Abstract
We estimate the intergenerational transmission of schooling in a country where the majority of the population was rationed in its access to education. By eliminating apartheid-style policies against blacks, the 1980 education reforms in Zimbabwe swiftly tripled the transition rate to secondary schools. Using a fuzzy regression discontinuity design, we find a robust intergenerational transmission. A one-year increase in the schooling of the mother raises her child’s attainment by 0.073 years; the corresponding father-to-child spillover is 0.092 years. Choices in the marriage and labor markets mediate the size of these schooling transmissions. Several smoothness and placebo tests validate our design.
I. Introduction
Worldwide, the schooling attainments of children are positively correlated with the schooling attainments of their parents.1 This intergenerational correlation is interesting for two different but related reasons: (i) it sheds light on inter-generational mobility2—for example, a higher correlation often spells lower mobility, and (ii) it evokes the debate over nature and nurture in child development. Here, it is important to distinguish intergenerational correlations from the causal effects of parental schooling on child schooling. Causal effects provide insight into the role of parental nurture in the production function of child human capital (Haveman and Wolfe 1995; Holmlund, Lindahl, and Plug 2011). These effects are a vital input into the design of policies for child development and intergenerational mobility.
Only a few papers in the literature on intergenerational schooling transmissions have been able to identify causal effects.3 Nearly all of these papers focus on developed countries and obtain identification using one of three approaches.4 One approach compares twins; examples include Behrman and Rosenzweig (2002) and Amin and Behrman (2014) for the United States, Pronzato (2012) for Norway, and Amin, Lundborg, and Rooth (2015) for Sweden. Another approach exploits quasi-randomness in adoption placements; examples include Sacerdote (2002, 2007) and Plug (2004) for the United States and Björklund, Lindahl, and Plug (2006) for Sweden. An exception to the focus on developed countries is de Walque (2009) who uses family recomposition in the aftermath of the Rwandan genocide to generate an adoptee-adopter sample; however, nonrandom orphan assignment complicates the interpretation of his finding as causal.5 A third approach identifies causal effects using changes in schooling laws in the developed world, including in the United States (Oreopoulos, Page, and Stevens 2006), the United Kingdom (Chevalier et al. 2013), Norway (Black, Devereux, and Salvanes 2005), and Germany (Piopiunik 2014). Despite the diversity of empirical methods and causal estimates, our understanding of the mechanisms underlying intergenerational transmissions is incomplete, even for developed countries.
Our paper contributes to this literature by identifying the causal intergenerational transmission of schooling using an education reform in a developing country and by exploring the underlying mechanisms. We focus on Zimbabwe, a fragile, low-income country in southern Africa, where the majority black population was severely rationed in its access to education. We observe a major turning point in 1980 when the post-independence reform introduced automatic advancement to secondary school for black students, a practice reserved hitherto for whites. Before the reform, black students had to complete primary school (Grades 1–7), pass a test, and hope for a seat in in the limited number of secondary schools available to them. Automatic advancement eliminated the school-rationing feature of the apartheid era and brought a swift and discontinuous change in the transition rate to secondary school. As Figure 1 shows, the transition rate climbed from 27 percent of the cohort graduating from primary school in 1979 to 86 percent of the cohort graduating in 1980.
Primary-to-Secondary Transition Rates, 1970/71–1988/89
Data source: Riddell and Nyagura (1991, Table 1.1).
Notes: The transition rate is the percentage of students graduating from the seventh grade of primary school who enroll in the eighth grade (also known as Form I of secondary school).
The timing of the reform provides us a source of exogenous variation in the schooling attainment of blacks, and hence, a fuzzy regression discontinuity design. Our results show positive and significant intergenerational spillovers among black Zimbabweans. One extra year of schooling for the mother is associated with 0.073 additional years for her child. Similarly, an extra year of schooling for the father is associated with 0.092 additional years for his child. These estimates are not statistically different from each other. Our findings are robust to alternative specifications applicable in a fuzzy regression discontinuity design and to controls for potential confounds, such as rainfall shocks at birth. A number of smoothness and placebo tests failed to detect discontinuities, thereby reinforcing the validity of our methodology.
As in all papers that use an education reform to identify causal effects, our estimation strategy provides a local average treatment effect (LATE) of the parameter measuring the intergenerational transmission of schooling. That is, the causal effect is estimated from people whose behavior is influenced by the policy change. However, our paper differs from the literature using compulsory schooling laws in three important ways. First, Zimbabwe’s rule of automatic advancement to secondary school creates a different and, arguably, larger set of compliers. With compulsory schooling laws, the set of compliers is characterized by those who would drop out in the absence of the laws, but must stay in school under the new regime. The law does not change the behavior of those who already wanted to remain in school. Under Zimbabwe’s reform, described in the next section, the set of compliers is formed by those who wanted to stay in school but couldn’t due to the apartheid-style regime. Second, the treatment under compulsory laws is the addition of an extra year of secondary education (or high school). In Zimbabwe, the treatment is gaining entrance to secondary school. Third, Oreopoulos (2006, p. 153) argues that most compulsory laws like the ones implemented in the United States, “typically affect fewer than 10 percent of the population exposed to the instrument.” Zimbabwe’s reform affected a much larger share of its population. When given the chance to advance to secondary school, 86 percent of the eligible students changed their behavior, more than tripling the transition rate of the previous year. This implies that our LATE is closer to an average treatment effect (ATE) as the share of nontakers in the reform is quite small. Thus, our LATE is highly relevant to developing countries and to the Sustainable Development Goal of removing barriers to secondary education—the bottleneck of many education systems (UNESCO 2011).
