Women’s empowerment across the life cycle and generations: Evidence from Sub-Saharan Africa

Does female empowerment evolve over the life cycle, and has it changed across generations? We use data from the Demographic and Health Surveys covering a sample of about 191,000 adult women to evaluate the age, period and cohort effects regarding individual attitudes to marital violence. Pseudo-panel data are constructed from repeated cross-sections from five African countries in the 2000s. The estimates show that, over the life cycle, women tend to think that marital violence is less and less justifiable, and that younger cohorts are less likely than older cohorts to view marital violence as justifiable, even controlling for education.


Introduction
In recent years, female empowerment in developing countries has attracted growing interest from policy-makers and researchers. This empowerment has become a policy goal per se and in acknowledgment of the long-run benefits it confers on women and their families. Empowerment can be viewed as a dynamic process that increases women's ability to make choices about their lives and environment (Malhotra and Schuler, 2005), or as the ability to access health, education, earning opportunities, rights and political participation (Duflo, 2011). It is a multidimensional concept which is not directly observable. A recent body of empirical literature has however made some attempts at measurement. Various approaches have been considered, such as participation in household decision-making, gender-relative status or control over money use (Beegle et al. 2001, Anderson and Eswaran 2008, Garikipati 2008, Ashraf et al. 2010, Allendorf 2012). The empowerment of women and reduced gender inequality have been shown to increase the use of prenatal and delivery care services (Beegle et al. 2001), influence investment in children and other family members (Reggio 2011), increase female labor-force participation (Hendy and Sofer 2009), and improve both productivity and efficiency (Alkire et al. 2013).
What are the main driving forces behind empowerment? Does female empowerment evolve over the life cycle and has it changed across generations? It is commonly-believed that young women now are more emancipated than their older counterparts. Does the fact that younger women are now more empowered mean that empowerment falls over the life cycle (i.e. the aging effect is negative)? This is probably not the case, especially if we view empowerment as a stock that is likely accumulated over time. An alternative reading is that young women being more empowered means that women from recent generations are more empowered than their counterparts from older generations (i.e. a cohort effect), with one explanation being changes in educational attainment.
Are women empowering as they age or are women from recent generations starting with a greater empowerment stock than their older counterparts? The use of panel or pseudo panel data is required to disentangle the life-cycle and cohort effects (Deaton, 1985). We here propose the use of pseudopanel data to estimate the respective roles of age, period and cohort on female empowerment.
One important dimension of female empowerment is the refusal of domestic or marital violence.
Domestic violence against women has received growing attention worldwide, and is considered not only as a violation of fundamental rights but also as a burden leading to considerable health and demographic damage (e.g. Sobkoviak et al. 2012). Considerable political effort has been mobilized against this form of violence. Violence against women can be analyzed by focusing on whether a woman has been a victim of violence, or alternatively whether women consider marital violence to be acceptable within their households, or against women in society in general. The knowledge of women's attitudes to wife-beating is fundamental for seeing how women perceive the status of their gender in society. Hindin (2003) notes that women's attitudes can serve as a marker for the social acceptability of wife-beating. Using a perception measure of whether marital violence can be justified as it provides insights into women's views of their place in society. Most of the work on the effect of age on female empowerment or views on marital violence has relied on cross-section data, and has produced contrasting results. In terms of female empowerment, Anderson and Eswaran (2008) conclude to a zero or negative effect of age on indicators of female autonomy, while Mahmud et al. (2012) found a significant effect for two out of five empowerment indicators, but with effects in opposite directions. In terms of the acceptance of domestic violence, Hindin (2003) revealed a strong positive association between acceptance and age in data from the 1999 Zimbabwe Demographic and Health Survey. However, it is difficult to identify age effects cleanly in cross-section data. In cross-sections, any estimated age differences may reflect actual age, or equally the effect of the characteristics of different birth cohorts (see e.g. Glenn 2005).
We here use pseudo-panel data to analyze attitudes to marital violence over cohorts based on a representative sample of about 191,000 women aged between 15 and 49. We use three successive cross-sections of data from the Demographic and Health Surveys in five Sub-Saharan African countries (Ethiopia, Malawi, Rwanda, Uganda and Zimbabwe). In each survey wave, the sample is random and representative. In repeated cross-section data, households and individuals are not tracked over time, but a pseudo-panel data can be constructed so as to follow groups of people who belong to the same birth cohorts, defined at the country-level.
