Moved to Poverty? A Legacy of the Apartheid Experiment in South Africa¶
Bladimir Carrillo; Carlos Charris; Wilman Iglesias (2023). Moved to Poverty? A Legacy of the Apartheid Experiment in South Africa. American Economic Journal: Economic Policy 15(4). American Economic Association.
| Type | journal article |
| Scope | Economics and economic history of apartheid |
| Method | econometric |
| Period studied | 1937–1991 |
| Regions | South Africa, homelands/Bantustans (Transkei, Bophuthatswana, Ciskei, Venda, others), White areas |
| Themes | bantustans removals influx, education human capital, inequality poverty unemployment, labour markets wages, demography urbanisation, methods data sources |
| Citation key | carrillo2023a |
Argument¶
The paper argues that forced relocation of Black South Africans from 'White' areas to the homelands in the 1960s-70s lowered children's human capital and adult labour-market outcomes. It uses a cross-cohort, age-at-policy difference-in-differences design that exploits the staggered formal establishment dates of homelands in an intent-to-treat framework, with homeland natives as the comparison group. Younger children at predicted arrival lose more schooling, earnings and employment, while cohorts past school age show no effect. A place-effects analysis, in the spirit of Chetty and Hendren, finds that schooling converges toward that of the destination district, and the authors attribute roughly 70 percent of the decline to the childhood environment in the homelands.
Evidence and method¶
Apartheid-era 100 percent census samples (1980, 1985, 1991) from Statistics South Africa via DataFirst, with the 1980 census as the preferred sample. The 1996 post-apartheid census 10 percent sample is used for timing of arrival, the October Household Survey 1994-95 for fertility histories, and the 1999-2000 Internal Migration Survey. Historical sources include Surplus People Project (1983) and Simkins (1983) for removals.
Place in the debates¶
- Legacy of apartheid spatial segregation for contemporary inequality (quantifying causal effects) (see Long-run legacy and persistence of apartheid). Provides rigorous causal evidence that homeland relocation reduced human capital and earnings, supporting the view that apartheid's spatial policies have long-lasting inequality consequences.
- Place effects versus direct disruption/eviction/discrimination as mechanisms of forced migration outcomes (see Long-run legacy and persistence of apartheid). Argues place effects (poorer destination environments) explain about 70 percent of the schooling decline; discrimination, fertility change and school interruption are judged less important, and the eviction-induced education-insurance effect found elsewhere (Becker et al. 2020) is not found here.
Findings¶
-
Relocation to homelands in childhood reduced educational attainment, with the baseline coefficient significant and a predicted move at age 9 costing about 0.55 years of schooling. (p. 3)
Our preferred estimates imply that a predicted move at age 9 would reduce educational attainment by about 0.55 years, or, equivalently, 11 percent of the sample mean.
-
Labour-market effects: sharply lower employment and modestly lower income in adulthood. (p. 3)
the relocation policy led to a 20 percent decline in the likelihood of working and a 5.6 percent reduction in adult income.
-
Scale of removals to the homelands between 1960 and 1980. (p. 7)
Between 1960 and 1980, approximately 3.5 million Black Africans were forcefully removed to the homelands (Simkins 1983).
-
Homeland establishment strongly predicted in-migration timing, with arrival lagging establishment by almost seven years on average. (p. 14)
On average, individuals arrived in the homelands 6.8 years after their formal establishment.
-
The baseline difference-in-differences schooling estimate. (p. 17)
The estimated coefficient of interest is −0.116 (standard error = 0.019) and highly significant at the conventional levels of significance.
-
Falsification: homeland-to-homeland migrants, whose moves were not induced by the policy, show no schooling effect. (p. 22)
It is estimated at −0.01 with a standard error of 0.009. Our baseline estimates allow us to rule out an effect five times as large as the placebo coefficient.
-
No selection on education among pre-policy movers and stayers. (p. 23)
educational attainment between both groups is strikingly similar, with a difference of only 0.05 years that is statistically insignificant.
-
Migrants moved to areas with markedly lower predicted schooling. (p. 30)
On average, migrants moved to areas where schooling predictions are 0.75 years lower—a difference of about 25 percent.
-
Place effects account for most of the schooling decline. (p. 35)
we estimate that changes in childhood environment can account for approximately 72 percent
-
Place effects are gender-specific, consistent with a causal place interpretation. (p. 34)
The own-gender exposure coefficient is −0.066 (standard error = 0.016), which is very similar to the baseline and three times as large as the other-gender exposure coefficient.
