WPS6946 Policy Research Working Paper 6946 Bridging the Gap Identifying What Is Holding Self-Employed Women Back in Ghana, Rwanda, Tanzania, the Republic of Congo, and Uganda Emily Nix Elisa Gamberoni Rachel Heath The World Bank Africa Region Poverty Reduction and Economic Management Unit June 2014 Policy Research Working Paper 6946 Abstract This paper explores the determinants of the gender while the gap for the most successful male and female gap in income earnings in five Sub-Saharan countries: entrepreneurs is largely driven by differences in returns the Republic of Congo, Ghana, Rwanda, Uganda, and to observable covariates in the majority of the countries. Tanzania. It shows that first, self-employment tends These results suggest a glass ceiling effect, wherein a to provide marginally lower average income (with the large portion of the income gaps between high-earning exception of Ghana and men in Rwanda) and much men and women cannot be explained by observable higher variability in income compared with wage work. characteristics. The paper concludes by looking at the Women on average earn less than men when they variables that account for a larger portion of the gender are self-employed and in wage employment, but also gap explained by observable characteristics and finds that have less volatile earnings. The analysis uses quantile hours of work and industry explain a higher fraction decomposition methods and finds that the differences in compared with standard human capital and demographic observable choices and endowments explain the gender factors such as age and education. gap in earnings for the self-employed who earn the least This paper is a product of thePoverty Reduction and Economic Management Unit, Africa Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at egamberoni@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Bridging the Gap: Identifying What Is Holding Self-Employed Women Back in Ghana, Rwanda, Tanzania, the Republic of Congo, and Uganda Emily Nix, Elisa Gamberoni, and Rachel Heath 1 JEL: J16, J31, D04 1 Nix: Yale University, emily.nix@yale.edu; Gamberoni:World Bank, egamberoni@worldbank.org; Heath: The University of Washington, rmheath@uw.edu. This work was made possible with support of the UN Foundation – ExxonMobil Foundation research collaboration on women’s economic empowerment (http://www.unfoundation.org/features/building-a-roadmap.html) and with support from the Bank Netherlands Partnership Program. We also thank Elena Bardasi, Louise Fox, and participants at the United Nations Greentree Convening on Women’s Economic Empowerment for helpful comments. We are grateful to Paolo Falco, Thomas Pave Sohneson and Jorge Huerto Munoz for help with the data used in the report. . 1 Introduction The existence of a gap between male and female income is a persistent reality across the world. In order to formulate policies to bridge these gaps, it is useful to explore what may be causing the gaps. There is a large literature examing the gender income gap in developed countries like the United States. There is also a growing literature that goes beyond simple mean decompositions. For example, quantile decomposition methods were developed and applied in Gosling, Machin, and Meghir (2000) and Machado and Mata (2005). Such techniques reveal whether the size and determinants of earnings gaps vary at different points across the income distribution. Our paper follows in the tradition of these two strains of literature by making use of a new decomposition approach developed in Firpo, Fortin, and Lemieux (2007) and applying this methodology to examine the gender income gap in five countries in Africa. To our knowledge this is the first paper that applies the Firpo, Fortin, and Lemieux (2007) methodology to examine the gender gap by quantile in these countries. Importantly, this methodology not only allows us to obtain aggregate decompositions of statistics beyond the mean, but further allows us to assess the contribution of individual covariates to the aggregate decomposition terms. We specifically decompose the gender gap in self-employment income. This is particularly important given that workers in the developing world, particularly in Sub- Saharan Africa, are frequently self-employed. In fact, of the adults in Sub-Saharan Africa whose primary employment is not family farming, more workers (51.4%) work in household enterprises than for wages in public or private firms or in agriculture. Fox and Sohnesen (2012).1 Indeed, in Sub-Saharan Africa, self-employment often constitutes the only alterna- tive outside the agriculture sector. For example, in Uganda, despite impressive growth rates, the majority of the nonagricultural employment was created in household en- terprises and microenterprises (Chuhan-Pole, Angwafo, Mapi Buitano, Korman, and Fox (2011)). Women are well- represented among these smaller size businesses: in Tanzania, for instance, 80 percent of micro entrepreneurs are female (Kweka and Fox (2011)). Thus, understanding the dynamics and the determinants of earnings among male and female entrepreneurs is crucial. This paper focuses on the role of individual traits in explaining income in five Sub- Saharan countries: the Republic of Congo, Ghana, Rwanda, Uganda, and Tanzania. While the literature has already emphasized the role of individual traits–Jovanovic (1982), for example, looks at the role of experience on the decision to enter self- employment while Fajnzylber, Maloney, and Rojas (2006) and Bates (1999) focus on the role of educational attainment on firm success– less attention has been paid to the contribution of specific traits to the gender income gap in these countries. To fill in this gap, we develop a new dataset that synthesizes data from National Household Surveys from the Republic of Congo, Ghana, Rwanda, Tanzania, and Uganda. 