Marital Shocks and Women's Welfare in Africa

Marital shocks are exceedingly common for women in Sub-Saharan Africa. The paper investigates whether women who have suffered a marital rupture experience lower welfare levels relative to married women in their first union. Conditional means for women's nutritional status are compared by marital status across 20 countries. Overall, the results indicate significantly lower nutritional status for Africa's widows and divorcees between ages 15 and 49. With some exceptions, this is found to be the case with country and household fixed effects and controls for HIV status. However, looking at country-specific associations underlines that disadvantage is by no means universal.


Introduction
The marital experience is remarkably gendered in Sub-Saharan Africa. Figure 1 gives the proportions of men and women in each marital status group by age. 2 Although marriage is practically universal, African men spend far more of their lives married than do African women.
From their early thirties to their early eighties, over 80% of all men are married. In contrast, the share of married women peaks at around 30-and lasts a much shorter period, dropping below 80% just after age 40. The drop is then precipitous and mirrored by a steady rise in the share of widows.
By age 65, there are as many widows as there are married women; by age 80, 80% of women are living in widowhood. All along the age distribution, the share of divorcees is also higher among women than men.
These patterns reflect several factors including far higher male remarriage rates following widowhood or divorce, large spousal age gaps, higher average life expectancy of women, the practice of polygamy, and the ravages of HIV. As a result, one in ten African women 15 and older are widows, and 6% are divorcees. And these women are likely to head their own households -72% of widows are heads of the family. In Africa, many are also quite young. 3 Across the region, 3 percent of all women aged 15-49 are widows at any one point in time, and accounting for those who have remarried, 5% are ever-widowed.
This paper asks whether African women who have been widowed or divorced are poorer than otherwise similar women who are in their first marriage. When today's rich countries were as poor as Africa today, widows were often identified as among the poorest and most vulnerable individuals, which led to the introduction of pension schemes and widow benefits in the late 19 th and early 20 th centuries in many of those countries (Frohman 2008;Hopkins 2016;Thane 2000). In more recent times, divorced women too have been singled out in the academic literature for the consumption and income losses they often suffer as a result of their marriage's collapse (Amato 2000). Though the evidence is scarce, one might expect a similar situation for women who experience a marital shock in Africa, where most countries are characterized by underdeveloped safety net and insurance mechanisms, and high levels of gender inequality in rights, human development and access to assets and livelihoods. For African women, such a shock may well entail a loss of economic means and support that are acquired through and conditional on marriage-including access to productive assets 2 See notes to Figure 1 for data sources. 3 Throughout we shorten Sub-Saharan Africa to Africa.
3 such as agricultural land. It may be conjectured that Africa today is ready for similar policies targeting widows to those introduced long ago in today's rich word.
Against this view, family, community and local institutions may provide assistance to women following a marital dissolution. A literature in economics has pointed to the existence of risk-sharing arrangements in poor societies which, although they appear to be far from perfect, might be expected to offer some protection (Fafchamps 1992;Jalan and Ravallion 1999). As will be discussed below, in some African countries, polygamy and the continuing practice of the levirate (defined below) appear to serve this purpose to some degree. The widespread tradition of child fostering in Africa has also been identified as a tool for risk-sharing (Akresh 2005). A large literature has emphasized the role of the moral economy in African peasant economies, comprising informal solidarity networks, risk-sharing institutions and mutual insurance systems. Policy makers in Africa may be reasonably confident that the extended family and village support systems will assist widows and divorcees and their offspring, and that such systems work well in most cases.
Yet there does not seem to be much support for that view in the sociological, anthropological and human rights literatures, which have pointed to the plight of widows. 4  Tragic anecdotes abound and undoubtedly occur, but there are no statistics on how common they are. Based on qualitative evidence and various sociological, anthropological and demographic writings, one might expect African women who have suffered a marriage shock to be highly disadvantaged relative to women who have not. Yet, here too there exists little quantitative analysis or evidence representative at the population level to indicate how widespread and generalizable this is within and across countries.
The consequences of marital ruptures for women in Africa have rarely been studied. Instead, the focus has been on female-headed households, whose heads are often abandoned, divorced or widowed. Among them, widow-headed households are frequently found to be particularly impoverished. Africa's widows have also figured in discussions of old-age poverty and more recently, in the context of gendered asset inheritance and the consequences of the HIV/AIDS 4 epidemic. 5 Although they are likely to be more heterogeneous across women, the welfare consequences of divorce also often appear to be negative and are likewise inadequately understood.
There has been interest from demographers (Locoh and Thiriat 1995;Renniers 1998;Clark and Brauner-Otto 2015) and a beginning of one among economists (Lambert et al. 2017;Cherchye et al. 2016). But, all in all, surprisingly little is known about the well-being of Africa's widows and divorcees.
The paper asks whether Africa's widows and divorcees are relatively disadvantaged, as has historically been the case for them in western economies. As noted, in the history of social policy, widows have been a recurrent target group and widows' pensions were important in reducing poverty. The issue was never about the causal effect of widowhood on welfare but rather the correlation with poverty. Widowhood was taken to be an indicator. However, this has received little systematic study for Africa. And "Africa" comprises 48 countries, with different historical and cultural setups. What is true in one may not hold elsewhere.
