Scholarly article on topic 'The effect of increased primary schooling on adult women's HIV status in Malawi and Uganda: Universal Primary Education as a natural experiment'

The effect of increased primary schooling on adult women's HIV status in Malawi and Uganda: Universal Primary Education as a natural experiment Academic research paper on "Sociology"

CC BY-NC-ND
0
0
Share paper
Academic journal
Social Science & Medicine
OECD Field of science
Keywords
{HIV/AIDS / "Sub-Saharan Africa" / "Universal Primary Education" / "Sexual behavior" / "Natural experiment"}

Abstract of research paper on Sociology, author of scientific article — Julia Andrea Behrman

Abstract This paper explores the causal relationship between primary schooling and adult HIV status in Malawi and Uganda, two East African countries with some of the highest HIV infection rates in the world. Using data from the 2010 Malawi Demographic Health Survey and the 2011 Uganda AIDS Indicator Survey, the paper takes advantage of a natural experiment, the implementation of Universal Primary Education policies in the mid 1990s. An instrumented regression discontinuity approach is used to model the relationship between increased primary schooling and adult women's HIV status. Results indicate that a one-year increase in schooling decreases the probability of an adult woman testing positive for HIV by 0.06 (p < 0.01) in Malawi and by 0.03 (p < 0.05) in Uganda. These results are robust to a variety of model specifications. In a series of supplementary analyses a number of potential pathways through which such effects may occur are explored. Findings indicate increased primary schooling positively affects women's literacy and spousal schooling attainment in Malawi and age of marriage and current household wealth in Uganda. However primary schooling has no effect on recent (adult) sexual behavior.

Academic research paper on topic "The effect of increased primary schooling on adult women's HIV status in Malawi and Uganda: Universal Primary Education as a natural experiment"

SOCIAL SCIENCE

Contents lists available at ScienceDirect

Social Science & Medicine

journal homepage: www.elsevier.com/locate/socscimed

The effect of increased primary schooling on adult women's HIV status in Malawi and Uganda: Universal Primary Education as a natural experiment

Julia Andrea Behrman

Department of Sociology, New York University, 295 Lafayette Ave, 4th Floor, New York, NY 10012, USA

CrossMark

ARTICLE INFO

Article history:

Available online 21 June 2014

Keywords: HIV/AIDS

Sub-Saharan Africa Universal Primary Education Sexual behavior Natural experiment

ABSTRACT

This paper explores the causal relationship between primary schooling and adult HIV status in Malawi and Uganda, two East African countries with some of the highest HIV infection rates in the world. Using data from the 2010 Malawi Demographic Health Survey and the 2011 Uganda AIDS Indicator Survey, the paper takes advantage of a natural experiment, the implementation of Universal Primary Education policies in the mid 1990s. An instrumented regression discontinuity approach is used to model the relationship between increased primary schooling and adult women's HIV status. Results indicate that a one-year increase in schooling decreases the probability of an adult woman testing positive for HIV by 0.06 (p < 0.01) in Malawi and by 0.03 (p < 0.05) in Uganda. These results are robust to a variety of model specifications. In a series of supplementary analyses a number of potential pathways through which such effects may occur are explored. Findings indicate increased primary schooling positively affects women's literacy and spousal schooling attainment in Malawi and age of marriage and current household wealth in Uganda. However primary schooling has no effect on recent (adult) sexual behavior.

© 2014 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND

license (http://creativecommons.org/licenses/by-nc-nd/3.0/).

1. Introduction

Over 35 million people infected with the human immunodeficiency virus (HIV) currently live in sub-Saharan Africa (UNAIDS, 2013). Given the magnitude of the epidemic there has been extensive focus on how to reduce infection rates for adolescents and young-adults between the ages of 15 and 24, a group identified as particularly at risk for new infection (UNAIDS, 2008). The literature suggests that adolescent African females face a special set of vulnerabilities. In part this is attributed to social norms that put women, particularly young women, at greater risk, for example the prevalence of relationships between older males and much younger females. In part the gender difference in HIV rates may reflect that women are biologically more susceptible to contracting the virus through sexual intercourse (Padian et al., 1991). In Africa today women comprise over 60 percent of HIV infections, a dramatic shift from the early days of the epidemic when infection rates were higher in men (UNAIDS, 2013).

Young females face heightened risk for infection just at the point when they leave school, thus increasing girls schooling has been

E-mail addresses: julia.behrman@gmail.com, Jab965@nyu.edu.

put forward as a mechanism to reduce HIV transmission. It is argued that more schooling will enable young women to lessen the possibility of transmission through greater knowledge of risks and more capacity for reducing risks through condom use or other preventative behavior. The focus on schooling in the HIV literature corresponds with a more general trend in the global development community on expanding access to schooling for children, particularly girls, throughout the developing world. Perhaps most visibly, the second Millennium Development Goal calls for universal completion of primary school and gender equity in school access (United Nations, 2001).

In this paper I explore the causal relationship between girls' primary schooling and their adult HIV status in Malawi and Uganda using the implementation of Universal Primary Education (UPE) policies in the mid-nineties as natural experiments. Using data from the 2010 Malawi Demographic Health Survey (DHS) and 2011 Uganda AIDS Indicator Survey (AIS), I model the relationship between primary schooling and adult HIV status using an instrumented regression discontinuity approach. While there has been substantial interest in the relationship between schooling and HIV status in Africa, very few studies have been able to infer causality. I further contribute to the literature by focusing on the relationship between primary, as opposed to secondary, schooling and HIV outcomes.

http://dx.doi.org/10.1016/j.socscimed.2014.06.034

0277-9536/© 2014 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/).

Studies focusing on the effect of secondary school on HIV status may suffer from selection bias since it is a different, potentially less vulnerable population, who is able to attend secondary school in the first place. Furthermore, I am able to extend existing studies, most of which have relatively short follow-up periods, by exploring the effect of schooling on young-adult HIV status. Comparing the experiences of Uganda and Malawi is important for establishing the external validity of findings given that the trajectory of the epidemic took different forms in each country.

