Is education a determinant of women’s decision-making power within the
household?
A case study of an educational reform in Kenya Ida Villemo Karnström
Supervisor: Joseph Vecci
Master’s thesis in Economics, 30 hec August 2019
Graduate School at the School of Business, Economics and Law, University of Gothenburg, Sweden
Key words: Empowerment, Educational reform, Kenya, Decision-making power ABSTRACT:
Female empowerment is important for economic development. In this paper, I will examine the relationship between education and decision-making power within the household.
Because of the issue of omitted variable bias, I will use a Difference-in-Difference method to test this relationship, using different exposure to the 1985 educational reform in Kenya. The two differences used are age cohorts and if one comes from a district with a low or high mean year of schooling in 1979. The findings show a fairly robust positive effect from the reform on two out of the five questions regarding final say on different decisions, namely the daily decisions. There is mainly no significant effect on the other dependent variables. I present several robustness checks for these results.
1
Acknowledgements
First, I want to express my gratitude towards Sida for the opportunity and scholarship to go on a Minor Field Study to Kenya to write my thesis. It was a great experience and rewarding to be able to see the country and culture I was writing about.
Second, I want to thank my supervisor in Kenya, Dr. Anthony Wambugu at the University of Nairobi, for the generous help I got throughout my stay in Nairobi. I got both access to all the university facilities and help finding the necessary data. Furthermore, I got great support and encouragement at every meeting.
Finally, I want to thank my supervisor in Sweden, Dr. Joseph Vecci, for encouraging me to go to Kenya and for all the feedback, such as comments on the paper for it to achieve its full potential. Furthermore, I want to thank him for the patience and quick responses to questions that arose during the process.
2
Table of contents
Acknowledgements ... 1
List of tables and figures ... 3
1. Introduction ... 4
2. The Educational Reforms in Kenya ... 7
3. Literature review ... 7
4. Theory ... 10
4.1. Hypothesis ... 12
5. Data ... 13
6. Methodology ... 17
6.1. Simple regression model ... 17
6.2. Difference-in-Difference approach ... 19
6.3. The parallel trend assumption ... 23
7. Results ... 24
7.1. Simple regression analysis ... 24
7.2. Difference-in-Difference estimates ... 26
7.2.1. DD estimates using OLS ... 26
7.2.2. DD estimates using other regression models ... 28
7.3. Sensitivity analysis ... 31
7.3.1. Changing the variable young ... 31
7.3.2. Excluding the sample to those always lived in same place ... 34
7.3.3. Changing the variable low mean to fraction finishing primary... 36
8. Pathway analysis ... 38
8.1. Data and Methodology ... 38
8.2. Results ... 40
9. Discussion ... 43
10. Conclusion ... 46
11. Reference list ... 48
12. Appendix ... 51
3
List of tables and figures
Title Page nr
Table 1: Summary statistics of main variables 16
Table 2: Effect on years of schooling 23
Table 3: Simple regression 25
Table 4: Main DD regressions, using OLS 26
Table 5: Main DD regressions, using ordered probit 29 Table 6: Ordered probit - Average Marginal Effects for the interaction term 30 Table 7: Sensitivity analysis: Young – smaller difference in age cohorts 33 Table 8: Sensitivity analysis: Young – later age cohorts 34 Table 9: Sensitivity analysis: Always lived in area 35
Table 10: Fraction finished primary school 37
Table 11: Summary statistics of pathway variables 40
Table 12: Pathway analysis 42
Table A1: Districts summary statistics 51
Table A2: Placebo test
Table A3: Main DD regressions, using multinomial logit
52 53
Table A4: Summary statistics - religion 53
Table A5: Summary statistics - ethnicity 54
Table A6: Summary statistics – childhood place of residence 54
Table A7: Summary statistics – hh owns structure 54
Table A8: Summary statistics – type of roof material 55
Graph 1: Autonomy and education in single years 18
Graph 2: Autonomy and education in single years Graph 3: Autonomy and year of birth
18 21
4
1. Introduction
There is still persistent gender inequality in the world today, especially in developing countries.
This can be seen through for example lower enrolment rates in secondary schools for girls, lower rates of female employment relative to males and the existence of unequal inheritance rights for women (Duflo, 2012). Empowerment of women is an important subject, which has gained focus due to the Millennium Development Goal of reducing gender inequality (Upadhyay et al., 2014). Primary or secondary school are for many African children still the highest level of education received. In Kenya, the enrolment rates have increased in the latest decades, leading to gross enrolment rates over 100 percent for primary schools in Kenya in 2016 (World Bank Data, 2018). However, for secondary school the gross enrolment rate was only 58 percent in 20091, which is a large increase from 40 percent in 1985 (World Bank Data, 2018). For girls, the enrolment rates are lower than for boys. This is especially true in secondary school where the gross enrolment rate was only 34 percent in 1985 and 54 percent in 2009, compared to 45 and 61 percent respectively for boys (World Bank Data, 2018). It is therefore important to consider school reforms that concern children at in primary school, since it may affect the main education many children receive during their lifetimes in developing countries.
An educational reform lengthening primary school by one year, such as the one of interest for this thesis (Kenya 1985), can therefore be crucial for how many years a child goes to school in developing countries.