The paper investigates whether women who have suffered a marital breakdown may also suffer lower welfare levels relative to married women in their first union. Associations are described in a consistent and systematic way across multiple African countries. The overall objective is to better understand the consequences of the loss of a spouse and what role policy may have in protecting women who have suffered a marital dissolution. Through changes in inheritance laws and their enforcement, cash transfer schemes, and preferential access to housing, training, employment and schooling for their children, social policies can potentially help compensate for misfortune stemming from marital shocks, if the evidence suggests that widows and/or divorcees are in fact poor.
Examining these issues is not straightforward even when causal attribution is not the aim.
Poverty and vulnerability are mostly measured with the household as the basic unit of observation as data on individual consumption for specific household members are not available. Thus, potentially disadvantaged individuals within the household -such as current or remarried widows and divorcees and their children -remain invisible in standard data sources. We can only ask whether women who have had a marital breakup are disadvantaged with respect to other welfare proxies. We use Africa's Demographic and Health Surveys (DHS) and indicators of nutritional statusgenerally considered an important dimension of individual welfare -to compare conditional means 5 for ever-widowed and ever-divorced women relative to those of once married women across 20 countries. We pool all the data, control for a vast number of individual and household characteristics, and alternatively, country and household fixed effects. The regressions are also estimated with marital status and country interactions to examine country-specific associations.
Our approach is subject to caveats. Nutritional outcome indicators are collected only for those aged 15 through 49, which may well miss the most vulnerable and disadvantaged women. In addition, small overall sample sizes in some cases imply relatively few observations to work with at the country level. We are also limited in the number of usable surveys. DHSs have not consistently collected marital histories. In most surveys, remarried widows and divorcees cannot be distinguished from married-once women. We restrict the analysis to the 21 DHSs fielded between 2004 and 2013 which collected this information. Another difficulty concerns the fact that in regions subject to the AIDS epidemic it may be hard to identify associations with widowhood and divorce from those of being sick and undernourished due to AIDS. To address this possibility, we rerun our estimations on the sample of countries for which HIV tests are reported, controlling for HIV status.
A final caveat relates to the possibility of selective mortality among widows and divorcees reflecting their treatment by society (Anderson and Ray 2016). Unfortunately, the paper is unable to address this issue. However, the existence of selective mortality itself constitutes evidence consistent with our findings that do not account for it.
Overall, the results indicate significantly lower nutritional status for Africa's widows and divorcees. This is generally found to be the case with both country and household fixed effects (though with some exceptions). Controls for HIV status reduce the magnitude of the estimates but statistically significant effects remain. However, when we study the country-specific associations we find that the disadvantage of a marital rupture is not universal. Significant disadvantage is found only for certain countries and for these, tends to endure under all specifications.
We begin with a brief review of the African context and literature as relevant to the potential vulnerability of African women to marital rupture. Section 3 discusses data issues and presents descriptive statistics. Section 4 describes our estimation methods while Section 5 presents the results testing for significant differences in women's welfare associated with marital status. Section 6 concludes. 6

Context and literature
African women tend to have significantly inferior human capital endowments to men which imperils their access to employment and public services. Despite recent progress in school attainment, gender gaps persist in school entry and remain considerable in attainment among poorer, older and rural women (Grant and Behrman 2010;Beegle et al. 2016). Fertility rates remain the highest of any region in the world at an average of close to 5 children per woman (United Nations 2015). Repeated child bearing and unreliable access to pre-natal and maternal care amplify the health risks faced by women.
Despite the wide variety of cultural groups and traditions, customs related to women and the rights afforded them in the case of marriage rupture are derived from customary laws that share basic similarities across African societies and patterns of kinship organization (Ndulo 2011). Across the continent, legal protection privileges men. This is particularly so with respect to the laws governing unions and their dissolution, child custody arrangements, property rights, and inheritance. Although constitutions, laws and international conventions have been adopted that forbid discrimination on the basis of gender, there continues to be a considerable chasm with actual practice. Civic law has been largely ineffective in displacing customary law which often denies women's rights. Two pillars of family law ─ inheritance and marriage ─ are still overwhelmingly controlled by customary law (Ndulo 2011, Richardson 2004. As the basis for production and women's avenue to social and economic rights, marriage and a surviving spouse remain crucial to a woman's access to resources and productive assets (Gray and Kevane 1999a, b;Fafchamps and Quisumbing 2005).
Traditional Islamic law as typically practiced in Africa dictates that daughters inherit half of what sons inherit and husbands are the sole owners of family property. Widows receive one-eighth of the inheritance, to be shared among any co-wives. Customary law also excludes women from property ownership and inheritance in much of the rest of Africa. Women's access to property and land use rights is obtained through marriage and contingent on marital status. In the case of divorce/separation or widowhood, the rights are lost. 6 Women who initiate divorce may need to return bride wealth, and are likely to get none of the jointly owned belongings. In all marital ruptures, women run the risk of losing custody of children.
Women's access and control over resources is thus limited. In West African cultures, men and their wives do not share incomes, keep separate budgets, and have different spending 7 responsibilities. Income from a personal plot acquired through marriage or petty trade is expected to provide wives and their dependents with any additional food and non-food needs. High fertility, responsibilities for child care, domestic tasks, and work on the household's communal land may result in little time to allocate to their own income earning activities. Different cultural traditions exist in East Africa, where (as a generalization) wives have historically had few possessions and no independent sources of income (Lesthaeghe 1989). Everywhere, women's limited assets and lower capacity for mobilizing resources are compounded by constraints linked to social norms, difficulties in accessing credit and public and private services (Gaddis et al. 2017). Access to own income sources as well as bargaining power within the extended household becomes crucial to welfare outcomes and to whether the loss of a husband turns out to be catastrophic.
