Demography, Vol. 31, No.2, May 1994
The Demographic Transition in Southern Africa:Reviewing the Evidence fromBotswana and Zimbabwe*
Duncan ThomasRAND1700 Main Street, Santa Monica, CA 90407andDepartment of Economics, UCLA405 Hilgard Avenue, Los Angeles, CA 90024
Ityai MuvandiCentre for African Family StudiesBox 60054Nairobi, Kenya
Part, but not all, of the observed decline in the number of children ever born reportedin the 1984 CPS and the 1988 DHS in Botswana and Zimbabwe can be attributed todifferences in sample composition: women in the 1988 survey appear to be bettereducated than women of the same cohort in the 1984 survey. Blanc and Rutstein arguethat differences in education levels in the pairs of surveys are not significant.However, weighted Kolmogorov-Smimov statistics, a comparison of average years ofschooling, and the proportions of women who complete primary school or attendsecondary school all indicate that the differences are, in fact, significant. This is truein both Botswana and Zimbabwe. Blanc and Rutstein also claim that these differencesdo not account for any of the observed decline in fertility between the surveys ofwomen age 15 to 49. Their methodology follows cohorts of women rather thanage-groups and thus cannot possibly address this issue. Furthermore, to interpret theirresults, response error and respondent education must be uncorrelated: this is a keyassumption which is violated by the data. We stand by our conclusions and argue forcaution when aggregate statistics from the CPS and the DHS are used to makeprojections about the course of fertility and population growth in Botswana andZimbabwe.
According to aggregate statistics from the Contraceptive Prevalence Survey (CPS) andthe Demographic and Health Survey (DHS), the number of children ever born to women age15 to 49 in Zimbabwe declined from 3.4 to 2.95 between 1984 and 1988. An even largerdecline is reported in Botswana for the same four-year period. Examination of microdata
* Discussions with Jim Smith have been very helpful. We are also grateful for the comments of the editors andof Janet Currie, Elizabeth Frankenberg, Linda Martin, Bill Mason, Anne Pebley and Dan Relies. Gary Bjorkprovided expert editorial advice.
Copyright 1994 Population Association of America
218 Demography, Vol. 31, No.2, May 1994
from these surveys reveals that in both Botswana and Zimbabwe, education levels of thesame cohort of women are significantly higher in the second survey. Differences in samplecomposition account for part, but not all, of the observed decline in the number of childrenever born. We argue, therefore, for caution when these aggregate statistics are used to makeprojections about the course of fertility and population growth in Botswana and Zimbabwe.
Blanc and Rutstein take issue with our conclusions on two grounds. First, they claim to"find compelling statistical evidence that the composition of the two samples with respect toeducation is the same." Second, they claim that our methodology for assessing the impactof differences in education levels on fertility is flawed.
We disagree with both claims. First, their statistical test for differences in educationlevels is not appropriate. They compute a Kolmogorov-Smirnov statistic but fail to take intoaccount the fact that sampling strategies in the CPS and the DRS are not the same. If thesample weights are incorporated in the calculation of the Kolmogorov-Smirnov teststatistics, then the distributions of education in the CPS and the DHS are significantlydifferent in both Botswana and Zimbabwe. It is, perhaps, more important to understandwhere the distributions differ. Women in the DHS not only have higher education levels onaverage but they are also more likely to have completed primary school or to have attendedsecondary school than exactly the same cohort of women in the CPS. All of thesedifferences are significant in both Botswana and Zimbabwe.
Second, Blanc and Rutstein's methodology for assessing the impact of thesedifferences on fertility decline does not address that question. They compare reportedfertility in the CPS with fertility as of 1984 in the DHS for exactly the same cohort ofwomen. If there are no differences in the samples in the two surveys, then these arereflections of precisely the same reality and the estimates should be identical. Key for ourpurposes is the fact that comparing cohort-specific fertility can say nothing about fertilitydecline among women age 15 to 49 between the two surveys, which is the substantive issueaddressed in our paper.