There is a need to analyse the factors which influence the empowerment of women in order to increase empowerment. In a study conducted in Gambia, Trommlerová et al. (2015) find that among others; age, marital status, economic activity and health are important determinants of empowerment at both the individual and communal level. Several studies analyse how different factors affect empowerment, such as if microfinance projects help women gain more influence in decision-making (Kim et al., 2007; Lakwo, 2006; Leach and Sitaram, 2002). My study aims to evaluate whether education is a determinant of women’s empowerment.2
Increased empowerment can be important for development because of its positive effect on health outcomes. There has been wide-ranging research examining the relationship between
1 2009 is the latest year the World Bank have data on for Kenya’s enrolment rates for secondary schooling, the same is true for other sites.
2 This relationship can be studied through a relative difference in education between the spouses, but since the reform I will study is a general reform one cannot study the relative effect of education, even though women might be more affected by the reform because of their lower enrolment rates in primary and secondary school.
Furthermore, the dataset I will be using to measure empowerment only asks women the questions I am evaluating and I will therefore focus on a general increase in education.
5 fertility and empowerment, where most studies find a positive association between higher empowerment and lower fertility (Upadhyay et al., 2014). Moreover, studies, such as Thomas (1990), finds that exogenous income additions to women lead to larger effects on children’s health than if you give cash transfers to men. Duflo (2003) studies cash transfers to grandparents after the end of the apartheid in South Africa and finds the same conclusion. For all the reasons mentioned, increasing the empowerment of women is important for increased development in low-income countries and it is therefore an important subject.
There are several channels for how education can lead to higher decision-making power within the household, which this study will examine. The bargaining model can be used to explain this through a change in the threat point, which is what the individual can receive if one would be on her or his own. The change can come from for example education affecting the ability to enter the labour market or through narrowing the gap between spouses in terms of education and age. Furthermore, formal education can equip women with new ideas and information, which can improve both their role in decision-making at home and their freedom of mobility (Mahmud et al., 2012). The theoretical framework behind the relationship leads to a hypothesis of a positive effect of education on women’s decision-making power within the household.
There is a large span of literature concerning the relationship between educational reforms and fertility and other measures of empowerment. However, the research on the effect of an educational reform’s on women’s decision-making power within the household is rather limited and this study aims to answer the question: Can more education lead to an effect on women’s decision-making power within the household? Looking at evidence from an educational reform in Kenya 1985.
A contribution with this thesis is to study this relationship in an African setting. Most studies concerning this relationship focus on Asian countries, which can lead to different results than an African setting because attitudes and other factors can vary. In Africa, there are several studies (e.g. Chicoine, 2012, and Ferré, 2009) which consider what effect an educational reform has had, but these studies ignore female empowerment. The previous studies investigating the relationship between decision-making power and education in Africa, concern educational reforms different from the one I will study and are conducted in other African countries than Kenya. Furthermore, another contribution of this study is to use five different questions relating to the decision-making power within the household as dependent variables, instead of just the
6 index of all of them3, which can show if there are different effects on the different measures. I will also categorise the dependent variable into three levels, instead of only using a binary response. This has not been done before and especially not in a Kenyan setting using a reform lengthening primary school by one year.
To study the causal effect of an educational reform on women’s decision-making power within the household I will study an education reform that took place in Kenya in 1985, which lengthened primary school by one year. I will use the Demographic Heath Survey (DHS) conducted in Kenya 2003, which asks married women different questions regarding to their autonomy for decisions within the household.
To study this reform, a Difference-in-Difference (DD) model will be used. Age cohorts are affected differently depending on if they finished primary school before the reform or not. One can therefore compare women born between 1953 to 1964 (untreated), with those born between 1972 and 1980 (treated by the reform), as used by Chicoine (2012). A second variance is how different districts were affected by the reform, depending on their initial mean of years of schooling, since if the district had a lower mean it should have been more affected by the reform.
The main results show a significant positive effect of the reform on two of the five question’s asked regarding the autonomy in the household – the final say on daily needs and what to cook each day. The sensitivity analyses mostly confirm the result. In the pathway analysis, I am not able to determine which channels create this increase in empowerment. The reason for the insignificant result on the other outcome variables may be due to parent’s not investing more in schooling for girls than boys, following the reform, and issues regarding causality.
The remainder of this paper proceeds as follows. In the next section, I will summarise the educational reforms in Kenya and then the existing literature closest to my study, to formulate my research question this study aims to answer. In section 4, the underlying theory will be presented, after which I present the DHS dataset used in the study in section 5. The empirical model will then be presented in section 6, as well as the issues regarding the method. In section 7, the results will be presented followed by a pathway analysis in section 8. Finally, a discussion and a conclusion of this thesis will be accessible in section 9 and 10.
3 Namely the final say on: own health care, large purchases, daily purchases, visits to relatives and what to be cooked each day.
7
2. The Educational Reforms in Kenya
A new educational reform was introduced in Kenya in January 1985. This reform was a change in the required amount of years in school to receive a primary school diploma. A new 8-4-44 system replaced the old 7-4-2-3 system from 1963, hence lengthening primary school by one year, while holding the required amount of years for secondary school intact. the reason for the new educational reform was high unemployment rates for school leavers. The new focus was on English, Mathematics and vocational subjects. The vocational training was meant for those students who would not continue to secondary school, to increase their skills before searching for a job and lowering the risk of unemployment after school (Wanjohi, 2011). The first 6 years of primary school was focused on literacy and numerical skills, while the last two were more focused on vocational training (Sifuna, 1992). Primary school is from the ages of 6 to 14 and secondary school from 14 to 18 years of age, but many starts at later ages (Sifuna, 2007).