Women thus face many restrictions that limit their possibilities for accumulating capital or earning income and have far lower access to various individual level coping mechanisms in times of downside shocks. They may be disproportionately vulnerable to shocks. There is evidence supporting this view in the results of Dercon and Krishnan (2000) for Ethiopia.
A frequent shock is union dissolution. Across Africa, early marriage to much older men is common. Eight of the world's 10 countries with the highest percentage of girls married before they are 18 were in SSA in 2010 (Walker 2012); and close to half (4 out of 10) of the women are married before they turn 18 (UNICEF 2016). Although many countries have set 15 as the legal minimum marriage age for girls and some have even managed to raise it further, special dispensations and customary law often intervene. Large age gaps between spouses are typical, as in Mali where they are around 12 and 14 years on average in urban and rural areas respectively, while in Senegal, a richer country than Mali, they are 11 and 13. 7 As a result, far more women than men experience the death of a spouse at some point in their lives; by the same token, significantly more elderly women are widows than men are widowers. Many also divorce and unlike men, stay divorced. As seen in 8 divorcees and widows to remarry quickly. A Senegal survey which collected such information reveals that of the 59 percent of divorcees who remarried, 47% did so in a polygamous relationship (Lambert et al. 2017). Among the 26% of widows who remarried, 72% did so to a polygamous husband and half of them in a leviratic union, for which 83% joined a polygamous union. Although they usually join as lesser ranked wives, such marriages offer women a status, and some protection and help with basic needs. By the same token, when a polygamous man dies, up to 4 new women become widows. Informal polygamy, without any legal basis in either customary or statutory law, is also rising across many African countries (Coast et al. 2011). In other countries, remarriage is forbidden or frowned upon.
Some form of leviratic union continues to be common in Africa. 8 By this custom, a widow is married to a relative of the deceased husband, thereby ensuring that her current and future offspring remain with the lineage. The levirate is made possible by polygamy and plays a similar role to it, providing support to widows and their children by ensuring that a male provider assumes responsibility for them and can make it easier for mother and children to stay together. In many kinship groups, a man's offspring are seen to belong to his lineage. Although a widow can refuse the levirate, she is not usually free to take with her the children she had with the deceased husband.
Furthermore, a new husband from outside the lineage may not accept her children with another man.
The AIDS epidemic has not only contributed to the prevalence of widowhood but has also accentuated the vulnerability of widows who are typically assumed to be infected such that, as a result, the levirate practice is thought to be dying down in affected areas (Kudo 2017;Tenkorang 2014). The epidemic has likewise been associated with a higher risk of divorce by individuals using it as a protective measure (Reniers 2008). AIDS widows may be shunned and dispossessed by inlaws, yet left with debts incurred during the deceased's illness (Ntozi 1997). In the absence of relatives willing to take them in, or the incomplete and limited protection afforded by legally sanctioned polygamy and the levirate, widows and orphans can be left homeless and destitute after the death of their husband or father. Using DHSs, Peterman (2012) documents that across African countries, only a small share of widows receives any assets following a husband's death. In places where polygamy is illegal, de-facto wives are likely to be even more vulnerable to destitution when a partner dies since any enforced rights associated with marriage apply only to the legal wife. Where 10 plowing and threshing in northern highland areas of Ethiopia, worsening their economic survival prospects (Loomba Foundation 2015).
Divorce on the other hand is more ambiguous in its welfare impacts; it could be an unwanted precursor to economic hardship or it could be desired. Although it remains easier for men to instigate divorce, in many countries the husband or wife can do so. A number of papers for various African countries have argued that young and educated women use divorce strategically as a way to improve their economic status (Locoh et al. 1995;Reniers 2003;Cherchye et al. 2017). Against that, there are also papers contending that those left by their husbands and their offspring can be dispossessed and impoverished (Clark and Hamplova 2013). Thus, there is likely to be both positive and negative selection into divorce.
As noted, union ruptures in Africa have not featured much in analytic work. There are a few important works on widows in India, who are found to be particularly discriminated against and disadvantaged (Chen 2000, Drèze and Srinivasan 1997, Jensen 2005 for West African widows (see Table 2). Table 1 also gives percentages of current widows by country based on the subsample of individually interviewed women aged 15-49. The share of current widows in this age group is much smaller, at about 3% of the entire African population of adult women with a mean age of approximately 40. As seen in Figure 1, these attributes are explained by an increasing incidence of widowhood with a simultaneously declining occurrence of remarriage as women age.
Keeping the focus on women aged 15 to 49, Table 2 shows the prevalence of ever-widowed women in the age bracket-a group that includes current widows and currently married, but previously widowed, women. This is a significantly larger group than that of current widows at 5% versus 3% of all women 15 to 49 Africa wide, suggesting that remarriage-although mostly hidden in surveys-is not uncommon among widows. Remarriage is higher in rural than in urban Africa: 45 (30) percent of ever-widowed women remarried in rural (urban) areas in SSA as a whole. Regional variation also emerges, with widows in Southern Africa least likely to remarry, and those in West and Central Africa most likely to do so. The frequency of remarriage is typically higher in Muslim populations. While data on previous marital status are available only for women in the 15 to 49 age band, we expect the number of ever-widowed women in the entire population to be much larger than indicated in Table 1.
