Evaluating Interventions for the Prevention of ... - Princeton University

Evaluating Interventions for the Prevention of Mother-to-Child Transmission of HIV: Evidence from Africa Caitlin McGugan⇤ November 18, 2016 Abstract Development assistance from rich countries to poor countries faces challenges to effective implementation. Aid in the form of technical recommendations and funding often fails to reach people in need or has unintended negative consequences when executed on a large scale. In this paper, I evaluate the impact of prevention of motherto-child transmission of HIV (PMTCT) services, a major contemporary example of development aid for health. PMTCT remains active, yet little is known about the outcomes of these services on a large scale. A key component of PMTCT efforts by WHO/UNICEF is inducing HIV-positive mothers to wean early to mitigate the risk of postnatal HIV transmission through breastfeeding. I propose a novel method to identify program adherence using survey data on breastfeeding durations and mother’s HIV status from 21 African countries. A difference-in-differences estimation strategy finds that HIV-positive mothers who know their status become 15-20 percentage points more likely than HIV-negative mothers to wean by the program benchmark of 6 months post-PMTCT availability, a 300% increase, while HIV-positive mothers who do not know their status do not change their breastfeeding behavior. Despite impressive program reach and adherence, I show that survival rates have not significantly changed for children of HIV-positive mothers after PMTCT. I also find evidence that early weaning increases mortality rates for children without access to piped water, even among children of HIV-positive mothers, indicating that the risk of malnutrition and disease from not breastfeeding in poor environments outweighs the risk of transmission. My work suggests that PMTCT services should be more tailored to reflect heterogeneous conditions on the ground, such as access to clean water. Department of Economics, Princeton University, Princeton, NJ 08544. Email: [email protected]. Web: http://scholar.princeton.edu/cmcgugan. I am greatly indebted to Anne Case for her advice. I would also like to thank Janet Currie, Angus Deaton, Thomas Fujiwara, Nikhil Gupta, Jeff Hammer, Ilyana Kuziemko, Fernanda Marquez, Ishita Rajani, Kai Steverson, Tom Vogl, Justin Weidner, as well as various seminar participants for their helpful comments. This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1148900. All remaining errors are my own. ⇤

1

1

Introduction

Development assistance from rich countries to poor countries faces challenges to effective implementation. Does aid achieve its goals? Critics argue that aid projects are often illconceived and can do much active harm (Easterly, 2006, 2009). Proponents emphasize the high potential benefits of aid and cite examples of its success, many of them related to disease eradication and health (Sachs, 2005). Given this evidence of widely varying impacts, it is crucial to continue to evaluate and learn from aid projects. Aid for health and specifically to combat HIV/AIDS has recently become a donor focus as part of the Millennium Development Goals. In this paper, I evaluate the impact of Prevention of Mother-to-Child Transmission of HIV (PMTCT) services, a major component of this policy effort that has drawn a considerable amount of aid in the form of technical assistance and funding to subSaharan Africa. PMTCT remains active, yet little is known about the outcomes of these services on a large scale. This aid intervention is an important case study of the collaboration of rich and poor countries to improve health outcomes, and also one that policymakers urgently wish to know how to improve. If we fail to learn from its successes and failures, we risk making costly mistakes in the future. While foreign aid for health has been touted as the most successful use of development aid dollars, the rollout of PMTCT services and the large sum of resources allocated to them are also controversial for two primary reasons. First, there is the question of how targeting resources toward HIV treatment has influenced general health service provision in affected countries. Critics are concerned that HIV services may have crowded out unrelated services, taxed the limited supply of trained health care workers, and deflected funding from other important interventions (Bongaarts and Over, 2010). Proponents counter that HIV treatments have had positive spillovers for the health care sector as a whole through investments in infrastructure and human capital. Second, the effectiveness of PMTCT at meeting the needs of affected mothers and infants is in question. PMTCT has continued to evolve, but what has been the legacy of the treatments provided, and how can they be improved? This paper will address the latter question, but the former demonstrates that there is much at stake beyond HIV prevention in this debate over resource allocation. Mother-to-Child Transmission of HIV (MTCT) is an important cause of new HIV infections. Each year nearly 1.25 million women living with HIV/AIDS give birth in sub-Saharan Africa (UNAIDS, 2014). These pregnancies pose a risk of MTCT through contact in utero, during delivery, and during breastfeeding and account for 90% of new HIV infections among children worldwide. Prior to interventions nearly 1,900 children per day acquired HIV from their mothers in Africa, and virtually no infants infected with the virus lived past age five 2

(Dabis and Ekpini, 2002; Newell et al., 2004). The total risk of transmission absent interventions is about 30-40%, and the risk attributable to breastfeeding alone is about 10-15%. As a result, HIV/AIDS has changed the basis for infant feeding decisions among affected mothers. While African mothers have traditionally followed norms of breastfeeding for 24 months or longer, this practice poses a heightened risk of MTCT through breastfeeding when the mother is HIV-positive. In response to this crisis, PMTCT became a focus of international policy activity and attracted large inflows of earmarked development aid with the goal of averting these infant deaths beginning in 2003. Most African countries depended on OECD donors to provide the majority share of funding for provision of these services. According to plans from the World Health Organization and UNICEF, expectant mothers who tested positive for HIV during prenatal care were to enroll in PMTCT to receive short-term antiretroviral therapy (ART) drug prophylaxis to reduce transmission risk around delivery. Another key component of PMTCT efforts was inducing HIV-positive mothers to wean early to mitigate the risk of postnatal HIV transmission through breastfeeding: HIV-positive mothers were supposed to receive counseling about the risks of breastfeeding with HIV and the new recommendation to wean as early as feasible, ideally by the benchmark age of 6 months. Recognizing that PMTCT explicitly targeted breastfeeding choices, I use breastfeeding durations as a novel measure of program adherence and outcomes to evaluate PMTCT’s performance. The evidence for the safety of PMTCT’s early weaning interventions in particular is mixed. While the recommendation to avoid breastfeeding reduces the risk of postnatal transmission, early weaning is not necessarily feasible or desirable in regions where breastfeeding protects against diseases that pose a mortality risk to infants, such as diarrheal and respiratory infections. Many African mothers cannot reliably afford infant formula, and how the risk of HIV transmission weighs against the nutritional risk of early weaning to infants under typical replacement feeding scenarios is unknown. Thus studying whether the recommendations have induced early weaning is crucial not only for assessing the reach and effectiveness of PMTCT, but also for identifying outcomes of early weaning among HIVpositive mothers that speak to the appropriateness of the recommendation itself outside of medical trials with carefully controlled replacement feeding. In this paper, I show that there was a sharp increase in the probability of weaning by age 6 months among HIV-positive mothers in the wake of aid inflows for PMTCT services, and I present suggestive evidence that the mortality risks of early weaning outweighed the benefits of reduced HIV exposure for infants without access to clean water. My empirical strategy uses recently available household survey data that conducts HIV testing for adults. This allows me to identify the HIV serostatus of individual mothers and to subdivide HIV3