We consider a number of approaches to the measurement of the relative roles of age, period and cohort (APC) on marital-violence attitudes. We first look at these pairwise, before setting out a full APC decomposition. Our age-period-cohort analysis aims to isolate the effects of each component on some outcome of interest net of the influence of the other two (Yang and Land 2008).
The period or wave effects in our analysis show the evolution in attitudes over the 2000s, as we use survey data collected at three points in time: the early 2000s, the mid-2000s and the late 2000s.
The results suggest that, when taken pairwise, all three components help explain the percentage of women considering marital violence as not justified under any circumstances. We have three main findings. First, in a given cohort, women are more empowered as they age. Second, women in the early 2000s are more likely to think marital violence justified than those in the mid-2000s, and especially those in the late 2000s, suggesting strong period effects. Third, women in more recent cohorts are less likely to accept violence than women born in the 1950s. In addition, the full ageperiod-cohort decomposition suggests that some of these findings were due to confounding factors, as cohort effects are no longer statistically significant, while the age and period effects continue to play an important role in explaining female empowerment.
We explore the mechanisms behind the age effects. First, these could reflect female labor-force participation rising with age (as compared to a young woman of 15). Labor-force participation may empower women in their households and communities, and enable them to communicate more with their peers, including with respect to gender roles and status. When we control for labor-force status, we find that working women are far more likely than others to refuse marital violence, and that the age effects become insignificant. A second pathway could work via household composition, whereby younger women are more likely to still live with their parents. As such, they are more influenced by older generations who are less critical of violence, and in addition they may not perceive what wife-beating means, as they are not in a couple. We thus control for whether the woman is living in a nuclear-type family in which she is the household head or the spouse of the household head. Our finding suggests that living in a nuclear family increases female empowerment, and somewhat attenuates the age effect. Last, age effects can reflect parenthood: older women are more likely to have had children, including sons, which might empower them in a patrilocal context.
Controlling for the number of sons in the estimation as well as the number of children does not change the strong effect of aging on empowerment.
The remainder of the paper is organized as follows. Section 2 sets out the conceptualization of empowerment. Section 3 presents the pseudo-panel approach and Section 4 the data and variables.
Section 5 then presents the statistical methods and comments on the empirical results. Some robustness checks and possible mechanisms behind the age and period effects are explored in Section 6. Last, Section 7 concludes.

Conceptualization of empowerment
Following Kabeer (1999), empowerment refers to a dynamic process through which individuals acquire the ability to make choices. More precisely, empowerment can be considered as the process of change by which individuals enhance their freedom to act and to achieve what they consider valuable, whereas they were denied this capacity beforehand. The acquired capacity is called agency. As such, empowerment can be defined as an increase in agency (Trommlerova et al. 2015).
Agency represents individuals' capacity to define their own life-choices and aspirations, and to act accordingly, even if these life-choices and related actions are in contradiction with their peers (spouse, family or community).
Applying these concepts of empowerment and agency to women's status, female agency can refer to what occurs at the end of the empowerment process. Women's empowerment leads to greater agency. Salem et al (2014) provide insights into the conceptualization of female agency by stating that this encompasses three domains: their influence in family decisions, their freedom of movement in public spaces, and their expressions of views favoring more equitable gender roles. These three domains are complementary. We can argue that a woman is not observed to be empowered if she does not take part in household decisions, if she is not free of movement, and/or if she shares views favoring male domination. The first two domains are linked to their bargaining power within the household or couple, while the last is related to values and attitudes, and can be seen as going beyond what a woman experiences in her own household, and is not restricted to married women.
Each of these three domains has specific indicators. Women's influence in intra-household decision-making refers to whether women themselves are involved in decisions about their own health care, large or small household purchases, visits to family or relatives, giving assistance to family members, and children's schooling (Anderson andEswaran 2009, Ashraf et al 2010). Freedom of movement refers to the relationship within the couple, and especially the relative control of the husband over his wife, and involves questions regarding whether their husband tries to limit his wife's contact with her family, whether she needs his permission to go to the market or visit friends, and whether he insists on knowing where she was at any time (Jensen and Oster 2009). Views on gender roles, which are the focus of the analysis here, could cover any attitudes in favor of an egalitarian society captured by views regarding girls' education, or about the refusal of wife-beating.