Our work is also related to a new set of papers that have attempted to estimate intergenerational associations in developing countries emphasizing the external validity of their analyses. For instance, Schady et al. (2015) estimate wealth gradients in five Latin American economies (Chile, Ecuador, Colombia, Nicaragua, and Peru) and focus on outcomes related to early childhood cognitive development. Behrman et al. (2017) explore a more geographically diverse set of countries from the Young Lives Project (Ethiopia, India, Peru, and Vietnam) and seek to estimate the association between parental resources (income and schooling) and a wide set of human capital outcomes of the next generation. A detailed exploration of the causal estimates of the intergenerational transmission of schooling sets us apart from these studies and represents an important contribution of our paper, especially for developing countries.
By using a population census, we are able to explore father-to-child schooling transmissions, rather than only the mother-to-child effects. Thus, we avoid a major limitation of women-centric surveys such as the Demographic and Health Surveys (DHS) for developing countries. While two other papers have used the reform to estimate the impact of education on health with data from the Zimbabwe DHS (Agüero and Bharadwaj 2014; Grépin and Bharadwaj 2015), ours is the first to exploit the natural experiment inherent in the reform to estimate its impacts on the schooling of women and men. This widens the external validity of our findings. It also allows us to investigate the indirect pathways of the schooling spillovers—namely, outcomes in the marriage and labor markets. We find that women with higher schooling attainments delayed childbearing and had fewer children, which suggests a quantity–quality tradeoff. We also find that additional schooling did not lead to higher labor force participation rates or a greater likelihood of paid work, but it did induce a change in occupation. Finally, we find a high degree of positive assortative mating: both men and women selected partners with schooling attainments similar to their own.6 A highlight of this paper is our analysis of assortative mating—how it arises and how it could hinder policies seeking to end the intergenerational transmission of poverty.
This paper is organized as follows. Section II discusses the education reform. Sections III and IV describe the data and methodology, followed by results in Section V. Section VI concludes and discusses the policy implications of our findings.
II. The Post-Independence Schooling Reform
In southern Rhodesia, as pre-independence Zimbabwe was known, blacks where heavily rationed in their access to schools and education opportunities.7 For instance, in 1976, for every 1,000 black school-aged children, 250 never went to school, 337 completed primary school, 60 enrolled in Form I, the eighth grade and first year of secondary school, and fewer than three finished high school (Riddell 1980). By contrast, whites obtained universal primary enrollment and near-universal transition to Form I. Racial disparities continued in secondary school. In 1975, as many as 3,000 white students were enrolled in secondary classes leading to university entrance, compared to only 790 black students (Chidzero 1977). New construction of black secondary schools was heavily restricted. Between 1961 and 1972, only one new public secondary school was built to accommodate nearly 8,000 new black students nationwide (Zvobgo 1981).
Elections in April 1980 brought the Republic of Zimbabwe into existence, with Robert Mugabe as Prime Minister. His party had campaigned with the goal of “establishing free and compulsory primary and secondary education for all Zimbabwean children regardless of their race, sex, or class.” (Nhundu 1992, p. 78). The ensuing reform, documented by Edwards (1995), Edwards and Tisdell (1990), and Dorsey (1989), implemented four initiatives: (i) the introduction of free and compulsory primary education, (ii) the removal of age restrictions to allow over-age children to enter school, (iii) building community support for education, and (iv) automatic promotion from primary to secondary school, that is, from Grade 7 to Form I. It is the last feature of the reform that we consider in our analysis. The automatic promotion rule made it illegal to deny Form I admission to any student graduating from primary school on the grounds of poor test results or classroom seating constraints. As a result, the reform year saw an unprecedented fraction of black primary school graduates entering secondary school.
Figure 1 shows yearly transition rates to secondary school. In 1979, the last year before the reform, this rate was 27 percent; in 1980, the first reform year, it jumped to 86 percent. In absolute numbers, Form I enrollment soared from 22,201 to 83,491. Throughout the 1980s, transition rates remained high, averaging about 75 percent.
Figure 2 shows the rise in secondary enrollment from under 100,000 in 1979 to more than 650,000 in 1989. Over the same period, secondary schools grew tenfold on the strength of an expanding education budget. Up to one-fifth of the national budget was allocated to the education sector between 1980 and 1985, the bulk of which was used to open new secondary schools, especially in rural areas (Dorsey 1989).
Annual Secondary School Enrollment in Zimbabwe: 1973–1995
Data Source: United Nations Statistical Year book, 1975, 1980, 1982, 1984, 1985–1989, 1992, 1994, 1995, and 1997.
Education in Zimbabwe is structured as a 7 + 4 + 2 system: seven grades of primary, followed by four forms of secondary, capped by two years of high school. Children had to be at least seven years old to enter primary school,8 which implies that the first cohort of students who could have taken advantage of the automatic transition rule were black Zimbabweans finishing Grade 7 in 1980. These students would have been 14–15 years old, and disproportionately more likely to benefit from the reform compared to their slightly older counterparts. The point of discontinuity in black schooling attainment that we infer has also been observed in the Zimbabwe DHS data by Agüero and Bharadwaj (2014), Fenske (2015), and Grépin and Bharadwaj (2015). In the next section, we show that the discontinuity is also observed in our data set, the 2002 population census.
III. Data
We use a 10 percent random sample of the 2002 Population Census. Because we are interested in linking the schooling attainments of the school-age generation at the time of the reform to the schooling levels of the children of that generation, we begin by selecting the sample based on two rules: adult black Zimbabweans, comprising individuals born between 1959 and 1974, and their children ages 6–15 years in 2002. This means that the schooling levels in the child sample are intermediate outcomes.9 We divide the sample along gender lines and drop the handful of observations missing schooling data.10 Additionally, we restrict the female sample to women who gave at least one live birth and report age at first birth. The census allows us to identify the parent of a child if the adult in question is either the household head or the head’s spouse, so we derive two analysis samples: 91,480 mother–child matches and 50,026 father–child matches.11
Table 1 reports descriptive statistics for our two analysis samples. The average black mother is 35 years old and has eight years of schooling. Fathers are slightly older and have slightly more schooling. The average child is about ten years old, and daughters and sons occur in equal proportion in both samples. In either sample, 97 percent of children currently attend school, while 100 percent have ever attended school. Clearly, the interesting question is not whether a child attends school but whether she attends a grade appropriate to her age. Our census sample shows, for example, that 98 percent of 11-year-olds are enrolled in school, but the average 11-year-old is already a full year behind in terms of grade for age. For simplicity and comparison with the rest of the literature, we choose the child’s current grade attainment as our outcome of interest and employ fixed effects for child age to absorb variations in grade attainment related purely to the child’s time in school.12 In the next section, we describe our identification strategy to obtain causal estimates of the intergenerational transmission of schooling among black Zimbabweans.