In 1973, there was an educational reform in Kenya which abolished the fees for children going to standard 1 to 4 in primary school and lower cost for the later years. This led to an extreme increase in school starters of standard 1 the first year after the reform – 1 million more children instead of the estimated increase of 400 000 children. The huge increase in children led to a shortage of trained teachers, which created pupil-teacher ratios up to 150:1 in some schools and high amounts of unqualified teachers. In the late 1980s, a cost-sharing policy was implemented that lowered quality on primary education and reduced gross enrolment rates. In 2003, primary education became free again in Kenya and also compulsory (Sifuna, 2007).
3. Literature review
There is a large literature concerning educational reforms in relation to fertility and other variables (e.g. Breierova and Duflo, 2004; Diamond et al., 1999 and Duflo et al., 2011).
However, for this study I will explore the relationship between education and empowerment.
Therefore, I will now focus on the literature evaluating this relationship and those closest to my study.
There are multiple studies that try to evaluate the relationship between empowerment and education by focusing on Asian countries. Samarakoon and Parinduri (2015) investigate the relationship between women’s empowerment and education through a fuzzy discontinuity regression design. They analyse the exogenous effect of a longer school year in 1978 on
4 8 years of primary school, 4 years of secondary school and 4 years of university studies.
8 empowerment of women in Indonesia. The study does not find evidence for education improving women’s decision-making authority, asset ownership or community participation.
According to Samarakoon and Parinduri (2015), the lack of significant results can be explained by the already high rates of some decision-making authority for women in Indonesia regarding expenditures and decisions regarding children, as well as high rates of asset ownership, in terms of jewellery and housing.
Another study evaluating this relationship in an Asian setting is Mahmud et al. (2012) who looks at several determinants of empowerment of women in rural Bangladesh. Significant positive relationships are found between years of schooling and one of their two self-esteem indicators and freedom of mobility. Jayaweera (1997) uses macro statistics on Asian countries to see if education affects empowerment of women, seen as for instance participation in decision-making in politics and income earnings, but did not find a significant relationship.
Malik and Courtney (2010) considers the effect of higher education on empowerment in Pakistan. The survey results show that some of the bigger benefits from higher education are increased status and recognition within the family and for achieving economic independence.
This paper will consider an education reform in Africa and how it affects decision-making power within the household, which I will now focus on. There are a number of other studies that focus on education reforms in Africa, of which the following studies focus on Kenya.
Chicoine (2012) investigates the same educational reform as I will evaluate (the reform in Kenya 1985). Several rounds of the DHS survey are used, as well as, other sources of information. The results show a positive effect of education on early use of modern contraceptives, a reduction in marital gap and a reduction in fertility. The method which is used to explore this exogenous variation is an instrumental approach, where year of birth predicts if the student was exposed to the reform using pre-treatment data. Ferré (2009) explores the same reform and also uses the DHS dataset to see if the extension of primary school has an effect on teenage fertility. Ferré (2009) uses a Regression Discontinuity Design (RDD) to explore this relationship and finds that increased primary education leads to a reduction in teenage fertility.
Duflo, Dupas, and Kremer (2010) looks at another school reform in Kenya, which lowered the cost of going to school through free school uniforms, where they found that the reform led to a reduction in dropouts from school and lower fertility rates. None of these studies explicitly assess the effect of the educational reform on empowerment.
There is also research on educational reforms in other countries in Africa, but which have a focus on empowerment. Empowerment can be defined in several ways. One can for instance
9 define empowerment as attitudes towards domestic violence and explore the effect an educational reform has on these attitudes, which some research has done (e.g. Abrahamsson, 2016 and Kishor and Johnson, 2005). Mocan and Cannonier (2012) explore how the educational reform in Sierra Leone 2001, leading to free and mandatory primary school, affect attitudes towards violence in the home and in terms of practice of female genital mutilation. Mocan and Cannonier (2012) uses the DHS dataset and an IV approach where they consider different levels of treatment effect through age during the reform and in terms of funding available for a specific district and find that women affected by the reform, hence have more education, are more intolerant to practices that can hurt their well-being. However, the results show no effect on men’s attitudes.
There are also papers which focuses on the same measure of empowerment as I will do in this thesis. Grépin and Bharadwaj (2015) also use an DHS survey to explore an educational reform in the 1980’s in Zimbabwe to see how maternal education effects child mortality. In addition to this relationship, they also consider other possible channels that can be affected by the education reform. An RDD is used to evaluate it and the study finds a significant effect on fertility. For empowerment, they create an index for the different questions regarding female autonomy in the household in the DHS, as I will use. They find a positive, significant effect on the index from the educational reform. Rakotondrazaka (2014) uses the 2008 Nigerian DHS survey and a DD method to exploit a variation of exposure to a Universal Primary Education program in Nigeria, introduced in 1976, in terms of year of birth and region. Variation in terms of intensity of federal capital funding for classrooms construction is used to see if the region is “high- intensity” or not, where eastern regions received more funding because of low initial primary enrolment rates and educational inputs. The results show an increase in women’s empowerment, but it is not significant. With an instrumental variable approach, using per capita state capital allocation for classroom construction as an instrument for years of schooling, the study finds a significant positive effect of education on women’s autonomy in decision-making (Rakotondrazaka, 2014).