Divorce is common in many SSA countries. Similarly to Table 1, Table 3 provides percentages of all women 15 years and older, and of women aged 15 to 49, who are current divorcees by country. In total, 5.5% of all adult women and 5.7% of those in the 15-49 age bracket are divorcees. In all regions and for both age groups, there are more divorcees living in urban than in rural areas although the differences are small. Prevalence in West Africa is less than half that found elsewhere in Africa. In both age brackets, the currently divorced are on average considerably younger than the currently widowed. Table 4 focuses on ever-divorced women aged 15-49 and provides detail on the breakdown between the remarried and the currently divorced. The total numbers are considerably larger than found in Table 3. This reflects the fact that remarriage is widespread and that beyond a certain age, 14 women no longer tend to divorce. Africa-wide, 14% of women in the age group are ever-divorced.
The shares are higher in rural areas in all regions, where remarriage also tends to be lower. As a generalization, a larger share of women under 50 get remarried than stay divorced across African countries. However, this pattern is reversed in Southern Africa. It should be noted that with respect both to widowhood and divorce, the numbers are underestimates given that multiple breakups may have happened. In Senegal, for example, Lambert et al. (2017) document that over 7% of evermarried women have had more than one marital dissolution.
Individual level welfare indicators: The analysis focuses on three nutrition-based indicators of individual welfare. The first is the body mass index (BMI), defined as a woman's weight (in kilograms) divided by her height (in meters) squared. DHSs exclude values of BMI smaller than 12 and greater than 60 on the grounds that these are almost certainly measurement errors. BMI is often used as a measure of health: low BMI may reflect heightened stress and undernourishment. 15 It has likewise been used to proxy for individual well-being (Steckel 1995;Brown et al. 2017). As noted by Sahn and Younger (2009), in addition to being measured for individuals, BMI has a number of important advantages as a measure of well-being: it captures consumption relative to needs (and so is better than caloric intake, for example), reflects command over food and non-food resources that affect health status, and is relatively easy to measure well. It is also found to be highly correlated with welfare measures based on consumption and income. One disadvantage is that beyond a certain threshold, higher BMI is associated with obesity and indicates an unhealthy state. However, obesity remains relatively rare in rural Africa and, significantly, evidence for South Africa shows that the relationship between BMI and more standard measures of economic resources is non-decreasing over the entire wealth/income range (Wittenberg 2011). Beegle et al. (2016) provide supportive Africa-wide evidence for this finding.
In addition to the level of BMI, we also examine the share of women whose BMI is below the underweight cut-off benchmark of 18. This transformation provides a continuous variable that places more weight on values that are further below the BMI cutoff. We then take the mean of * across all i, which we dub the Watts measure of underweight, following its original use by Watts (1968) in measuring income poverty. 16 This index has a number of desirable properties as a poverty measure, as shown by Zheng (1993). It is theoretically possible for mean log BMI to be lower for women of a given marital status, while underweight (say) is no different and the Watts measure of underweight is lower for them. Table 5 presents sample mean values and standard deviations for the welfare indicators aggregated across countries by marital status and urban or rural residence. Unconditional average differences in BMI favor current widows, with previously widowed, married once, and everdivorced women not too far behind. Patterns are somewhat different with respect to underweight where the lowest average rates are found for married-once women. Never married women, however, have by far the worst nutrition outcomes. Age plays a role here. At an average age of 19, the never married are also much younger than the other women and include adolescents undergoing growth spurts and menarche. Given our focus on marital shocks, it could make sense to exclude never married women from the analysis. However, the disadvantages of doing so are twofold: the loss of a substantial share of observations and the risk of creating sample selection bias. For these reasons, we keep never married women in the analysis, although we will not focus on them in the discussion.
Covariates including household wealth: The estimations below control for observed individual and household level characteristics that could affect nutritional status. Individual level attributes include age (represented by a full set of age dummies for maximum flexibility), years of education, whether pregnant, marital status dummies (with married once omitted), and relation to the head: whether she herself is head; the spouse of the head; a parent, child or sibling of the head; or no relation; with the left-out option being 'other relation'. Although there may be endogeneity concerns with the latter variables, leaving them out raises omitted variable concerns. 17 At the household level, covariates include log household size, the composition of the household by age and gender: shares of members aged 0-6, 7-15, and 65 and over (with the 16-64 age group omitted) all by gender; attributes of the head including gender, age, age squared, whether Muslim and years of education.
Mean statistics by marital status are given in Table 6.
As a proxy for household wealth, the DHS Wealth Index is also entered linearly and in its squared form. Constructed separately for each country by the DHS using factor analysis on a 16 household's assets and amenities, the index is then rescaled to be centered on zero with a standard deviation of one. It provides a within-country relative measure and is not comparable across countries. However, as explained in Section 4 we believe that under the regression structure used, where country fixed effects will pick up the country-specific wealth effect, it can be included as a covariate.
