2. Education reform in Uganda

Uganda is a developing country located in East Africa. Its GDP per capita was 1,868 USD in 2017. The population is about 41 million, and the fertility rate is 5.5 births per woman. There is a large gender gap in the literacy rate, with 70 percent of females and 82 percent of males being literate. Life expectancy at birth for women is 65 years, while life expectancy at birth for men is 60 years. The education system has not changed fundamentally since Uganda gained its dependence in October 1962 [Uganda Ministry of Education and Sports (1999)].

In 1987, the government of Uganda created the Education Policy Review Commission (EPRC), which was mandated to recommend policies for all levels of education. In 1989, the Commission recommended using the universal primary education program to achieve the transformation of the society and for accelerated growth of the economy. This program, however, was not pursued at that time. In the spring of 1996, the universal primary education program was brought up again during the presidential election by the current president Yoweri Museveni. After being elected, president Museveni announced the implementation of the universal primary education program in January 1997. Several campaigns were carried out before the 1997 school year started in February to notify parents about the program.

The government of Uganda pledged several policies with the introduction of the universal primary education program. One of the major policies was the removal of school fees for all primary grades simultaneously [Grogan (Reference Grogan2008)]. The government was committed to paying 5000 Ugandan Shillings (Ush) for each child per year in primary grades 1–3, and 8100 Ugandan Shillings for each student per year in grades 4–7 (about 5 USD and 8 USD in 1997 Dollars, respectively) [Uganda Ministry of Education and Sports (1999); Noreen and Khalid (Reference Noreen and Khalid2012)].Footnote ⁶ Prior to the reform, the average primary school fee was about 5000 Ugandan Shillings [Deininger (Reference Deininger2003)]. Other costs related to schooling such as transportation, uniforms, books, and school supplies remained the responsibility of parents. The universal primary education program also included other policies such as the construction of classrooms and teacher training [Uganda Ministry of Education and Sports (1999); Grogan (Reference Grogan2008)].

According to Uganda Ministry of Education and Sports (1999), in the first year following the reform, the number of students enrolled in primary schools increased by 58 percent. The gross enrollment rate increased dramatically, from 77 percent in 1996 to 137 percent in 1997. There was a substantial decrease in resources available per student and a large increase in the student–teacher ratio [Grogan (Reference Grogan2008)]. The increase in the number of schools did not keep up with the increase in the number of students. In 1980, there was one school for every 305 students, while in 1999, there was one school for every 722 students. The student–teacher ratio also increased substantially, from 37.62 in 1996 to 63.63 in 1999. Even though the number of trained teachers has improved after the introduction of the reform, the quality of teaching has been negatively affected by the significant increase in the student–teacher ratio [Uganda Ministry of Education and Sports (1999)]. Deininger (Reference Deininger2003) also suggested that there was a general decline in the quality of education following the introduction of the universal primary education program.

The public education system in Uganda consists of 7 years of primary school, 4 years of lower secondary school, and 2 years of upper secondary school. The academic year begins in February and ends in December, and the official school entrance age is 6 years. However, many children enter school at age 7 or 8, with the majority starting school at age 8. Appendix Figure 4 shows the fraction of girls who are enrolled in school by age group prior to the reform using the 1995 wave of Ugandan DHS. Less than 40 percent of girls aged 6 are enrolled in school. Most girls aged 15 and older had dropped out of school.

Appendix Figure 5 presents the percentage of girls by age group with incomplete primary education who are still enrolled in school in 1995. The gap in primary school attendance for girls between the ages of 14 and 15 is 20 percentage points (86 percent vs 66 percent). While the majority of girls aged 6–14 are still enrolled in primary school, there is a sharp drop in the likelihood of attending primary school for girls aged 15 and older. Given this pattern of primary school attendance prior to the reform, it is plausible that the reform had a differential effect on the educational attainment of girls by age, especially at the margin between the ages of 14 and 15. Thus, children aged 14 and younger would have most likely benefited from the abolition of primary school fees. Those aged 15 and older were less likely to have benefited from the reform. This pattern has also been confirmed by Keats (Reference Keats2018).