1 Overall, 18 percent of respondents worked in household enterprise. 1 In the first part of the paper we show that self-employment tends to provide marginally lower average income (with the exception of Ghana and men in Rwanda) and much higher variability in income compared to wage work. Importantly, women on average earn less than men both when they are self-employed and in wage employ- ment but also have less volatile earnings. This suggests either a difference in returns to traits or differences in average traits among self-employed men and women. We further find that the self-employment income gap is not constant across the distribution across the five countries. This motivates our main analysis where we employ quantile decomposition methods developed in Firpo, Fortin, and Lemieux (2007). A quantile decomposition allows us to decompose the gap in income at each quan- tile into the composition effect and the structural effect. The composition effect is the difference in income due to observable differences in the included set of covariates: marital status, experience, education, number of children, average monthly hours worked, and industry. The structural effect is the difference in income due to dif- ferences in returns to the same set of covariates. The structural effect can also be considered the “unexplained” portion of the gap. With the quantile decomposition method we use we are further able to decompose the composition and structural ef- fects by covariates. This allows us to assess the relative importance of each covariate in explaining the income gap, and how these factors vary based on low earners and high earners. Because the type of work, input and output markets, and non-work con- straints faced by low and high income earners is very different, an examination of the contributions to the gender wage by income quantile is key to the policy implications we provide. We find that the majority of the gap across the distributions for all countries is due to the structural effect. Still, the composition effect plays an important role for all countries but Uganda. In Ghana we find that the share of the gender gap explained by differences in covariates remains relatively constant across income quantitles, at around 30%. In contrast, in the Republic of Congo, Rwanda and Tanzania we find that the composition effect is largest for the bottom 20%. In the Republic of Congo, the composition effect explains 50% of the gap, in Rwanda it explains 48% of the gap, and in Tanzania it explains 31% of the gap at the 10th quantile. The composition effect then decreases in importance relative to the structural effect as we move across the income distribution towards the higher earners. By the 90th quantile it explains only 16% of the gap in all three countries. Thus, while differences in observable choices and endowments play a role in caus- ing the gap between women and men earning the least in self-employment, the gap for the most successful male and female entrepreneurs is largely driven by differences in returns to observable covariates in the Republic of Congo, Tanzania, and Rwanda. Differences in returns to a given covariate could be indicative of discrimination. How- ever, in order to unequivocally state that the structural effect is due to discrimination, one must include all possible determinants of income in the set of covariates. We are 2 unable to do this given our data. We next decompose the aggregate composition and structural effects into the contributions of the individual covariates. We find that monthly hours drives the composition effect in Tanzania and Rwanda. More specifically, while gaps in monthly hours worked explains a significant portion of the income gap between low earning men and women, it explains much less of the gap between higher earning men and women. Industry drives the composition effect in the Republic of Congo and savings drives the composition effect in Ghana. Consistent with the aggregate decomposition results, the savings composition effect remains fairly constant for Ghana while the portion of the gap explained by industry differences decreases in importance for the Republic of Congo as we move to the right in the income distribution. We conclude by highlighting the policy implications of our findings. Specifically, we argue that our results highlight the importance of observable determinants of income – namely, hours and industry – beyond standard human capital and demo- graphic factors such as age and education. We also point out that our results from several countries suggest a glass ceiling effect, wherein a large portion of the income gaps between high-earning men and women cannot be explained by observable char- acteristics. We argue that these descriptive results can be tested in future research that causally identifies determinants of earning gaps between men and women. 