positive mothers into those who plausibly know their status and those who do not. Exploiting these distinctions in difference-in-differences estimates, I find that HIV-positive mothers who know their status become 15-20 percentage points more likely than HIV-negative mothers to wean by the program benchmark of 6 months post-PMTCT availability. In contrast, the breastfeeding behavior of HIV-positive mothers who do not know their status remains indistinguishable from that of HIV-negative mothers, alleviating concerns that these unusually short durations are driven by another factor correlated with contracting HIV or morbidity from HIV infection. As further evidence that exposure to PMTCT impacted infant feeding behavior, I investigate plausible exposure of these mothers to the program. The World Health Organization advocated personalized counseling about feeding decisions as the main channel of educating women about the risks of breastfeeding with HIV. I find that HIV-positive mothers who report receiving counseling about breastfeeding or PMTCT during prenatal care are many times more likely to wean by 6 months. Also, counseled mothers are more likely to answer questions about the basic facts of MTCT correctly, so counseling is correlated with HIV-related knowledge. Contrary to program goals, difference-in-differences estimation for infant mortality does not show statistically significant improvements for the children of HIV-positive mothers after PMTCT interventions. Moreover, the magnitude of the estimated survival gains is far smaller than the program estimates of 25% and higher. The estimated survival change is actually negative for the 7-12 month age interval immediately following the program’s early weaning benchmark. I also find that early weaning is significantly correlated with infant mortality over this same interval, and this risk is if anything higher for the children of HIV-positive mothers, contrary to what I would expect if early weaning were averting infant deaths on net. In fact, when the household does not have access to piped water the mortality risk from early weaning over the 7-12 month horizon is similar to the total MTCT risk from breastfeeding. To the best of my knowledge, my paper is the first to study the effectiveness of PMTCT at realizing its stated goals without resorting either to rough estimates using overall program disbursements or specific clinic-level or regional outcomes. The direct impact of PMTCT is challenging to identify due to lack of outcome data on adherence to prescribed drug regimens and HIV status of treated infants. Additionally, many HIV-exposed infants are lost to followup due to high rates of attrition at PMTCT clinics, beginning with roughly half even before delivery, which likely biases clinic-level results (Sibanda et al., 2013; Finlayson and Downe, 2013). Thus I believe that making use of the widely-available breastfeeding duration outcome variable represents a significant advance in the empirical study of PMTCT interventions. My 4

empirical results are optimistic about the potential impacts of this program, as there is strong evidence that women changed their infant feeding behavior in cooperation with breastfeeding recommendations. Weaning by 6 months is very rare in the African context, but postPMTCT, it suddenly becomes commonplace for HIV-positive mothers. It is interesting that long-term adherence to this information-based health intervention is so robust despite overall program attrition. This finding bodes well for potential cooperation with medical advice to reduce the spread of HIV/AIDS. However it is equally important to acknowledge the suggestive evidence of unintended consequences of this intervention in the form of excess deaths from early weaning offsetting deaths averted from reduced HIV transmission. The infant mortality results underscore mistakes made through misunderstanding the heterogeneity of the risks of early weaning and inadequate support for replacement feeding. Any studies of excess mortality among HIVexposed infants in this cohort must take into account the role played by short breastfeeding durations. Moreover, mortality and morbidity from the demographic shift toward early weaning may have profound economic impacts. In the short run, this generates additional claims on already taxed health care systems for treating illnesses. In the long run, increased morbidity could influence the human capital, cognition, labor earnings, and wellbeing of those who survive in the affected cohort. A growing literature in economics acknowledges the impact of prenatal and early childhood health on these outcomes (Almond and Currie, 2011; Case and Paxson, 2009; Bleakley, 2007). My findings complement previous work addressing the potential spillover effects of HIV care into other health services. They also provide an interesting counterpoint to the findings of Wilson (2015) that PMTCT expansion in Zambia increased child mortality in geographic proximity to PMTCT clinics. The rest of the paper is organized as follows: Section 2 reviews the medical literature on MTCT and breastfeeding and related policy interventions. Section 3 outlines the data set and the construction of key variables. I present the main results for early weaning in Section 4. Section 5 addresses the link between early weaning and exposure to PMTCT services by exploring the channel of counseling during antenatal care. The outcomes for infant mortality are shown in Section 6. Section 7 concludes.

5

2

Mother-to-Child-Transmission Health Policy Debates and Medical Knowledge

2.1

HIV Transmission through Breastfeeding

Medical research confirms that breastfeeding is an important cause of mother-to-childtransmission, but estimates of the total risk from this channel are variable and must be understood in relationship to several postnatal risk factors. Without the benefit of drug interventions, postnatal transmission through breastfeeding accounts for an infection rate of 5-20%, which first depends on the total duration of breastfeeding, since the risk of transmission is cumulative (De Cock and Fowler, 2000). However, there is evidence that the greatest threat of infection is in the first four to eight weeks of life, although these figures may be confounded with infection via delivery rather than breastfeeding (Nduati et al., 2000). The infection rate for confirmed late postnatal infection via breastfeeding among infants who tested negative for HIV at 4 weeks old is 9.3% at 18 months, based on estimates from several randomized controlled trials in sub-Saharan Africa (BHITS et al., 2004). Mother-to-child transmission through breastfeeding depends on the interplay of several factors beyond duration. HIV virions are present in the breast milk of infected mothers to varying degrees, where larger viral loads are more likely to transmit HIV. The levels of HIV detected in breast milk correlate with the mother’s systemic viral loads, which in turn rise with new infection in the postnatal period on the one hand and advanced infection with progression toward AIDS on the other (Willumsen et al., 2003; Embree et al., 2000; Leroy et al., 2003). Alternatively, the infant may ingest the virus through any cuts and sores around the mother’s nipples, so cases of mastitis, abscess, and infant oral thrush are positively correlated with transmission risk (Embree et al., 2000). Regardless of the method of contact, the virus may still pass through the child’s digestive tract without causing infection. Lifelong HIV infection only occurs when the virus both enters the host’s bloodstream and embeds itself in her immune system via a CD4 cell, where mouth sores and punctures in the mucous membranes of the gut are the most likely entry sites into the bloodstream for infants who ingest the virus. A further complicating factor is that breast milk itself variably contains an antibody that protects against HIV infection specifically, which also helps to explain why the risk of transmission during breastfeeding is not higher (John-Stewart et al., 2004). Consistent with this theme, during the 2000s medical research found that antiretroviral therapy (ART), a drug regimen that suppresses viral replication but does not eliminate HIV infection, successfully reduces the probability of transmission during pregnancy, delivery and breastfeeding. A full ART regimen is capable of reducing total transmission rates 6

to less than 5% with breastfeeding or less than 2% without breastfeeding (World Health Organization (WHO), 2010). ART is most effective when taken from the third trimester of pregnancy until weaning, but even short-course ART prophylaxis around the time of birth leads to significantly improved MTCT rates after breastfeeding at 24 months of age (Leroy et al., 2002). Clinical trials of breastfeeding with HIV conducted in Botswana concluded that breastfeeding with HIV poses little additional risk when combined with ART—at 18 months, there was no significant difference in cumulative HIV-free survival between the breastfeeding and formula feeding arms (Thior and Lockman, 2006). A separate study in Mozambique also found that with ART until weaning, transmission rates were low regardless of breastfeeding choice, and breastfed babies were no more likely to contract HIV than those who received formula replacements, although selection into breastfeeding was voluntary in this trial (Palombi et al., 2007).

