.038) 16–25 years 0.175*** 0.178*** 0.163*** 0.175*** (0.037) (0.036) (0.046) (0.043) 26–35 years 0.192*** 0.189*** 0.188*** 0.188*** (0.061) (0.060) (0.061) (0.060) 36–45 years 0.197*** 0.197*** 0.190** 0.195*** (0.075) (0.074) (0.077) (0.075) 46–55 years 0.256*** 0.270*** 0.253*** 0.270*** (0.086) (0.084) (0.086) (0.083) 56–65 years 0.232** 0.238*** 0.231** 0.238*** (0.093) (0.091) (0.093) (0.090) 66þ years 0.313*** 0.302*** 0.310*** 0.300*** (0.114) (0.112) (0.114) (0.111) Constant À0.137** À0.125** (0.061) (0.059) Cragg-Donald 11.86 9.33 Sargan statistic 10.59 11.44 Sargan p-value 0.06 0.04 Number of observations 3,227 3,227 3,227 3,227 Standard errors are in parentheses. Significance at ***1%, **5%, *10%.