In regard to the existing literature within the field, this paper deviates in several ways. Firstly, another setting will be considered, namely Kenya, which is different from the other research investigating the relationship between education and empowerment. Secondly, I will consider an educational reform that lengthened primary school by one year, not for example free or mandatory primary school reforms which others have used. Thirdly, I will present five dependent variables, to show the various effects on the different questions relating to autonomy
10 in the household, instead of just showing the index. Furthermore, I will show three levels of autonomy in my dependent variable, not only create a binary variable from the responses, which other researchers have chosen to do.
To evaluate the causal effect of an educational reform on empowerment, the research question is as follows:
- Can more education lead to an effect on women’s decision-making power within the household? Looking at evidence from an educational reform in Kenya 1985.
4. Theory
Existing literature within economic research has tried to explain the theory behind the relationship between education and empowerment. However, there is no consensus regarding the channels by which education can affect the empowerment of women.
One theoretical model that can be used to explain this relationship is the bargaining-power model developed by Lundberg and Pollak (1993); Manser and Brown (1980) and McElroy and Horney (1981). In this model, the spouses cooperate within the household according to a bargaining model, which determines the distribution within the household. It is assumed that spouses can have different preferences, such as desired number of kids or amount of food given to the children, which have been proven in several empirical studies (e.g. Duflo, 2003). The household decisions are then decided by bargaining and the outcome is determined by the amount of the spouses’ bargaining power. This bargaining power is said to be a function of the threat point, which can be divorce (Manser and Brown, 1980; McElroy and Horney, 1981) or a noncooperative equilibrium within marriage (Lundberg and Pollak, 1993). The latter was an extension of the model taking into account that one cannot legally divorce in several countries and traditional gender norms are therefore considered instead.
If the spouses fail to agree, they will each receive the utility given at the threat point. Hence, none of them will accept a decision resulting in lower utilities than those they can receive at their individual threat points. The threat point in the divorce threat bargaining model depends on the individual’s separate income, as well as, extra-household environmental parameters described by McElroy and Horney (1981). These are factors affecting divorced men and women, but not the marital utility, e.g. public resources available to divorced women and men and factors affecting the divorce market, such as if women can re-marry and status of divorced women.
11 Education can affect bargaining power within the household since it can affect the threat point for the women and therefore also change the decision-making power within the household. One way in which education can affect the threat point is through higher salary and more opportunities in the labour market, since one more year of schooling can lead to obtaining skills relevant for the job market. The utility at the threat point would then increase, leading to more autonomy within the household when the woman requires more education. Duflo (2001) shows that an educational reform in Indonesia, where an increase in the amount of schools lead to a higher average year of schooling in the affected regions, also led to an increase in salaries.
Another channel could be that education can lead to a decrease in fertility. A decrease in fertility can create more free time for women, since women are responsible for the majority of household work. It can also lead to less childbearing years. This can result in an increased number of women in formal work and higher salaries for women (Duflo, 2012). As mentioned above, this can then increase the bargaining power within the household for women.
Furthermore, a lower teenage fertility could result in fewer dropouts from school, which could affect the labour market opportunities for those girls (Duflo, 2012).
Education can also affect the autonomy within the household through increased access to information (Mocan and Connonier, 2012). An increase in the amount of years studied can lead to higher literacy rates. Being able to read can give access to information through for example magazines and books. It can also help to better process information and lead to an improvement in the skill of interacting with others, hence enhancing their bargaining power (Samarakoon and Parinduri, 2015). Access to information can result in women having better health, through for example learning about how to protect oneself from AIDS and diseases, as well as family planning (Duflo, 2012). It can also lead to women learning about their rights and for them to be able to vote, as well as, becoming aware of the injustices and discrimination women face. In other words, education can give women the tools necessary for a change and hence increase empowerment (Gosh et al., 2015).
A final channel can be through a change in the marital market, through changing who one chooses to marry. An increase in education can lead to women marrying later and finding a partner closer in age, since young married women are often abided in marriage with older men.
It can decrease the risk of girls marring older partners if they finish school early (Duflo, 2012).
An increase in education for women could reduce the educational gap between the spouses, as well as, the age gap. This can be because of women having more say in whom they marry or because they wait with marriage until after school, leading to more equal ages of marriage as
12 well as education. The reduction in marital gaps can lead to higher autonomy for women in the household decision-making, since the spouses are then more equal in terms of skill level and maturity. It would decrease the likelihood of the husband to have a superior bargaining position, since they are then less likely to have gathered more resources than the women (Malhotra and Mather, 1997).
To summarize, there are several channels in which education can affect bargaining power within the household. All of these can affect the threat point for women and therefore lead to increased autonomy within the household according to the power-bargaining model by Lundberg and Pollak (1993); Manser and Brown (1980) and McElroy and Horney (1981).
Education can affect empowerment through: better access to the formal job market, higher salaries, lowering fertility rates, better access to information and lower gaps in spouses age and education levels. All of the channels presented would result in women gaining more autonomy within the household, hence result in an increase in their decision-making power within the household.
4.1. Hypothesis
Considering the underlying theory concerning the relationship between education and decision- making power within the household, as well as the previous literature presented, my hypothesis is that this educational reform had a positive effect on women’s decision-making power in Kenya. Furthermore, I expect the reform to result in: increased salaries; a higher participation in the formal labour market; lower fertility levels; lower rates of teenage fertility; lower marital gap in ages between spouses; higher literacy rates; higher participation rates in political votes and increased knowledge regarding for example transmittable diseases.