2.2

The Benefits of Breastfeeding

Breastfeeding confers immunological benefits to infants even in the developed world, protecting against illnesses such as diarrhea, ear infections, influenza, and respiratory infections.1 These risks are particularly pertinent to a high-mortality environment such as Africa, where over 50% of under-5 mortality is attributable to these kinds of infections (Bryce et al., 2005). In fact, a study of a severe diarrhea outbreak caused by contaminated water in Botswana found that not breastfeeding was by far the leading predictor of diarrhea and death among infants and young children, highlighting the vulnerability of this group (Creek et al., 2010). Lack of access to piped water and sanitation interacts poorly with lack of breastfeeding for infant survival in the developing world (Habicht et al., 1988). Medical studies in Africa that have addressed HIV-free survival with and without breastfeeding involve providing mothers with ideal replacement feeding scenarios of no-cost infant formula and ongoing medical support and monitoring. As a result, little is known about the risk of HIV transmission relative to mortality linked to early weaning in typical developing country settings. Still, even with best-case replacement feeding, various studies indicate a real risk to early weaning. Multiple studies have found little difference in HIV-free survival at 18 or 24 months between formula fed and breastfed infants, but signs of increased malnutrition and/or morbidity among children of non-breastfeeding mothers. Two randomized trials of breastfeeding versus formula feeding with HIV were conducted in Kenya and Botswana, respectively. The Kenya study occurred in an urban setting and found that morbidity and mortality at 24 months of age were not significantly different in 1

See Nicoll and Williams (2002) for an overview of this topic.

7

breastfed and formula fed infants, although the infants in the breastfeeding arm exhibited significantly better nutritional status (Mbori-Ngacha et al., 2001; Nduati et al., 2000). The MASHI trial in Botswana also concluded that mortality between the two groups was not significantly different at 18 months, but there was significantly higher mortality in the formula arm between birth and 7 months, which was mostly attributable to diarrhea and pneumonia rather than HIV (Thior and Lockman, 2006). There is speculation that the difference in results may be due to better access to clean water in the urban Kenya trial, but this remains unsubstantiated. The quality of replacement feeding is vital for preventing adverse health outcomes. A study in Cote d’Ivoire followed infants born to HIV-positive mothers for 24 months, some of whom formula fed from birth and some of whom breastfed short-term until 4 months. Its main finding was high rates of morbid events such as diarrhea, acute respiratory infection, and malnutrition in both groups, and also risks of severe health events including death. In fact, the breast-milk substitutes provided to mothers were updated mid-study due to evidence of malnutrition in the participating children (Becquet et al., 2007). In summary, there is strong evidence that both breastfeeding and not breastfeeding present dangers for the children of HIV-positive mothers in low-resource environments. Identifying safe alternatives to prolonged breastfeeding that are appropriate to this setting is a complicated task.

2.3

PMTCT Policy and Recommendations for Breastfeeding with HIV

Policy strategies to eliminate MTCT through breastfeeding begin with technical recommendation about best-practice treatments, and these have evolved over time to incorporate new medical research. In 1998, the World Health Organization, UNICEF, and UNAIDS issued an initial statement acknowledging that breastfeeding is a proven channel of HIV transmission and advising that infected mothers should receive counseling about the risks of breastfeeding with HIV in order to make informed feeding decisions, rather than offering general advice about breastfeeding durations (WHO et al., 1998). These guidelines were revised in 2003 with the stronger recommendation that HIV-positive mothers should avoid all breastfeeding whenever it is “acceptable, feasible, affordable, sustainable and safe” or otherwise exclusively breastfeed for “the first few months [of life],” stopping as soon as it is feasible (WHO et al., 2003). The next update to the guidelines occurred in 2006, in response to new research that exclusive breastfeeding may be preferable to mixing breastfeeding and other foods for averting HIV transmission. Similar to before, it recommended avoiding breastfeeding if feasible, 8

but conceded that exclusive breastfeeding to 6 months, the age when solids are typically introduced, would be the best substitute given the health risks of early weaning (WHO, 2006). It also strongly recommended continued breastfeeding for infants known to be HIV-positive to the general population recommendation of 24 months. The final update to the feeding guidelines in 2010 coincided with a change to the ART guidelines that allowed breastfeeding mothers to continue to take antiretrovirals until one week after weaning, whereas in 2006 women were only eligible to continue them after shortterm prophylaxis surrounding delivery if they were already sick enough to need them for their own health. The 2010 guidelines recommend that women who have access to ART should breastfeed to 12 months, unless it is feasible to stop earlier (WHO et al., 2010). However, this update occurs at the end of my sample period, so the early weaning recommendations and ART provision rules only undergo minor changes after PMTCT rollout for the purposes of this paper, most notably the emergence of 6 months as a weaning benchmark. Note that comparatively few mothers had access to ART during breastfeeding prior to 2010, so my mortality results for breastfed infants do not benefit from the reduction in predicted transmission risk from taking ART until weaning. All iterations of the feeding guidelines emphasize the importance of blood testing and infant feeding counseling from trained professionals in integrating PMTCT with antenatal care. Knowledge of HIV status is key to initiating PMTCT treatments, but many HIVinfected mothers are unaware of their status, and considering the long timeline for progression from HIV to AIDS, many are asymptomatic. In 2005, only 10% of those living with HIV in sub-Saharan Africa knew their status (WHO et al., 2007). Testing rates have increased substantially over the last decade with the expansion of services, but nevertheless in 2013 over half of those infected still did not know that they are HIV-positive, and testing remains a programmatic goal (UNAIDS, 2013). All policy documents stress that women who do not know their status should not be counseled to wean early. In keeping with the urgency surrounding PMTCT and ART rollout, Official Development Assistance for realizing these goals rapidly increased during the mid-2000s. Figure 1 shows donor commitments for HIV assistance from OECD countries and international organizations to all of Africa between 1998 and 2010. Aid flows remain below $1 billion in 2002 and earlier, and then skyrocket upwards from 2003 onward to reach over $5 billion in 2008 and 2009. These surges in funding coincide with the formation of The Global Fund in 2002 and The United States President’s Emergency Plan for AIDS Relief (PEPFAR) in 2004, the two major OECD donors supporting HIV treatment interventions. Since PMTCT protocol calls for initiating HIV testing and counseling about feeding as early as a mother seeks treatment during pregnancy, counseling about the feeding guidelines should work with a lag with respect 9

to the birth year of the child. Thus disbursements beginning in 2003 place the post-period for implementing the feeding guidelines at birth years of 2004 and later, or 2005 if we allow additional time for initial program rollout. The infant feeding guidelines have been responsive to new knowledge and emphasized personalized counseling and choices on the part of the mother, but what choices do mothers actually make, and are they appropriate to their circumstances? My work adds to literature on this topic that has addressed programmatic issues with the infant feeding guidelines on a small scale. In general, past findings highlight difficulties with access to reliable replacement foods. In two small interview-based sociological studies in Malawi and Zimbabwe, mothers cited food unavailability as their primary barrier to early weaning, yet some weaned at 6 months regardless, and in Zimbabwe this had negative consequences for their children’s measured nutritional status (Levy et al., 2010; Lunney et al., 2008). A study of three PMTCT sites in South Africa found that two-thirds of mothers without access to piped water or electricity nonetheless chose to formula feed from birth, but their infants exhibited the highest risk of HIV transmission and death (Doherty et al., 2007). Despite the emphasis on informed choices in feeding, in practice recommendations may lack that nuance, and it appears that African mothers may be persuaded to wean early despite facing replacement options that are not acceptable, feasible, affordable, sustainable, and safe. These studies raise questions about the effectiveness of the guidelines, and the remainder of this paper will contribute to answering them.