Descriptive Statistics
IV. Identification Strategy
The intergenerational schooling relationship may be captured by the equation13
(1)
where yi is the schooling outcome of child i, and si is the schooling attainment of her parent; 
is a vector of child and parental characteristics, such as age, gender, and location.14 The parameter of interest capturing the intergenerational spillover is β. However, omitted unobserved variables in εi correlated with parental schooling would bias the least-squares estimate of β. To minimize this possibility, we seek a source of exogenous variation in parental schooling.
As detailed in Section II, the reform made the transition from Grade 7 to Form I automatic. Although the reform did not go so far as to impose an age cutoff on its potential beneficiaries, the timing induced very different probabilities of secondary school enrollment among black Zimbabweans. Specifically, those who were younger than 15 years of age in 1980 were disproportionately more likely to achieve the primary-to-secondary school transition than those who were slightly older. The probabilities diverge at age 15 because this was the typical transition age to Form I at the time. Transitioning to Form I was also the natural next step for those who had not been rationed out of a seat in Grade 7 before 1980 (Dorsey 1989; Nhundu 1992).
The reform-induced jump in the probability of secondary school enrollment creates a fuzzy regression discontinuity design. This provides an instrumental variable for parental schooling si in the point of discontinuity, 
. We can now estimate the inter-generational schooling relationship via two-stage least squares (2SLS) with the following equations:
(2)
(3)
where f(·) takes the form
(4)
Following van der Klaauw (2002), we choose a piecewise linear representation for f(·) with 
. However, as we show later, our results are unaffected by approximations where K > 1 in f(·).
Based on Hahn, Todd, and van der Klaauw (2001), a fuzzy regression discontinuity design implies that a consistent estimation of β by 2SLS requires two assumptions. First, the reform needs to have discontinuously altered schooling levels of the target population at the threshold to avoid a weak-instrument problem. Second, the reform needs to have affected children’s schooling only through the schooling attainments of their parents.
Figure 3 supplies visual support for the first assumption, which we formally test in Section V. Following Imbens and Lemieux (2008), we plot the conditional expectation function (CEF) of parental schooling detrended in age15; that is, we display the residuals from a regression of parental schooling on a linear polynomial in the running variable (parent age in 1980) and an interaction between the threshold and the polynomial term. In Figures 3a and 3b, we observe a clear discontinuity in completed years of schooling around the threshold for black mothers and black fathers, respectively.16
The Schooling Attainment of Black Mothers and Black Fathers by Age in 1980
Notes: Mean years of schooling are plotted against age in 1980 separately for black mothers and black fathers.
The treatment threshold under the reform is marked by the vertical line at age 15.5.
The second assumption that the discontinuity satisfies the exclusion restriction is not directly testable; however, we attempt to validate it in several ways. First, we examine “covariate smoothness.” Although the census is limited by way of variables that could be considered predetermined with respect to the reform, several key covariates are smooth around the point of discontinuity: these include parent age (Online Appendix Figures A1a and A1b), the race composition of the adult sample (Online Appendix Figure A2), and the sex composition of children in the mother and father samples (Online Appendix Figure A3). It also includes parent height (from the Zimbabwe DHS), which is largely determined by early-life factors. As Online Appendix Figure A4 shows, height is smooth around the threshold age of 15.
Second, we conduct a placebo test. We explore if Zimbabwe’s schooling discontinuity at age 15 in 1980 is mirrored in other Sub-Saharan African countries, particularly those where no major reform occurred in 1980 or targeted the same age cohorts as in Zimbabwe. This investigation is shown in Online Appendix Figures A5 and A6 for women and men, respectively. The data were taken from IPUMS (https://international.ipums.org/international/). We looked for countries with a population census circa 2002, so the placebo samples would contain individuals who were contemporaries of the parents in our Zimbabwe sample.17 As expected, the graphs for these countries show a smooth relationship of schooling with the running variables at the point of discontinuity.18
Third, we explore if Zimbabwe undertook any other policies, pre- or post-independence that may have directly affected the schooling of the children via parent wealth or parent outcomes in the labor market. Because we use a fuzzy regression discontinuity design, our identification strategy is threatened only if there is another policy affecting those who were 15 in 1980 discontinuously compared to those who were 16 in 1980. Our identification is not threatened by policies that targeted or affected the black population according to age unless those policies had discontinuous impacts on 15-year-olds and 16-year-olds in 1980. We found no such policy in the labor market. However, a land redistribution program, a major post-independence policy initiative, could have affected the nature of rural labor markets (see Oryoie, Alwang, and Tideman 2017; Deininger, Hoogeveen, and Kinsey 2004; Kinsey 2004). In the early 1980s, Mugabe’s government conceived the land redistribution program as a form of redress for rural impoverishment under apartheid. The program’s resettlement activities peaked for a brief time in the mid-1980s, but implementation fizzled soon after (Kinsey 2004). Clear eligibility criteria for resettlement were not developed until the program was reevaluated in the following decade, and there is no indication that the treated or control cohorts used in our paper were targeted for resettlement in any systematic way.19 We discovered no other reforms at the national or provincial level that could have affected the intergenerational schooling relationship, and at the same time, differentially impacted those aged 15 in 1980 compared to those aged 16. This is not surprising given how fragile Zimbabwe was politically and economically as it emerged from apartheid. The state was in no position to finance additional large programs at the same time that it launched the very popular and ambitious reforms of school expansion and land resettlement. In the next section, we present the regression counterparts of these graphical analyses and consider additional robustness tests for the intergenerational spillovers in schooling.