2 Data and Summary Statistics For the main part of the analysis, we use nationally representative household survey data for the following countries (survey year in parenthesis): the Republic of Congo (2009, urban only), Ghana (2005), Rwanda (2005), Tanzania (2009) and Uganda (2009). Our main dependent variable is the logarithm of monthly self employment income for the non-agricultural self-employed individuals. We calculate self-employment in- come as wages to the owner plus profits. Profits are calculated as the difference between revenues and expenses with adjustments as recommended in the recent liter- ature (De Mel, McKenzie, and Woodruff (2009).2 Since the survey provides minimal information, profit figures are less rigorous only in the case of Tanzania. Concerning the explanatory variables, on the one hand, our main interest is on the role of the individual traits identified by the literature in explaining firm performance (see for example Minniti (2005)). On the other hand, to ensure consistency in the interpretation of the results across countries, we selected and harmonized as many variables as possible. The main explanatory variables include marital status, age, education, number of children, average monthly hours worked, savings, and industry. Education consists 2 Unfortunately, the surveys do not directly ask for profits, which is the recommended measure by De Mel, McKenzie, and Woodruff (2009) 3 of 8 categories: none, some primary school, completed primary school, some middle school, completed middle school, some secondary school, completed secondary school, and some college. We use age as a proxy for experience. Industry consists of the fol- lowing categories: agriculture/fishing, mining/energy, manufacturing, construction, retail, trade, finance, public services, and other. Sample means for these variables by gender and country can be found in the Online Appendix in Tables C.1-C.5.3 Table 1 provides summary statistics for these variables, pooling all cross-sectional data sets. The data are presented for the three main occupational categories available: agricultural employees, wage employees, and the self-employed. Tables A.1 - A.5 in the Online Appendix provide the same descriptive statistics for each country. Table 1: Summary Statistics by Employment Type mean sd min max count Agriculture Primary Income in USD 27.92 57.66 0.00 664.14 1711.00 Education (Categorical, 1-8) 3.09 1.40 1.00 8.00 29321.00 Age in Years 35.96 13.25 15.00 65.00 37012.00 Male Dummy 0.48 0.50 0.00 1.00 37012.00 Marriage Dummy 0.67 0.47 0.00 1.00 37012.00 Paid Employment Primary Income in USD 92.91 117.71 0.00 1138.52 13696.00 Education (Categorical, 1-8) 4.23 2.21 1.00 8.00 13346.00 Age in Years 33.73 11.49 15.00 65.00 14242.00 Male Dummy 0.69 0.46 0.00 1.00 14242.00 Marriage Dummy 0.51 0.50 0.00 1.00 14242.00 Self-Employed Primary Income in USD 79.95 111.59 -29.20 1138.52 10999.00 Education (Categorical, 1-8) 3.80 1.71 1.00 8.00 12384.00 Age in Years 35.33 10.93 15.00 65.00 13383.00 Male Dummy 0.47 0.50 0.00 1.00 13383.00 Marriage Dummy 0.64 0.48 0.00 1.00 13383.00 Total Primary Income in USD 83.30 113.31 -29.20 1138.52 26406.00 Education (Categorical, 1-8) 3.52 1.77 1.00 8.00 55051.00 Age in Years 35.34 12.46 15.00 65.00 64637.00 Male Dummy 0.52 0.50 0.00 1.00 64637.00 Marriage Dummy 0.63 0.48 0.00 1.00 64637.00 Observations 64637 3 The Online Appendix can be found at http://faculty.washington.edu/rmheath/research.html. 4 3 Preliminary Analysis We start the analysis by looking at the distribution of individuals across primary occupation (agriculture, non-agricultural self employment, and non-agricultural wage work) and gender (Figure 1). Clearly, agriculture is still a very important sector in these economies. However, self-employment is a significant source of employment for each of these countries.4 This is in line with Fox and Sohnesen (2012)) who point out that self-employment is not only a large source of employment, but also a growing sector in countries across sub-Saharan Africa. There are distinct gender and country differences in occupational choices for pri- mary employment. In Ghana, Tanzania, and Uganda very few women report paid employment as their primary form of employment. Most women in Tanzania and Uganda work in agriculture, but close to 20% report self-employment as their primary form of employment. In Ghana and the urban sample from the Republic of Congo, the majority of women report self-employment as their primary form of employment. In Rwanda, over 80% of women report agriculture and wage employment as their primary employment, with a much smaller percentage reporting self-employment as their primary form of employment. Men are more likely than women to report