Prior to the reform, 68 percent of girls aged 15–24 who dropped out of primary school stated that the main reason for dropping was because they could not afford the fees [Uganda Bureau of Statistics and ORC Macro (1995)]. The gender differences in educational achievement was steep: 56 percent of women have ever attended primary school as opposed to 67 percent of men [Uganda Bureau of Statistics and ORC Macro (1995)].

3. Defining women's empowerment

Because of the contested nature of the concept, there are many definitions of empowerment in the literature.Footnote ⁷ These definitions can generally be grouped into three sets of thought. The first set of thought views empowerment in terms of the ability to make choices. Kabeer (Reference Kabeer1999) describes empowerment as “the expansion in people's ability to make strategic life choices in a context where this ability was previously denied to them.” The World Bank also stresses the importance of choices and defines empowerment as “the process of enhancing an individual's and group's capacity to make purposive choices and to transform those choices into desired actions and outcomes.” Alsop et al. (Reference Alsop, Bertelsen and Holland2006) perceives empowerment as “a group's or individual's capacity to make effective choices, that is, to make choices and then transform those choices into desired actions and outcomes.”

The second way of thinking about empowerment is in terms of an increase in power, which refers to control. According to Batliwala (Reference Batliwala1994), empowerment includes control over resources (physical, human, intellectual, and financial) and over ideology (beliefs, values, and attitudes). Rowlands (Reference Rowlands1997) suggests that power can be divided into four categories: power from within (the psychological strength residing in each person that is a basis for self-confidence), power to (the productive power, that is the capability to exert agency without domination), power with (the ability to act in a group), and power over (controlling power). Rowlands further shows that empowerment can be demonstrated within three dimensions: personal, relational, and collective. Personal empowerment refers to developing a sense of self and confidence. Relational empowerment is concerned with the ability to negotiate and influence the decisions made within one's network of relationships. Finally, collective empowerment is about cooperative actions to effect more changes than one could alone.

There are broader ways of understanding empowerment, which focus not only on the ability to act on one's choice, but also the relationship between individuals and institutions. Narayan-Parker (Reference Narayan-Parker2002) defines empowerment as “the expansion of assets and capabilities of poor people to participate in, negotiate with, influence, control, and hold account institutions that affect their lives.” Narayan emphasizes four elements of empowerment, including access to information, inclusion and participation, accountability, and local organizational capacity. This definition has two components: the first component focuses on the expansion of agency (the ability to act on behalf of what you value and have reasons to value) and the second component is related to the institutional environment, which offers people the opportunity to exert agency fruitfully [Ibrahim and Alkire (Reference Ibrahim and Alkire2007)].

In this paper, I define empowerment as having the knowledge to gain control over one's own life. Based on this definition, women's empowerment can be thought as the process by which women, through enhanced knowledge, become aware of their rights and power relations and how these can be altered. My definition is similar to Pinto [2001, as cited in Srivastava (Reference Srivastava2005)], who thinks of empowerment as an internal transformation of one's consciousness that enables one to overcome external barriers to accessing resources or changing traditional ideology. My definition of women's empowerment has two components. The first component refers to having the power to develop a critical awareness of the matters that impact one's well-being. This component focuses on psychological power, that is changes taking place within the person (similar to Rowlands’ power from within). The other component is related to having one's voice heard through involvement in decision making within the household. Essentially, this component is about relational power (similar to Rowlands’ power over), that is changes taking place in power dynamics between a woman and other people in her household.