Due to lack of data I cannot test all the possible channels in which education can affect women’s decision-making power within the household. Moreover, the effect of education on fertility rate has already been considered for this specific reform by Chicoine (2012) and Ferre (2009), which show a decrease in fertility levels due to more years of schooling. For this reason, this specific channel will not be tested in this thesis. Based on the discussion in the theory, section 4, the main hypothesis in this thesis is the following:
1. Being affected by the 1985 educational reform has a significant, positive effect on women’s decision-making power within the household.
13 The possible testable channels, in this thesis, for how education affects decision-making power within the household are the following5:
a. Being affected by the 1985 educational reform has a significantly positive effect on women’s wealth6, their participation in the formal labour force7 and the portions of household expenditures the respondent’s earnings pay.
b. Being affected by the 1985 educational reform has a significantly positive effect on literacy and frequency of reading the paper/listening to the radio/watching TV.
c. Being affected by the 1985 educational reform has a significantly negative effect on the marital gap in age and education between spouses.
5. Data
In my thesis, I will be using the Demographic and Health Survey (DHS), which measures women’s empowerment through gender attitudes, women’s decision-making power, education and employment of women (DHS, 2018a). I will focus on questions relating to female autonomy in household decisions. The variables I will consider are questions asked to married women in the DHS survey regarding which person who usually decides on: respondent's health care; large household purchases; daily household purchases; visits to family and relatives and food to be cooked each day.8 The answers to the questions concerning the autonomy within the household, used as my dependent variables, are categorised into five different options in the DHS.9 I have chosen this specific measure of autonomy since I consider it to be an important part in females’ empowerment in developing countries. If the woman can get more autonomy at home, it can lead to positive effects for her, such as being able to work more and have more incentives to work if she can decide more about her money. There can also be benefits for the society, such as lower fertility levels since women’s desired number of children has been shown to be lower than the husbands and a larger amount of food for the children (Duflo, 2003).
5 For further discussion regarding exact variables used, see section 8.1.
6 It would have been preferred to consider women’s salaries instead of wealth, since this is more likely to be their own income and not their husband and would therefore be more relevant for their threat point, but unfortunately this information is not available in the DHS for women.
7 Measured as if the woman works; if this is from home or away and what type of earnings the respondent earns.
8 Some of these variables are only available for some of the surveys. In the survey in 2003, a sixth question of what to do with the husband’s money, which is included in other surveys, is not available and I will therefore not use it in this thesis.
9 The answers that can be given to the questions are the following: 1 “Respondent alone”, 2 “Respondent and husband/partner”, 3 “Respondent and other person”, 4 “Husband/partner alone”, 5 “Someone else”, 6 “Could not answer the question” and 9 “no answer” (I will treat answers 6 and 9 as missing values).
14 In the studies by Rakotondrazaka (2014) and Grépin and Bharadwaj (2015) respondents’
answers were transformed into a binary variable: this binary variable was equal to 1 if the respondent made household decisions by herself or jointly with husband, and 0 otherwise. This means that one cannot see the effect of a change of the women’s autonomy from deciding together with her partner to deciding alone, which is an increase in empowerment. Furthermore, one can argue for the same lack of autonomy if the husband alone makes decisions or if someone else makes the decisions. In other words, the answers on if husband alone makes the decisions or if someone else makes them can be grouped together. The same can be said for answers where it is the woman together with someone else (either husband or other person) making the decisions. In this study, I will therefore categorize respondents´ answers into three categories.
This variable will take the value of 3 if the respondent has indicated that she makes a decision alone (answer 1), 2 if respondent has indicated that she makes a decision together with someone else (answers 2 and 3), and 1 if someone else makes the decision for her (answers 4 and 5).10 The five questions on empowerment will be my different dependent variables, but I will also create an index with all of them to have as a dependent variable, which Grépin and Bharadwaj (2015) choose when evaluating this relationship. I will drop the individuals with missing answers on any of the five questions asked in the survey regarding DHS for the index variable and then calculate the mean of the five questions to create the index.11
The questions mentioned are asked in the DHS survey in Kenya in 2003, 2008 and 2014. I will use the survey from 2003 which contains 8195 respondents12. The reason for not using the 2008 and 2014 surveys is that the 2008 survey does not present the different districts’ names because of confidentiality issues, which is necessary for my methodology. The 2014 survey has another division of districts than the one in 2003 and I choose the one in 2003 because it contains women born before 1964 in its sample. The sample is limited to women born between 1953 to 1988.
Each survey contains women between the ages of 15 to 49 years old (DHS, 2018a).
For my independent variables, information about the respondent´s district of residence and what year the participant was born will be used. These variables are provided in the DHS dataset. In the 2003 DHS dataset, Kenya is divided into eight different regions13 and 69 districts.
10 This indicates that I will not only consider the empowerment within the household, but could also mean within the village or tribe since other is not defined. Although, one can consider general empowerment as an increase in the decision-making power within the household.
11 There are 8195 respondents in the 2003 survey, of who 8022 answered all five questions.
12 Of which at least 8081 has responded to at least one of the questions of interest.
13 Nairobi, Central, Coast, Eastern, Nyanza, Rift Valley, Western and North-eastern.
15 Furthermore, in a sensitivity analysis the variable for how many years the respondent has lived in the specific region will be used. This variable ranges from 1-45 years and contains the answer
“always”14.