wage employment as their primary form of employment in all five countries. Figure 1: Percent in Agriculture, Paid Employment, and Wage Work by Country Concerning the type of businesses under analysis, Figure 2 illustrates the percent- age of male and female owners of household (defined as a business without employees), micro (1 to 4 employees), small (5 to 19 employees), medium (20-99 employees), and large (100 + employees) businesses. As the figure shows, the sample primarily con- sists of household and micro enterprises. Additionally, women engage in smaller size businesses compared to men. 4 Note that the Republic of Congo is a nationally representative urban household survey. As a result, the low percent of individuals in agriculture activities only reflects urban realities. 5 Figure 2: Enterprise Size by Country and Gender We next look at the distribution of primary incomes across countries (Figure 3). As shown in these Figures and in Table 2, self-employment is characterized by far greater variability in earnings, with the exception of women in Ghana, Rwanda and Tanzania. For women, self-employment offers higher average earnings relative to wage employment only in the case of Ghana and Uganda. Men have substantially higher average earnings than women, both in wage employ- ment and in self-employment, but men also have greater variability in their earnings. This can be seen in Figures 4 and 5 which show income distributions by gender and country for wage workers and the self-employed, respectively. Table 2 provides the exact figures. This seems in line with the idea that women are more risk averse than men: Croson and Gneezy (2009), reviewing recent studies, suggest that women are more risk averse than men but gender differences in financial risks, for example, disappear among high level professionals (such as mutual fund managers). 6 Figure 3: Distribution of Log U.S. Monthly Income by Country and Employment Type Figure 4: Distribution of Wage Workers’ Log U.S. Monthly Income by Country and Gender 7 Figure 5: Distribution of Self-Employed Log U.S. Monthly Income by Country and Gender Table 2: Mean and Standard Deviation of Monthly Income in U.S. Dollars Wage Employment Self-Employment mean sd mean sd Congo, Rep. Male Primary Income in USD 234.546 166.427 202.624 168.128 Female Primary Income in USD 163.512 123.836 120.115 130.626 Ghana Male Primary Income in USD 122.028 130.171 133.009 154.082 Female Primary Income in USD 96.032 106.076 81.966 94.586 Rwanda Male Primary Income in USD 56.031 82.960 51.661 92.554 Female Primary Income in USD 41.037 76.070 28.223 55.883 Tanzania Male Primary Income in USD 86.680 100.788 83.225 104.557 Female Primary Income in USD 69.843 87.497 44.753 69.434 Uganda Male Primary Income in USD 82.825 96.577 85.130 132.718 Female Primary Income in USD 61.035 75.979 56.695 104.856 In all countries, the gender gap in self-employment earnings is large in magnitude and statistically significant: on average self-employed men in the Republic of Congo make $203 a month compared to $120 for women; in Ghana self-employed men make $133 a month compared to $82 for self-employed women; in Rwanda, self-employed 8 men on average make $52 a month compared to $28 for self-employed women. In Tanzania, self-employed men on average make $83 a month compared to $45 for self- employed women. Finally, in Uganda self-employed men make $85 a month compared to $57 for self-employed women. Looking at the gender gap along the income distribution of self-employmed indi- viduals (Figure 7), we observe a positive gap between male and female self-employed income across all income quantiles for the Republic of Congo, Ghana, Rwanda, Tan- zania and Uganda.5 Figure 6 shows the gender income gap by quantile for wage workers, which is also positive for all countries for all quantiles. To sum up, with the exception of Ghana, the gap is significantly higher across all quantiles for self-employed workers compared to the gap for wage workers. In the next section, we thus focus on the gender income gap among self-employed men and women. Figure 6: Raw Male-Female Wage Employment Income Gap by Country 5 To understand what Figure 7 shows us, think of lining all men up in order of their income, and the same for women. Then think of dividing the line into 10 groups, where the first group is the bottom 10% in terms of their income, the second group is the next 10% in income, and so on. Figure 7 takes the maximum income for men in each of these 10 quantile groups and subtracts the equivalent number for women. 