2. Literature review

In a comprehensive review of the literature Jukes et al. (2008) develop a framework to theorize three different pathways through which formal schooling can affect HIV infection. Firstly, schooling may affect the "socio-cognitive" determinants of sexual behavior, which include knowledge, attitudes and perceived control over behavior. Because schools are often the site of information dissemination campaigns, increased schooling may increase knowledge about HIV and risk reduction. The cognitive and non-cognitive skills developed in school may also change the way individuals process risk and increase perceived control over behavior. Additionally, schooling may affect HIV infection by influencing social and sexual networks. Finally, schooling may lead to improved economic circumstances, which could prompt changes in sexual behavior. For women, this might include delays to age of first sex and age of marriage. However it is not clear that increased socio-economic status necessarily decreases risk of infection. In the early years of the HIV/AIDS epidemic higher socio-economic status was associated with riskier sexual behavior for men and also to some extent for women (Lopman et al., 2007; Weltz et al., 2007; Mishra et al., 2007).

Evidence on whether changes in knowledge and attitudes about HIV risk lead to behavioral change is mixed. In Tanzania, an adolescent sexual health program in 20 communities was found to significantly affect knowledge and attitudes about HIV but did not reduce HIV infection in either the short or long term (Ross et al., 2007; Doyle et al., 2010). On the other hand, a randomized experiment in Kenya finds the provision of information on relative risk of infection by age led to a reduction in teen pregnancy (a proxy for unprotected sex) and a shift from older to younger partners (Dupas, 2011). However, the same study found the Kenyan government's HIV risk campaign, which focused on abstinence as a strategy for risk reduction, had no effect on adolescent sexual behavior, thus indicating that the type of information or messaging may matter for behavioral change.

There is some evidence to suggest that schooling delays the onset of sexual activity, marriage and/or childbearing, all of which potentially put women at risk for infection. In a randomized experiment in Western Kenya girls who received an educational subsidy for the last three years of primary school were significantly less likely to drop out of primary school, marry, or start child-bearing even following the end of the subsidy (Duflo et al., 2012). Using an instrumental variable approach, Osili and Long (2008) exploit variation in the introduction of Universal Primary Education in Nigeria. They find increasing female schooling by one year reduces early pregnancy by 0.26 births. Ozier (2011) uses performance on the national Kenyan secondary school entrance exam to conduct a regression discontinuity analysis, finding that secondary schooling leads to a reduction in adolescent pregnancy for females. In Uganda, Alsan and Cutler (2013) instrument secondary school enrollment with homestead distance from school and find girls' enrollment in secondary school significantly increases their likelihood of abstaining from sexual activity.

The association between schooling and HIV in the African context has been documented in a number of cross-sectional

studies (Glynn et al., 2004; Pettifor et al., 2005; de Walque, 2007; Hargeeves et al., 2008). However, provision of credible evidence of a causal relationship between schooling and HIV status is limited to a handful of studies. In a cluster randomized trial of a cash transfer program designed to keep adolescent girls (ages 13—22) in school, Baird et al. (2012) find HIV prevalence in the combined treatment group, which included conditional and unconditional cash transfer beneficiaries, to be 1 —2 percent as compared to 3 percent in the control group at the end of an 18-month exposure period. Baird and colleagues attribute their findings to changes in respondent self-reported sexual behavior. Transfer beneficiaries report decreased frequency of sexual activity and younger sexual partners then non-beneficiaries. Interestingly, the program does not affect other aspects of sexual behavior, including age of sexual debut or prevalence of unprotected sex.

In another experimental study, Duflo et al. (2012) evaluate the effects of a randomized HIV intervention amongst 13 year olds from 328 schools in Western Kenya over a seven-year period. The intervention compared three approaches to behavioral changes; (1) educational subsidies; (2) an HIV information campaign focused on abstinence; or (3) a combination of both approaches. The second approach was found to be least effective, failing to reduce pregnancy or sexually transmitted infections, including HIV and HSV-2. The first approach reduced adolescent girls' rates of school exit, pregnancy and marriage but not the rates of sexually transmitted infection. The third approach had larger effects on reductions in sexually transmitted infections, but less pronounced effects on schooling and self-reported sexual behavior than the first approach. Additional analysis of partnership data leads the authors to conclude that pregnancy and STI infection are related to the level of commitment of adolescent relationships, not just prevalence of unprotected sex.

3. Universal Primary Education in Malawi & Uganda

The primary school systems in both Malawi and Uganda follow the legacy of the British colonial system. In both countries primary school is officially supposed to run from ages 6 to 13, though as in much of Africa grade repetition is common and many children start late and end early. Both countries were early implementers of Universal Primary Education Policies aimed at increasing primary school enrollment principally through the elimination of school fees. In 1991 the primary school Gross Enrollment Ratio (GER), the number of children who are actually in school divided by the number of children who are of school age, was 71 percent for Uganda and 75 percent for Malawi (UNESCO, 2011a, 2011b). By 2011 these figures had risen to 110 percent for Uganda and 140 percent for Malawi (UNESCO, 2011a, 2011b). The percent for this ratio can be greater than 100 because students outside of the primary school age range may still be in school.

Starting in 1991 Malawi adopted a sequential approach to elimination of school fees, which entailed provision of fee waivers for grade one in the first year of implementation and for grade two in the second year (Kattan, 2006; World Bank, 2009). In 1994 this plan was replaced by a "big bang" approach in which primary school fees were eliminated across all grades effective starting September 1994. The government also eliminated "indirect" fees, such as contributions to school development funds, and the requirement to wear uniforms to school. The Ministry of Education, the agency charged with carrying out implementation, launched a mass media campaign in July of 1994 to ensure that the public was aware of the policy. In addition to elimination of fees a number of other changes were put into place to deal with the influx of new students; local languages become the dominant language of instruction for early grades, 20,000 new teachers were recruited and

received expedited training and the budgetary allocation to the education sector significantly increased. In the year following the 1994 "big bang" implementation 1 million new children entered primary school, an increase in enrollment of over 50 percent (Kattan, 2006; World Bank, 2009). Because entry was allowed at any grade increases were actually largest in the older grades. In the final grade of schooling (standard 8) enrollment increased by 79 percent (Kattan, 2006; World Bank, 2009).