3 3.1

Data Demographic and Health Surveys

The data used for analysis in this paper come from Demographic and Health Surveys (DHS), which are nationally representative household surveys conducted roughly every 4-5 years in low and middle-income countries. A typical survey administers questionnaires to the household, adult women ages 15-49 within the household, and men ages 15-59 within a subset of households. The questions focus on population, health, and reproductive history. For eligible women who have given birth within 5 years of the survey date, the survey responses include details of antenatal/birth delivery/postnatal care and child health, as well as responses about HIV/AIDS knowledge and attitudes. In the early-to-mid 2000s, many countries also began conducting HIV testing on a subsample of adults as part of the survey. This additional information makes the DHS well-suited to this study, because it allows me to match mothers with their respective HIV serostatuses in order to associate a mother’s 10

HIV status with her breastfeeding choices at the individual level. In this paper, I use data from 26 surveys of 21 African countries, spanning all geographic regions of the continent and levels of HIV prevalence. I selected these surveys because they represent all those available that meet the crucial criteria of containing both individuallylinkable HIV test results and final breastfeeding duration data. While the DHS survey questionnaires overlap substantially across countries and survey rounds, they are not identical. In particular, some survey countries either did not conduct HIV testing or did not collect breastfeeding durations for weaned children, excluding them from this analysis.2 My main results on early weaning and mortality encompass all 21 countries, but some ancillary results incorporate other variables that are not available for certain country-round datasets, forcing me to eliminate a subset of surveys from the specification. The notes below each table or figure indicate any surveys that were omitted from the results that it presents.

3.2

Country and Survey-Round Characteristics

Table 0 outlines the characteristics of each of the countries and surveys in the sample, including survey dates, birth years represented in the survey, HIV prevalence rates and sampling restrictions, sample size, and number of regions. Of the 21 countries included, five have two relevant surveys: Cameroon, Kenya, Lesotho, Malawi, and Tanzania. These are displayed in Panel A of Table 0. Panel B presents the remaining countries, those for which I only have one representative survey. The countries are arranged chronologically, from earliest survey date to most recent. It is important to note that each DHS covers births spanning the five years preceding the survey date, as the “Birth years” row highlights, yielding a sample that includes observations for all birth years from 1998 through 2010. Thus even the onesurvey countries can exhibit within-country time trends in breastfeeding durations based on the interval between a mother’s most recent birth and the survey date. The Panel A countries are especially valuable for this analysis given that, with the exception of Cameroon, they each cover a full decade of birth years without gaps. They provide longer-term like-with-like comparisons starting from the late 1990s or early 2000s, just before or contemporaneous with early policy information regarding the risks of breastfeeding interacted with HIV, through the ramp-up of PMTCT services in the mid-2000s and on to the end of the decade. Many of the one-survey countries also cover a window in the mid-2000s that allows for some comparisons between pre-intervention and post-intervention behavior. However, Burkina Faso, Ghana, 2 Some countries are not consistent in their collection of breastfeeding data across survey rounds, flipping from asking for the final breastfeeding duration for weaned children to asking only whether the child is currently being breastfed. This means that I have had to eliminate additional survey rounds with HIV testing for some countries in my sample due to a lack of usable infant feeding data.

11

Guinea, Rwanda, and Senegal are earlier surveys that serve as pre-period reference points, while Uganda provides only post-period data. All birth years from 2000 through 2007 contain data drawn from at least ten countries. The years on the fringe of the sample period have fewer contributing surveys but are still comprised of multiple countries, ranging from three countries represented in 1998 to seven in 2008. 2011 has only a handful of observations from three different countries, so it will not figure in the results. The adult HIV prevalence rates shown in the table are calculated using the DHS’s own HIV testing data for men and women ages 15-49 from each survey. These 21 countries have widely varying HIV/AIDS burdens, with estimated infection rates ranging from a low of 0.7% in both Senegal and Niger to a high of 26% in Swaziland. Nine countries have relatively high prevalence rates of above 5%: Cameroon, Kenya, Lesotho, Malawi, Swaziland, Tanzania, Uganda, Zambia, and Zimbabwe. The remaining twelve countries are low-prevalence: Burkina Faso, Ghana, Guinea, Democratic Republic of Congo, Ethiopia, Liberia, Mali, Niger, Rwanda, Sao Tome & Principe, Senegal, and Sierra Leone. Most countries do not conduct HIV testing on all female respondents, but rather on a representative subsample of households that is also slated for the men’s questionnaire, most frequently every other household. Table 0 reports each survey’s HIV sampling frame relative to all survey households. The collection of a blood sample for HIV testing is also conditional on the consent and availability for testing of the respondent, further narrowing the field of women with matching HIV test results. Table 0 shows that the testing response rates among eligible women range from 70% in the Malawi 2004 survey to 97% in Rwanda and Uganda. In general, these response rates have increased over time. Respondents are informed that the test results are anonymous, and they do not learn their serostatus from this test,3 but instead receive referrals to free voluntary testing and counseling. Interviewers ask respondents if they have previously been tested for HIV but not about the results of any tests, except in rare cases. The DHS summary reports for each survey report on sample selection issues among the HIV testing respondents. For my purposes, it is unimportant if the testing procedures produced an unbiased population mean for overall infection rates, as long as the breastfeeding decisions of the women who rejected testing do not systematically differ from those who accepted it.

3.3

The Breastfeeding Sample

The final sample for regression analysis is at the level of the most recent live birth within the last 5 years to mothers with HIV testing results. I limit the sample to most recent births 3 With the exception of Uganda, which offered the HIV results upon request, but did not automatically reveal them.

12

for several reasons. Importantly, many countries collect antenatal care variables only for the most recent birth. Furthermore, I cannot identify when mothers who tested HIV positive on the survey date contracted the virus, but they are more likely to have already been infected at the time of their most recent birth as compared to more distant births. Other survey responses regarding the mother’s current knowledge of HIV and past testing opportunities also have better proximity to the most recent child. There is also concern that, because breastfeeding can act as an important form of birth control in developing countries, giving birth to many children in the survey’s five-year window may be mechanically correlated with shorter breastfeeding durations for children born at the beginning of these short birth intervals. My sample includes children who have died but necessarily excludes children whose mothers have died, since all survey information is obtained from interviewing the mother. Thus breastfeeding duration is never determined by the death of the mother, and I exclude cases where the child’s age-at-weaning is equal to her age-at-death in the breastfeeding duration results, since here weaning was unlikely to have been the mother’s choice. The child deaths data will prove useful for the final section of this paper to address the impact of early weaning on child survival. While many children have been orphaned due to the AIDS pandemic, a harmful outcome in its own right, note that these data do not pertain to orphans but instead speak to separate impacts of being born to an HIV-positive mother that disadvantage even children who remain in their mother’s care. The key behavioral outcome of interest in this analysis is breastfeeding duration. Ideally, I would observe the age-at-weaning for each of the children in my sample, but this is impossible in the context of the DHS. Each survey instead collects a cross-section of breastfeeding behavior by asking the mothers if they are currently breastfeeding, and if not, when they stopped.4 This results in right-censored breastfeeding data for the most recent birth, as the sample includes children of various age groups, some of whom have been weaned by the interview date and some of whom have not. 5 Censoring complicates the interpretation of means and changes in means of breastfeeding durations, but PMTCT evaluation does not require examination of the full distribution of times to weaning. The infant feeding guidelines for PMTCT present early weaning by 6 months or earlier as a benchmark, so I 4

The breastfeeding variable does not describe whether or for how long the child was exclusively breastfed. Supplemental feeding with traditional foods and liquids is common in Africa, and counseling has been mandated to instruct mothers to breastfeed exclusively for the first 6 months of life. Most children older than 6 months who are still breastfeeding will also be eating solid foods, and the child’s reliance on breast milk within the overall diet will usually continue to decline with age. 5 Although values for breastfeeding are also truncated at 60 months by the survey design, this is not of much practical concern, as breastfeeding beyond 5 years is seemingly rare, with over 99% of women in the breastfeeding sample weaning by 36 months, and may become impracticable.

