Online Appendix to "Moved to Poverty? A Legacy of the Apartheid Experiment in South Africa"¶
Bladimir Carrillo; Carlos Charris; Wilman Iglesias (2023). Online Appendix to "Moved to Poverty? A Legacy of the Apartheid Experiment in South Africa". Online appendix to journal article (For Online Publication).
| Type | other |
| Scope | Economics and economic history of apartheid |
| Method | econometric |
| Period studied | 1937–1996 |
| Regions | South African homelands (Ciskei, KwaZulu, Gazankulu, Lebowa, QwaQwa, KaNgwane, KwaNdebele, Bophuthatswana, Venda), White South Africa, Magisterial Districts |
| Themes | bantustans removals influx, education human capital, inequality poverty unemployment, labour markets wages, methods data sources, demography urbanisation |
| Citation key | carrillo2023 |
Argument¶
The appendix backs the main paper's claim that forced relocation of people from White areas into homelands lowered their educational attainment, employment and income relative to locals. The identification compares migrants from White areas with homeland natives across cohorts whose childhood exposure to the homeland differs, using the date each homeland was established. The appendix documents the census data and then tests robustness: alternative inference, falsification, fertility and cohort-size selection, outmigration, alternative specifications, and bounds for selection into the income sample. It also estimates that reduced schooling explains about a quarter to over a third of the income effect.
Evidence and method¶
Individual-level apartheid censuses (1980 complete count, 1985, 1991), the 1996 post-apartheid census (10 percent sample), the 1994/1995 October Household Surveys for fertility histories, homeland establishment dates from Dinkelman (2017), GIS maps of homeland boundaries, and Surplus People Project (1983) interview testimony on living conditions.
Place in the debates¶
- Long-run human-capital and welfare costs of apartheid homeland policy and forced removals (see Long-run legacy and persistence of apartheid). Treats relocation as a quasi-experiment showing causal negative effects on education, employment and income, and argues the results are robust to inference, selection and specification checks.
- Mechanisms: schooling versus other channels (school quality, inputs) in the income penalty (see Long-run legacy and persistence of apartheid). Finds reduced years of schooling account for roughly 25-37 percent of the income effect, with the rest attributed to unmeasured quality or other inputs.
Findings¶
-
A permutation test with 500 placebo formalization dates yields p-values below 0.01, suggesting the results are not a data artifact. (p. 13)
The implied p-values are always below 0.01, a result that is consistent with that computed using conventional inference
-
Trend-break tests place the end of the sensitive childhood exposure window at ages 13-15, so age 15 is used as the cutoff. (p. 11)
The test suggests a structural break between ages 13 and 15. Based on this evidence, we consider age 15 as the upper age limit of exposure
-
Cohort sizes do not change significantly with homeland establishment, so fertility responses are unlikely to drive selection. (p. 22)
While we do not observe significant changes in cohort sizes correlated with homeland establishment, we further explore this issue using the October Household Survey
-
Birth-rate trends of migrants and locals are similar around establishment, and fertility selection is not a major issue. (p. 23)
Summarizing, there is no strong indication that sample selection due to changes in fertility is a major issue.
-
Excluding migrants from districts near international borders leaves the education estimate unchanged, ruling out selective outmigration/immigration as a driver. (p. 26)
Simulated Exposure × Migrant -0.116 -0.1156 -0.1211 -0.1169 -0.1147
-
Education effects persist in the 1996 post-apartheid census, though smaller. (p. 19)
Simulated Exposure × Migrant -0.0683 -0.0439 [0.0270] [0.0190]
-
Controlling for schooling or using external returns-to-schooling estimates, education explains about a quarter to a third of the income effect. (p. 37)
Reduced educational attainment as a result of the homeland policy accounts for about 25-37 percent of the actual impacts on income.