In contrast to Malawi's initial sequential approach, Uganda adopted a "big bang" approach to elimination of school fees from the start (Deininger, 2003; Kattan, 2006; Grogan, 2008). All school fees were eliminated at all levels for all primary school children effective January 1st 1997. The policy eliminated both direct school fees and indirect school fees such as donations to parents funds. School uniforms, another impediment to school access for the poor, were also eliminated. The implementation of this policy was a highly decentralized process with individual districts across the country taking charge of carrying out implementation. A large nation-wide information dissemination campaign was launched and included target messages specifically focused on educating girls. In an evaluation of the program Deininger (2003) finds evidence that UPE led to dramatic increases in primary school attendance and reduced gender and income inequities in primary school attendance across the country.

In both countries an important critique of UPE policies is that improving the access to education came at the expense of deteriorating quality of primary schooling. In both countries the student-to-teacher ratio rose fairly significantly and critics lamented that new teachers were poorly trained or unprepared (Deininger, 2003; Kadzamira and Rose, 2003). In some regions the government did not follow through on promise to provide promised materials and parents were forced to shoulder this cost (Deininger, 2003; Kadzamira and Rose, 2003). All of this has important implications for school quality.

4. Data

Data for this paper comes from the 2010 round of the Malawi Standard Demographic Health Survey (DHS) and the 2011 round of the Uganda AIDS Indicator Survey (AIS). The DHS and AIS are nationally representative cross-sectional household-based surveys collected by ICF International in collaboration with host country governments. Standardized questionnaires facilitate cross-national comparisons.

In the 2010 Malawi survey a nationally representative cross-sectional sample of 27,345 households was interviewed from urban and rural areas. In one third of these households, HIV tests were also administered to all men and women in the household between the ages of 15 and 49. The female HIV sample for Malawi consists of 19,363 women, of whom nine percent (n = 1834) tested positive for HIV in 2010 (five respondents were excluded due to indeterminate test results). The analysis in this paper focuses on a sub-sample of 1445 Malawian women born between 1978 and 1983 fifteen percent (n = 220) of whom tested positive for HIV in 2010 (Table 1).

The Ugandan AIS collects much of the same demographic information as in the standard DHS survey, but with an additional focus on sexually transmitted infections (STIs). All women and men ages 15—49 in the household are eligible to be interviewed and tested as part of the AIS. The full Ugandan sample includes 11,967 women of whom eight percent (n = 944) tested positive for HIV in 2011 (one respondent was excluded due to indeterminate test results). The analysis in this paper focuses on a sub-sample of 2309 Ugandan women born between 1981 and 1986, nine percent (n = 205) of whom tested positive for HIV in 2011 (Table 1). In both

Table 1

Descriptive statistics.

Not exposed to UPE Exposed to UPE at 13 at 13

N Mean SD N Mean SD

Malawi

Years schooling completed 626 5.08 3.97 819 5.87 3.67

Catholic 626 0.20 0.40 819 0.18 0.38

Christian (non-Catholic) 626 0.68 0.47 819 0.72 0.45

Muslim 626 0.11 0.31 819 0.10 0.30

Other religion 626 0.01 0.12 819 0.01 0.08

Chewa ethnicity 626 0.28 0.45 819 0.30 0.46

HIV positive in 2010 626 0.18 0.39 819 0.13 0.34

Age at first intercoursea 581 16.61 2.72 758 16.68 2.72

Age at first marriageb 613 17.76 3.32 803 17.48 2.93

Age difference between spousesc 503 5.58 4.74 676 5.63 4.56

Condom used last intercoursed 548 0.06 0.24 736 0.07 0.25

Number sexual partners last 623 0.89 0.34 819 0.90 0.31

12 monthse

Poor on DHS wealth index 626 0.42 0.49 819 0.38 0.49

Spouses years schoolingf 606 6.75 3.96 787 7.27 3.90

Literateg 625 0.65 0.48 818 0.73 0.44

Uganda

Years schooling completed 1080 5.32 4.05 1229 6.58 4.21

Catholic 1080 0.42 0.49 1229 0.41 0.49

Christian (non-Catholic) 1080 0.31 0.46 1229 0.32 0.47

Muslim 1080 0.01 0.09 1229 0.02 0.13

Other religion 1080 0.26 0.44 1229 0.25 0.43

Baganda ethnicity 1080 0.15 0.36 1229 0.17 0.38

HIV positive in 2011 1080 0.10 0.31 1229 0.07 0.26

Age at first intercourseh 1070 16.89 2.83 1211 16.98 2.58

Age at first marriagei 1028 17.51 4.05 1112 17.95 3.46

Age difference between spousesj 868 6.77 6.03 967 6.76 6.17

Condom used last intercourse1' 988 0.08 0.27 1115 0.08 0.27

Number sexual partners last 1080 0.95 0.36 1226 0.96 0.67

12 monthsl

Poor on DHS wealth index 1080 0.41 0.49 1229 0.36 0.48

a 4 respondents excluded because never had sex; 102 respondents excluded by DHS due to inconsistent response. b 29 respondents excluded due to missing data.

c 244 respondents excluded due to missing data; 22 respondents excluded due to inconsistent response. d 161 respondents excluded due to missing data. e 3 respondents excluded because responded "don't know". f 16 respondents excluded because don't know; 36 respondents excluded due to missing data. g 2 respondents excluded due to missing data.

h 21 respondents excluded because never had sex; 7 respondents excluded by DHS due to inconsistent response. i 169 respondents excluded due to missing data.

j 430 respondents excluded due to missing data; 44 respondents excluded due to inconsistent response. k 206 respondents excluded due to missing data. l 3 respondents excluded due to missing data.

countries test refusal rates were very low (less than one percent) in the full sample of men and women. Thus there is minimal concern about non-response bias.