13

expect to see an increased frequency of weaning by this cutoff among HIV-positive mothers if the policies achieved their goals. With this in mind, I define an indicator variable for very early weaning that equals one if the child was weaned by 6 months of age, consistent with the 2006 guidelines benchmark, and another less restrictive indicator for early weaning that equals one if the child was weaned by 12 months of age. Further, by limiting the very early weaning and early weaning samples to children who reached at least 6 months old and 12 months old by the survey date, respectively, I can exactly identify whether or not each child was weaned by these markers, even when final breastfeeding duration is missing. Throughout this paper, I use these conditional samples when evaluating very early and early weaning. This strategy does require discarding data for the youngest children in the sample but solves the issues of duration comparability for the majority of observations. To delve further into breastfeeding behaviors, Section 4.1 turns to survival analysis as another means to address the data censoring issues that can estimate the full distribution of weaning times. Table 1 presents weighted sample means for the DHS data separately for HIV-positive mothers, HIV-negative mothers, and all mothers regardless of availability of HIV test results. Several facts deserve attention. The unconditional means of all breastfeeding durations, while inadequate, indicate that the average duration is quite long, and the very small proportions of HIV-negative mothers who wean by 6 months and by 12 months underscore this observation. The proportion of HIV-positive mothers weaning by 6 months is more than double that of HIV-negative mothers at 13%, and the proportion weaning by 12 months is almost double again at 22%. HIV-positive mothers have a 50% overall probability of having a past HIV test, but this statistic must be viewed in concert with Figure 2, which shows that the proportion of mothers reporting having received a previous HIV test climbed from below 20% in the early 2000s to more than 70% by the end of the decade. Nevertheless, in each birth year the mothers that I can identify as HIV-positive belong to two groups: those who have never been tested and do not know their serostatus, and those who have been tested and may know that they are HIV-positive. This distinction will have important ramifications for the empirical strategy in Section 4. Consistent with prior research about the demographics of the HIV pandemic, Table 1 also shows that in this sample HIV/AIDS is more common among mothers who reside in urban areas, are better educated, and in turn are more likely to be literate.

14

4 4.1

Weaning Results Kaplan-Meier Survival Functions

To give a picture of the full distribution of breastfeeding durations in the data, I follow (Kaplan and Meier, 1958) to construct a nonparametric estimate of the population timeto-weaning function that is well-equipped to correct for right-censoring in the data. The Kaplan-Meier estimator produces a declining step function that shows piecewise survival curves of the fraction of children still breastfeeding at each age. This approach relies on sufficient sample size and the assumption that the distribution of weaning times for the unweaned children in the sample will not systematically differ from the distribution observed among the already-weaned children. To construct the Kaplan-Meier estimator, let m1  m2  ...  mN denote the final duration of breastfeeding in months for each of the N weaned children observed in the sample, sorted from youngest to oldest. Associate with each mi an ni , the number still “at risk” of weaning just prior to month mi , and wi , the number of weaning events at mi . Note that with censoring, ni accounts for the number of subjects lost. Let S(m) be the probability that a mother will still be breastfeeding after month m. The Kaplan-meier maximum likelihood estimate of S(m) is given by ˆ S(m) =

Y

mi 0, such that the impact of weaning by 6 months has survival benefits for the children of HIV-positive mothers. Past research regarding the risks of early weaning in low-resource environments predicts 2 < 0, but if the risk of death due to HIV infection exceeds this danger the magnitude of 3 should be great enough that 2 + 3 > 0, lending support to very early weaning. This sum can also be compared to 1 , 26

which can be interpreted as an estimate for the risk of death from the HIV transmission channel over a given age interval. In this case, I will evaluate the 7-12 month and 1324 month horizons. Other important controls that address channels through which early weaning could be risky is pirt , an indicator for whether the child’s household has access to piped water, and its interaction with early weaning. Despite a lack of formal cause-of-death, a significant moderating impact of clean water would implicate diarrhea as a likely factor. Table 7, Panel A contains the regression estimates for the 7-12 month survival interval immediately following the early weaning benchmark. Unsurprisingly, having an HIV-positive mother has a universally negative and significant relationship to survival. However, these estimates place 3 < 0 at -4.5 percentage points, albeit not significant: there is no evidence that weaning by 6 months is associated with a survival benefit for the children of HIVpositive mothers. In fact, the confidence intervals rule out that 3 could be large and positive enough to overcome the negative main effect on early weaning. Moreover, 2 is estimated as negative and significant, such that weaning by 6 months is itself associated with nearly a 5.5 percentage point decrease in the probability of survival within the interval. Summing the coefficients, the mortality risk is 10 percentage points higher over this interval for children of HIV-positive mothers who were weaned according to the feeding guidelines (without access to piped water) than for children of HIV-positive mothers who were breastfed beyond the benchmark. This means that predicted mortality associated with following the guidelines over the 6-month interval immediately following weaning is already comparable to the total estimated risk of AIDS-related death up to 24 months of age from Section 6.1 (10 vs. 11 percentage points). Piped water and its interaction with HIV both show positive and significant coefficients, and the estimate for its impact on survival is much larger in magnitude for very early weaners. Note that the sum of 4 and 5 , the total estimated impact of having access to piped water for children weaned by 6 months, very nearly cancels out the negative point estimate on very early weaning. The estimates are not significantly different from each other. The results for the 13 to 24 month interval are shown in Table 7, Panel B. 3 remains negative at -2.2 percentage points as shown in column 4, and the estimates for weaning by 6 months are now -2.1 percentage points and insignificant. There is again no evidence of a survival gain for children of HIV-positive mothers who follow the guidelines. Interestingly, point estimates for access to piped water again balances out the negative estimates for weaning by 6 months. These estimates are highly suggestive of unintended negative consequences to following the infant feeding guidelines. Both sets of difference-in-difference survival estimates turned negative over the 7-12 month interval, and this is precisely the interval over which following 27

the guidelines is associated with excess mortality. This evidence also points to the centrality of clean water to facilitate safe early weaning.

7

Discussion

In this paper, I studied the impact of Prevention of Mother-to-Child Transmission of HIV services in Africa. I specifically evaluated efforts to induce HIV-positive mothers to wean early to mitigate the risk of postnatal HIV transmission through breastfeeding. I proposed a novel method to identify program adherence using survey data on breastfeeding durations and mother’s HIV status from 21 African countries spanning birth years before and after PMTCT was made available. Through difference-in-difference estimation, I found that PMTCT is associated with a 300% increase in the probability of weaning by the policy benchmark of 6 months among HIV-positive mothers who know their status. This is evidence that PMTCT overcame logistical and human resources challenges to achieve large-scale implementation of the infant feeding guidelines. It is a significant achievement that mothers altered their behavior substantially compared to cultural norms in compliance with the advice of newly-trained PMTCT counseling staff. However, I also showed that survival rates did not significantly improve for children of HIV-positive mothers after the introduction of PMTCT, and they may have decreased over the 7-12 month age interval immediately post-weaning. Further, I found that early weaning is associated with increased mortality rates for children without access to piped water in this same age interval, in particular for children of HIV-positive mothers, suggesting that the weak link between PMTCT and child outcomes may be due to unintended consequences in the form of excess deaths from early weaning offsetting deaths averted from reduced HIV transmission. These results have important implications for policymakers. My work demonstrates the need for evaluation of PMTCT services to account not only for HIV transmission rates and HIV-related deaths, but also mortality and morbidity associated with undernutrition and diseases linked to early weaning. While the guidelines acknowledged the nutritional risks of early weaning, it appears that they were larger than anticipated and that the advice given was too one-size-fits-all, persuading women to wean when it was not “acceptable, feasible, affordable, sustainable, and safe.” Improving the feeding advice offered has the potential to improve outcomes without increasing total funding to the program. I recommend that policymakers consider: (1) relaxing the early weaning recommendation where it is still binding (for mothers without ART access through weaning), or at least conditioning it on demonstrated access to clean water; (2) addressing potential reluctance about breastfeeding 28