V. Results
A. First Stage: Impact of the Reforms on Parental Schooling
Table 2 reports estimates of the first stage, with standard errors clustered on the discrete assignment variable, age in 1980 (Lee and Card 2008). In Online Appendix Table A4, we consider six other clustering options and find that inference is similar across clustering protocols.20
The Effect of the Reforms on Parent Schooling: First Stage
In Panel A of Table 2, Column 1 shows that at the discontinuity, black mothers have 0.819 more years of schooling compared to their slightly older counterparts. In Panel B, Column 1 shows that black fathers have 0.683 more years of schooling at the threshold. Both estimates are significantly different from zero, and the F-statistics attest to a strong first stage. Compared to the program estimates in Duflo (2001), where an additional primary school per thousand Indonesian children delivered an additional 0.12–0.19 years of schooling, the impact of Zimbabwe’s reform is at least three times as large. In Column 2, we control for background parental characteristics using rainfall in parent year of birth. We find that this does not change the estimated reform effect in either Panel (we return to this issue in the next subsection).
In Column 3, we run the same specification as Column 1 in a pooled sample of 11 Sub-Saharan African countries. Limiting each country’s sample to its native-born in the ages of 6–21 years in 1980, we test for discontinuity in years of schooling at age 15 in 1980. Consistent with the visual evidence presented earlier, we find no discontinuity in either the women’s sample or in the men’s sample. The estimates (0.154 for women and 0.109 for men) are smaller but more precise than the estimates for Zimbabwe (each regression sample has more than 1.5 million observations); they are positive but not statistically different from zero. This result is reinforced by the country-by-country results shown in Online Appendix Table A2 and the scatterplots discussed above. These placebo tests bolster the validity of our identification strategy.
Table 3 reports smoothness tests on covariates. As Columns 1–3 show, predetermined variables like the child’s sex, the parent’s height (in centimeters), and the parent’s height-for-age Z-score are not discontinuous around the age threshold. In the women’s sample in Column 4, the coefficient on the discontinuity is significant at the 1 percent level, but it is tiny—it suggests that at age 15, the probability that a woman is black increases by 0.1 percentage points from a base of 99 percent. This coefficient is statistically zero in the sample of men.
Continuity in Exogenous Characteristics
In sum, we base our empirical strategy on the evidence that the education reform discontinuously affected cohorts that were close in age, namely, those who were just under age 15 in 1980 and those just older than 15. While independence brought much economic change as well as political and social reform to Zimbabwe—as in many other African countries—our identifying assumption is that these changes did not discontinuously affect the cohorts around the cutoff age of 15 years in 1980. Indeed, on the strength of previous research and the new evidence presented here, we conclude that there is strong evidence of a fuzzy regression discontinuity in the schooling of black Zimbabweans and that it provides an exogenous source of variation in the schooling attainments of the parent generation in our sample.
B. Estimates of the Intergenerational Spillover
We now present our estimates for the intergenerational transmission of schooling in the sample of black Zimbabweans. We start by showing the reduced-form effects graphically. Figure 4 plots the CEF of child grade attainment against parent age in the two samples of mothers and fathers, respectively. For mothers, we observe a clear jump at the discontinuity (Figure 4a). For fathers, the discontinuity is less pronounced (Figure 4b), but in both cases, the reduced form estimates show a difference at the threshold that is statistically different from zero and, as discussed below, robust to several specifications.21
Mean Child Grade Attainment by Parent Age in 1980
Notes: Because children in the sample are still in school, grade attainment data was purged of time-in-school variation by running a piecewise linear regression in child age and parent age, and extracting the residuals. Above, the means of these residuals are plotted against parent age.
Table 4 complements these figures with the regression estimates of the intergenerational transmission of schooling. In Column 1, we show ordinary least squares (OLS) estimates. For both mothers (Panel A) and fathers (Panel B), we find a positive and statistically significant intergenerational association. Because these estimates could be biased by the omission of unobservable covariates of parental schooling, such as ability, in Column 2, 2SLS we present estimates using the fuzzy discontinuity design. We find that a one-year increase in the schooling of the mother results in 0.073 additional years in the schooling of her child. A similar increase in the schooling of fathers brings about 0.092 additional years of child schooling. For both samples, the estimates are statistically different from zero at the 1 percent level, and we cannot reject the null hypothesis that the mother and father effects are equal to each other. The 2SLS estimates are smaller than the OLS, which suggests that unobserved covariates like ability account for the gap between them.
The Intergenerational Effects of Schooling
To put our 2SLS estimates in perspective, we estimated intergenerational associations in schooling in each of the 11 Sub-Saharan African countries we invoked in the placebo tests discussed.22 Limiting the placebo samples to native children aged 6–15 with parents aged 6–21 in 1980, we obtained estimates for mothers in the range 0.068–0.288 and for fathers in the range 0.073–0.215 (see Online Appendix Table A3 for details). Thus, our Zimbabwe estimates are consistent with schooling spillovers for other countries in the region. This is also consistent with the evidence presented by Beegle et al. (2016), who show that Africa has greater intergenerational educational mobility than Latin America.