This analysis focuses on women rather than men due to women's increased biological and social vulnerability to HIV acquisition. Indeed, in the full DHS sample of HIV positive respondents in Malawi 63 percent (n = 890) are female and 37 percent (n = 530) are male. Similarly in the full Uganda sample 63 percent (n = 944) of HIV positive respondents are female and 37 percent (n = 551) are male. These gender patterns hold for the sub-samples of interest as well. Women's HIV status also has important implications for intergenerational transmission of HIV via mother to child transmission. Furthermore, female children have historically faced more difficulties in accessing school than their male counterparts. Thus we would expect the jump in schooling that accompanies UPE policies to be higher for girls than boys.

5. Empirical strategy

In this paper I use a regression discontinuity design (RDD) to assess the causal effect of primary schooling on adult HIV status. The RDD is a special case of an observational study design in which assignment to treatment (exposure to UPE) is decided solely based on values of one measured variable, Z (birth cohort), with all values of Z on one side of a specified cutoff, c (13 and younger at policy implementation), given the treatment and all observations with values on the other side denied the treatment (Angrist and Pischke, 2009). The RDD takes advantage of the fact that girls just above primary school age and just below primary school age at the year of policy implementation will be comparable on both observed and unobserved characteristics and differ only in their exposure to the UPE policy. Thus, girls 13 and younger may be able to continue or extend their schooling thanks to elimination of fees, but girls over 13 will not have this opportunity. This empirical strategy offers a number of advantages over Ordinary Least Squares (OLS) regression because it is likely that there are characteristics such as socioeconomic status, cognitive ability or personal preferences that predict both schooling attainment and probability of HIV infection. If so, then OLS estimates overstate the schooling impact because schooling is partially proxying for these omitted characteristics (see OLS estimates in Tables A2 and A3).

However, in East Africa grade repetition is common and it is not unusual for children to start school late or end school early. Some girls who are beyond primary school age will nonetheless be exposed to UPE because they are still in primary school when UPE is introduced. Likewise, some girls who are primary school age will not attend primary school even with the elimination of school fees because truancy is not enforced, child labor is often needed for household livelihood activities, or education for girls is not valued by households. In order to deal with this noncompliance I use an instrumental variable to estimate the effect only for those who complied with primary school age limits (e.g. were exposed to UPE if primary school age or younger and were not exposed to UPE if above primary school age). I model the relationship using a two stage instrumental variable probit estimator where the outcome is a dichotomous indicator of HIV status at survey and the treatment is a continuous variable that measures number of years of school completed. The treatment in instrumented using a dichotomous indicator of exposure to UPE while age 13 or younger. In Malawi exposure is assigned to girls born in 1981 and later and in Uganda exposure is assigned to girls born in 1984 and later. In the first stage, equation (1), I regress D, the treatment, on Z, the randomly assigned instrument. In the second stage, equation (2), I regress Y, the outcome, on the predicted value of D from the first stage.

Di = a0 + a1Zi + ■■■akXk + vi

Yi = b0 + b\Di + ...fikXk + Ei

Malawi and Uganda are religious and ethnically heterogeneous countries. Religion is represented by four indicator variables for (1) Christian (non-Catholic); (2) Catholic; (3) Muslim; and (4) Other religion. I use Christian (non-Catholic) as the reference category. Uganda and Malawi have over 20 different ethnicities recorded in the DHS and AIS. To capture ethnicity I create an indicator to represent whether the respondent is a member of the "dominant" ethnic group of each country, or the group with the largest response rate. In Uganda this is the Baganda tribe and in Malawi this is the Chewa tribe. Family wealth also likely determined who benefited most from the policy change, with girls from poorer families benefitting more than wealthier peers. However, I am unable to control for family wealth because the DHS does not

collect retroactive information on wealth. I do not control for current levels of household wealth or any other post-treatment variables as they may be endogenous to the model. For example, women with increased schooling may have greater wealth because they have better labor market opportunities or improved marriage matches due to higher school attainment.

The estimand of interest for this analysis is the complier average causal effect (CACE), a special case of the Local Average Treatment Effect. In equation (3) Z denotes the randomly assigned instrument, Y(0) and Y(1) stand for potential outcomes, Y denotes the observed outcome, D(0) = D(Z = 0) and D(1) = D(Z = 1) denote potential "program exposure" under each assignment of the instrument, D denotes observed participation behavior (in other words the treatment).

E[Y (Z = 1)-Y (Z = 0)] E[D{Z = 1) -D(Z = 0)]

Inference is made over the interval of birth cohorts specified for compliers, who are girls who would have adhered to their treatment assignment no matter which side of the age threshold they fell on. I must assume that there are no additional pathways through which exposure to UPE could affect HIV status. That is to say, I must assume that Zi is uncorrelated with both ei and vi. If this is not the case, then it will be impossible to know whether schooling, as opposed to these alternative pathways, affects HIV status. This approach relies heavily on the assumption of igno-rability of the instrument, in other words that the instrument, exposure to UPE, is randomly assigned. The validity of this assumption depends considerably on the width of the interval of birth cohorts I consider around the specified cutoffs of age at exposure. Correctly specifying that this interval is adequately narrow is important to ensure treatment and control groups are comparable and differ only in their extent of exposure to UPE. At the same time, the interval must also be large enough to allow for adequate sample sizes, particularly in light of concerns that not all girls adhere to their treatment assignment. In the main analysis I focus on the sub-sample of girls born three years above and below the specified cutoff. As a robustness check, I also run models with birth cohort intervals two and five years above and below the cutoff and find comparable results (see Table A1).

In order to use UPE exposure as an instrument for years of schooling UPE must actually increase primary school attendance, in other words cov(Zi, Di) s 0. Graphically, evidence of the discontinuity can be seen in Figs. 1 and 2. Since I have a single instrument and single endogenous regressor the instrument can be considered relevant if the t-value for the instrument is larger than 3.2 or the corresponding p-value below 0.0016 or the F for the excluded instrument is greater than 10 (Stock et al., 2002). I test this empirically in the first stage and find results to be highly significant at the p < 0.001 level (Table 2). I also explore the possibility that UPE may most affect girls at the start, as opposed to the end, of their schooling career by running a separate second set of analyses for a younger sub-sample of girls using a different instrument that indicates exposure to UPE for the entire schooling career (ages 7 or younger) as opposed to exposure for only part of the schooling career (ages 8 and older). First stage results indicate this is a weak instrument thus I do not pursue this strategy (results available upon request from the author).