among women who were previously treated with the early weaning recommendation; and (3) offering assistance with replacement foods and water purification for affected mothers who wish to wean early. An important question that fell outside the scope of this paper is whether early weaning has implications for the long term health of surviving children. Studying the impact of early-weaning on long-term morbidity outcomes presents an interesting direction for future work that has the potential to reinforce the conclusions in this paper. In this context where data on HIV healthcare inputs are sparse and aggregated, this paper’s insight that PMTCT targeted and increased early weaning behavior may also assist with future work. Identifying areas with high concentrations of early weaning as those where PMTCT was active could be applied to study its relationship to other outcomes of interest, such as spillovers to unrelated health services.

References Almond, D. and Currie, J. (2011). Human capital development before age five, volume 4. Becquet, R., Bequet, L., Ekouevi, D. K., Viho, I., Sakarovitch, C., Fassinou, P., Bedikou, G., Timite-Konan, M., Dabis, F., and Leroy, V. (2007). Two-year morbidity-mortality and alternatives to prolonged breast-feeding among children born to HIV-infected mothers in Cote d’Ivoire. PLoS Medicine, 4(1):0139–0151. BHITS, Coutsoudis, A., Dabis, F., Fawzi, W., Gaillard, P., Haverkamp, G., Harris, D. R., Jackson, J. B., Leroy, V., Meda, N., Msellati, P., Newell, M.-L., Nsuati, R., Read, J. S., and Wiktor, S. (2004). Late postnatal transmission of HIV-1 in breast-fed children: an individual patient data meta-analysis. The Journal of infectious diseases, 189(12):2154–66. Bleakley, H. (2007). Disease and Development: Evidence from Hookworm Eradication in the American South. Quarterly Journal of Economics, 122(1):73–117. Bongaarts, J. and Over, M. (2010). Global HIV / AIDS Policy in Transition. Science, 328:1359–1360. Bryce, J., Boschi-Pinto, C., Shibuya, K., and Black, R. E. (2005). WHO estimates of the causes of death in children. Lancet, 365(9465):1147–1152. Case, A. and Paxson, C. (2009). Early life health and cognitive function in old age. In American Economic Review, volume 99, pages 104–109.

29

Creek, T. L., Kim, A., Lu, L., Bowen, A., Masunge, J., Arvelo, W., Smit, M., Mach, O., Legwaila, K., Motswere, C., Zaks, L., Finkbeiner, T., Povinelli, L., Maruping, M., Ngwaru, G., Tebele, G., Bopp, C., Puhr, N., Johnston, S. P., Dasilva, A. J., Bern, C., Beard, R. S., and Davis, M. K. (2010). Hospitalization and mortality among primarily nonbreastfed children during a large outbreak of diarrhea and malnutrition in Botswana, 2006. Journal of Acquired Immune Deficiency Syndrome, 53(1):14–19. Dabis, F. and Ekpini, E. (2002). HIV-1/AIDS and maternal and child health in Africa. The Lancet. De Cock, K. and Fowler, M. (2000). Prevention of mother-to-child HIV transmission in resource-poor countries. JAMA: the journal of . . . , 283(9). Doherty, T., Chopra, M., Jackson, D., Goga, A., Colvin, M., and Persson, L.-A. (2007). Effectiveness of the WHO/UNICEF guidelines on infant feeding for HIV-positive women: results from a prospective cohort study in South Africa. AIDS (London, England), 21(13):1791–1797. Easterly, W. (2006). The White Man’s Burden: Why the West’s efforts to aid the rest have done so much ill and so little good. Oxford University Press, Oxford. Easterly, W. (2009). Can the West Save Africa? Journal of Economic Literature, 47(2):373– 447. Embree, J. E., Njenga, S., Datta, P., Nagelkerke, N. J., Ndinya-Achola, J. O., Mohammed, Z., Ramdahin, S., Bwayo, J. J., and Plummer, F. a. (2000). Risk factors for postnatal mother-child transmission of HIV-1. AIDS (London, England), 14(16):2535–2541. Finlayson, K. and Downe, S. (2013). Why Do Women Not Use Antenatal Services in Lowand Middle-Income Countries? A Meta-Synthesis of Qualitative Studies. PLoS Medicine, 10(1). Habicht, J., DaVanzo, J., and Butz, W. (1988). Mother’s Milk and Sewage : Their Interactive Effects on Infant Mortality. Pediatrics. John-Stewart, G., Mbori-Ngacha, D., Ekpini, R., Janoff, E. N., Nkengasong, J., Read, J. S., Van de Perre, P., and Newell, M.-L. (2004). Breast-feeding and Transmission of HIV-1. Journal of Acquired Immune Deficiency Syndrome, 35(2):196–202. Kaplan, E. L. and Meier, P. (1958). Nonparametric estimation from incomplete samples. J. of the ASA, 73(282):457–481. 30

Leroy, V., Karon, J. M., Alioum, A., Ekpini, E. R., Meda, N., Greenberg, A. E., Msellati, P., Hudgens, M., Dabis, F., and Wiktor, S. Z. (2002). Twenty-four month efficacy of a maternal short-course zidovudine regimen to prevent mother-to-child transmission of HIV-1 in West Africa. AIDS (London, England), 16(October 2001):631–641. Leroy, V., Karon, J. M., Alioum, A., Ekpini, E. R., van de Perre, P., Greenberg, A. E., Msellati, P., Hudgens, M., Dabis, F., and Wiktor, S. Z. (2003). Postnatal transmission of HIV-1 after a maternal short-course zidovudine peripartum regimen in West Africa. Aids, 17(10):1493–1501. Levy, J. M., Webb, A. L., and Sellen, D. W. (2010). "On our own, we can’t manage": experiences with infant feeding recommendations among Malawian mothers living with HIV. International breastfeeding journal, 5:15. Lunney, K. M., Jenkins, A. L., Tavengwa, N. V., Majo, F., Chidhanguro, D., Iliff, P., Strickland, G. T., Piwoz, E., Iannotti, L., and Humphrey, J. H. (2008). HIV-positive poor women may stop breast-feeding early to protect their infants from HIV infection although available replacement diets are grossly inadequate. The Journal of nutrition, 138(2):351–357. Mbori-Ngacha, D., Nduati, R., John, G., Reilly, M., Richardson, B., Mwatha, a., NdinyaAchola, J., Bwayo, J., and Kreiss, J. (2001). Morbidity and mortality in breastfed and formula-fed infants of HIV-1-infected women: A randomized clinical trial. JAMA : the journal of the American Medical Association, 286(19):2413–2420. Nduati, R., John, G., Mbori-Ngacha, D., Richardson, B., Overbaugh, J., Mwatha, A., Ndinya-Achola, J., Bwayo, J., Onyango, F., Hughes, J., and Kreiss, J. (2000). Effect of breastfeeding and formula feeding on transmission of HIV-1: a randomized clinical trial. JAMA, 283(9):1167–1174. Newell, M.-L., Brahmbhatt, H., and Ghys, P. D. (2004). Child mortality and HIV infection in Africa: a review. AIDS (London, England), 18 Suppl 2:S27–S34. Nicoll, A. and Williams, A. (2002). Breast feeding. Arch Dis Child, 87:91–92. Palombi, L., Marazzi, M., Voetberg, A., and Magid, N. (2007). Treatment acceleration program and the experience of the DREAM program in prevention of mother-to-child transmission of HIV. AIDS, Suppl 4:S65–71. Sachs, J. D. (2005). The End of Poverty: How we can make it happen in our lifetime. Number 4. 31