C. Robustness Checks
We conduct a number of robustness and specification checks as shown in Tables 5 and 6. Table 5 reproduces our 2SLS estimate of the intergenerational spillover in Column 1. In Column 2, we check if the original estimate is biased by the omission of parental background characteristics correlated with the timing of the reforms. We know from Almond, Currie, and Herrmann (2012) and Currie and Vogl (2013) that early-life shocks often have long-lasting health effects. In a predominantly agrarian society like Zimbabwe, at-birth exposure to drought or flood could slow or permanently reduce human capital accumulation.23 Using a 40-year time series, Richardson (2007) finds that the growth rate of Zimbabwe’s per capita GDP is strongly correlated with annual rainfall. From Figure 5, which displays standardized rainfall data24 for the years 1959–1985, it is clear that rainfall is smooth in the neighborhood of 1965, the birth year of 15-year-olds in 1980. This explains why adding rainfall to the first-stage regression did not change the reform’s effect on parental schooling (see Column 2 of Table 2). For the same reasons, adding rainfall to the 2SLS estimation does not alter our main findings (see Column 2 of Table 5).
Annual Rainfall in Zimbabwe, 1959–1985
Data Source: Zimbabwe Meteorological Service Department.
Notes: Time series data on annual rainfall comes from a sample of 38 stations across Zimbabwe for the years, 1959–2001. Above, we standardize and plot this data for the period, 1959–1985. A given year, such as 1970, refers to the 1970–1971 crop-year. The vertical line represents the year of birth for the cohort aged 15 in 1980.
Robustness Checks (2SLS)
Additional Specification Tests: Alternative Parent Samples and Polynomials in Age Splines (2SLS)
In Column 3 of Table 5, we deal with province-level unobservables potentially confounding our 2SLS estimates by adding province-of-birth fixed effects to our original specification. We find that the intergenerational estimates remain highly significant, but become marginally smaller. We find similar results when we control for parents born in an urban area (Column 4). In Column 5, we omit units at the threshold (that is, parents aged 15 in 1980) and rerun the original 2SLS specification. Once again, the estimated effects (0.066 in the mothers sample and 0.079 in the fathers sample) are statistically different from zero and belong within the confidence interval of the baseline 2SLS. In all our robustness checks, we continue to find a very strong first stage based on the F-statistics and p-values.
We also investigate the sensitivity of our estimates to four alternate specifications of the running variable and for two age spans (6–21 and 0–30 in 1980). The specifications differ according to the degree of the polynomial in parent age, and whether ages 14 and 15 in 1980 are included in or omitted from the estimation sample. In Table 6, Column 1 displays the baseline 2SLS estimates for ease of comparison; Columns 2 and 4 use quadratic polynomials in parent age with the difference that Column 4 omits parents at or near the threshold. The implied point estimates for mothers and fathers are not statistically different from the baseline estimate. This is also the case in Column 3, where we omit 14- and 15-year-olds in 1980 while retaining a linear polynomial in parent’s age. When we expand our estimation sample to include parents aged 0–30 years in 1980, we find that the estimates are more sensitive but continue to show a statistically significant intergenerational transmission of schooling for fathers and mothers. Overall, these sensitivity checks suggest that our baseline specification is conservative in the sense that alternative specifications yield larger estimates of the intergenerational schooling effect.25
D. Heterogeneous Effects
We now examine heterogeneity in the intergenerational spillovers. Table 7 shows the effects on daughters and sons separately. In all four parent–child pairings, we find a strong and statistically significant transmission of schooling at the 1 percent level. While the mother’s effect is larger for daughters than for sons, and the opposite holds for fathers, we cannot reject the null hypothesis that the mother (and father) effects are the same for daughters and sons. This suggests that there are no systematic gender preferences in the schooling transmission.
The Intergenerational Effects by Sex of the Child (2SLS)
Figure 6 explores whether the transmission varies with the child’s age. The top panel shows that mothers clearly raise child grade attainment when children are in the ages of 7–10 and, again, for children in the ages of 12–14. The bottom panel shows that fathers affect child schooling positively when children are between 9 and 12 years of age. Until children enter the teenage years, in point estimate terms, the father’s schooling effect outweighs the mother’s effect. The two panels suggest a complementarity in the timing of the parental schooling effects: it is consistent with a model where the marriage market plays an important role in shaping the intergenerational spillover. In the next section, we provide evidence in favor of this mechanism as we uncover significant positive assortative mating on schooling attainment by parents.
Heterogeneity in the Effects of Parental Schooling
Notes: In each figure, the solid line connects the 2SLS estimates of the effect of parent schooling on child schooling at child ages 6–15 years. The dashed lines represent the 95 percent robust confidence bounds clustered by parent year of birth. Each 2SLS regression controls for the child’s sex, uses linear splines in parent age in 1980 on both sides of the discontinuity and instruments parent schooling with the discontinuity at age 15 in 1980.
E. Additional Child Outcomes
In Table 8, we explore the intergenerational spillovers on three additional child outcomes. In Column 1, we ask if parental schooling affects whether the child is currently attending school. We find that it does not. Since nearly all children have attended school at some point, we infer no selection based on the level of parental schooling. In Columns 2 and 3, we investigate impacts on child labor.26 In Column 2, we define child labor as market work participation: the binary outcome equals one if the main activity of a child in the last 12 months was one of paid market work, own-account work, unpaid family work, or unemployed as opposed to student, homemaker, or other. In Column 3, we widen the definition of child labor to include domestic chores; that is, homemaker children are now counted in child labor. We see no effect of parental schooling on either measure of child labor prevalence.27 In sum, a more-educated parent helps raise the child’s grade attainment, but affects neither the child’s likelihood of enrollment nor her labor force participation.
The Intergenerational Effects on School Attendance and Child Labor (2SLS)
F. Possible Mechanisms
We now explore some of the pathways through which the intergenerational schooling spillovers operate. Towards this end, we replace child grade attainment in Equation 2 with a new set of outcome variables. Figures 7 and 8 provide a visual analysis of these pathways for mothers and fathers.