One potential concern with this approach is that children may be HIV positive prior to starting school and may leave school early because they are ill or because families are less interested in investing in schooling for sick children. Given the trajectory of the epidemic, rates of mother-to-child transmission should be low for

Table 2

First stage results predicting women's total years of schooling using 2010 Malawi DHS and 2011 Uganda AIS.

Fig. 1. Average years of completed schooling by birth cohort in Malawi. Source: Created by the author

children born in the late 1970s to the mid 1980s, though it is difficult to know precisely due the paucity of data for this period. Nonetheless, the poor quality of treatment in both countries well into the 21st century means that it would have been nearly impossible for an HIV positive child born in the 1980s to survive to be interviewed as an adult in 2010/2011. Thus for the women in this sample the prime pathway of HIV transmission should be through sexual intercourse which, barring instances of childhood sexual abuse, would occur in adolescence or adulthood. In the Malawi sample the average age of first intercourse is 16.61 (SD 2.72) and in the Uganda sample the average age of first intercourse is 16.89 (2.83). Amongst the treated sample, I find very small numbers of girls who report sexual initiation prior to exposure to the policy (n = 12 for Malawi and = 14 for Uganda). I run analyses including and excluding these respondents and find no difference in results; thus in the final analysis they are included.

6. First-stage results

In the first stage I regress years of schooling completed on exposure to UPE. In Malawi, I find girls exposed to UPE have an average of 0.8 more years of schooling than girls not exposed

Variables (1) Malawi (2) Uganda

Exposure age 13 0.80*** (0.20) 1.18*** (0.17)

Catholic 1.08*** (0.25) -0.63** (0.21)

Muslim -1.54*** (0.33) 0.56 (0.75)

Other religion -0.65 (0.78) 0.00 (0.22)

Chewa ethnicity -0.70** (0.23)

Baganda ethnicity 3.34*** (0.27)

Constant 5.24*** (0.20) 5.07*** (0.19)

Observations 1445 2309

R-squared 0.05 0.12

Robust standard errors in parentheses (clustered at the enumeration area level). ***p < 0.001, **p < 0.01, *p < 0.05.

(Table 2, column 1). This result is highly significant at the p < 0.001 level and represents a fairly large jump in a country where overall levels of schooling are low. A graphical representation of this discontinuity can be seen in Fig. 1. In Uganda, I find girls exposed to UPE have an average of 1.18 more years of schooling than girls not exposed (p < 0.001) (Table 2, column 2). A graphical representation of this discontinuity can be seen in Fig. 2.

7. Second-stage results

In the second stage, I regress adult HIV positive status on the predictions from the first stage. I use probit regression (ivprobit) to model HIV status and other dichotomous outcomes. Because coefficients from probit models are difficult to interpret, I present the marginal effects for the probit model (for coefficients from the probit model see Table A1). In Malawi, the marginal effect (-0.06) is negative and significant at the p < 0.01 (Table 3, column 1). This indicates that a one year increase in schooling for a girl in Malawi decreases the probability of testing positive for HIV in 2010 by 0.06 compared to if she had not attended for that year. In Uganda the effect is also negative (-0.03) and is significant at the p < 0.05 level (Table 3, column 2). These results indicate that a one year increase in schooling for a girl in Uganda decreases the probability of testing positive for HIV in 2011 by 0.03 as compared to if she had not attended the extra year. These results are robust to the specification of alternative birth cohort intervals (see Table A1).

8. Extensions: exploration of impact pathways

I explore a number of pathways through which schooling could affect HIV status using the same strategy to make causal inference. I start by exploring the effect of schooling on a number of indicators of adolescent/early adult sexual behavior related to risk of acquiring HIV. In Malawi I find no significant effect of school on age of sexual debut, age at marriage or age difference between the respondent and current spouse (Table 4, columns 1—3). However, I do find a

Table 3

ivprobit (marginal effects) second stage estimates predicting women's HIV positive status using 2010 Malawi DHS and 2011 Uganda AIS.

Variables (1) Malawi (2) Uganda

Years schooling -0.06** (0.02) -0.03* (0.01)

Catholic 0.06 (0.04) -0.02 (0.02)

Muslim -0.08* (0.04) 0.03 (0.09)

Other religion 0.01 (0.11) -0.02 (0.02)

Chewa ethnicity -0.10*** (0.02)

Baganda ethnicity 0.14* (0.06)

Observations 1445 2309

Fig. 2. Average years of completed schooling by birth cohort in Uganda. Source: Created by the author

Robust standard errors in parentheses (clustered at the enumeration area level). ***p < 0.001, **p < 0.01, *p < 0.05.

Table 4

ivprobit (marginal effects) and ivreg second stage estimates predicting additional pathways using 2010 Malawi DHS.