Instead, Blanc and Rutstein's methodology can, in principle, address the issue ofrespondent error in demographic recall data. They compare estimates of fertility as of 1984,based on the 1988 DHS, with those of the CPS. The DHS estimates are constructed usingretrospective birth histories and, thus, are likely to be contaminated by respondent error.Comparing the CPS with the DHS estimates may provide information about these errors.However, as an empirical matter, reporting errors are related to the respondent's education,and so it is not even possible to disentangle these errors from the effects of differences ineducation levels in the samples. Thus, it is unclear what interpretation should be given toBlanc and Rutstein's estimates of the impact of differences in sample composition in theCPS and the DHS. In contrast, we believe that our own methodology, which comparesage-specific fertility of women, is appropriate. We thus view the conclusions in our paper ascorrect.
DIFFERENCES IN MEASUREMENT OF EDUCATION
Let us first put the magnitudes of differences in education in perspective. Becauseeducation levels have increased over time in both Botswana and Zimbabwe, it is notappropriate to compare the education of women of the same age group. Instead, to isolatethe differences in education between the samples, it is important to compare the samecohorts of women.' Table 1 reports education levels of women age 25 to 44 in 1984 (andthus 29 to 48 in 1988). In the DHS, these women report about half a year more schoolingin Botswana than exactly the same cohort of women in the CPS. In Zimbabwe, the
Demographic Transition in Southern Africa
Table 1. Education Levels, Cohort of Women Age 25-44 in 1984,by Data Source (CPS and DHS)
CPS DHSCountry and Education Measure (1) (2)Bostswana
Average Number of Years of Education 3.53 4.02(0.06) (0.05)
Percentage of WomenNo education 37.3 36.5
(1.2) (1.1)Completed primary 27.1 30.2
school/more (1.1) (1.0)Attended secondary 10.5 15.3
school (0.8) (0.9)Completed more than 1.8 6.6
Form 3 (0.3) (0.6)Zimbabwe
Average Number of Years of Education 4.36 4.68(0.08) (0.08)
Percentage of womenNo education 22.9 21.2
(1.1) (1.0)Completed primary 30.0 34.0
school/more (1.2) (1.1)Attended secondary 10.5 13.5
school (0.8) (0.8)Completed more than 4.4 6.6
Form 2 (0.6) (0.6)Note: Standard errors in parentheses.
difference is about one-third of a year of schooling. Given the average level of schooling isabout four years, there can be little argument that these magnitudes are substantial.
Before we show that the differences are also statistically significant, it is as well toaddress the specific concerns raised by Blanc and Rutstein regarding measurement; each oftheir concerns is discussed briefly below. While it is clearly important to be cognizant ofdifferences in survey design when comparing two surveys, it is our judgment thatdifferences in the measurement of education in the CPS and in the DHS are both small andunlikely to affect our inferences. In addition, as a check on the robustness of theseinferences, we report a battery of different statistical tests below.
Questionnaire WordingBlanc and Rutstein point out that questions on educational attainment in the CPS and
the DHS are not identical. The CPS asks about the highest grade completed, whereas theDHS asks about the highest grade passed. On the basis of our knowledge of the educationsystem in Zimbabwe (with which we are both very familiar, having been schooled thereourselves) and Botswana, the distinction between completion and passing is largely one of
220 Demography, Vol. 31, No.2, May 1994
semantics." In Zimbabwe, the question in the DHS asks about formal school, whereas theCPS does not; since there is virtually no informal schooling there, this distinction isirrelevant. Thus, differences in the wording of these questions are likely to be of littleimport.
Changes in the Education SystemOur examination of women's educational attainment in the CPS and the DHS compares
women of exactly the same cohort in the two surveys. In the middle panel of Table 4, forexample, the education of women age 25 to 34 in 1984 (Column 1) is compared with that ofwomen age 29-38 in 1988 (in Column 2). These women faced exactly the same educationalsystem when they were in school. Thus we fail to understand why Blanc and Rutstein claimthat changes in the education system in Botswana in 1965 affect our inferences.
Real Increases in EducationAlthough the National Literacy Programme in Botswana may be associated with
increases in literacy rates among women with little or no education, it is not clear why, asBlanc and Rutstein argue, those women who had participated in the program would reportthat they had completed more years of schooling. Even if they did, this cannot account forthe magnitude of the differences between the education levels in the CPS and the DHS.3More importantly, as noted above and as discussed in detail in our footnote 17, the keydifferences in the education distribution between the CPS and the DHS are found amongwomen with secondary schooling. This fact cannot possibly be explained by the impact ofliteracy programs.