Pathways in the Intergenerational Transmission of Mother’s Schooling
Notes: Mean values of a marriage or labor market outcome are plotted against the age of the mother in 1980. In Panels A, B, and F, the outcomes are dichotomous; Panel B displays the fraction of mothers in each of the 16 age cohorts whose partners completed schooling up to at least Form I. In Panels C, D, and E, the outcomes are continuous; Panel C displays the average age gap (excess of male partner’s age over woman’s age) in each age cohort of mothers. The vertical line at 15.5 indicates the treatment cutoff under the reform.
Pathways in the Intergenerational Transmission of Father’s Schooling
Notes: Mean values of a marriage or labor market outcome are plotted against the age of the father in 1980. In Panels A, D, E, and F, the outcomes are dichotomous. Panel C displays the average age gap (excess of man’s age over female partner’s age) in each age cohort of fathers. The vertical line at 15.5 indicates the treatment cut-off under the reform.
We start by examining if schooling changed household formation by altering the marital status of parents or their likelihood of living together. Papers in the literature that give considerable weight to this mechanism include Behrman (2010) and Fafchamps and Shilpi (2014). In Norwegian data, Kalil et al. (2015) find that the father’s presence alters the magnitude of the intergenerational transmissions from both parents. In Table 9, Column 1, we find that schooling does not alter the marital status of mothers (Panel A), but it does make fathers more likely to be married (Panel B). Exploring if educated parents are more likely to live with their partners, measured by coresidency on the night before the census, we do not find an effect for mothers or for fathers (Column 2). This finding contrasts with results from developed countries but echoes Fenske (2015), who, using the DHS, finds no effect of schooling on polygamy in Zimbabwe. Thus, schooling appears to have no effect on the probability of entering the marriage market.
Marriage Market Pathways of the Parental Schooling Effect (2SLS)
Next, we investigate the quality of the match in the marriage market. Studies that have discussed match quality as a potential mechanism in developed economies include Charles, Hurst, and Killewald (2013) and Edwards and Roff (2016). Foster (2002) develops a model of the marriage market in which prospective mates seek a match to maximize own private consumption, as well as the human capital of their future child in the marriage, given incomes and tastes for schooling. Applying this model to Bangladesh, Foster finds that marital selection accounts for a significant portion of the cross-sectional correlation between parental and child schooling. We extend his finding with estimates for Zimbabwe. In Table 9, Columns 3 and 4 report how a parent’s own schooling relates to the schooling of the coresident partner. The dependent variables are the partner’s completed years of schooling (Column 3) and an indicator that the partner’s attainment is at the secondary schooling level or higher (Column 4). We find strong evidence of positive assortative mating. In Panel A, an additional year of maternal schooling is associated with 0.56 more years in her partner’s education and a 7.2 percentage-point (or 13 percent) higher probability that her partner is educated beyond primary school. In Panel B, an additional year of paternal schooling is associated with 0.49 more years in his partner’s education, and again, a 13 percent (that is, 0.079/0.6) greater probability that his partner’s educational attainment exceeds primary school. All of these estimates are statistically significant at the 1 percent level.
Arguably, the assortative mating is behavioral, not mechanical. If parents in our sample married within the same age cohort, assortative mating would be a mechanical consequence of the reform. However, there is evidence that parents do not marry within the same age cohort. On an average, mothers have partners who are seven years older (Figure 7) while fathers have partners who are six years younger (Figure 8). Most importantly, there is no discontinuity in partner age gap at the threshold (see Column 5 in Table 9, Panels A and B). We can therefore reject the possibility that parents married within the same age cohort. Educated parents—much like their less-educated counterparts—do not deviate from the “social norm” of women (men) selecting an older (younger) partner. Yet, offered a pool of possible partners from a different age cohort, educated parents seem to choose more-educated partners.
Using only the sample of mothers, we examine how, if at all, schooling affected fertility. In policy circles and academic writing, fertility is regarded as a key driver of the impact of maternal schooling on child welfare (Summers 1992; Behrman 1997; Schultz 2007). The last two columns of Table 9 provide evidence in support of this view. In Column 6, we find that an additional year of schooling increases women’s age at first birth by 0.56 years, a 3 percent increase from the mean. In Column 7, we show that each extra year of education reduces the number of children women bore by 4 percent. Thus, schooling appears to have caused women to postpone childbearing and to make a quantity–quality tradeoff by having fewer but more-educated children.28
The intergenerational schooling transmission could also be explained via income effects in the labor market. Unfortunately, neither the population census nor the DHS contain data on earnings, which would have allowed us to test for an education premium in Zimbabwean labor markets.29 The census does allow us to estimate the effects of schooling at the extensive margin, such as labor force participation and the status of being a paid worker. Paid employment is an important labor market outcome, especially in the context of Sub-Saharan Africa (Vijverberg 1992). Yet, in our analysis (Table 10), neither mothers nor fathers show any schooling effects on their workforce participation (Column 1) or probability of being a paid worker (Column 2). This is not very surprising because for the fathers in our sample, participation in the labor force and in paid labor is near universal (98 percent and 96 percent, respectively), leaving a very small margin to be affected by schooling. It is also well documented that women in low-income countries are more likely to be employed relative to women in middle-income countries (Mammen and Paxson 2000). In our sample, 75 percent of mothers are in the labor force, and 95 percent of working mothers are paid for their labor.
Labor Market Pathways of the Parental Schooling Effect (2SLS)
We do find an effect, however, on the type of occupation where the educated are employed. As Columns 3 and 4 of Table 10 show, both educated fathers and mothers are less likely to be in primary sector employment or manual labor (for example, farming, hunting, fishing, logging, mining, quarrying, brick-laying, masonry, painting, cleaning, or subsistence work in general) and more likely to be in jobs requiring higher skills (for example, programmers, academics, accountants, lawyers, doctors, engineers, artists, executives, librarians, and bankers, among others). It is this margin of the labor market that matters for the intergenerational transmission of education.