9 Figure 7: Raw Male-Female Self-Employment Income Gap by Country 4 Gender Decomposition of Profits In this section, we look at the role of individual traits in explaining observed gender gaps in self-employment. We present quantile decompositions of self-employment income, using the decomposition procedure developed by Firpo, Fortin, and Lemieux (2007). The goal of the decomposition is to look at differences in an outcome (here self-employment income) at different points in the income distribution across two groups and separately identify how much of the gap is explained by differences in endowments and how much of the gap is explained by differences in returns to those endowments. For example, consider education. A decomposition could be used to establish what part of the gender gap in income is due to observable differences in education levels across men and women (which would arise if men have more schooling than women and there are positive returns to schooling) and what part of the gender difference is due to different returns to a given education level for men and women. Different returns to men and women with the same characteristics could be indicative of discrimination in the labor market, but could also be due to omitted variable bias if these characteristics are correlated with unobserved determinants of income, such as ability. Given that it is impossible with the given data set to include all determinants of income in our analysis, differences in returns to the same endowments for men and women should not be taken as evidence of discrimination. Instead, it shows that discrimination may be happening, or there may be an omitted variable that is driving the results. In keeping with Firpo, Fortin, and Lemieux (2007), we define the part of the gender income gap explained by differences in endowments to be the composition effect and the part of the gender income gap explained by differences in returns to 10 endowments to be the structural effect. 4.1 RIF Decomposition Methodology In the traditional Oaxaca-Blinder mean decomposition, the difference in average in- come is decomposed into structural and compositional components in the following way. First, assume that income, Y , depends linearly on a set of covariates, X . This implies that we can write Ygi where g stands for gender and i stands for observation in the following way. K Ygi = αg + Xik βgk + νgi , g = F, M (1) k=1 Where νgi is the error term and the expected value of the error, conditional on X , is 0 (E [νgi |Xi ] = 0). Define the difference in average income to be D = E [Ym ] − E [Yf ] (2) where Yf is female income and Ym is male income. We can decompose this differ- ence into structural and compositional components by estimating Equation 1 , which allows us to rewrite Equation 2 as K K D = E [Ym ]−E [Yf ] = (αm − αf ) + E [Xmk ] βm − βf + (E [Xmk ] − E [Xf k ]) βf k=1 k=1 ˆ s =unexplained ∆ ˆ c =explained ∆ (3) ˆ ˆ Here, ∆s gives the aggregate structural portion of the wage gap while ∆c tells us what portion of the gender income gap can be explained by differences in the covariates X . We can also estimate more detailed results that give the contribution of each covariate to the aggregated structural and compositional effects. However, this approach only works for mean decompositions. To obtain both an aggregate and detailed decomposition by quantile we use a new econometric approach described in Fortin, Lemieux, and Firpo (2011). The approach proceeds in two steps. In the first step we estimate recentered influence functions (RIF, see Firpo, Fortin, and Lemieux (2009)) for the quantiles. The RIF we use is: τ − 1 {y ≤ Q τ } RIF (Y, Qτ ) = Qτ + (4) fy (Qτ ) where 1{} is an indicator function, fy (.) is the density of the marginal density of Y , and Qτ is the population τ -quantile of the unconditional distribution of Y, as in (Fortin, Lemieux, and Firpo (2011), 77). 11 In the second step we decompose the gender income gap in the same way that we would do for the mean, but with income, Yf and Ym replaced with the RIF parameters, obtaining the following equations. In the second step we decompose the gender income gap in the same way that we would do for the mean, but with income, Yf and Ym replaced with the RIF parameters, obtaining the following equations. K K c D = E [Ym ] − E [Yf ] = (αm − αf ) + E [Xmk ] βm − βf + E [Xmk ] − E [Xf k ]βf k=1 k=1 ˆ s =structural ef f ects ∆ ˆ c =composition ef f ects ∆ (5) We do not compute a reweighted-regression decomposition since reweighting pro- cedures are not simple to extend to detailed decompositions, which we are particularly interested in for this paper. Reweighting should be used if the conditional mean is nonlinear. Thus, as pointed out above, we assume a linear wage equation holds in this context. In summary, the RIF technique allow us to perform detailed compositions using the estimated coefficients. 4.2 Quantile Decomposition Results As described in the data section, we include the following covariates in our decom- position analysis: marital status, experience (proxied by age), education, number of children, average monthly hours worked, savings, and industry.6 We use all of the co- variates for Ghana. For Rwanda, Tanzania, and the Republic of Congo we are unable to include savings, as there are insufficient observations on savings. For Uganda we do not have data on hours worked so we do not include average monthly hours worked. The base group used for the decomposition is unmarried men in agriculture/fishing who did not complete primary school and who do not have a savings account. 