Abstracting from issues of intra-household inequality, a household's wealth can be expected to have bearing on its members' nutritional status. To investigate whether women who have experienced a marital dissolution live in wealth poorer households on average, we run OLS regressions by country of a woman's household wealth index on a constant term and a set of marital status dummies excluding married-once women on the sample of ever-married women. Figure 2 plots the estimated country-specific coefficients for each marital status with 90% confidence intervals, interpretable as differences from that for married-once women. 18 Across countries, with just a few exceptions in West Africa, and more so in urban than in rural areas, ever-widowed women are more likely to live in wealth poor households than married-once women. The estimated coefficients are negative for a majority of countries and statistically significant in 9 countries (urban and rural areas) for widows; and in the urban areas of 17 countries and rural areas of 11 countries for ex-widows. Similar patterns are apparent for urban ever-divorced women, although the picture is more mixed for those residing in rural areas. To more concisely summarize these findings, Table 7 averages the coefficients across countries by re-estimating the regression on the pooled data. Never married women live in significantly richer households than married-once women, while as we have seen in Figure 2, urban ever-widowed or ever-divorced women reside in significantly wealth poorer households. We make no causal interpretation here. The wealth of one's household may reflect prerupture living standards as well as trajectories since. Table 6 also shows that, as expected, never married women have the most years of schooling on average, followed by widows, divorcees and married-once women. 19 From this point of view, exwidows and ex-divorcees are the most disadvantaged on average. These mean values are formed over aggregated data for a heterogeneous group of countries and without conditioning on key covariates such as age. 18 A similar approach is taken by Case et al. (2004) to compare the living arrangements of orphans and non-orphans. We use OLS, weighted data and correct the standard errors for cluster level heteroskedasticity. 19 We use 'married-once' or 'married women' as a shorthand for currently married women in their first union. Table 8 presents percentages of all tested women aged 15-49 who are HIV positive by marital status for 14 out of the 20 countries. 20 As testing is voluntary, not all sampled women were tested in the surveys that do report HIV incidence. Table 8 shows the shares that were tested in each country. Analysis of non-response conducted for most DHS surveys with HIV testing shows minimal bias. 21 However, this missing information results in reduced sample sizes and more imprecise estimates.

HIV/AIDS:
Prevalence rates vary substantially across countries but, as might be expected relative to married-once and never-married women, they are everywhere considerably higher for women who have had a marriage dissolution. Among them, the highest prevalence is generally evidenced for widows, followed by ex-widows, divorcees and lastly, ex-divorcees. It is clearly important to take account of HIV status in our estimations. Women who have suffered a marital dissolution and are HIV positive are doubly disadvantaged. Having HIV may have caused their widowhood or not. We cannot know. But having this often debilitating and life-threatening disease together with the stigma and ostracism that accompany it, can be expected to lead to even worse outcomes for women who have known a dissolution.

Methods
Across 20 African countries, we investigate the association between three nutrition-based welfare indicators and marital status for ever-married women aged 15 through 49. Our aim is to test the hypothesis that women who have suffered a marital breakdown also suffer disadvantage and lower well-being as reflected in their nutritional status.
We test this hypothesis using various regression specifications, each with its pros and cons.
We begin with the following Africa-wide model in which the micro data are pooled across countries to enhance precision: , and c c w    . The difference between the household's and the country's wealth index ends up in the country fixed effect.
Differences in nutritional status between married-once women and those who have suffered a marital breakdown may reflect how households allocate resources to different women. This and other sources of latent heterogeneity across households can be addressed by including household fixed effects in the regression, so that one compares the welfare of women by marital status within the same household. This is only possible for the sub-sample of households that contain women of different marital statuses. So we run a household fixed effects model only on the sample of women who live in households where there are at least two women in the 15 to 49 age group. This reduces sample size but far less than if we restricted the sample to households with a married woman as well as at least one other of a different marital status. Thus, a second specification of the pooled model replaces the country fixed effects with household fixed effects jc  as follows: In this formulation, the household effects can be said to control for all differences in the dependent variables due to both observed and unobserved household characteristics. (Of course, the wealth index can no longer be included as a separate regressor since its effect is already included in jc Although the sample is restricted relative to (1), the specification in (2) provides a more powerful test in that (unlike (1)) it is robust to any latent household characteristics. Restricting the sample in this way may result in sample selection bias to the extent that the households with at least two adult women are not a random sub-sample. We will test and correct for this using a control 22 Recall that there are two surveys for Nigeria. We include a dummy variable for the later year of the two surveys. 19 function given by a cubic function of logit estimates of the predicted probabilities of being in the sub-sample using the same set of covariates.
The pooled model assumes constant parameters across countries. To examine whether marital status parameters vary by country, we take each marital status dummy one at a time, and replace it by its interaction with a complete set of country dummies (and no leave out country). This enables us to capture differences between women of that marital status and married-once women within each country since the regression also controls for other marital statuses. By stratifying one parameter by country and imposing constancy of the other parameters, this approach ensures more efficiency and estimation precision than if we ran separate country-specific models. As before, we estimate this model version separately with country and household fixed effects. With country fixed effects the model is as follows: where c  are the marital status-specific country effects, and the reference is a married-once woman in the same country. All other things equal, the difference between the BMI of a widow in country c and that of a married woman in the same country is c  when the interaction is for widows.
Finally, we test the degree to which being HIV positive affects our results. To do this we rerun the regressions with dummy variables for 'tested positive' and 'HIV test missing' on the sample of countries for which HIV test results are available (Table 8). 23 As noted, six surveys for which women's HIV test results are not available are dropped from our sample. In addition, not all sampled women were tested in the surveys that do report HIV incidence. For consistent comparison of the results controlling for whether a woman has tested HIV positive, we also re-estimate the regressions without such controls on the smaller sample of countries that report HIV tests.