G. Quantity and Quality of Education
When describing the 1980 education reform in Zimbabwe, Kanyongo (2005) notes that “[t]he emphasis [of the reform] was not so much on quality and cost effectiveness of the education system, but on accessibility to education.” This narrow focus helped Zimbabwe achieve universal primary enrollment as early as 1990 and the highest adult literacy rate in Sub-Saharan Africa (UNICEF 2010). Yet, as several descriptive studies report, the quality of education in Zimbabwe declined even as the number of schools increased. For instance, Edwards (1995) calls 1984 the last year with “good” quality outcomes, at least in primary education. Nhundu (1992, p. 87), reviewing documents from the Ministry of Education on the decline in quality after 1984, observes that school enrollment in the 1980s occurred “faster than classrooms and teacher’s houses could be built.” According to Dorsey (1989) and Nhundu (1992), growth in teaching staff failed to keep pace with enrollment. The share of untrained teachers in secondary schools rose from near zero in 1980 to 28 percent in 1988. Because a significant number of secondary schools had been built as extensions of existing primary schools, schools resorted to “hot seating,” a practice of reducing the length of a school day to accommodate more students. Nhundu (1992) further reports that between 1981 and 1988, the number of students taking exams to gain admission to higher levels of secondary education (O-level exams) grew by 2,253 percent, and the failure rate increased by 7,220 percent.
Though it is clear that the quality of education declined over the 1980s, we lack the data to determine if the decline occurred discontinuously, but it is rather likely. As a phenomenon, decline in education quality is not necessarily unique to Zimbabwe. In several low- and middle-income countries, the rapid expansion of education has been marked by serious concerns about a decline in quality (Grisay and Mahlck 1991; World Bank 2018). Even developed countries enacting minimum compulsory schooling laws have experienced problems with education quality. Unfortunately, the literature has not evaluated such impacts. For Zimbabwe, our 2SLS estimates should be regarded as lower bounds to the “true” effects that could have been obtained if quality was held constant. Understanding whether and how much these reforms impact on the quality of education is an important item for future research.
VI. Conclusion
Although the literature on intergenerational transmissions of schooling has grown in recent years, developing countries remain understudied. Our paper presents causal estimates of the schooling transmission in the context of a low-income country and a population that was systematically deprived of access to education for several generations. In particular, the scale of Zimbabwe’s natural experiment, facilitating the transition to secondary school, bestows a large degree of external validity on our estimates, as the set of compliers (those whose behavior is affected by the reform) was large. Exploiting the regression discontinuity design spawned by the 1980 reform, we estimate that the intergenerational spillover from an additional year of parent schooling is approximately 0.07 years of child schooling from the average mother and 0.09 years of child schooling from the average father.
An important contribution of this work is the exploration of a number of mechanisms to explain the intergenerational schooling effects. We find that both labor and marriage markets are important channels in the transmission of parental schooling. We find that more educated couples exhibit assortative mating on schooling, delayed childbearing, and quantity–quality tradeoffs in family size, all of which suggest that they reap larger marriage surpluses than less-educated couples. However, in a dynamic setting, assortative mating reduces mobility. Thus, an important question for future research concerns the long-term implications of higher schooling transmissions and assortative mating for intergenerational mobility.
Footnotes
The authors thank staff at the Central Statistical Office of Zimbabwe for providing a 10 percent sample of the 2002 population census and Craig Richardson for sharing the rainfall data. They thank Chris Barrett, Prashant Bharadwaj, Marianne Bitler, Resul Cesur, Pascaline Dupas, Steven Helfand, Mindy Marks, Todd Sorensen, and Aman Ullah for comments on early drafts of this article, as well as seminar participants at Cornell-IBEPP, Cal-State Fullerton, PACDEV, Oxford-CSAE, NBER Program on Children, Claremont McKenna, the PAA Economic Demography Workshop, and the SEA annual conference. All views and outstanding errors are those of the authors. The paper uses confidential data from the Zimbabwe National Statistics Agency, which can be obtained by filing a request directly with the Director General (http://www.zimstat.co.zw, dg{at}zimstat.co.zw). The authors are willing to assist any scholar seeking access to the data (Jorge M. Agüero, jorge.aguero{at}uconn.edu). The data from the Demographic and Health Survey can be obtained for free after registration at https://dhsprogram.com.
Supplementary materials are freely available online at: http://uwpress.wisc.edu/journals/journals/jhr-supplementary.html
↵1. See Björklund and Salvanes (2011) for examples from high-income countries and Beegle et al. (2016) and Ferreira et al. (2013) for examples from low- and middle-income countries.
↵2. As Black and Devereux (2011) note, education allows mobility to be measured at an earlier moment in the lifecycle because most people tend to end their education in their early twenties. Unlike earnings, education is measured with lower error and can be observed for both the unemployed and those not in the labor force.
↵3. See a review by Haveman and Wolfe (1995) and, more recently, by Black and Devereux (2011) and Björklund and Salvanes (2011).
↵4. There is a growing literature summarized by Grossman (2015) on the effects of parental schooling on child health that includes research on developing countries. While important, and related to the effects of parental schooling on child schooling, we do not investigate this intergenerational transmission in our paper. There is another related literature that examines the intergenerational transmission of assets and poverty dynamics; see for example Fafchamps and Quisumbing (2005).
↵5. See also Zimmerman (2003) for a related study about fostered children in South Africa.
↵6. As discussed later in more detail, this assortative mating is not mechanical. Zimbabweans do not tend to marry within their age cohort; rather men marry women who are, on average, six years younger. Other patterns in partner age gap suggest a deliberate behavior of assortative mating.
↵7. For a history of the apartheid-era education system and the policies dictating the quantity and quality of schooling permitted to Africans, see Atkinson (1972) and O’Callaghan and Austin (1977).
↵8. The entry age was lowered to six years in 1989.