Variables (1) Age first sex (2) Age marriage (3) Spouse age (4) Spouse years (5) Condom used (6) Number (7) Poor (8) Literate

ivreg ivreg difference school last sex partners ivprobit (ME) ivprobit (ME)

ivreg ivreg ivprobit (ME) ivreg

Years school 0.09 (0.20) -0.34 (0.25) 0.05 (0.32) 0.63** (0.21) 0.00 (0.01) 0.02 (0.02) -0.05 (0.03) 0.07*** (0.02)

Catholic 0.32 (0.26) 0.62+ (0.35) -0.38 (0.48) 0.41 (0.32) 0.02 (0.02) -0.06* (0.03) 0.03 (0.05) -0.01 (0.03)

Muslim -0.60 (0.40) -1.20* (0.48) 1.37* (0.66) -0.77 (0.42) -0.06*** (0.01) -0.01 (0.05) 0.06 (0.07) -0.05 (0.06)

Other religion 0.90 (0.97) 0.67 (1.27) -1.25 (1.32) -1.75 (1.05) 0.01 (0.07) 0.03 (0.06) 0.08 (0.12) 0.02 (0.06)

Chewa ethnicity 0.69** (0.22) -0.01 (0.26) -0.20 (0.32) -0.70** (0.23) -0.04** (0.01) 0.04 (0.02) 0.04 (0.04) 0.03 (0.02)

Constant 15.96*** (1.15) 19.46*** (1.40) 5.32** (1.79) 3.81** (1.18) 0.80*** (0.13)

Observations 1339 1416 1179 1393 1284 1442 1445 1443

Robust standard errors in parentheses (clustered at the enumeration area level). ***p < 0.001, **p < 0.01, *p < 0.05.

positive effect (0.63) of schooling on the respondent's spouses years of completed schooling that is significant at the p < 0.01 level (Table 4, column 4). In Uganda, I find a positive effect (0.43) of schooling on age of marriage that is significant at the p < 0.01 level (Table 5, column 2). However, there is no significant effect of school on age of sexual debut or spousal age difference (Table 5, columns 1 & 3). Information on spousal schooling attainment is not collected in the Uganda AIS. Next I explore the effect of schooling on recent adult sexual activity including: (1) whether the respondent used a condom during last sexual encounter; and (2) number of sexual partners (including spouse) in the last 12 months. Results are not significant in either country, suggesting a limited effect of schooling on these later adult behaviors (Tables 4 and 5).

Finally, I explore the possibility that schooling affects HIV status by increasing wealth. In Uganda, I find that the schooling has a significant negative effect (-0.04) on the respondent's current household being characterized as "poor" on the DHS generated household wealth index (p < 0.05) (Table 5 column 6). In Malawi, school has no such effect on the household wealth index (Table 4 column 7). However, in Malawi I do find that schooling has a significant positive effect (0.07) on respondent's literacy in 2010 (p < 0.001) (Table 4 column 8). Literacy is relevant as a crude measure of skills or cognitive development that transpires in school. Literacy data were not collected in the 2011 Uganda AIS.

9. Discussion

Building on the earlier work of Pettifor et al. (2005), de Walque (2007), Hargeeves et al. (2008) and others on the negative association between girls schooling and HIV status I use an instrumented RDD to causally demonstrate the negative effect of schooling on adult women's HIV status. I extend the findings of Baird and colleagues (2012) by demonstrating the effect of primary school, as opposed to secondary school, on adult HIV status. This is an

important distinction because many students, particularly female students, do not reach secondary school due to financial or family constraints. In addition, primary school ends just at the point that many young adolescent females are about to engage in sexual activity for the first time.

My investigation of impact pathways provides some preliminary insight into how schooling affects women's adult HIV status. In Uganda there is a positive effect of schooling on age at marriage, supporting the perspective that increased schooling delays relationships that may put women at risk. Likewise, in Malawi school has a positive effect on the respondent's spouses schooling attainment, signaling that increasing girls schooling leads to marriages with different types of partners. However, there is no effect of schooling on age of sexual debut, age difference between spouses or recent adult (self-reported) sexual behavior in either country.

I also explore the possibility that schooling affects HIV status by increasing wealth. In Uganda, schooling has a negative effect on the respondent's current household being characterized as "poor" on the DHS generated household wealth index. In Malawi, school has no effect on wealth as measured by the DHS index. However, school does have a positive effect on literacy, a crude measure of school-based skill/cognitive development. This suggests that in spite of concerns about poor school quality, schooling may be imparting girls with skills that will aid in other life domains. Ultimately, this is an incomplete exploration of impact pathways as there are likely a multitude of pathways through which schooling affects adult HIV status. Further research should better explore the complicated pathways through which schooling affects HIV status, with particular attention to variation between contexts.

Interestingly, OLS and probit regression estimates of the relationship between schooling and self-reported adolescent and adult sexual behavior provide quite different results from the instrumented RDD (Tables A2 and A3). In both countries, there is a significant positive association between schooling and age at sexual

Table 5

ivprobit (marginal effects) and ivreg second stage estimates predicting additional pathways using 2011 Uganda AIS.

Variables (1) Age first sex ivreg (2) Age marriage ivreg (3) Spouse age difference ivreg (4) Condom used last sex ivprobit (ME) (5) Number partners ivreg (6) Poor ivprobit (ME)

Years school 0.07 (0.09) 0.43** (0.14) -0.06 (0.28) -0.00 (0.01) 0.01 (0.02) -0.04* (0.02)

Catholic 0.24 (0.15) 0.33 (0.21) -0.18 (0.38) -0.01 (0.02) -0.04 (0.03) 0.10*** (0.03)

Muslim 0.65 (0.42) 0.38 (0.71) 0.32 (1.17) 0.06 (0.07) -0.06 (0.08) -0.07 (0.09)

Other religion -0.37** (0.14) -0.40* (0.20) 0.26 (0.39) -0.02 (0.01) -0.03 (0.03) -0.03 (0.03)

Baganda ethnicity 0.04 (0.35) -0.72 (0.49) 0.54 (1.00) 0.11 (0.06) -0.07 (0.08) -0.32*** (0.04)

Constant 16.51*** (0.51) 15.38*** (0.81) 6.99*** (1.56) 0.92*** (0.09)

Observations 2281 2140 1835 2103 2306 2309

Robust standard errors in parentheses (clustered at enumeration area level). ***p < 0.001, **p < 0.01, *p < 0.05.

Note: Spouse schooling and literacy questions not collected in 2011 AIS.

debut, age at marriage and condom use during last intercourse and a significant negative relationship between schooling and age difference between spouses. Likewise, probit regression provides somewhat different estimates of the association between school and HIV status. In Uganda there is a significant negative association between schooling and HIV positive status; however the marginal effect is considerably smaller then that generated by the RDD (Table A3 column 1). In Malawi this association is not significant (Table A2 column 1). The differences in findings between the RDD and OLS estimates demonstrate how past explorations of the relationship between schooling, sexual behavior and HIV status may be biased due to unobserved heterogeneity.