In this paper, the refusal of wife-beating is adopted as the measure of women's empowerment, as ask whether gender attitudes and behaviors vary across villages according to whether and when they gained access to cable television. They show that after the introduction of cable to a village, women are less likely to consider that it is acceptable for a husband to beat his wife. At the same time, Jensen and Oster (2009) present evidence that women's status also changes with television: women report increased autonomy (as measured by the ability to go out without permission and to participate in household decision making), lower fertility and reduced son preferences. On the other hand, there is evidence that the refusal of marital violence is positively associated with women's intra-household decision making. Hindin (2003) suggests that women are more likely to think that marital violence is justified if their husband or partner makes most of the household decisions on his own, and less likely if decisions are jointly made. Rani et al (2004) analyze the probability of accepting wife-beating in a sample of Sub-saharan African countries, and show that the greater the number of household decisions in which women do not take part, the more likely they are to consider marital violence as being acceptable.
In addition, the use of the perception of wife-beating as a measure of empowerment has a number of advantages over the use of actual violence. Views about wife-beating may be easier to elicit via surveys than women's own experiences of household violence, which may be subject to misreporting for social-judgment reasons (see Amoakohene, 2004). Views regarding marital violence will provide an indication of its extent, given the high correlation between the two. Using DHS data, WHO that women who tolerate wife-beating are more likely to have experienced marital violence. This may show that either women learn to rationalize violence when they are themselves victims, or that women are at a greater risk of violence in communities where more people accept this form of violence. The potential reverse-causality bias is less of an issue when analyzing the impact of socio-demographic characteristics on perceptions than on the actual experience of violence. Kabeer (1999) and Rani et al. (2004) put this measure in perspective as a preliminary step towards societal change. Following Kabeer (1999), being able to refuse marital violence is crucial for the emergence of a critical consciousness, the process by which women move from a position of unquestioning acceptance of the social order to a critical perspective on it. According to Rani et al. (2004), the level of acceptance of wife-beating can provide insights into the stage of social, cultural and behavioural transformation of a society in its evolution towards a more gender-egalitarian society.

A pseudo-panel approach
For each woman i surveyed at time t in country c, we observe her opinion on whether marital violence is justified under some circumstances. The probability of not thinking marital violence as justified under any circumstances can then be estimated via a linear probability model such that: where X i includes a full set of explanatory variables (marital status, education and wealth), γ c is a country-specific effect, and ε ic the error term. The age effects λ are not well-identified in this cross-section model, as they are confounded in the cohort effects. In the context of the current paper, there may well be unobserved cohort heterogeneity such as the social environment each cohort faced during childhood and adulthood, or the exposure to human rights while at school or from the media.
The use of panel data allows us to follow individuals over time and estimate the effect of changing age while controlling for both unobserved individual characteristics via individual fixed-effects (δ i ) and time fixed effects (θ t ), as in Equation 2.
However, individual panel data is rarely available in developing countries. Following Deaton Hereȳ gct is the proportion of women in generation (cohort) g in country c surveyed at time t who say that marital violence is not justified under any circumstances. The δ g can be viewed as unobserved cohort fixed effects, and the variables inX gct are the means or proportions of the individual characteristics in each gct group (e.g. the average number of years of education).
Equation (3) has two advantages. First it allows us to correct for unobserved cohort heterogeneity via cohort fixed effects. Second, the use of cohort means helps to "average out" any individual We hence control for country specific-effects γ c . We could argue that there are intra-country differences too, as regions within countries may be very different from each other. A more disaggregated analysis at the regional level is undertaken later in the paper.
Dargay (2007) includes a useful discussion on how to define the cohorts or groups of individuals to be followed in pseudo-panel analysis, and the trade-off between the number of observations included in each group and the number of groups. In our case, the DHS includes data over a large sample of individuals, so that we can define groups in a fairly fine way and not pool birth cohorts over all countries. As a result the variation within cohorts is quite small compared to that between cohorts.
- Table 1 here -We carry out an age, period and cohort decomposition. In our approach the three effects can be simultaneously identified. The identification relies on the definition of the periods, which avoids the problem of strict multi-collinearity between the three components. Instead of using the exact survey year -which is equal to the sum of birth year and age -dummy variables for whether the observations refer to the women surveyed in the first, second or third rounds of the DHS are employed. As shown in Table 1, the three rounds of data were roughly collected at the same period of time over the five countries in our sample. We approximate changes over time by  Table 1). Appearing in three successive DHS survey waves is actually fairly rare: in addition to the countries we analyze here, this only occurs for Peru, Armenia and Indonesia. A detailed description of how we constructed the pseudo panels using the DHS data can be found in Appendix 1.