↵9. In this regard, we differ from studies that estimated intergenerational associations for adult children, such as Beegle et al. (2016) and Behrman, Gaviria, and Székely (2001).
↵10. We also drop polygamous households. This is a minor restriction because polygamy occurs in only 2 percent of the households in our sample, and a recent paper finds no schooling effect on polygamy in Zimbabwe (Fenske 2015). Furthermore, we find that the schooling outcomes of children excluded from our analysis because they are not the offspring of household heads are very similar to the schooling outcomes of the children in our analysis. This information is available upon request.
↵11. As shown in Online Appendix Table A3, it is common in Sub-Saharan African countries to be able to match more mothers to children than fathers to children. Indeed, South Africa and Botswana have a similar sample size ratio of mother–child matches to father–child matches as Zimbabwe.
↵12. Our results are robust to differences in recall precision, so long as recall error is zero-mean for educated and less-educated parents.
↵13. See Black and Devereux (2011) for alternative ways to measure intergenerational mobility and Holmlund, Lindahl, and Plug (2011) for the specific case of schooling.
↵14. In our models, we do not control for the spouse’s education. As Holmlund, Lindahl, and Plug (2011) explain, including the schooling of the spouse complicates the interpretation of the intergenerational effects. There is also a statistical difficulty in our case because we estimate the intergenerational coefficients as instrumental variable estimates. In the presence of assortative mating, spouse schooling levels and reform exposures will tend to be strongly correlated. With highly correlated instruments, we are in danger of ending up with imprecise 2SLS estimates of the intergenerational effects. A partial workaround is suggested by Oreopoulos, Page, and Stevens (2006) who use the sum of maternal and paternal schooling as the endogenous regressor of interest. Their approach overcomes the multicollinearity problem at the cost of having less informative intergenerational estimates, a tradeoff that is worthwhile if the relative sizes of maternal and paternal effects are not particularly relevant. We prefer to estimate them separately so we can later test for assortative mating.
↵15. Negative correlations between schooling and age have been noted before in cross-sectional data. Detrending is a simple way to accommodate the increase in schooling attainments over time. Using household data over multiple decades from South Africa, Nigeria, Côte d’ Ivoire, Kenya, Burkina Faso, and Ghana, Schultz (2004) describes the progress in female attainments over time as slow but continuous. Zimbabwe’s reform delivered universal primary education and literacy in a span of just 11 years.
↵16. The unconditional functions, which contain age trends, show similar patterns and have been used elsewhere (Agüero and Bharadwaj 2014; Fenske 2015; Grépin and Bharadwaj 2015).
↵17. The 11 countries in our placebo are Botswana, Ghana, Malawi, Mali, Rwanda, Senegal, Sierra Leone, South Africa, Uganda, Tanzania, and Zambia. See Online Appendix Table A2 for details about the census years.
↵18. In principle, one could consider the behavior of white Zimbabweans as an alternative placebo test. However, the white samples are very small (N < 250), and it is not clear that their response to the reform in Zimbabwe was necessarily unique. Thus, we do not use them for our placebo tests. We thank a referee for this insight.
↵19. Health campaigns, including family planning, seem to have followed similar trajectories. Boohene and Dow (1987) find that health programs had no impact until the second half of the 1980s, and even then, it was very small and not related to the age groups in our analysis.
↵20. The variations we consider include no clustering, clustering by parent’s age in 1980 (that is, the parent’s year of birth), clustering at parent province of birth, clustering at parent district of birth, two-way clustering on age and province, and wild–cluster bootstrap–t procedures with age and province clusters, respectively. The bootstrap procedure with 16 age clusters also incorporates finite–cluster corrections. See Online Appendix Table A4 for more details.
↵21. The reduced-form estimate is 0.061 for mothers (standard error is 0.018) and 0.062 for fathers (standard error is 0.011).
↵22. These countries provide a better comparison to Zimbabwe than Scandinavia (Holmlund, Lindahl, and Plug 2011) or Latin America (Behrman, Gaviria, and Székely 2001), where the level of economic development and quality of institutions differ markedly from those in African countries.
↵23. For example, Maccini and Yang (2009) find that Indonesian women born in years of plentiful rain were taller, completed more grades of schooling, and lived in better homes. Alderman, Hoddinott, and Kinsey (2006) report that early-life exposure to drought in Zimbabwe was associated with delay in school enrollment, lower schooling attainment, and poorer health, while Hoddinott (2006) finds negative impacts on household assets.
↵24. The rainfall data come from 38 stations across Zimbabwe. We standardize annual rainfall using the mean and standard deviation of precipitation over the period 1959–2001.
↵25. This conclusion is reinforced by additional tests in Online Appendix Table A5. Here, we start by comparing those aged 14–15 with those aged 16–17 in 1980 and then increase the age interval smoothly. We find that the 2SLS estimates grow larger as we narrow the sample around the discontinuity. Because the number of clusters shrink greatly as we narrow the sample, the estimates from the smallest age interval warrant caution. In light of this, our baseline estimate is a conservative estimate of the intergenerational spillover.
↵26. Because Zimbabwe’s census, like most questionnaires, collects main activity data only on persons ten years or older, our child labor results were obtained using those parent–child pairs where children were in the ages of 10–15 years.
↵27. Edmonds (2007) argues that the relevant margin for child labor is the number of hours worked. However, this is not measured in most population censuses, including Zimbabwe’s.
↵28. The difference of one-half year in age at first birth suggests that we can rule out the possibility that children of educated parents are overrepresented in the younger ages. It is unlikely that our results have a composition bias.
↵29. The lack of labor force surveys and longitudinal household data in Zimbabwe is also discussed in Montenegro and Patrinos (2014). Among the 819 surveys from 139 countries used in their paper, there are no surveys from Zimbabwe to estimate Mincerian returns to schooling.
- Received August 2016.
- Accepted May 2018.