-
Returns to schooling are not different for White-place migrants, so the income penalty is not due to lower returns for them. (p. 39)
Years of schooling × Migrant -0.000 -0.007 [0.002] [0.006]
-
Homeland residents faced scarce work and severe underfunding of services, per Surplus People Project interviews. (p. 7)
“Not enough supply of work”, “in this place unemployment is very rife”, “our problem is that we have not got enough money as a consequence of scarcity of work”
Evidence ledger¶
Sources used¶
| Source | Kind | Period | How used |
|---|---|---|---|
| 1980 Census complete count file (individual-level) | population census | 1980 | Main analysis sample for education outcomes; uses Magisterial District of birth to identify migrants from White areas and compare them with homeland natives. |
| 1980 census separate Bophuthatswana data file | population census | 1980 | Used only in a robustness check that adds Bophuthatswana to the sample (Table G.4), where province of birth for those born in White areas is unavailable. |
| 1985 Census | population census | 1985 | Used to examine basic patterns of education among migrants from White areas, with only province/homeland of birth available. |
| 1996 Census, 10 percent sample | population census | 1996 | Supplementary analysis of education (Table G.1) and out-migration/in-migration event studies; migrants from White areas identified from previous residence. |
| October Household Survey 1994/1995 | household or labour survey microdata | 1994-1995 | Constructs a pseudo-panel of women's fertility histories to test whether families changed fertility around homeland establishment. |
| South African Internal Migration Survey 1999-2000 | household or labour survey microdata | 1999-2000 | Cited to show few individuals move more than once, supporting accuracy of childhood exposure assignment in the 1996 census. |
| Homeland establishment dates from Dinkelman (2017) | secondary literature compilation | 1959-1977 establishment years | Gives the year of homeland formalization, which determines each cohort's age at homeland establishment and hence childhood exposure. |
| High-resolution GIS map of South Africa during the apartheid era | maps spatial data | Apartheid era | Matches 1996 Magisterial Districts to former homelands using a 60 percent territorial overlap rule. |
| Surplus People Project report (1983), interviews with relocated people | interviews oral history | 1980-1983 | Qualitative interview quotations illustrating that relocated White-place migrants faced scarce jobs and poor opportunities in the homelands. |
| Human Sciences Research Council (2007) on 1991 census coverage | secondary literature compilation | 1991 census | Cited for the fact that poor and remote homeland areas were not enumerated in 1991, limiting the reliability of those data. |
Inference¶
Quantification: Econometric (causal design).
Quasi-experimental difference-in-differences/event-study design in individual-level census microdata. It compares migrants from White areas (forcibly relocated to homelands) with homeland natives, by cohort exposure to homeland establishment during childhood. It uses fixed effects for homeland-by-birth-year, province of birth, and Magisterial District, plus flexible age-at-arrival estimates, falsification tests, permutation and wild bootstrap inference, and bounding for income selection.
Key claims¶
| Claim | Basis | Strength |
|---|---|---|
| The education result is not driven by particular homelands, provinces, districts, or the omission of Bophuthatswana. (p. 30) | 1980 census; alternative samples in Table G.4 and Figures G.11-G.12. | Systematic data |
| The key results are not explained by anticipatory fertility changes or changes in cohort sizes around homeland establishment. (p. 23) | 1980 census cohort-size regressions (Table G.2); October Household Survey 1994-95 fertility event study. | Partial data |
| The policy's negative effect on income is partly mediated by reduced education, accounting for about 25-37 percent of the impact. (p. 37) | 1991 census income regressions (Table G.6) combined with returns to schooling from Duflo (2001). | Partial data |
| The age-15 upper limit of childhood exposure is supported by structural-break tests on age-at-shock coefficients. (p. 11) | Trend-break F-tests on age-at-arrival coefficients from the 1980 and 1991 censuses (Figure E.1). | Systematic data |
| Homeland residents faced scarce jobs and poor living conditions, with severe underfunding of education and health. (p. 7) | Surplus People Project interview quotations and secondary literature (De Beer, Horrell, Dinkelman). | Illustrative or anecdotal |
| Post-1986 out-migration from the homelands was not significant, so selective out-migration is not a major threat; results are also found in the 1996 post-apartheid census. (p. 8) | 1996 census residential histories (Figure C.1); Table G.1 estimates. | Partial data |
The author's caveats¶
The authors note that the 1991 census has severe incomplete coverage in poor, remote homeland areas and omits Transkei, so results should be viewed with caution. Ciskei is excluded from the 1991 sample because place-of-birth information was inconsistent. The 1985 and 1991 censuses lack Magisterial District of birth. The core 1980 file omits Transkei, Venda and Bophuthatswana. The 1996 census has no place-of-birth question, so previous residence is used instead.
Gaps in the evidence¶
The study infers forced relocation from migrant status (White-area birth) rather than from records of actual removals, so measurement error in treatment is likely. Linking removal records (e.g., Surplus People Project files or state archives of relocations) to individuals or districts would test the mechanism better. Direct school quality and funding data for homeland schools would help explain the unexplained share of the income effect; some of this existed historically in official reports but is not used. Panel data tracking individuals over time would also allow stronger checks on selection.
Replicability: Data public and well described.