4.2.1 Aggregate Decomposition Results Figures 8 and 9 show the results for the aggregate decomposition. For each country, the total log monthly income gap is graphed by quantile along with the amount explained by the composition effect and the amount explained by the structural effect. Numerical results are presented in Tables D.1 - D.5 in the Online Appendix. These tables also include the percent of the overall explained by the composition effect versus 6 An additional covariate we were not able to include in the analysis but that may be of interest for future research is risk aversion. Average self-employment earnings for men are larger but more volatile than self-employment earnings for women. It would be interesting to apply the decomposition methodology we present here for quantiles to decompose the variance of self-employment earnings with this variable. We do not decompose the variance of earnings in this paper, though it is possible to do so using the decomposition method outlined in the previous subsection. 12 the structural effect. The standard errors reported in these tables are computed via a nonparametric bootstrap with 100 replications. In what follows, we will describe the results by country highlighting both patterns common across countries and significant divergences from these patterns. In the Republic of Congo, the gender income gap in self-employment is increas- ing across the distribution. The majority of the gap is explained by the structural effect, which is also increasing. By contrast, the fraction of the gap explained by the composition effect is smaller overall and decreasing across the distribution. In fact, the fraction explained by the composition effect decreases from 50.5% in the 10th quantile to 16.2% in the 90th quantile. In contrast, the gender income gap is decreasing across the distribution for Ghana. Both the composition effect and the structural effect are also decreasing across the distribution. As a consequence, the share of the gap explained by the composition effect remains relatively constant across the distribution, at around 30%. Decomposition estimates for Rwanda are less precise. Similarly to the Republic of Congo, the gap is increasing across the distribution. Also similar to the Republic of Congo, the share explained by the composition effect decreases as we move to the right of the distribution, from 46.04% at the 10th quantile to 15.95% at the 90th quantile. In Tanzania, the overall gap is decreasing across the distribution, as was the case in Ghana, from a gap of 0.73 at the 10th quantile to a gap of 0.61 at the 90th quantile. However, unlike in Ghana, the structural effect remains relatively steady while the composition effect decreases across the distribution. As a result, the share of the gap explained by the composition effect decreases across the distribution, from 31% at the 10th quantile to 16.2% at the 90th quantile. In this regard, Tanzania is similar to both the Republic of Congo and Rwanda. Last, in Uganda the income gap between self-employed men and women is u- shaped from the 10th to the 80th quantile, at which point it increases sharply. This is quite different from the other four countries. Interestingly, in Uganda the entire gap is due to the structural effect and is effectively “unexplained”. In fact, the composition effect is actually marginally negative across the distribution, though it is statistically indistinguishable from zero. 13 Figure 8: Decomposition of Total Gender Self-Employment Income Gap into Com- position and Structural Effects 14 Figure 9: Decomposition of Total Gender Self-Employment Income Gap into Com- position and Structural Effects 15 4.2.2 Detailed Decomposition Results Figures 10 and 11 show the contributions of the individual covariates to the compo- sition effect. In the Republic of Congo, differences in choice of industry across men and women is the single most important driver of composition effects for the bottom 50%. In keeping with the aggregate results for the Republic of Congo, where we saw the overall gap attributable to the composition effect decrease over the distribution, industry decreases in importance as we move across the distribution. Differences in education levels and monthly hours worked are the next two most important contrib- utors to the composition effect, though both variables are far less important compared to industry. In Ghana, savings drives the largest component of the composition effect,7 followed by monthly hours worked and industry. While the contribution of savings to the gap increases across the distribution, the contribution of monthly hours and industry decrease marginally. This could indicate that savings is more important for self- employed at the higher end of the distribution, when capital may be necessary to expand businesses. However, an important caveat to these results is reverse causality, since many models of intertemporal optimization would predict that savings rates increase in business earnings. Still, it is interesting that relationship between savings and earnings is stronger at higher quantiles, possibly indicating that lower earners are liquidity constrained and spend all their business earnings. In both Rwanda and Tanzania, monthly hours is by far and away the most impor- tant covariate. In Rwanda, married