(a) The cohort data The pseudo-panel appeals to the women's age at the time of the survey to establish her birth cohort: if a woman is x years old in year t, then she will be x + 1 years old in year t + 1, x + 2 years old in year t + 2, and so on. year, even for surveys collected over two calendar years: the retained survey year is that in which the most data-collection months occurred.
- Table 2 here - The cohorts are defined for the birth years from 1951 to 1996, using data from surveys for 1999 through 2011. Note that the 1962-1984 cohorts are observed over three successive surveys in all five countries. The averages for each birth year are generated by country and survey year. We have the advantage here of a large sample size: only 2% of the observations are averages based on country-cohorts with fewer than 100 observations. The average number of women in each cell is between 229 and 577 (see Table 2).

(b) Attitudes toward wife-beating
The surveys collect information on individual attitudes toward wife-beating, which we use as a measure of empowerment. Women were asked whether a husband is justified in beating his wife under a series of circumstances. The five circumstance questions are as follows in the questionnaire 2 : "Sometimes a husband is annoyed or angered by things which his wife does. For each circumstance, the woman can answer that wife-beating is justified, is not justified, or she does not know. There are only few "does not know" replies. 3 We henceforth recode these as missing values (although this assumption will later be relaxed as a robustness check). We create an individual-level dummy variable for not justifying wife-beating under any circumstances (as in Hindin 2003, Rani et al 2004. Table 3 shows the number of women per wave and country survey and the percentage not justifying wife-beating. - Table 3  is often used to proxy wealth via an index from a principal components analysis, as suggested in Filmer and Pritchett (2001). The data provider has applied this methodology to generate a wealth index, and ranks households according to the value of this continuous index.
- Table 4 here -  Second, regarding cohort effects, the refusal rate at every age is higher for the younger compared to the older cohorts. For instance, the proportion of 22 year-old Ethiopian women who think marital violence is not justified under any circumstance is higher in the 1986-1990 than in the 1981-1985 cohort, and much higher than that in the 1976-1980 cohort. All of these graphs exhibit large movements between the segments, although this is more pronounced in Ethiopia and Malawi.
These displacements suggest large cohort (or generational) effects at a given age. Such large cohort effects may reflect changes in legislation against marital abuse, or the recent rise in education. in the paper, we estimate the percentage of women not justifying wife-beating using weighted leastsquares method, with the weights being the size of the sampled cohort. We first look at relative roles of age, period and cohort pairwise, before setting out a full APC decomposition. We include the following set of covariates in each regression: years of education, marital status, relative wealth categories, number of children, urban residence and country-specific effects.
Once the core results are established, we perform some robustness checks modifying the definition of the dependent variable, the specification of the cohort effects and the analytical sample.
As the age and period effects dominate, we consider different interpretations of these two effects. First, we investigate the mechanisms behind the age effects and present suggestive evidence by exploring the role of female labor-force participation, household composition and parenthood.
These models are illustrative and cannot detect causal relationships due to possible endogeneity, particularly reverse causality. Second, we explore the role of media access as fundamental change that affects all age groups simultaneously.
The last step of the analysis consists in extending the model by allowing for regional disparities within the five countries of our sample.
(c) Estimation results Table 5 shows the results from weighted least-squares estimation in the pseudo-panel data.
Columns 1-4 display the estimates from the age-period, age-cohort, period-cohort and age-periodcohort models respectively. The results show that the percentage of women not justifying wifebeating is rising and somewhat concave in age in every specification. In the first specification, the age variable captures both the life-cycle and generation effects, while in the last specification, the effect of age is net of the cohort effects. As such, the column 4 results suggest that within-cohort, women are less accepting of violence as they age.
There are strong period effects in columns 1, 3 and 4. Saying that wife-beating is not justified under any circumstances has become more common over time, even controlling for age and cohort. - Figure IV here - The estimated cohort effects appear in Table A1 in Appendix 2 and in Figure IV. The cohort effects are statistically significant when considered pairwise in addition to age or period effects, but become insignificant in the age-period-cohort model. In this latter (column 4), the three effects are combined, and the life-cycle and period effects are found to be the most powerful predictors of empowerment. Within each cohort, women are less likely to justify wife-beating as they get older and as they are interviewed more recently.