Sibanda, E. L., Weller, I. V. D., Hakim, J. G., and Cowan, F. M. (2013). The magnitude of loss to follow-up of HIV-exposed infants along the prevention of mother-to-child HIV transmission continuum of care: a systematic review and meta-analysis. AIDS (London, England), 27(17):2787–2797. Thior, I. and Lockman, S. (2006). Breastfeeding Plus Infant Zidovudine Prophylaxis for 6 Months vs Formula Feeding Plus Infant Zidovudine for 1 Month to Reduce Mother-toChild HIV Transmission in Botswana. JAMA: the journal of . . . , 296(7). UNAIDS (2013). GLOBAL REPORT: UNAIDS report on the global AIDS epidemic 2013. UNAIDS (2014). The Gap Report: Children and Pregnant Women Living with HIV. (2):20. UNAIDS, UNICEF, PEPFAR, and WHO (2016). On the Fast-Track to an AIDS-Free Generation. WHO (2006). HIV and Infant Feeding Update. World Health Organization. WHO, UNAIDS, and UNICEF (2007). Access to HIV Therapy Grew Significantly in 2006, but Significant Obstacles Remain to Approaching Universal Access to HIV Services. WHO, UNICEF, and UNAIDS (2010). Guidelines on HIV and infant feeding. Technical report. WHO, Unicef, UNAIDS, and Unfpa (1998). HIV and Infant Feeding: Guidelines for Decision Makers. WHO, UNICEF, and UNFPA (2003). HIV and Infant Feeding: Guidelines for Decision Makers. Willumsen, J. F., Filteau, S. M., Coutsoudis, A., Newell, M.-L., Rollins, N. C., Coovadia, H. M., and Tomkins, A. M. (2003). Breastmilk RNA viral load in HIV-infected South African women. AIDS, 17(3):407–414. Wilson, N. (2015). Can Disease-Specific Funding Harm Health? in the Shadow of HIV/AIDS Service Expansion. Demography, 52(5):1671–1700. World Health Organization (WHO) (2010). Antiretroviral drugs for treating pregnant women and preventing HIV infections in infants. Recommendations for a public health approach. Technical report.

32

NOTES: Sourced from the Organization for Economic Cooperation and Development Creditor Reporting System. Data is in constant 2014 US dollars. HIV aid is the sum of sector codes 130.4 and 160.64. Data reflects aid commitments rather than total disbursements as these figures are not available prior to 2002.

0.00

Fraction Still Breastfeeding 0.25 0.50 0.75

1.00

Figure 3: Time to Weaning by Mother's HIV Status

0

6

12

18

24

30

36 42 Months

HIV -

48 HIV +

54

60

Figure 4: Kaplan-Meier Survival Func4ons by Birth Year

0.00

Fraction Still Breastfeeding 0.25 0.50 0.75

1.00

Figure 5: Time to Weaning, 2000 vs. 2008

0

6

12

18

24

30

36 Months

HIV -, 2000 HIV +, 2000

42

48

54

HIV -, 2008 HIV +, 2008

60

Figure 6: Estimates for Previously Tested HIV+ Mothers

-.2

HivXBirthYr + HivXTestedXBirthYr -.1 0 .1 .2 .3

.4

Linear Prob on Weaning by 6 Mos

1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Birth Year

Figure 7: Estimates for Untested HIV+ Mothers

-.2

-.1

HivXBirthYr 0 .1 .2

.3

.4

Linear Prob on Weaning by 6 Mos

1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Birth Year

Figure 8: Estimates for Previously Tested HIV+ Mothers

HivXBirthYr + HivXTestedXBirthYr -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 .5

.6

Linear Prob on Weaning by 12 Mos

1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Birth Year

Figure 9: Estimates for Untested HIV+ Mothers

HivXBirthYr -.4 -.3 -.2 -.1 0 .1 .2 .3

.4

.5

.6

Linear Prob on Weaning by 12 Mos

1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Birth Year

Survey dates Birth years Adult HIV prevalence rate HIV sampling frame within overall DHS HIV test response rate for eligible women N Number of regions

PANEL B: Countries with One Survey

Number of regions Burkina Faso 6/2003-12/2003 1998-2003 1.8% 1 in 2 hhs 92% 2001 14

8 Ghana 7/2003-10/2003 1998-2003 2% All hhs 89% 2176 10

12

Guinea 2/2005-6/2005 2000-2004 1.5% 1 in 2 hhs 93% 1734 8

10

10/2009-1/2010 2004-2009 23% 1 in 2 hhs 94% 1264

Rwanda 2/2005-7/2005 2000-2005 3% 1 in 2 hhs 97% 2092 5

3

6/2010-11/2010 2005-2010 11% 1 in 3 hhs 91% 3721

Senegal 2/2005-5/2005 2000-2004 0.7% 8 in 21 hhs 85% 1702 11

26-30*

12/2011-5/2012 2007-2011 5% All hhs 90% 4563

1/2011-8/2011 2006-2011 4.3% 1 in 2 hhs 94% 2894

Survey dates Birth years Adult HIV prevalence rate HIV sampling frame within overall DHS HIV test response rate for eligible women N

11/2008-2/2009 2003-2008 6% 1 in 2 hhs 86% 1531

Tanzania 10/2007-2/2008 2002-2007 6% All hhs 90% 3786

TABLE 0: Sample Details for Data from 26 Demographic and Health Surveys of 21 African Countries PANEL A: Countries with Two Surveys Kenya Cameroon Lesotho Malawi Survey dates 4/2003-9/2003 2/2004-8/2004 9/2004-1/2005 10/2004-1/2005 Birth years 1998-2003 1999-2004 1999-2004 1999-2004 Adult HIV prevalence rate 7% 5.5% 24% 12% HIV sampling frame within overall DHS 1 in 2 hhs 1 in 2 hhs 1 in 2 hhs 1 in 3 hhs HIV test response rate for eligible women 76% 92% 81% 70% N 1317 2169 1054 1500