In my DD analysis, I will include control variables to avoid a non-randomness in terms of different mean years of schooling in the districts and to avoid omitted variable bias. I will include ethnicity, to control for different enrolment rates that may arise because of a certain district having a specific culture in education, which may influence years of schooling in the regions. For the same reason, I will control for religion, since this might influence if one goes to school as well. Moreover, I will control for wealth through a wealth index that is created in the DHS survey using data on for example household ownership of specific assets (DHS, 2018b), since wealth my influence the years of schooling in a specific region. I will also control for the individual’s childhood place of residence, for example if it was in an urban or rural area, since this may also affect the decision of schooling. These control variables are also included in the paper by Mocan and Cannonier (2012). Since wealth might not correctly control for the assets the individual has, I will also try controlling for access to electricity, type of roof and if the household owns the house it lives in.
To conduct my DD approach, I also need information about the initial means of years of schooling in the different districts. Unfortunately, the first Kenyan DHS was in 1989, after the reform of interest. However, a Kenya Population and Housing Census was conducted in 197915 and I have used this census to calculate the means for the different districts. This census is similar to the DHS surveys in terms of questions and execution. To calculate the means for the districts I used the women aged between 15-49, as used in the DHS survey, containing 286 506 observations. The same data and ages were used when considering the fraction who finished primary school. Initial mean years of schooling for the different districts and the fraction finishing primary school is presented in Table A1 in the appendix.
Unfortunately, the district boundaries were not the same in 1979 as in 2003, and I have therefore matched the 69 districts in the 2003 DHS survey to the 41 districts in the 1979 census and used these 41 districts in my methodology. Many of the 41 districts were temporarily divided into several districts during the time period between 1980 and 2000 or changed names, these were
14 2230 individuals have answered always, which will be used in one sensitivity analysis.
15 A survey from 1984 or 1985 would have been preferred, but these surveys are only conducted every 10 years and the next one is therefore 1989, like the DHS. However, since there were no major educational reforms during the period between 1979 to 1985, one can argue that it should be a close measure to the mean years of schooling in the different districts at that time.
16 later divided up into 47 districts which exists today. I have therefore matched the new districts origin to its former name in 1979.16
Table 1: Summary statistics of main variables
Variable Obs Mean Std. Dev. Min Max
Dependent variables
Index of questions of autonomy 8022 1.86 0.69 1 3
Final say on own health care 8150 1.90 0.94 1 3
Final say on making large household purchases
8081 1.51 0.78 1 3
Final say on making hh purchases for daily needs
8088 1.84 0.92 1 3
Final say on visits to family/relatives 8097 1.80 0.85 1 3 Final say on food to be cooked each day 8121 2.27 0.93 1 3
Independent variables
Respondent's year of birth 8195 1974.34 9.34 1953 1988 Mean year of schooling for districts in 1979 819517 4.06 2.14 0.12 7.92 Fraction 7 years of schooling in 1979 819518 0.13 0.06 0.004 0.23
Respondent’s years of schooling 8190 7.10 4.30 0 26
Religion 8185 1.95 0.70 1 6
Wealth index 8195 3.32 1.47 1 5
Has electricity 7904 0.22 0.42 0 1
In Table 1, summary statistics are presented19. For the possible answers to the five autonomy questions, the means ranges between 1.51 and 2.27, with the most decision-making power within the daily decisions and decisions regarding own health care. Final say on large purchases has the lowest mean, indicating that the woman has the least to say about this, compared to the other decisions. All except for the final say on what is to be cooked daily has a mean below 2, meaning that the individual alone is not making the decision.
The respondent’s year of birth ranges between 1953 to 1988 with a mean of 1974. The mean year of schooling ranges between 0.12 to 7.92 between the different districts in 1979, which is quite a large range. This also means that the average woman did not finish secondary school, even in the district with the highest mean year of schooling in 1979. The mean is about 4 years of education, indicating that the average did not finish primary school when considering women
16 This was done through comparing geographical maps and reading about the history of developments of districts, as well as, by help from my supervisor in Kenya.
17 There are 41 districts in the 1979 census, hence 41 means/observations. For each individual, the value is taken from the district she is living in now, creating 8195 observations. For mean values, see table A.1.
18 Same as above.
19 For a description and summary statistics regarding the variables used in the pathway analysis, see section 8.1.
17 between the ages of 15 to 49 in 1979. For the second method, which will be presented in the sensitivity analysis, the fraction of individuals who stopped their education right after finishing primary school has a mean of 0.13 in 1979. In other words, on average 13% stopped their education right after finishing primary school. The district with the lowest fraction has 0.4%
and the highest 23% who just finished primary school. For the individuals in the survey 2003, the average years of education is 7.1 years, so about 3 years higher than in 1979, considering ages between 15-49 years of age.
The wealth index ranges between 1 to 5, with 5 being the richest. The mean is over 3, which stands for the middle income. Furthermore, only about 20 % of the interviewed individuals has electricity at home. The rest of the control variables are not presented in this table, since values such as sum does not give more information when the variables are strictly categorical with more than two values, but are presented in the appendix (Table A4-A8). There are 5 different categories of religions, with over 62% being protestant. There are 15 categories of ethnicity, with the two largest being kikuyu (24%) and luhya (15%). For childhood place of residence, about 77 % grew up on the countryside, while 12 % grew up in the capital and the rest abroad or in a town. For the other asset and income indicators, 70% owns their house, 26% lease it and the rest does not pay rent. For types of roof, the most common is corrugated iron (65%) and then grass (22%). The control variables used in the simple regressions are not presented here, since they are not of particular interest for this thesis since one cannot trust these results even though I have included controls.