In the age-cohort and period-cohort specifications, the estimated cohort effects are of opposite sign. Women from more recent generations than 1951 (the reference category) are more empowered when controlling for age but not for period (column 2) but less empowered when we control for period but not for age (column 3). In the pairwise decomposition, the omitted effect is captured in one of the other two included effects. In the age-cohort model, the positive cohort effects mean that at a given age, women born recently are more likely to refuse wife-beating. This feature captures in a sense the period effects. In the period-cohort model, the cohort effects being negative means that in a given survey wave, women born more recently are less likely to refuse wife-beating. In other words, young women are more likely to accept wife-beating, which is the life-cycle effect found in the age-period and the age-cohort models.
For a given cohort, the age and period effects can also be mixed together. In the age-cohort model the age effects are larger than in the age-period model, suggesting that the period effect is captured in the estimated age coefficient. Women from a given cohort are getting older and at the same time the society is changing over the 2000s. In the period-cohort model, the age effects are captured by the period effects in the same way.
The estimated country effects show that women in Uganda, Rwanda, Zimbabwe and Malawi are significantly more likely than women in Ethiopia to think that marital violence cannot be justified.
There is country heterogeneity in terms of empowerment. These systematic differences can result from differences in school quality or the political effort to back awareness campaigns, for example.
The formal explanation of this heterogeneity is beyond the scope of this paper.
Education plays a significant role in explaining attitudes to marital violence, and violence refusal rises with education. Urbanization is also associated with greater refusal. Married women are however less likely than single women to reject wife-beating, and previously-married women are the most accepting of marital violence. Relative wealth plays no statistically significant role in our regressions.

Robustness checks and underlying mechanisms (a) Robustness checks
We now carry out some robustness checks to test the validity of the results from the age-periodcohort decomposition model (Table 5, column 4).
- Table 6 here - The first check consists in modifying the definition of the dependent variable. We first follow Mahmud et al. (2012) and consider a finer way of looking at refusal using the five original circumstance questions that women were asked. The number of circumstances under which marital violence is not accepted is then averaged over all of the women in a given country-specific birth cohort for each survey year. Age is not significant in this specification of the APC model (column 1 of Table 6).
A second check regarding the dependent variable refers to the "does not know" category. To date, these have been recoded as missing values. However, it can be argued that women who do not want to reveal their preferences may say that they have no opinion or do not know. As such, "does not know" may hide either a Yes or a No. In the Table 6, column 3 (column 2) estimates, the does not know answers are considered as (not) justifying wife-beating. The results are robust, as the age effects remain positive, significant and slightly concave, and the period effects remain statistically significant.
The second type of robustness check consists in restricting the sample used in the estimations (see column 4 of Table 6). As noted above, some cohorts are followed and observed over three successive surveys, while others are seen only twice or even only once. Restricting the sample to the cohorts appearing in three successive surveys confirms our previous findings, as the age effects remain significant, positive and slightly concave, while the cohort effects do not significantly predict attitudes to marital violence (column 4 of Table 6). The period effects are not statistically significant at conventional levels, although the p-values are only just above the 10% level.
The last robustness check concerns the definition of the birth cohorts. The cohort dummies may be insignificant as we use single-year birth cohorts, making the analysis too fine. We could argue that women born in year n have similar characteristics to those born in n + 1, as they grew up in similar environments, were married on average in roughly the same year etc. We keep the same observation-level but now group cohorts in 5-year intervals (see Table 6, column 5). In other words, women born between 1960 and 1964 are compared to those born before 1955 whatever their exact birth year. The age and period effects are robust to this change in specification. The cohort effects remain statistically insignificant (not reported here).

(b) Explaining the age effects
Our results suggest that women's attitudes to marital violence change over the life cycle towards greater refusal. We explore three pathways in Table 7 to provide some insights into the mechanisms underlying these age effects.
- Table 7 here -(i) Household composition Younger women may be more likely to accept wife beating due to their living arrangements and the people they live with. Women who are still living with their parents may not realize what marital violence means (as they are not confronted with it) and may be more influenced by their mothers, who are from the earlier cohorts that are more accepting of violence. We explore this by introducing the percentage of women by cohort and wave who live in a nuclear family where the woman is the household head or his spouse. Column 1 in Table 7 shows the benchmark estimation results (from Table 5, column 4) and column 2 then adds our new variable. The results suggest that household composition does play a role in violence refusal, as women living in a nuclear family are more likely to reject wife-beating. The age effects are stable in terms of sign and significance.