4. Data

This paper uses data from four waves of Ugandan Demographic and Health Survey (UDHS), administered in 2000, 2006, 2011, and 2016. The UDHS belongs to a series of demographic and health surveys implemented in developing countries.Footnote ⁸ The UDHS is a repeated cross-sectional and nationally representative household survey.Footnote ⁹ Each round of the survey contains information regarding fertility, family planning, maternal health, child health and nutrition, and women's status for women aged 15–49. Each round of the UDHS also provides information on men aged 15–54 and children aged 5–25 who still live at home. The UDHS sample is selected in two stages. In the first stage, enumeration areas are randomly selected from census files. The second stage of sampling involves the random selection of households from the list of households within each enumeration area.

The UDHS surveys are nationally representative, although, in the 2000 wave, some districts in the Western and Northern regions of Uganda were not surveyed due to security problems.Footnote ¹⁰ The 2006, 2011, and 2016 waves cover the entire country. In the 2006 wave, portions of the Northern region of Uganda were oversampled to provide estimates for two special areas of interest: Karamoja and internally displaced persons camps. For these reasons, DHS sample weights are used in all regressions in this paper.

From each of the four surveys, data for all women born between 1975 and 1990 are extracted (i.e., these women were between the ages of 7 and 22 in 1997 when the reform was implemented). I discard women who were younger than 19 years old at the time of the survey, as it is difficult to determine their completed years of schooling. Because questions about decisions within the household are asked to partnered women only (those who are married or living with a man), I further restrict the sample to women who answered questions about decision making.Footnote ¹¹ The description of the pooled data as well as data for treatment and control groups is reported in Table 1. Pooling across all four surveys, the sample consists of 14,503 women who are between the ages of 19 and 41. The average years of education completed for the women in the sample is 5.7 years. Women in the treatment group have more education (6.5 years) than women in the control group (5 years). Approximately 38 percent of women have completed at least primary education, 79 percent of women are in rural areas, and 73 percent of women in the sample are Catholic or Protestant.

Table 1.

Sample description - female sample

Note. The sample is restricted to women who answered questions about decision making (partnered women). These women are aged 19 or older at the time of the survey. The cohorts of women in this table were born between 1975–1990. The 2000 UDHS includes cohorts born between 1975–1982. The 2006 UDHS includes cohorts born between 1975–1987. The 2011 and 2016 UDHS have cohorts born between 1975–1990. Standard errors are in parentheses. Survey weights are used in all estimations.

4.1. Measuring women's empowerment

Alkire et al. (Reference Alkire, Meinzen-Dick, Peterman, Quisumbing, Seymour and Vaz2013) point out a number of issues to be taken into consideration when selecting indicators of empowerment, including whether to focus on measures that are direct or indirect; intrinsic or extrinsic; universal or context-specific; the level of application; and whether to include objective or self-reported measures. In this paper, I focus on direct measures of empowerment at the individual level that are more valid in the context of Uganda.

Eight variables are constructed to reflect the two aspects in my definition of women's empowerment: women's preferences regarding matters related to their well-being and women's voice within the household. The first five indicators seek to assess women's attitudes toward gender-based violence. Violence against women is a violation of basic human rights, and has serious impacts on well-being of women [OECD (2015)]. I use a set of questions asking women if they think it is justified for a man to beat his wife in various situations, such as when she goes out without telling him, neglects the children, argues with him, refuses to have sex with him, and burns the food. For each of these circumstances, I create a binary indicator that takes a value of one if a women disagrees that domestic violence against women is justified, and zero otherwise. To gauge women's involvement in decision making within the household, I use a set of questions asking women whether they have some final say over important decisions within the household, such as issues related to their own heath, large household expenses, and visits to family or relatives. I create a binary empowerment indicator for each of these three decisions and consider all women who, either solely or jointly with their husband, have final say over the decision process as being empowered, and other women as not being empowered.

The summary statistics of the indicators of women's empowerment are displayed in Table 2. Women in the treatment group are more involved in household decisions on family and relative visits, large purchases, and their own health care than women in the control group. As for attitudes toward domestic violence, a lower fraction of women in the treatment group agrees with men's rights to beat their wives in all five situations. For example, about 63 percent of women in the treatment group think that it is not justified for a man to beat his wife if she goes out without telling him, as opposed to 56 percent for women in the control group.