has a slight contribution and industry actually contributes negatively to the composition effect while all other covariates’ contribu- tions are essentially zero. In Tanzania all other covariates’ contributions are essen- tially zero. Consistent with the aggregate results, where we saw the contribution of the com- position effect to the overall gap decrease across the distribution, we see that the contribution of monthly hours decreases across the distributions for both Rwanda and Tanzania. This implies that while differences in hours worked accounts for siz- able share of the overall gap for the lower quantiles, at higher quantiles there is only a very marginal composition explanation, given our covariates, that can explain the gender income gap at the top of the distribution in Rwanda and Tanzania. As discussed above, in Uganda the composition effect is marginally negative overall across the distribution, though it is not statistically different from zero. In Figure 11 we see that this is driven by differences in industry choices between self-employed men and women and differences in the number of women who have savings accounts compared to men. Figures 12 and 13 break down the aggregrate structural effect into the contribu- 7 It is important to note that savings could also play an important role in the Republic of Congo, Tanzania and Rwanda, but unfortunately we had too much missing data on savings for these coun- tries to include it is a covariate in the decomposition analysis. 16 tion of each covariate. In the Republic of Congo and Ghana, education appears to contribute the most to the overall structural effect. In Rwanda and Tanzania, dif- ference in returns to the same industry contribute the most to the structural effect, particularly at the top of the distribution of self-employment income gaps. In Uganda, differences in the returns to savings accounts contributes the most to the structural effect. Figure 10: Detailed Decomposition of Composition Effects by Country 17 Figure 11: Detailed Decomposition of Composition Effects by Country 18 Figure 12: Detailed Decomposition of Structural Effects by Country 19 Figure 13: Detailed Decomposition of Structural Effects by Country 20 5 Conclusion This paper analyzes the role of individual traits and choices on self-employment in- come gaps between men and women who own businesses. The analysis based on the Republic of Congo, Ghana, Rwanda, Tanzania and Uganda shows that a classifica- tion of entrepreneurs based on broad categories (such as the size of the business) provides a useful categorization. However the relevance of individual specific traits demonstrated so far in this paper shows that additionally considering the traits of individual entrepreneurs helps researchers and policy makers identify the sources of income gaps. Looking at business profitability, the analysis shows that differences in endow- ments of traditional traits such as education, experience, and marriage (which can be seen as a proxy for time availability) do not do a good job in explaining income gaps. This is despite the fact that there are clear differences in relative endowments in these traits when we look at the raw data.8 However, while composition effects overall do not explain the majority of the gap, they do play an imporant role in explaining aggregate income gaps (with the exception of Uganda).9 In place of traditional traits such as education, experience, and marriage, we find that the composition effect acts primarily through covariates such as industry, savings, and monthly hours worked. In Ghana, we find that the share of the gap accounted for by differences in endow- ments is constant across the distribution. Interestingly, while differences in monthly hours and industry choices play a role in explaining the composition effect, the most important covariate is savings. The contribution of savings to the overall gap in- creases across the distribution. This may indicate that women at the top of the female self-employment distribution lag behind the top earning self-employed men due to financial constraints in Ghana. In contrast, we find that the composition effect explains a decreasing share of the income gap for the Republic of Congo, Rwanda and Tanzania. The fact that the gender income gap is increasing (or moderately decreasing in Tanzania) while the share of the gap explained by the composition effect is decreasing may indicate a “glass ceiling” for women in the Republic of Congo, Rwanda, and Tanzania which is associated with discrimination against women at the top end of the earnings distri- bution. Unexplained differences in returns to the same endowments could be due to discrimination; and in the Republic of Congo, Rwanda, and Tanzania, more of the gap is due to unexplained differences in returns to the same endowments at the top of the income distribution.10 We view our primary policy implications as two-fold. Firstly, our results highlight 8 See Tables C.1-C.5 in the Online Appendix. 9 In Uganda the entire gap is due to the structural effect. 