6. Methodology
6.1. Simple regression model
To evaluate the association between education on autonomy, I will first implement an OLS regression considering the relationship between years of education on decision-making power.
The regression I wish to estimate is the following:
𝐴𝑢𝑡𝑜𝑛𝑜𝑚𝑦𝑖 = 𝛽0+ 𝛽1𝑦𝑒𝑎𝑟𝑠 𝑜𝑓 𝑠𝑐ℎ𝑜𝑜𝑙𝑖𝑛𝑔𝑖+ 𝛽2𝑋𝑖+ 𝑢𝑖 (1)
where autonomy is the dependent variable for the different empowerment variables mentioned in section 5 and also the index of them. 𝛽0 is the intercept and X indicates the included control variables, for instance different individual characteristics such as year of birth and age at first marriage. 𝑢𝑖 indicates the error term for each individual. I will use the DHS sample of women born between 1953 and 1988.
18 This method will not be able to establish the causal effect of the relationship, since a simple regression in this setting does not control for reverse causality or omitted variable bias. For example, if women choose more years of education when they have strong decision-making power, this would generate reverse causality. There can also be unobserved individual characteristics that affect both the years of education and decision-making power within the household, such as attitudes, using a simple OLS would then present a spurious relationship.
However, it can give an indication of the relationship.
In the two graphs above, I show the relationship between the different dependent variables and years of schooling. In graph 1, the first three dependent variables are presented. For the first 5 years of education, the lines are quite flat. It is then more volatile, but with an upward trend.
After one finishes secondary school, at 12 years, there is a small incline in autonomy, but it then rises with starting university. All of the six dependent variables have a significant dip in autonomy at about 17 years of education, around when the individual finishes university. In graph 2, one can see that the index follows the final say on visits to relatives quite closely, but all of them follow the same trend until after 20 years of education. Just as in graph 1, there is a decline in autonomy after finishing secondary school, but which then starts to increase again of one continues with university studies. In summary, there is no clear pattern between years of education and the six different autonomy variables, but it depends on which level of education the individual finishes. However, there are issues regarding causality.
The RDD method can be an alternative to evaluate this relationship, since the reform only treats a specific age cohort. However, for developing countries, this method has problems since it is not clear at what age the child actually goes to a specific grade, because of repetition of classes, Graph 1: Autonomy and education in single years Graph 2: Autonomy and education in single years
19 lack of knowledge of the exact age of the child and late start in school. One should therefore avoid using RDD in this setting (Chicoine, 2012).20
For these reasons, this study will consider one other empirical method to evaluate the causal effect of education on the woman’s decision-making power within the household. In the next section, I describe how the approach of using an DD addresses the problems related to reverse causality and omitted variable bias mentioned above.
6.2. Difference-in-Difference approach
The study will use the following regression model to estimate the effect of the educational reform on the different dependent variable outcomes:
𝐴𝑢𝑡𝑜𝑛𝑜𝑚𝑦𝑖,𝑐 = 𝛽0+ 𝛽1𝑦𝑜𝑢𝑛𝑔𝑖+ 𝛽2𝐿𝑜𝑤 𝑚𝑒𝑎𝑛𝑐+ 𝛽3𝐿𝑜𝑤 𝑚𝑒𝑎𝑛𝑐∗ 𝑦𝑜𝑢𝑛𝑔𝑖+ 𝛽4𝑋𝑖+ 𝑢𝑖 (2) Young is equal to 1 if the individual is born between 1972 and 1980, hence affected by the reform, and 0 if the individual is born between 1953 and 1964 (for the sensitivity test 1 if individual is born between 1972 to 1976 and 1 if born between 1960 and 1964, along with other age cohorts in the sensitivity analysis). Low mean is equal to 1 if the individual lives in a district which had a mean year of schooling below the general mean for Kenya in 1979, 0 otherwise.
In other words, it is assumed that districts with lower initial amount of years of schooling are more affected by the reform. C is an index indicating which district the respondent is a residence in, while i indicates an index for the individual and 𝑢𝑖 and 𝑋 are defined as before.
In the approach, an ordered probit regression can be used, since the dependent variables are ordinal variables with more than two outcomes. Using a linear regression, like OLS, would assume equal distances between the answers in the dependent variable, which it may not be in the case of categories such as the one in this study. Ordered probit is therefore a better option in this case, since it facilitates this issue and relaxes that assumption (Stock and Watson, 2007).
When considering multiple categories, one can also use multinomial logistic regression. This would not take into account that the dependent variables have ordered answers, but compares each category to a base category. The multinomial logistic regression has an advantage, because
20 Another method used is IV, such as the one used by Chicoine (2012). However, because of the lack of data on grade-one enrolees and repetition at each grade, I will not use this method in this study. Furthermore, I do not have access to data for Kenya regarding the difference in funds dispersed to finance the reform, as used in an IV by Rakotondrazaka (2014) and Mocan and Cannonier (2012). Furthermore, I tried using the same method as e.g.
Abrahamsson (2016) where one use the DD interaction in an IV, but it resulted in a too weak instrument (F-stat below 10).