Table 2.

Summary statistics of outcome variables - female sample

Note. The sample is restricted to women who answered questions about decision making (partnered women). These women are aged 19 or older at the time of the survey. The cohorts of women in this table were born between 1975–1990. The 2000 UDHS includes cohorts born between 1975–1982. The 2006 UDHS includes cohorts born between 1975–1987. The 2011 and 2016 UDHS have cohorts born between 1975–1990. Column (4) indicates statistically significant differences between control and treatment groups. Standard errors are in parentheses. Survey weights are used in all estimations.

*p < 0.10, **p < 0.05, ***p < 0.01.

Each sub-index is built by combining indicators that measure the same dimension of women's empowerment. The statistical correlation between the indicators is tested using Cronbach alpha coefficient, which indicates how well a set of indicators describes a single dimension. Cronbach alpha coefficient of 0.65 or higher ensures that indicators are highly correlated and can be aggregated into a summary indicator [Nardo et al. (Reference Nardo, Saisana, Saltelli, Tarantola, Hoffman and Giovannini2005)]. Panel A of Appendix Table 1 reports the Cronbach alpha coefficients of each dimension of women's empowerment. The Cronbach alpha coefficients for women's involvement in household decisions and women's attitudes toward gender-based violence are 0.72 and 0.79, respectively. I also construct a composite index aimed at measuring women's empowerment by combining all eight indicators of empowerment. The Cronbach alpha coefficient for these eight indicators of women's empowerment is 0.74, as shown in Panel B of Appendix Table 1.

The first aggregation method used to formulate the sub-indices and composite index is principal component analysis (PCA).Footnote ¹² The PCA allows me to aggregate the indicators using a data-driven weighting scheme [OECD (2015)]. The first principal component (FPC) of a set of variables is the linear combination of all variables that captures the largest amount of information that is common to all of the variables [Filmer and Pritchett (Reference Filmer and Pritchett2001)]. I perform PCA for each of the four surveys to check for consistency in terms of factor loading. Next, I create a sub-index for each dimension of women's empowerment as well as a composite index to measure the empowerment of women as a whole in each survey separately. The sub-index for women's involvement in household decisions is the first principal component of the three indicators of women's involvement in decision making within the household. Similarly, the sub-index for women's attitudes toward gender-based violence is the first principal components of five indicators measuring women's acceptability of domestic violence. The women's empowerment index is the first principal component of the eight indicators of women's empowerment. The results from the PCA showing factor loading on the first component for two sub-indices and the composite index are reported in Appendix Tables 2 and 3, respectively.

Although FPC is helpful in extracting common information contained in the indicators, the estimate of the regression using FPC as a dependent variable is hard to interpret. To ease interpretation, I construct z-scores for FPC sub-indices and composite index. In addition to PCA, as a robustness check, I construct sub-indices using Alkire-Foster (AF) method. In the first step, I transform the indicators into binary form, each of which takes a value of one if a woman is considered to be empowered, and zero otherwise. Each indicator is given a weight, and then an aggregated measure is created by computing the weighted sum of the indicators. In the second step, I choose a cut-off point to transform the aggregated indicator into an index to measure empowerment. The cut-off gives the index a value being either zero or one, thus making it easier for interpretation. To create the sub-index for women's involvement in household decisions, I give an equal weight to each indicator of this dimension and then sum across the three indicators. If the weighted sum is greater or equal to 33.33 percent, the sub-index takes the value of one, and zero otherwise. Essentially, this means that if a woman involves in one of three decisions, she is empowered, otherwise she is not. Similarly, if a woman says that it is not justified for a man to beat his wife in two out of five situations, the sub-index for women's attitudes toward gender-based violence is equal to one, and zero otherwise. The composite index for women's empowerment takes the value of one if the weighted sum of eight indicators is greater than or equal to 50 percent, and zero otherwise.