10 However, we stress again that while discrimination could be the root cause of the structural effect, we cannot definitively attribute the structural effect to discrimination. This is the case unless we are able to include all possible determinants of income in the set of covariates, which is not possible with our data. 21 the fact that determinants of earnings gaps between men and women vary across the earnings distribution. Accordingly, policymakers who seek to close the gap be- tween high earning businesses should target different variables than if they are trying to help low-earning businesses become more successful. Secondly, while our results are descriptive, we hope they will motivate future research on the causal determi- nants of earnings gaps between men and women. For instance, we argue that the increasing role of structural factors across the earnings distribution in the Republic of Congo, Rwanda, and Tanzania suggests that a glass ceiling is holding back female entrepreneurs in these countries. Future research could test whether this glass ceiling is indeed in place by examining differences in deteriminants of business success such as access to inputs, business networks, and consumers. Additionally, future data collec- tion efforts could collect information on determinants of profits such as risk aversion, ability (both cognitive and non-cognitive), and discount rates, allowing researchers to confirm that the structural differences hold up to controls for a wider range of observable differences in profits. Overall, we hope that our results contribute to both academics and policymakers interested in understanding and closing earnings gaps between male and female entrepreneurs. References (2005-2013): “World Development Indicators,” . Bates, T. (1999): “Existing self-employment: an analysis of Asian immigrant-owned small businesses,” Small business economics, 13(3), 171–183. Chi, W., and B. Li (2008): “Glass ceiling or sticky floor? Examining the gender earnings differential across the earnings distribution in urban China, 1987–2004,” Journal of Comparative Economics, 36(2), 243–263. Chuhan-Pole, P., M. Angwafo, A. D. Mapi Buitano, V. Korman, and M. L. Fox (2011): “Africa Pulse: An analysis of issues shaping Africa’s economic future, vol. 4,” Washington, DC: The World Bank. Croson, R., and U. Gneezy (2009): “Gender differences in preferences,” Journal of Economic Literature, pp. 448–474. De Mel, S., D. J. McKenzie, and C. Woodruff (2009): “Measuring microen- terprise profits: Must we ask how the sausage is made?,” Journal of Development Economics, 88(1), 19–31. DiNardo, J., N. M. Fortin, and T. Lemieux (1995): “Labor market institutions and the distribution of wages, 1973-1992: A semiparametric approach,” Discussion paper, National Bureau of Economic Research. 22 Dupas, P., and J. Robinson (2009): “Savings constraints and microenterprise de- velopment: Evidence from a field experiment in Kenya,” Discussion paper, National Bureau of Economic Research. Fafchamps, M., D. McKenzie, S. R. Quinn, and C. Woodruff (2011): “When is capital enough to get female microenterprises growing? Evidence from a ran- domized experiment in Ghana,” Discussion paper, National Bureau of Economic Research. Fajnzylber, P., W. Maloney, and G. M. Rojas (2006): “Microenterprise dy- namics in developing countries: How similar are they to those in the industrialized world? Evidence from Mexico,” The World Bank Economic Review, 20(3), 389–419. Firpo, S., N. Fortin, and T. Lemieux (2007): “Decomposing wage distributions using recentered influence function regressions,” University of British Columbia (June). (2011): “Occupational tasks and changes in the wage structure,” . Firpo, S., N. M. Fortin, and T. Lemieux (2009): “Unconditional quantile re- gressions,” Econometrica, 77(3), 953–973. Fortin, N., T. Lemieux, and S. Firpo (2011): “Decomposition methods in eco- nomics,” Handbook of labor economics, 4, 1–102. Fox, L., and T. Sohnesen (2012): “Household enterprises in Sub-Saharan Africa: why they matter for growth, jobs, and livelihoods,” World Bank Policy Research Working Paper, (6184). Gosling, A., S. Machin, and C. Meghir (2000): “The changing distribution of male wages in the UK,” The Review of Economic Studies, 67(4), 635–666. Jovanovic, B. (1982): “Selection and the Evolution of Industry,” Econometrica: Journal of the Econometric Society, pp. 649–670. Kilic, T., A. Palacios-Lopez, and M. Goldstein (2013): “Caught in a Produc- tivity Trap,” . Kweka, J., and L. Fox (2011): “The household enterprise sector in Tanzania: why it matters and who cares,” World Bank Policy Research Working Paper Series, Vol. Machado, J. A., and J. Mata (2005): “Counterfactual decomposition of changes in wage distributions using quantile regression,” Journal of applied Econometrics, 20(4), 445–465. Minniti, M. (2005): “Entrepreneurial activity and economic growth,” Global Busi- ness and Economics Review, 1(1), 31–42. 23 Oostendorp, R. H. (2009): “Globalization and the gender wage gap,” The World Bank Economic Review, 23(1), 141–161. Schoar, A. (2010): “The Divide between Subsistence and Transformational En- trepreneurship,” in Innovation Policy and the Economy, Volume 10, pp. 57–81. University of Chicago Press. 24