We use OLS for the estimations. 24 As noted, when estimating with household effects we limit estimation to the sample of households containing at least two women, as reflected in considerably smaller sample sizes. Standard errors are corrected for heteroskedasticity within clusters. 23 An alternative way to test the impact of HIV would be to estimate the regressions on the sample restricted to women who have tested HIV negative. We chose not to do so as this approach reduces the sample considerably. 24 To analyze the likelihood of being underweight, we also estimated a logit and a conditional (fixed-effects) logistic regression model (clogit) which enables introducing household fixed effects in a binary dependent variable. In the clogit case, there is the additional requirement that there be variation in the incidence of underweight in included households. Because this resulted is so many dropped observations, we settled on the linear probability model (LPM). We found that the estimated coefficients are generally smaller with the LPM but not qualitatively different. 20

Results
We begin with our Africa-wide model using pooled DHS data, first with country fixedeffects then (on a more limited sample) household fixed-effects. We then study some countryspecific aspects of our results, and the role of HIV status at the individual level.

Africa-wide results using the pooled model
Our estimated Africa-wide coefficients on marital status and women's other individual level attributes are presented in Tables 9 and 10 for equations (1) and (2), respectively. 25 Relative to the reference group of married-once women, current African widows and divorcees have worse nutritional status indicators on average, with lower levels of BMI, and higher rates of underweight and of the Watts UW measure. The differences are statistically significant, and the same is true across nutrition indicators when the sample is disaggregated by urban and rural areas. Larger estimated coefficients in the former possibly reflect rural-born women who have sought refuge and livelihood opportunities in urban areas post dissolution. The coefficient on ex-widows is also statistically significant for log BMI but the effect is less than half that estimated for current widows, while the parameter for ex-divorcees is not significantly different from zero.
A negative parameter on BMI implies lower nutritional well-being, while it is the opposite for both underweight measures. As one might expect given heterogeneity across countries, and despite highly significant coefficients for widows and divorcees across the board, average effects for Africa as a whole are not large. Across the pooled sample, and conditional on the covariates, being a widow aged 15-49 is correlated with 1.9 percent lower BMI on average, reducing a BMI of 18.5 to 18.2, and to 18.1 in urban areas; it is associated with an increase in the likelihood of being underweight of 1.8 percentage points on average, rising to 2.4 percentage points in urban areas. The Watts measure indicates that among those who are underweight, the mean (negative) percentage change in BMI associated with being a widow is 0.15%. (Recall that to make results easier to read, the Watts measure is multiplied by 100.) Relative to the reference married woman, being a divorcee reduces BMI by an average of 1.3 percent; raises the probability of underweight by 1.4 percentage points; and increases the proportionate change in underweight by 0.15%. As found for widows, all the estimates pertaining to divorcees are more pronounced in urban areas. 26 25 The full regressions are available in the Addendum. 26 Dropping the wealth index from the regressions made practically no difference to the coefficients of interest. Omitting the living arrangement covariates results in somewhat lower coefficients on widows and divorcees although these remain statistically significant. Both regressions are in the Addendum.

21
One way to judge the importance of the effects of marital status on nutritional status is to compare them with those estimated for schooling by looking at the ratio of the coefficients. 27 The overall gain from not becoming a widow and remaining married is equivalent to about 6 to 7 years of schooling depending on the nutritional indicator. Around 4 to 6 years of education would be needed to compensate for divorce. Here, the impacts of education may be attenuated by the fact that the regressions also control for household wealth. When the wealth index is omitted, fewer years of education are required to make up the difference at 4 to 5 years for widows and 3 to 4 for divorcees.
In Table 10, the results from equation (2) are reported. 28 Once household fixed effects are introduced, the significant and negative effects on log BMI of being a current widow or divorcee persist with somewhat larger estimated parameters. On average, relative to married-once women, African widows have 2.4% lower BMI, while for divorcees the hit is equal to 1.9% reduction. The impacts on the Watts measure of UW are not statistically significant. Widowhood is associated with an overall (statistically insignificant) 1.9 percentage points increase in underweight, and a (statistically significant) 3.2 percentage points increase in urban Africa. The parameter estimates on divorce are an overall 1.6 and a statistically significant 3 percentage points increase in urban areas relative to being once married. The years of schooling needed to compensate for widowhood or divorce are even higher in the specification with household fixed effects, ranging from 10 to 12 years. It should be noted that where people live is endogenous. The household fixed effect estimations compare women within specific households. If women chose to go live with richer relatives after a marriage dissolution, they may look nutritionally worse off than women in that household but may still be better off than if they had moved to a poorer relative's house. Against that, we know from Figure 2 and Table 7 that widows and divorcees have a higher probability of living in wealth-poorer households, particularly in urban areas.
The estimated parameters on other individual-level controls accord with expectations.
Pregnancy and years of education are positively associated with nutritional measures. The woman's link to the household's head is of clear importance. Relative to living in a household headed by a distant relative, women who are themselves head, his spouse or not related to its head have better nutritional outcomes all around.
27 Based on equation (1) the gain in school years needed to compensate for widowhood ( 0 is / . 28 As noted earlier, we also implemented a selection bias correction. The estimated coefficients of interest and their statistical significance levels are somewhat higher when the correction is applied but the changes are small. The results are available in the addendum.

Country specific marital-status effects
Tables 11 and 12 provide the country-specific parameters for widow disadvantage for the models controlling for country and household fixed effects, respectively, as represented by Eq. (3).
As noted, the country-specific parameters can be interpreted as the differences in nutritional indicators between widows and married-once women in each specific country. The corresponding parameters for divorcees are given in Tables 13 and 14.