The largest limitation in this analysis is the possibility that older women may have higher rates of HIV infection simply because they have had longer time to be exposed to sexual activity and the risk of HIV transmission. Some of this is accounted for by narrowing the interval around the cutoff age to ensure that respondents are not so far apart in age and comparable on a variety of dimensions; the robustness of results across different specifications of birth cohort intervals lends credibility to findings (see Table A1). More causal

work is needed to expand the evidence base on the effect of schooling on adult HIV status and other important health outcomes. Nonetheless, these results suggest that schooling has important impacts on health and other outcomes outside of the labor market that should be considered by policy makers and researchers in evaluating the returns to increasing schooling.

Acknowledgments

I am grateful to Jere Behrman, Amber Peterman, Florencia Torche, Paula England and Jennifer Hill for helpful commentary on earlier versions of this paper. Support for this study was provided by the grant Team 1000+ Saving Brains: Economic Impact of Poverty-Related Risk Factors for Cognitive Development and Human Capital "0072-03" provided to the Grantee, The Trustees of the University of Pennsylvania by Grand Challenges Canada.

Appendix

Table A1

ivprobit second stage results predicting women's HIV positive status across different birth cohort intervals using 2010 Malawi DHS and 2011 Uganda AIS

Variables (1) Malawi (2) Malawi (3) Malawi (4) Uganda (5) Uganda (6) Uganda

1976-1985 1978-1983 1979-1982 1979-1988 1981 -1986 1982-1985

Years schooling -0.19*** (0.03) -0.18*** (0.04) -0.22*** (0.04) -0.12*** (0.03) -0.14*** (0.04) -0.15* (0.06)

Catholic 0.14+ (0.08) 0.18+ (0.10) 0.18 (0.11) -0.15* (0.07) -0.10 (0.09) -0.14 (0.11)

Muslim -0.26** (0.10) -0.30* (0.13) -0.42** (0.14) -0.04 (0.29) 0.13 (0.40) 0.28 (0.44)

Other religion -0.32 (0.29) 0.02 (0.34) -0.07 (0.41) -0.11 (0.08) -0.11 (0.10) -0.20+ (0.12)

Chewa ethnicity -0.36*** (0.07) -0.32*** (0.09) -0.26** (0.10)

Baganda ethnicity 0.49*** (0.11) 0.59*** (0.17) 0.59* (0.23)

Constant 0.43+ (0.25) 0.39 (0.39) 0.73 (0.47) -0.56* (0.23) -0.37 (0.36) -0.20 (0.57)

Observations 2445 1445 942 3774 2309 1567

Robust standard errors in parentheses (clustered at the enumeration area level). ***p < 0.001, **p < 0.01, *p < 0.05.

Table A2

OLS & probit (marginal effects) estimates predicting HIV status and additional pathways using 2010 Malawi DHS

Variables (1) HIV positive (2) Age first sex (3) Age marriage (4) Spouse age difference (5) Condom used last sex (6) Number partners (7) Poor

OLS OLS OLS OLS Probit (ME) OLS Probit (ME)

Years school 0.00 (0.00) 0.25*** (0.02) 0.33*** (0.02) -0.16*** (0.04) 0.00** (0.00) -0.00 (0.00) -0.05*** (0.00)

Catholic -0.01 (0.02) 0.17 (0.19) -0.04 (0.18) -0.16 (0.32) 0.02 (0.02) -0.05+ (0.02) 0.03 (0.03)

Muslim 0.00 (0.03) -0.33 (0.23) -0.16 (0.26) 1.06* (0.48) -0.06*** (0.01) -0.03 (0.03) 0.07 (0.05)

Other religion 0.07 (0.10) 0.98 (0.92) 0.99 (1.12) -1.34 (1.22) 0.01 (0.07) 0.02 (0.07) 0.08 (0.12)

Chewa ethnicity -0.06** (0.02) 0.81*** (0.16) 0.40* (0.16) -0.32 (0.29) -0.04*** (0.01) 0.03 (0.02) 0.04 (0.03)

Constant 15.02*** (0.15) 15.67*** (0.17) 6.51*** (0.30) 0.90*** (0.02)

Observations 1445 1339 1416 1179 1284 1442 1445

R-squared 0.14 0.17 0.03 0.01

Robust standard errors in parentheses (clustered at the enumeration area level). ***p < 0.001, **p < 0.01, *p < 0.05.

Table A3

OLS & probit (marginal effects) estimates predicting HIV status and additional pathways using 2011 Uganda AIS

Variables (1) HIV positive (2) Age first sex (3) Age marriage (4) Spouse age difference (5) Condom used last sex (6) Number partners (7) Poor

OLS OLS OLS OLS Probit (ME) OLS Probit (ME)

Years school -0.00** (0.00) 0.24*** (0.02) 0.36*** (0.02) -0.14*** (0.04) 0.01*** (0.00) -0.00 (0.00) -0.05*** (0.00)

Catholic -0.00 (0.01) 0.36** (0.14) 0.29 (0.19) -0.23 (0.33) -0.00 (0.01) -0.05 (0.03) 0.09*** (0.03)

Muslim 0.01 (0.07) 0.52 (0.39) 0.43 (0.70) 0.37 (1.19) 0.05 (0.07) -0.05 (0.07) -0.06 (0.09)

Other religion -0.02 (0.01) -0.36** (0.14) -0.40* (0.20) 0.24 (0.39) -0.02+ (0.01) -0.03 (0.03) -0.03 (0.03)

Chewa ethnicity 0.04+ (0.02) -0.52** (0.17) -0.51* (0.21) 0.82+ (0.43) 0.06** (0.02) -0.03 (0.03) -0.29*** (0.02)

Constant 15.54*** (0.15) 15.74*** (0.19) 7.47*** (0.37) 1.00*** (0.03)

Observations 2309 2281 2140 1835 2103 2306 2309

R-squared 0.13 0.14 0.01 0.00

Robust standard errors in parentheses (clustered at the enumeration area level). ***p < 0.001, **p < 0.01, *p < 0.05.