5. Empirical strategy

The objective of this paper is to investigate the impact of education on women's empowerment. Using OLS to estimate this relationship is likely to produce biased estimates due to unobserved factors that can affect both years of schooling and women's empowerment. In order to address this endogeneity issue, the paper exploits an exogenous shift in women's educational attainment induced by the education reform that eliminated primary school fees in 1997. Note that this reform only abolished primary school fees without making primary education mandatory. Thus, this type of reform generates the local average treatment effect (LATE) among women who chose to receive more education due to the lower costs of attending school. These women had a high desire to obtain additional education but could not do so because they were financially constrained.

I identify cohorts of women who were exposed to the reform using their birth year. As stated earlier, the official school entrance age is 6 years. However, many children delay school enrollment till the age of 8. Since it takes 7 years to complete primary education, women who were 14 years or younger at the time of the reform would have most likely benefited from the reform. Therefore, I consider those who were born between 1983 and 1990 as part of the treatment group. Women aged 15 and older were less likely to be treated by the reform. Thus, those who were born between 1975 and 1982 constitute the control group. Essentially, I compare the outcomes of women born between 1983 and 1990 who were exposed to the reform to those of older women born between 1975 and 1982 who did not benefit from the reform.

I use an instrumental variable approach to estimate the effect of education on women's empowerment, as shown in equations (1) and (2) below:

(1)$$\eqalign{Schooling_{ibtr}& = \beta_{0} + \beta_{1}Treated_{b} + \beta_{2}Treated_{b} \times( Yob_{i} - 1983) \cr & \quad + \beta_{3}( 1-Treated_{b}) \times( Yob_{i} - 1983) + {\bf X_{ibtr}^{\prime}\Pi} + \eta_{t} + \mu_{r} + \upsilon_{ibtr}}$$

(2)$$\eqalign{Y_{ibtr}& = \alpha_{0} + \alpha_{1}\widehat{Schooling_{ibtr}} + \alpha_{2}Treated_{b}\times( Yob_{i} - 1983) \cr& \quad + \alpha_{3}( 1-Treated_{b}) \times( Yob_{i} - 1983) + {\bf X_{ibtr}^{\prime}\Pi} + \eta_{t} + \mu_{r} + \epsilon_{ibtr}}$$

where i indexes women, b is birth cohort, t stands for year of the interview, and r is region of residence. Schooling _ibtr denotes the number of years of schooling completed for woman i from cohort b residing in region r. Y _ibtr represents one of the eight indicators of women's empowerment for woman i born in cohort b, as shown in Table 2. Treated _b is a dichotomous variable which captures whether a woman was exposed to the reform (i.e., it takes a value of 1 if a woman was born between 1983 and 1990 and 0 if a woman was born between 1975 and 1982). The vector X_ibtr consists of personal characteristics of woman i, including rural residence and indicators of religion. η _t and μ _r are year of survey and region of residence fixed effects, respectively. While the indicators of the region of residence (Northern, Central, Eastern, and Western) account for unobserved attributes of the region that are correlated with women's empowerment, the year of survey fixed effects control for effects from pooling data across four surveys.

I use a bandwidth of eight years, meaning eight cohorts on both sides of the cutoff cohort (born in 1983), although using other bandwidths does not significantly change the results, as shown in the Robustness section. The bandwidth is selected to be small enough to minimize the bias due to time trend, while still allowing for an adequate sample size. Yob _i is the woman's year of birth. Treated _b × (Yob _i − 1983) and (1 − Treated _b) × (Yob _i − 1983) control for potentially differential trends in the outcomes of treatment and control groups. Standard errors are clustered by the enumeration census area-by-year of birth.

