Childlessness and Development in Sub-Saharan Africa: Is There Evidence for a U-shaped Pattern?

In high-income countries, women increasingly remain permanently childless. Little is known about the relationship between childlessness and socioeconomic development in non-Western societies and particularly sub-Saharan Africa. At lower levels of development, poverty-driven (i.e., involuntary) childlessness may decrease with increases in levels of development, while at higher levels of development opportunity-driven (i.e., voluntary and circumstantial) childlessness may rise with development. Thus, we expect a U-shaped relationship between childlessness and development overall. We examine this idea for sub-Saharan Africa. We further contribute by differentiating between female and male childlessness; and between involuntary, voluntary and circumstantial childlessness. Moreover, we construct new indicators of subnational historical development to assess both inter- and intra-country variation, and distinguish between three components (health, education and income) to investigate the drivers behind the hypothesized U-shaped relationship. Using 291 Demographic and Health Surveys between 1986 and 2018 from 38 countries and 384 regions, we find a U-shaped relationship between female childlessness and development, and a linear relationship for men. The U-shape for women results from negative associations of female involuntary childlessness with health and educational advancements, combined with positive correlations of voluntary and circumstantial childlessness with education and income improvements. While these positive associations are stronger among men than women, the negative relationships of involuntary childlessness with health and education observed for women are absent for men, resulting in an overall positive and linear relationship between development and childlessness among men. Our findings have implications for how we might expect childlessness rates to evolve with future levels of development. Supplementary Information The online version contains supplementary material available at 10.1007/s10680-022-09608-5.


List of
shows a scatter plot of sample sizes against childlessness percentages for unique subnational region-year combinations, and shows that there are relatively few (16% for women and 18% for men) zeros in our dependent variable and that these are not necessarily due to aggregation of data from small samples, as the zeros range between sample sizes of 1 and 265.

Supplementary Information I Computation of Orthogonal Polynomials
For a vector x of length n, we can generate monic orthogonal polynomials of degree d, P 0 (x), P 1 (x), . . . , P d (x), as follows (pp.343-4 Kennedy and Gentle, 1980):

Supplementary Information J Pooling Strategy for Hierarchical Regression Results after Multiple Imputation
We impute our data M = 36 and M = 29 times respectively for men and women, approximately equalling the percentage of incomplete observations as recommended by Bodner (2008); White et al. (2011). We run each of the aforementioned models on each of these M imputed data sets and we pool the results by taking the mean of the resulting M k-length vectors of coefficients and by computing the pooled variance aŝ whereŴ * is the average within-imputation variance; U * i is the within-imputation variance (i.e. the variance of π that would be observed if the imputed data would be the true observed data); andB * is the between-imputation variance.
Note thatV * ,Ŵ * andB * are vectors of length k. We further adjust the distribution ofπ * for small sample sizes as recommended by Barnard and Rubin (1999) by basing the inference for the coefficient of variable j = 1, . . . , k on a t-distribution rather than asymptotic normal distribution using degrees of freedom, where denotes the estimated observed-data degrees of freedom with ν comp = n−k −1 the complete-data degrees of freedom andγ M,j = M M +1B j * Vj * the estimated proportion of missing information for the j th variable. Figure S9: Correlation between independent variables in male data       Table S5 shows the goodness-of-fit measures for the subnational-level univariate basic models used to build the hierarchical model. The intercept model predicts childlessness with only a fixed intercept; the randomIntercept model predicts childlessness with an intercept allowed to vary across countries; the shihdRI model adds a first-order SHIHD orthogonal polynomial term to the randomIntercept model; the shihdRS model additionally allows the effect of a first-order SHIHD orthogonal polynomial term to vary across countries; and the shihdQ model is the same as shihdRS model but then with second-order orthogonal polynomials of the SHIHD term. As the random effects specifications for the models in Table S5 differ, we have compared model fits between two models A and B by executing a χ 2 -test with two degrees of freedom and as parameter the difference between the −2·log-likelihood of models A and B.

Supplementary Information N Model Comparison for Basic Models
As adding the second-order orthogonal polynomial term for SHIHD does not significantly improve the model for men, we fit our hierarchical models linearly for men.

Supplementary Information O Random Effects for Hierarchical Models
The random effects of the total male and female models (M5) in Table S6 show that while the baseline childlessness level hardly varies across countries, the effects of the linear (and, in the case of women, quadratic) SHIHD terms vary quite substantially across countries. Further, there is a positive correlation between the baseline childlessness level and the effect of the first-and, in the case of females, second-order orthogonal polynomial terms, suggesting that the effect of SHIHD is larger for countries where the baseline childlessness level is higher. Finally, there is a positive correlation between the effects of the linear and quadratic terms of SHIHD among women.  Note: + p < 0.1; * p < 0.05; * * p < 0.01; * * * p < 0.001