References

Alsan, M., Cutler, D.M., 2013. Girls education and HIV risk: evidence from Uganda. J. Health Econ. 32, 863-872.

Angrist, J., Pischke, J.S., 2009. Mostly Harmless Econometrics: an Empiricists' Companion. Princeton University Press, Princeton USA.

Baird, S., Garfein, R.S., McIntosh, C., Ozler, B., 2012. Effect of a cash transfer programme for schooling on prevalence of HIV and herpes simplex type 2 in Malawi: a cluster randomised trial. Lancet 379,1320-1329.

Deininger, K., 2003. Does cost of schooling affect enrollment by the poor? Universal primary education in Uganda. Econ. Educ. Rev. 22, 291-305.

de Walque, D., 2007. How does the impact of an HIV/AIDS information campaign vary with educational attainment? Evidence from rural Uganda. J. Dev. Econ. 84, 686-714.

Doyle, A.M., Ross, D.A., et al., 2010. Long-term biological and behavioural impact of an adolescent sexual health intervention in Tanzania: follow-up survey of the community-based MEMA kwa Vijana trial. PLoS Med. 7 (6), 1-14.

Duflo, E., Dupas, P., Kremer, M., 2012. Education, HIV, and Early Fertility: Experimental Evidence from Kenya (Unpublished results).

Dupas, P., 2011. Do teenagers respond to HIV risk information? Evidence from a field experiment in Kenya. Am Econ. J. Appl. Econ. 3(1), 1-36.

Glynn, J.R., Carae, M., Buve, A., Anagonou, S., Zekeng, L., Kahindo, M., Musonda, R., 2004. Does increased general schooling protect against HIV infection? A study in four African cities. Trop. Med. Int. Health. 9 (1), 4-14.

Grogan, L., 2008. Universal primary education and school entry in Uganda. J. Afr. Econ. 18 (2), 183-211.

Hargreaves, J.R., Bonell, C.P., Boler, T., et al., 2008. Systematic review exploring time trends in the association between educational attainment and risk of HIV infection in sub-Saharan Africa. AIDS 22, 403-414.

Jukes, M., Simmons, S., Bundy, D., 2008. Education and vulnerability: the role of schools in protecting young women and girls from HIV in southern Africa. AIDS (Suppl. 4), S41-S56.

Kadzamira, E., Rose, P., 2003. Can free primary education meet the needs of the poor? Evidence from Malawi. Int. J. Educ. Dev. 23, 501-516.

Kattan, R.B., 2006. Implementation of Free Basic Education Policy. In: The World Bank Education Working Paper Series, vol. 7. World Bank, Washington DC.

Lopman, B., Lewis, J., Nyamukapa, C., Mushati, P., Chandiwana, S., Gregson, S., 2007. HIV incidence and poverty in Manicaland, Zimbabwe: is HIV becoming a disease of the poor? AIDS 2, S57-S66.

Mishra, V., Assche, S.B., Greener, R., Vaessen, M., Hong, R., Ghys, P.D., Boerma, J.T., Van Assche, A., Khan, S., Rutstein, S., 2007. HIV infection does

not disproportionately affect the poorer in sub-Saharan Africa. AIDS 21,

517-S28.

Osili, U.O., Long, B.T., 2008. Does female schooling reduce fertility? Evidence from Nigeria. J. Dev. Econ. 87 (1), 57-75.

Ozier, O., 2011. The Impact of Secondary Schooling in Kenya: a Regression Discontinuity Analysis (Unpublished results).

Padian, N.S., Shiboski, S.C., Jewell, N.P., 1991. Female-to-male transmission of human immunodeficiency virus. J. Am. Med. Assoc. 266,1664-1667.

Pettifor, A.E., Rees, H.V., Kleinschmidt, I., et al., 2005. Young people's sexual health in South Africa: HIV prevalence and sexual behaviors from a nationally representative household survey. AIDS 19,1525-1534.

Ross, D.A., Changalucha, J., et al., 2007. Biological and behavioural impact of an adolescent sexual health intervention in Tanzania: a community-randomized trial. AIDS 21,1943-1955.

Stock, J.H., Wright, J.H., Yogo, M., 2002. A survey of weak instruments and weak identification in generalized method of moments. J. Bus. Econ. Stat. 20 (4),

518-529.

United Nations (UN), 2001. Road Map towards the Implementation of the United Nations Millennium Declaration: Report of the Secretary-general. United Nations Department of Public Information, New York.

United Nations Programme on HIV/AIDS (UNAIDS), 2008. Report on the Global AIDS Epidemic. UNAIDS, Geneva.

United Nations Educational Scientific and Cultural Organization (UNESCO), 2011a. Malawi Country Statistics. http://stats.uis.unesco.org/unesco/TableViewer/ document.aspx?ReportId=121&IF_Language=en&BR_Country=4540 (accessed 04.12.13.).

United Nations Educational Scientific and Cultural Organization (UNESCO), 2011b. Uganda Country Statistics. http://stats.uis.unesco.org/unesco/TableViewer/ document.aspx?ReportId=121&IF_Language=en&BR_Country=8000 (accessed 04.12.13.).

United Nations Programme on HIV/AIDS (UNAIDS), 2013. UNAIDS Global Report on the AIDS Epidemic. United Nations, New York. http://www.unaids.org/en/ media/unaids/contentassets/documents/epidemiology/2013/gr2013/UNAIDS_ Global_Report_2013_en.pdf (accessed 04.12.13.).

Welz, T., Hosegood, V., Jaffar, S., Batzing-Feigenbaum, J., Herbst, K., Newell, M.L., 2007. Continued very high prevalence of HIV infection in rural KwaZulu-Natal, South Africa: a population-based longitudinal study. AIDS 21 (1), 467-1472.

World Bank, 2009. Abolishing School Fees in Africa: Lessons from Ethiopia, Ghana, Kenya, Malawi and Mozambique. Research Report. The World Bank, Washington D.C.