20 it is better suited when one only has three categories, like in this case, since ordered probit is generally considered to work best with more categories then three.
However, it is harder to interpret the magnitude of the effect when using multinomial logit or ordered probit regression models, since the coefficient in these only represent the parameter estimates and cannot be interpreted ceteris paribus. Usually, the significance levels and signs of the coefficients remain the same in all these models. Because of these reasons, I will mainly consider OLS results. All the results have been replicated using ordered probit and multinomial logistic models as well though.
This empirical approach will give the intention to treat effect (ITT), since all individuals will be analysed that are in a specific age cohort or living district, even though all of them were not affected by this reform. The average treatment effect (ATT) measures the average effect on the treated individuals. For this reason, ITT will be smaller than ATT, hence it will avoid overestimating the results (Gupta, 2011).
My main approach to address the question of whether the educational reform is increasing the empowerment of these women affected, is through a DD strategy. Since the reform affected primary schools, children who had finished primary school did not experience the reform, while those children in primary school when the reform was implemented did. Furthermore, some regions were probably more affected by this reform since they had lower mean years of schooling before the reform, which creates a variation in implementation of the reform that can be explored in my analysis. If a district had a low mean years of schooling before the reform, it is more likely that these individuals was more effected since they would otherwise probably not have done 8 years of school. If instead the mean was above for example 9 years, then the educational reform might not change the decision of how many years one goes to school as much, since most finish primary school anyway.
At the time of the implementation of the reform, school start was possible, but not mandatory, at the age of six. In addition, grade repetition and late starters were frequent. For these reasons several age cohorts were affected by the reform (Chicoine, 2012). This paper will therefore, as have been done in previous studies (Chicoine 2012), have the post-treatment cohorts consist of women born in or after 1972. The ones born in 1972 are 13 years of age at the time of the reform and should therefore be at the end of primary school at the start of the reform if they started school at the age of 6 and did not repeat grades. If one is born before or in 1964, the treatment effect is assumed to be zero, since the women born before this date should have finished primary
21 school before this reform took place. I follow the age cohorts used by Chicoine (2012) studying the same reform.
Optimally, one would want to look at a reform that was carried out in different regions randomly, but since this is not possible in this setting, I have chosen the following method. The first difference will be seen through different age cohorts. I will compare individuals born in 1972 to 1980 to women born in 1953 to 1964, to compare women before and after the reform.
I will exclude the group in the middle, since the women born between 1964 and 1972 might be partly affected by the reform, because of reasons mentioned earlier of overage children and high repentance of grades, which makes it hard to conclude if those women were affected by the reform or not. This method is also used by Chicoine (2012) studying the same reform and for example Mocan and Cannonier (2012) use the same strategy.
Furthermore, in Chicoine (2012) the sample of women is limited to those born before 1980, which I have also limited my sample to, to make age groups more comparable to each other which will increase the probability that the parallel trend assumption holds. Also, I will exclude individuals born after 1980, because these might not have finished school in 2003.
In graph 3, the relationship between the six autonomy variables and year of birth is presented.
One can see a clear downward trend for all of the six variables when the respondent is born later, with some fluctuation between the years. Being younger seems to have a negative effect on autonomy, which is to be expected. If one only compares the difference of young and old, one would most likely then see a negative effect on autonomy from being young (and therefore Graph 3: Autonomy and year of birth
22 more affected by the reform). However, this would lead to bias results since it shows the effect of an older woman probably having higher bargaining power then a young woman, not the actual educational reform. This is why one uses a DD approach, using two differences and focusing on the interaction between them.
The second difference will explore the differences in exposure to treatment through the individual’s place of residence. Different districts have different initial mean of years of schooling before the educational reform, hence the intensity of treatment will vary depending on place of residence, assuming not too high rates of movement across regions as discussed in the sensitivity analysis in section 7.3.
Previous studies have also explored difference in intensity of programs in different regions (e.g.
Duflo 2001, Osili and Long 2008 and Abrahamsson 2016). Usually the reforms for primary schools have been reduced fees or mandatory attainment, which is assumed to lead to increased enrolment rates. Regarding this reform, it most likely affects years of schooling more than enrolment, since it affects the amount of years one must attend in primary school to get a degree.
This will not automatically lead to an increase in enrolment and education. This because some may decide not to go because of the reform, since it now requires 8 instead of 7 years in primary to get the degree. For this reason, I cannot use the same measure of intensity as Abrahamsson (2016) who used the percentage change in enrolment rates from the initial rate, which is why I have chosen to use mean years of schooling instead.
One could also argue for the individuals with a high initial mean year of schooling to be most effected by this reform, since the mean years of schooling in 1979 was 4 years, well below the year 7 of finishing primary school. I will therefore present results for how the interaction term (presenting the effect of the reform) affected years of schooling.
In Table 2, the results for the DID approach on years of schooling is presented both when using below the mean as treatment group (column 1) and above the mean as treatment group (column 2). The results for the interaction term are weak in both cases and only significant at the 10%
level. However, one can see that being young at the time of the reform and coming from a district with low initial mean years of schooling are associated with a positive effect on years of schooling, while it is negative for the ones coming form a district with a high initial mean year of schooling. One reason for this may be that the average years of schooling increased perceptibly in most regions during the years after 1979 (such as seen by the mean year of schooling having increased 3 years in the survey in 2003). Therefore, it could be that the