The countries where we find that widows tend to have significant nutritional disadvantage are DRC, Ethiopia, Guinea, Lesotho, Malawi, Rwanda, Swaziland, Tanzania, Uganda, Zambia and Zimbabwe (Table 11). However, nutritional deprivation vis-à-vis married-once women is by no means universal across African countries. There is little sign in Table 11 of such a disadvantage for widows in Mali, Namibia, Niger, Nigeria, Senegal and Sierra Leone. Still, the estimated coefficients, although not always significant, indicate relatively worse outcomes for widows in most non-West African countries (where some correlations favor widows). In 12 countries, the estimated correlations suggest that widows are significantly worse off with respect to one or more nutritional indicators and at least one sector. In such cases, the associations are considerably more pronounced than the Africa-wide averages estimated on the pooled models for Eqs.1 and 2. Some of the strongest effects are found for Ethiopia, Rwanda and Zambia, where all estimates except for the Watts UW for urban widows (rural for Zambia) are statistically significant. For example, widowhood in Ethiopia is associated with BMI levels that are on average 7.5 percentage points lower than for married-once women and with a probability of being underweight that is 11.4 percentage points higher on average, rising to over 13 percentage points for urban widows. There are also strong significant effects on the Watts measure of UW.
Adding household fixed effects to this model results in some coefficients losing statistical significance, while those that maintain it (in 9 countries), increase in magnitude ( Turning to the estimated correlations between divorcees and nutritional outcomes in Table   13, we find similar results to those for widows with a few exceptions (Uganda exhibits no significant associations for divorcees, while Malawi and Namibia now do). Fewer of the initially statistically significant coefficients survive the inclusion of household fixed effects (Table 14).
We estimated the equivalent models with the dummies for ex-widowed and ex-divorced women replaced with their interaction with country dummies and found fewer significant effects (see Addendum). That said, it is also true that for both remarried widows and remarried divorcees there are large and significant effects for many of the same countries-Benin, DRC, Ethiopia, Malawi, Rwanda, Uganda, Zambia and Zimbabwe, with and without household fixed effects.
In summary, we find considerable overlap across specifications and the four possible marital dissolution statuses across countries: significant negative associations with nutrition repeatedly emerge for the same set of countries and never do so for the other set. One of the distinguishing attributes of the first set is that they are located in parts of Africa that have been much more affected by the AIDS epidemic. We next examine whether the widowhood and divorce associations are attributable to HIV status.

Allowing for HIV status
To test that our results are not simply the result of HIV infection, we rerun models (1) Tables 9 and 10 for the country and household fixed effects models respectively, but estimated on the significantly reduced sample and set of countries. 29 These are followed by the estimated coefficients accounting for HIV status.
Results for equations (1) and (2) on the reduced sample ignoring HIV do not appreciably differ from the previous ones (Tables 9 and 10). As before, widows and divorcees have significantly lower BMI and higher rates of undernutrition and Watts undernutrition measures on average. In many cases, the magnitudes of the coefficients rise. However, reflecting the considerably reduced sample, statistical significance is somewhat lower in some cases. 29 We focus attention on current widows and divorcees rather than the remarried among them for space reasons and because they appear not to be disadvantaged on average.

24
Our key findings are robust to controlling for HIV status. A further lesson is the relevance of HIV/AIDS to nutritional status. As expected, testing positive for HIV is strongly related to lower nutritional status and generally has a larger adverse effect than widowhood (in the country fixed effects model). There is not much effect of not having been tested. Yet controlling for HIV status and despite the smaller sample sizes, the finding that, on average, widows and divorcees have statistically significantly worse nutritional outcomes persists. This is particularly so for those residing in urban areas.
Relative to married-once women, a decline in BMI-of 1.1 percentage points on average and of 2.2 percentage points in urban areas (implying a reduction of a BMI level of 18.5 to 17.9) is associated with widowhood. The probability of being underweight for this same group is 2.3 percentage points higher and the proportionate change in underweight is increased by 0.14 percentage point (although the last is not statistically significant). Being a divorcee also reduces BMI relative to the reference, by an average of 1.4 percentage points rising to 1.8 percentage points for urban divorcees. It is associated with a probability of underweight that is higher by significant 1.  30 All in all, the conclusions are essentially as before.

Conclusions
The circumstances of women naturally vary across Africa, including across geographic areas, urban and rural residence, the legal context, income, education levels, ethnic, religious and cultural identity. Some countries are much more urban; some practice polygamy and encourage remarriage following spousal loss, while in others, marriage is strictly monogamous and remarriage may be frowned upon; some have been hard hit by AIDS and/or conflict. Yet, there are also some pronounced similarities across the continent: high levels of gender inequality in human development, access to assets and livelihoods and in rights (including inheritance and property rights) characterize all countries, together with underdeveloped or non-existing formal safety net and insurance mechanisms. Women's economic means and support-including access to productive assets, are typically acquired through, and conditional on, marriage.
One might thus expect African women to be vulnerable when faced with the shock of spousal loss. The neglect, economic marginalization, social stigma, property dispossession and demeaning cultural practices to which widows are submitted in various African countries have been popular topics in the sociological, anthropological and human rights literatures and a focus of various specialized NGOs. Less has been written about separation and divorce, although there are also suggestions of negative welfare consequences.