The size of the coefficient is slightly reduced, suggesting that a small part of the age effect reflects household living arrangements. Note however that the period effects become insignificant here, suggesting that the prevalence of nuclear families has changed over time.
(ii) Parenthood experience and number of sons In columns 3 and 4, we test whether having had sons can help explain the age effects. Column 3 controls for the average total number of sons women have had, and column 4 for the number of sons still at home and the number of sons living elsewhere. The age effects remain positive, concave and significant in both cases, and increase in size in the last specification. The sign, significance and size of the period effects are robust in these two specifications.
(iii) Labor-Force status Teenage women are less likely to work, so that part of the age effect could reflect labor-force status. Controlling for the proportion of women currently working in column 5 reveals that work does indeed significantly predict attitudes, and moreover renders the age effects insignificant, suggesting that a large part of the age effects reflect labor-force status. Work not only helps women to obtain bargaining power as they bring money home, but also helps them to stay in touch with their peers and communicate regarding gender roles and status.
(c) Explaining the period effects Period effects represent changes over time that affect all age groups simultaneously. These changes can be related to shifts in the social, cultural or physical environment, and can be structural or conjunctural. They may be related to the business cycle but also to fundamental changes in society. In our context, in addition to the household composition changes (shown above), another potential change is access to media and in particular television, where the TV exposure may have  Table 7, we include the frequency of watching TV, listening to radio and reading magazines or newspapers as additional explanatory variables. The period effects become smaller and insignificant, so that at least part of the increase in refusal over time seems to have been due to greater media access, and potentially via unobserved changes in program content.
(d) Regional pseudo-panel analysis We now extend our model by allowing for regional disparities within the five countries in our sample.
In order to do so, we take a regional pseudo-panel approach in which the cohorts are no longer defined by country and year of birth, but rather by region and year of birth. A detailed description of how we constructed regional pseudo panels using the DHS data can be found in Appendix 1. The sample size and estimation results are summarized in Appendix 2 (Tables A2-A5). We estimate the following model:ȳ grt = α + λage grt + βX grt + δ g + θ t + γ r +ε grt In equation (4), the index changes, as we now consider groups of women defined over the region in which they live, denoted by r, instead of their country of residence as in Equation (3). Hereȳ grt is the proportion of women in generation (cohort) g living in region r surveyed at time t who say that marital violence is not justified under any circumstances. The variables inX grt are the means or proportions of the individual characteristics in each gr group surveyed at time t.
This strategy has a number of advantages. First, it allows us to control for unobserved regional heterogeneity, which is captured in the parameter γ r : this can be viewed as the unobserved regional fixed effect. The sources of regional heterogeneity may be either observable or unobservable. Regarding observable heterogeneity, in the present analysis, due to data limitations, we cannot control for a full range of socio-demographic covariates (religion is one example). Regional heterogeneity may also come from unobservables such as culture and traditions that may influence the views women express about gender-roles. Second, it increases the sample size on which the estimation results are based from 524 to 3,454.
We however believe that this approach is not to be preferred. First, as the number of regions varies across countries, the number of times each country is observed will vary across countries. In the pseudo-panel at the country and birth-year level, the number of observations per country was the number of birth years times the three rounds. By way of contrast, in the regional pseudo-panel, the number of times each country is observed is the number of regions times the number of birth- year cohorts times the number of rounds. As such, Malawi with the smallest number of regions (3) will be observed 315 times, and Ethiopia with its 11 regions 1,154 times.
Second, and more importantly, each observation defined by the survey year, region and birth year is an average coming from fewer women. As noted above, the number of observations used to calculate any average should be greater than 100. While this requirement was largely satisfied in the country pseudo-panel (see Table 2), this is no longer the case in the regional approach, where most observations are based on fewer (and far fewer) than 100 observations. Table A2 lists the average number of women used to generate the regional pseudo-panel observations. It turns out that, except for Malawi, the median is always below 100, meaning that more than half of observations are based on information elicited from fewer than 100 respondents. 90% of the observations are based on less than 82 women in Ethiopia, 118 in Rwanda and 101.5 in Uganda. In Zimbabwe, no regional pseudo-panel observation is based on over 100 observations, as the highest number of respondents per wave and group is 94.