There are two assumptions that ensure the validity of the estimates of the instrumental variable approach. The first assumption is that the instrument is highly correlated with the number of years of schooling completed for women. I provide the F-statistics obtained from the first stage (Equation 1) to show the strength of the instrument. The second assumption, which is exclusion restriction, assumes that, conditional on other observables, the instrument only impacts the outcome variables through increasing education for women. I examine possible violations of this assumption in the Robustness section.

7. Potential mechanisms

In this section, I explore potential channels that link women's education and women's empowerment. Existing literature offers a number of pathways through which increased schooling for women may affect women's empowerment. Education can improve women's cognitive ability, which enhances their ability to seek and understand information regarding their rights and the benefits of standing up for themselves [Glewwe (Reference Glewwe1999); Smith-Greenaway (Reference Smith-Greenaway2013)]. Alternatively, education can influence women's empowerment by providing women with access to information, either formally through books and teachers or informally through interactions with peers. As for peers, attending school helps create friendships with other girls, who can then share information. Education can empower women via other mechanisms, such as time preference [Becker and Mulligan (Reference Becker and Mulligan1997)] and increased labor market opportunities and wages [Psacharopoulos and Patrinos (Reference Psacharopoulos and Patrinos2004)]. In the area of decision making, education can increase the bargaining power of women within the household by improving their economic resources and by endowing them with the knowledge and skills to make informed choices that enhance their welfare [Duflo (Reference Duflo2012)]. Iyigun and Walsh (Reference Iyigun and Walsh2007) develop a model to show that the marital bargaining power over the household matters depends on the relative earnings of husband and wife. Moreover, an educated woman has a higher ability to prove her credibility in terms of being able to make good and informed decisions.Footnote ¹⁴ I explore the existence of some of these mechanisms.

First, I examine the impact of education on women's cognitive ability, proxied by women's literacy. A dichotomous variable is generated based on the question that evaluates the reading ability of women by DHS. It takes a value of one if a woman was able to read the entire sentence or parts of the sentence and zero otherwise. As mentioned in the Education Reform in Uganda section, the sudden access shock triggered by the introduction of the reform has generally decreased the quality of education. Thus, it is plausible that the reform might not necessarily improve learning. For example, Cannonier and Mocan (Reference Cannonier and Mocan2018) find that the reform in Sierra Leone has no effect on women's literacy. Table 1 shows that 59 percent of women in the sample are literate. Column (1) of Table 5 reports the result of the effect of education on literacy. One extra year of education increases women's propensity to be literate by 10 percentage points. This translates to an increase in the probability of being literate by 17 percent.

Table 5.

Potential mechanisms - female sample

Note. The sample is restricted to women born between 1975 and 1990 who answered questions about decision making (partnered women). All regressions control for age trend differentiated by treatment status, rural residence, indicators of religion, and survey year and region of residence fixed effects. Standard errors, clustered by the census enumeration area-by-year of birth, are in parentheses. Survey weights are used in all estimations.

*p < 0.10, **p < 0.05, ***p < 0.01.

Second, I investigate the effect of education on women's labor market opportunities. I construct three variables to capture women's labor market outcomes. First, a binary variable, Working, takes a value of one if a woman was employed in the last twelve months and zero otherwise. Second, among women who were employed in the last twelve months, I create a dichotomous variable, Working for others, that equals one if a woman was employed by others and zero if a woman worked for a family member or was self-employed. Third, to measure women's types of earnings, I generate a dummy variable which takes a value of one if a woman who worked in the last twelve months received cash for her work and zero if a woman was paid in kind or not paid. Columns (2)–(4) of Table 5 show that women's schooling has no effect on their employment in the last twelve months, and that it has no effect on women's propensity to be employed by others, and that it has no impact on the probability that a woman is paid in cash for her work. It is possible that an increase in women's education, where the base level is low, has no impact on women's labor market opportunities. International Monetary Fund (2019) shows that secondary and higher education matter more when it comes to the probability of getting a job and having a paid job in Uganda.Footnote ¹⁵

8. Robustness