Evidence ledger¶
Sources used¶
| Source | Kind | Period | How used |
|---|---|---|---|
| South African Census 1980, 100 percent individual-level sample (Statistics South Africa, via DataFirst) | population census | 1980 | Preferred estimation sample: years of schooling, place of birth and residence used to identify White-place migrants and homeland locals and to estimate the main DiD effects and place effects. |
| South African Census 1985 (individual-level, 100 percent sample) | population census | 1985 | Replicates the schooling DiD with older cohorts and younger cohorts at formalization. |
| South African Census 1991 (individual-level, 100 percent sample) | population census | 1991 | Replicates schooling results and provides the employment and log income outcomes (Table 3, Figure 7). |
| South African Census 1996 (10 percent sample), incl. previous residence and year of arrival questions | population census | 1996 | Used for retrospective migration histories to estimate the event study of the timing of moves, the 6.8-year average lag, actual vs simulated exposure, and a post-apartheid replication. |
| October Household Survey 1994 and 1995 | household or labour survey microdata | 1994-1995 | Retrospective birth histories for women are used to build a pseudo-panel testing for pre-formalization fertility changes. |
| South African Internal Migration Survey 1999-2000 | household or labour survey microdata | 1999-2000 | Used to check that out-migration after 1986 and multiple moves across districts are similar across migrants and natives and low (about 5 percent). |
| Homeland formalization dates and laws (Bantu Authorities Act 1951, Bantu Resettlement Act 1954; online Appendix Table A.1 of establishment dates) | legislation gazette | 1951-1970s (homeland establishment 1959 onward) | Staggered homeland establishment dates give the policy-induced variation in age at predicted move that identifies the effects. |
| Surplus People Project (1983), Forced Removals in South Africa, Vols 1-5 | commission or official report | 1960s-1980s | Alternative removals data cited as limited in dates; used to support that all documented removals were from White areas (falsification logic). |
| Simkins (1983) estimate of removals to homelands | secondary literature compilation | 1960-1980 | Source of the figure of about 3.5 million forced removals to the homelands. |
| Political map of South Africa 1986, Perry-Castañeda Library Map Collection | maps spatial data | 1986 | Source of the homelands map in Figure 1. |
| Human Sciences Research Council (2007), Using the 2001 Census | commission or official report | published 2007 (on 1991 census coverage) | Cited for the severe incomplete coverage of the 1991 census in the homelands, justifying use of 1980 as the benchmark. |
| Life expectancy series for South Africa 1960-1980 (theglobaleconomy.com) | national accounts macro | 1960-1980 | Used to argue mortality attrition among older cohorts is minor. |
New data¶
Authors built estimation samples from the 1980, 1985, 1991 and 1996 census microdata (e.g. about 2.49 million observations in 1980) matched to homeland formalization dates, with simulated and actual childhood exposure measures, magisterial-district predicted schooling gaps from pre-1950 cohorts, a 1996 individual-year migration panel, and a women-by-year fertility pseudo-panel from the OHS. Replication data deposited at ICPSR.
Inference¶
Quantification: Econometric (causal design).
Cross-cohort difference-in-differences (age-at-policy) design in an intent-to-treat framework, exploiting staggered homeland establishment dates. It compares White-place migrants with homeland natives using homeland-by-birth-year, province-of-birth and magisterial-district fixed effects. Validation uses event-study and flexible age-of-exposure coefficients, placebo/falsification samples, movers-vs-stayers checks, selection and attrition tests, and alternative inference (wild bootstrap, permutation). Mechanisms are tested with a Chetty-Hendren-style origin-destination schooling gap design.
Key claims¶
| Claim | Basis | Strength |
|---|---|---|
| Predicted moves to homelands in childhood reduce completed schooling: a predicted move at age 9 lowers attainment by about 0.55 years (11 percent of the mean). (p. 3) | 1980 census microdata, DiD with homeland x birth-year FE (Table 1, column 1; coefficient -0.116), supported by the 1985 and 1991 censuses and the 1996 census. | Systematic data |
| Childhood exposure also reduced adult employment (about 20 percent) and income (about 5.6 percent among workers). (p. 27) | 1991 census with the same DiD design (Table 3, Figure 7), with bounds for selection into the income sample. | Systematic data |
| Homeland establishment strongly predicts the timing of arrival of White-place migrants in homelands, with an average lag of about 6.8 years, and simulated exposure predicts actual exposure (slope 0.8). (p. 14) | 1996 census retrospective residence histories; event study of equation (1), Figures 2-4. | Partial data |
| The design is credible: no effects for cohorts predicted to move after school age, no effects for homeland-to-homeland migrants (placebo), and no sign of selection of movers in pre-policy cohorts. (p. 22) | Flexible age-at-arrival coefficients (Figure 5), falsification (Table 2), movers vs stayers pre-1950 cohorts (Figure 6), cohort-size and fertility checks. | Systematic data |
| Changes in childhood environment (place effects) account for roughly 70 percent of the overall decline in migrant schooling. (p. 35) | 1980 census; interaction of simulated exposure with origin-destination gap in district-level predicted schooling from pre-1950 cohorts (Table 4); back-of-envelope ratio of coefficients (-0.08x5 / -0.11x5). | Partial data |
| Discrimination, fertility change and school interruption are not the main explanations for the decline; there is no differential return to education for migrants and no increased migrant fertility after the move. (p. 36) | Online Appendix Table G.7 on returns to education; Appendix G.5 on OHS fertility; age pattern of effects. | Partial data |
| The census-based identification of migrants is accurate because mobility between homelands and out of them was strictly controlled during apartheid. (p. 7) | Historical accounts (Lemon 1984, Reed 2013), census-based out-migration estimates in Appendix C, and the 1980 census share of migrants in their ethnic homeland (65 percent). | Partial data |
The author's caveats¶
The authors note that there is no complete official information on the number or timing of removals. Place of birth is detailed only in the 1980 census, and the 1991 census has incomplete homeland coverage, which attenuates its estimates. The 1996 census lacks place of birth, so previous residence is used, and post-1986 out-migration may bias estimates. Ethnicity is missing in 1991, temporarily absent labor migrants may be mismeasured, and the ITT design relies on simulated rather than actual arrival ages. Some homelands are excluded from the core files, and the standard errors rest on few clusters (24).
Gaps in the evidence¶
The argument lacks direct administrative records of removals (dates, origins, destinations) to measure actual exposure rather than simulating it from homeland formalization dates and a lag estimated from the 1996 census. Direct data on school quality, health, parental employment and household assets in the homelands and White areas would test the mechanisms better than a district-level schooling gap proxy. Archival or Surplus People Project removal registers partly existed then and could be digitized now. Apartheid-era census quality, especially undercounting of Blacks, remains hard to assess.
Replicability: Data public and well described.