With these caveats in mind, the regional pseudo-panel confirms the previous findings established above. In the Age-Period-Cohort decomposition, as in the country pseudo-panel, the age and period effects dominate the cohort effects (see Table A3). These findings are robust to a number of tests, except for the specification in which the number of circumstances under which women refuse marital violence serves as the dependent variable. In this equation, the period effect vanishes (see Table   A4). Last, the attempt to explain the age and period effects yields similar findings (see Table A5).
It is worth noting that the region-specific effects are very significant in predicting the refusal of marital violence, suggesting that regional disparities really do exist.

Conclusion
This paper has used Demographic and Health Survey data to analyze female empowerment, as measured by women's beliefs about whether wife-beating is justified. We focus on five countries We distinguish life-cycle from birth-cohort effects, and in an age-period-cohort decomposition it is the age effect which prevails. Over the five countries in our sample, women become increasingly less accepting of marital violence as they age. As such, empowerment is a dynamic process in women's lives. These attitudinal changes over the life cycle are shown to be related to laborforce status and women's family and household positions. Considering empowerment as fixed seems to be a mistake. Although background (e.g. education, relative status with respect to the husband, parents' socio-economic characteristics compared to those of her parents-in-law, and assets at marriage) are powerful determinants of initial empowerment, this initial level evolves over the life course. This dynamic process means that current empowerment cannot be approximated by family background.
Societal changes have a considerable effect on the propensity to accept marital violence, as time dummies play an important explanatory role. Empowerment is then a dynamic process at both the societal and the individual level. Societal modernization can take different forms affecting empowerment, such as media access, changes in attitudes and social norms, the awareness of other lifestyles via the internet, soap operas, etc. Given the increasing speed of adoption of new technologies, we can expect a concordant change in societal behaviors and attitudes. If individuals are now more willing to adopt societal changes than they were some decades ago, we can expect the accumulation of empowerment from one generation to another to rise at an increasing rate.
Empowerment will appreciate either through societal change or intergenerational transmission. As daughters are influenced by their mothers, that mothers accept violence less over time and new mothers accept less than older mothers will both lead to rising empowerment. This paper has provided a picture of the situation in Eastern and Southern Africa during the 2000s, where we observe a sharp drop in the acceptance of marital violence, and thus rising female empowerment.
We would therefore predict that future Demographic and Health Surveys will continue to reveal falling rates of justification of wife-beating over the next decades.
In addition to this dynamic process, our findings suggest that policies to curb the acceptance of wife-beating among the population should essentially target younger women.
One weakness of this study is that it relies on the aggregation of the individual characteristics in order to create a pseudo-panel. The ideal dataset would follow women over time to disentangle at the individual level the age effect from the cohort effect in the degree of women's empowerment.
Furthermore when the mechanisms behind the age effects are investigated, the analysis is illustrative and does not detect causal relationships, as labor-market participation and the number of children very likely suffer from reverse causality bias.
Future research could usefully explore the respective effects of age, period and cohort on other measures of empowerment, namely participation in intra-household decision making or relative freedom within the couple. This could be carried out at the expense of restricting the sample to married women, whereas one of the strengths of our measure here was to explore life-cycle and cohort effects across all women, independently of their marital status, and to identify the strong effects of being or having been married. It would also be particularly helpful to test for multiplicative effects in changing female status from the intergenerational transmission of values from mothers to daughters. Last, one of our results suggests that the changes over the life-cycle concern the extensive margin and only insignificantly the intensive margin when we consider the age, period and cohort effects on the number of circumstances in which wife beating is not justified. Future research could compare the elasticities of the intensive and the extensive margins, and explore which circumstances are more likely to change over time and over the life-cycle.  3 Over the full sample of individual observations, the proportion of "does not know" varies from 1.11% for circumstance (v) to 3.45% for circumstance (iv). 4 As information on religious affiliation is not collected in Rwanda 2000, we do not use this variable as a control.      Notes: a Age squared is divided by 100. * p < 0.10, * * p < 0.05, * * * p < 0.01. Robust standard errors in parentheses. Omitted categories: Ethiopia, early 2000s, being born in 1951. The estimated coefficients on the control variables are not shown. The same controls are used as in Table 5. Beating not justified (%) 10 20 30 40 50 age 1941-1955 1956-1960 1961-1965 1966-1970 1971-1975 1976-1980 1981-1985 1986-1990 1991-1996 Raw life-cycle profiles: Ethiopia .6 .7 .8 .9