Data Processing RulesThe education variables that we use are drawn directly from the survey questions and
are reported in exactly the same format in the CPS and the DHS: each respondent reports thehighest grade she attained at the final level of schooling." An assumption does have to bemade about the years of schooling completed iIi previous levels, but because we have madeexactly the same assumptions in our processing of both surveys, this fact is unlikely toexplain discrepancies in education levels between the two. 5
TESTING FOR DIFFERENCES IN EDUCATION LEVELSIn view of the discussion above, there is no compelling reason to discard the
information contained in years of education reported in the survey and exploit onlyinformation on whether a woman completed preschool, primary school, or secondaryschool. However, our inferences about differences in education levels do not rely on thisdistinction.
Using a Kolmogorov-Smirnov test, Blanc and Rutstein use only information oneducation levels and compare the distributions in the CPS and the DHS. They acknowledgethat the distributions are different in Botswana for women age 35 to 44 but find that thedifference is not significant in Zimbabwe.
The sampling strategies in the Zimbabwe CPS and DHS are quite different. The DHS
Demographic Transition in Southern Africa 221
is a proportional probability sample and thus self-weighting. The CPS is not; it oversampledurban and better-educated women." The Kolmogorov-Smimov test statistics calculated byBlanc and Rutstein appear not to take account of the sample weights. But if the test statisticsare recalculated, incorporating the sample weights, then the hypothesis that the educationdistributions are the same is rejected:" the distribution in the DHS is to the right of that in theCPS for Botswana and Zimbabwe. Using data on years of education, the p-values for thistest are less than 0.01 in both countries; the differences are clearly significant. Wheninformation on years is thrown away, and only grouped levels of education are used, thep-values are 0.02 and 0.06 respectively. Higher p-values are to be expected because it iswell known that when data are grouped, the Kolmogorov-Smimov test statistic is overlyconservative (Noether 1976).8
It is also well known that the Kolmogorov-Smimov test lacks power when differencesin the distributions are in the tails (Conover 1980). This is precisely the case in these data.Rather than rely solely on a Kolmogorov-Smimov test, it seems to us to be important tocarefully examine each component of the distributions to determine where the differenceslie. For this reason, in Table 4 of our paper, we discuss both average years of schooling andproportions of women who complete particular levels.
A summary of these data is presented in Table 1 for the cohort of women age 25 to 44(in 1984) in the CPS (Column 1) and the DHS (Column 2) in Botswana and Zimbabwe.Differences between the surveys are reported in Column 3. Since these numbers are forexactly the same cohort of women, the differences should all be zero if there are nodifferences between the surveys.
The first row of each panel shows that, on average, the same cohort of women in theDHS reports more years of schooling than those in the CPS; this difference is significant inboth Botswana and Zimbabwe. In Botswana, for example, the t-statistic on the difference of0.5 years is 3.8.
Where, within the education distribution, are these differences concentrated?Examining the grouped data in the remaining rows of Table 1, slightly fewer women reportno schooling in the DHS than in the CPS. But this discrepancy is not significant. However,the probability that a woman reports herself as having completed primary school issignificantly greater in the DHS. Similarly, a significantly higher proportion of DHS womenreport having attended secondary school. (The t-statistics on the differences are 4.0 inBotswana and 2.5 in Zimbabwe.) This inference also carries through to the proportionsreporting completion of Form 3 in Botswana and Form 2 in Zimbabwe.
Blanc and Rutstein claim that there is "compelling statistical evidence" indicating thatthe education distributions in the CPS and the DHS are the same. Our interpretation of thestatistical evidence is the reverse. For both Botswana and Zimbabwe, reported education ofwomen in the two surveys is significantly different, the differences are large in magnitude,and are concentrated in the upper tail of the distribution, particularly among women withsecondary schooling. This last fact is important because, as demonstrated in Table 6 of ourpaper, there is a significant negative correlation between education and fertility only amongbetter-educated women, namely those who completed primary school or attended secondaryschool.
THE IMPACT OF EDUCATION DIFFERENCES ON ESTIMATESOF FERTILITY DECLINE
How much of the decline in aggregate fertility reported in the CPS and the DHS can beexplained by differences in education levels? Very little, according to Blanc and Rutstein.
222 Demography, Vol. 31, No.2, May 1994
However, their methodology follows cohorts of women...