However, much of the evidence is qualitative, or based on individual case-studies, country or even village specific. With a few exceptions, economists have not studied the welfare consequences of marital ruptures for women in Africa and there have been no studies representative at the population level to indicate how widespread and generalizable the events and outcomes noted in the existing literature are within and across countries. 30 The Addendum presents the equivalent regression results for remarried widows and remarried divorcees. Once again, disadvantage is apparent for those who have experienced a marital shock in many of the same countries in which widows and divorcees have been found to be nutritionally worse off than married-once women.

26
Based on DHSs for 29 African countries, we find that one in ten women aged 15 and older are current widows, and 5% are divorced. These numbers obscure the fact that some women remarry after marital dissolution. Among the 15-49 age group for the 20 countries for which such detail is available, 5.3% are ever-widowed and 15% are ever-divorced. Across African countries, everwidowed women are found to live in wealth-poorer households on average, as are urban everdivorced women. A further dimension of disadvantage is revealed by disaggregating HIV prevalence by marital status among women for whom test results are available: the probability of being HIV positive is highly correlated with having had a marital rupture.
We have investigated the association between marital status and well-being as reflected in three different measures of nutritional status for women aged 15-49. Pooling data for 20 countries and conditioning on a wide array of individual and household characteristics, and either country or household fixed effects, we examined whether women who have suffered a marital breakdown are nutritionally deprived relative to married women in their first union. The same regressions are reestimated with marital status and country interactions to investigate how the relationships vary across countries. Finally, we tested the sensitivity of our results to HIV status.
A few analysis-related limitations should be flagged. For data reasons, our investigation has been limited to relatively young women, aged 15 to 49. We are unable to throw light on older widows and divorcees who may well face worse conditions. This age and other methodologyimposed restrictions also mean that we have worked with relatively small samples, especially in the specifications with household fixed effects. Although there are a fair number of widows, they are less prevalent than divorcees in the 15-49 age group, and exhibit a generally lower prevalence in countries where remarriage is expected for women of child-bearing age. We also examined remarried widows and divorcees but each individual group can be on the small side. Finally, our investigation has focused on correlations with nutritional status. We are unable to speak to other aspects of disadvantage. Furthermore, nutritional status may be a less appropriate or relevant measure of relative deprivation in countries where one expects intra-household inequalities to be more prominent with respect to non-food consumption (Lambert and de Vreyer 2017).
While acknowledging some exceptions, we find a reasonably general pattern whereby young widows and divorcees are significantly nutritionally disadvantaged in the Africa region as a whole.
This finding is generally robust to differences in how nutritional status is measured and the controls used. Our estimates suggest that the negative health impacts of widowhood (in terms of being malnourished) are on average equivalent to the health gap associated with 6-7 years less schooling 27 when controlling for country fixed effects and at least 10 with household fixed effects. Considering the mean level of schooling, being widowed basically removes any protective health effects associated with education for women. Similar results are found for the impact of becoming divorced.
Our findings are also robust to HIV positive status, which is a strong negative factor in its own right.
The effects of marital dissolution on nutritional status are about the same magnitude as that of being HIV positive when controlling for household fixed effects.
However, there is clearly variation across countries. Nutritional disadvantage is by no means universal for women in the 15-49 age group who have undergone a marital shock. Much more pronounced negative correlations than are found on average for Africa are apparent under all specifications for certain countries in Central and East and Southern Africa. Ethiopia, Rwanda and Zambia particularly stand out. Significant negative nutritional associations also emerge for many of these same countries for remarried widows and divorcees. Interestingly, the countries where our analysis suggests significant nutritional disadvantage for widows and divorcees often coincide with those that are prominent in the sparse literature.
These findings are unlikely to be just about the welfare of widows and divorced women. For example, the paper has said little about their children who are necessarily also adversely affected.
We have focused on what might be termed the first-round effects. The existence of such damaging first-round effects may well have far broader societal implications for children, girls, and gender inequalities more generally. As argued by Drèze (1990), "Combating the neglect of widows must be seen as an integral part of the broader struggle against gender inequalities." While Drèze is writing about India, the results of the paper suggest the point applies to Africa.    Note: Percentages are given. Sub-totals and totals are country population weighted. Total widowed refers to ever-widowed women. These are disaggregated into those who are currently widowed and those who remarried post widowhood and are hence currently married. Source: Authors' calculations based on 20 DHSs with marital status detail.     Note: Statistics are population weighted averages across countries of the estimated household wealth difference between married once women and those of other marital statuses, as summarized in Figure 2. For example, urban widows tend to live in households with wealth indices that are 2.2 lower than that for married women. In the two last columns, we control for region in addition to marital status. The standard deviation averaged across country-level estimations is given in parentheses. The wealth index is normalized to be centered on 0 with a standard deviation of 1. Source: Authors' calculations using DHSs.  (2) Rural (3) All (4) Urban (5) Rural (6) All (7) Urban (8) Rural ( (11) Rural (12) All (13) Urban (14) Rural (15) All (16) Urban (17) Rural (   Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. OLS is used for all dependent variables. Estimations also control for country FEs, a full set of age dummies and HH characteristics including rural/urban residence, demographic size and composition by gender, head's age and education, the Wealth Index and Wealth Index squared, dummy for Nigeria survey of 2013. Source: Authors' computations using DHS data.  Standard errors clustered at the village level are given in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. OLS is used for all dependent variables. Estimations also control for country FEs, a full set of age dummies and HH characteristics including rural/urban residence, demographic size and composition by gender, head's age and education, the Wealth Index and Wealth Index squared, dummy for Nigeria survey of 2013. Source: Authors' computations using DHS data.