Zimbabwe 8/2005-3/2006 2000-2005 18% All hhs 76% 2895 10

Niger 1/2006-5/2006 2001-2005 0.7% 1 in 2 hhs 91% 2269 8

Mali 4/2006-12/2006 2001-2006 1.3% 1 in 3 hhs 92% 2263 9

Swaziland 7/2006-3/2007 2001-2006 26% All hhs 87% 1634 4

Liberia 12/2006-4/2007 2002-2006 1.5% All hhs 87% 3010 6

Dem Rep of Congo Zambia Sierra Leone Sao Tome & Principe Uganda Survey dates 1/2007-3/2007 4/2007-10/07 4/2008-6/2008 9/2008-3/2009 2/2011-9/2011 Birth years 2002-2007 2002-2007 2003-2007 2003-2008 2006-2011 Adult HIV prevalence rate 1.3% 14% 1.5% 1.5% 7.3% HIV sampling frame within overall DHS 1 in 2 hhs All hhs 1 in 2 hhs All hhs All hhs HIV test response rate for eligible women 90% 77% 88% 88% 97% N 2040 2675 1472 1184 5152 Number of regions 4 9 4 4 10 *In 2012, Tanzania added 4 new regions (Geita, Katavi, Njombe, & Simiyu), raising its total from 26 to 30 and modifying the boundaries of 5 of its existing regions (Kagera, Mwanza, Shinyanga, Rukwa & Iringa). The boundaries of these 9 regions are irreconcilable between the 2007-8 & 2011-2 surveys. Thus all regression results exclude these 9 regions in order to make within-region comparisons, for a total of 21 contributing regions. Notes: “Adult HIV prevalence rate” is defined as the fraction of men and women age 15-49 who tested positive for HIV among those who consented to be tested as part of the DHS. In general, only women living in households who were also selected for the DHS men’s survey were eligible for HIV testing, as outlined in the “HIV sampling frame within overall DHS” row. In cases where only a fraction of households were eligible for testing, these sub-samples were selected to be nationally representative. Reasons for non-response among those eligible for HIV testing include unavailability for or refusal of the interview, refusal of the HIV test portion of the interview, absence of the respondent upon callbacks for testing, and technical problems with drawing or processing of the blood sample. “N” is defined as the number of women in the survey who (i) have DHS-provided HIV test results, (ii) have provided a breastfeeding duration for their most recent birth, and (iii) whose youngest child has survived until at least 6 months of age.

PANEL B: Countries with One Survey, Continued Ethiopia Survey dates 4/2005-8/2005 Birth years 2000-2005 Adult HIV prevalence rate 1.4% HIV sampling frame within overall DHS 1 in 2 hhs HIV test response rate for eligible women 83% N 2362 Number of regions 11

TABLE 1: DHS Sample Means by Mother’s HIV Status HIV+ Mothers Infant Feeding and Family Size Breastfeeding duration for most recent birth (mos) 14.271 (8.724) Proportion of most recent births weaned by 6 mos1 0.127 (0.333) Proportion of most recent births weaned by 12 mos2 0.220 (0.414) Proportion of mothers previously tested for HIV 0.518 (0.500) Number of surviving children 2.854 (1.885) Number of children who have died 0.611 (0.959) Mother and Family Characteristics Urban residence 0.348 (0.476) Household asset count 1.592 (1.465) Mother’s age at time of most recent birth (years) 26.919 (6.445) Mother’s years of education 6.390 (3.564) Mother’s literacy (from DHS reading test): Illiterate 0.225 (0.417) Some Literacy 0.120 (0.325) Literate 0.656 (0.475)

HIV- Mothers

All Mothers

15.133 (9.159) 0.058 (0.233) 0.132 (0.338) 0.350 (0.477) 3.403 (2.199) 0.554 (0.998)

15.099 (9.339) 0.065 (0.246) 0.141 (0.348) 0.327 (0.469) 3.343 (2.175) 0.576 (1.024)

0.234 (0.424) 1.369 (1.355) 27.190 (7.154) 4.229 (4.104)

0.243 (0.429) 1.364 (1.374) 27.068 (7.094) 4.085 (4.151)

0.553 (0.497) 0.089 (0.285) 0.358 (0.479)

0.541 (0.498) 0.088 (0.283) 0.371 (0.483)

Number of countries 21 21 21 Number of regions 204 204 204 Number of observations 5,636 69,182 139,364 1 Sample limited to children who were at least 6 months old at the interview date and who survived until at least 6 months of age. 2 Sample limited to children who were at least 12 months old at the interview date and who survived until at least 12 months of age. All results are sample weighted. Standard errors are reported in parentheses. The “All Mothers” category includes mothers who were not selected for HIV testing. “Household asset count” is a count of indicators for whether the household has each of the following: electricity, a radio, a television, a refrigerator, a bicycle, a motorcycle or scooter, a car or truck, and a telephone. Source: DHS Surveys of 21 African countries. The 21 countries included are: Burkina Faso, Cameroon, Democratic Republic of Congo, Ethiopia, Ghana, Guinea, Kenya, Lesotho, Liberia, Malawi, Mali, Niger, Rwanda, Senegal, Sierra Leone, Sao Tome & Principe, Swaziland, Tanzania, Uganda, Zambia, and Zimbabwe.

TABLE 2: “Treatment” with the Infant Feeding Guidelines by Type of Mother and Counseling Status No Counseling Counseling HIV-, Untested HIV- Control HIV- Control HIV-, Tested HIV+, Untested HIV+, Tested

HIV+ Control Potential spillovers

HIV+ Control Treatment

TABLE 3: Antenatal Counseling and Comprehensive MTCT Knowledge compMTCT Counseled

0.116*** (0.016)

HIV+

-0.003 (0.020)

HIV+ * Counseled

0.042 (0.027)

Primary School

0.055***

Secondary School

0.054***

(0.013) (0.015)

Higher Education

0.082+ (0.045)

N

36,440

Birth Year and Country-Region FEs? Individual Controls? Number of Countries Number of Country-Regions

Yes Yes 17 154

+ p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001

All results are sample weighted. Standard errors clustered by country-region are reported in parentheses. Sample excludes data from Burkina Faso, Ethiopia, Ghana, Liberia, Cameroon 2004 and Malawi 2004-5. The dependent variable is an indicator equal to one if the mother answered all four DHS questions about MTCT correctly and zero if she responded incorrectly or ’I don’t know’ to at least one question. ’Counseled’ is an indicator equal to one if the mother received counseling about either breastfeeding or MTCT during antenatal care. Other controls include urban residence, household assets, child’s parity, and linear and quadratic controls for the mother’s age.

No Yes

N 31654 Birth Years 1998-2011 Number of Countries 17 Number of Regions 154 + p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001

Country-Region FE? Household Controls?

31654 1998-2011 17 154

Yes Yes

5834 1998-2011 3 41

No Yes

5834 1998-2011 3 41

Yes Yes

TABLE 4: Linear Probability on Weaning by Prenatal Counseling Exposure and HIV Status Dependent Var: Pr(weaning by 6 mos) Counseling Spec. Counseling 1 2 3 4 HIV+, Untested, Not Counseled 0.019 0.015 0.026 0.027 (0.024) (0.023) (0.036) (0.034) Post * (HIV+, Untested, Not Counseled) -0.018 -0.018 -0.042 -0.043 (0.028) (0.029) (0.044) (0.046) HIV+, Tested, Not Counseled 0.036 0.033 0.048 0.046 (0.040) (0.041) (0.073) (0.075) Post * (HIV+, Tested, Not Counseled) 0.105 0.094 0.124 0.123 (0.106) (0.109) (0.194) (0.195) HIV+, Untested, Counseled 0.043+ 0.049* 0.132+ 0.133+ (0.024) (0.025) (0.066) (0.066) Post * (HIV+, Untested, Counseled) -0.032 -0.032 -0.157* -0.167* (0.028) (0.026) (0.071) (0.071) HIV+, Tested, Counseled 0.003 0.009 -0.050* -0.051* (0.011) (0.010) (0.020) (0.020) Post * (HIV+, Tested, Counseled) 0.191*** 0.177*** 0.337*** 0.339*** (0.030) (0.030) (0.070) (0.070)

TABLE 5: Di↵erence-in-Di↵erences for Child Survival, HIVComparison HIV+ Post * HIV+ N Birth Year and Country-Region FEs? Individual Controls? Number of Countries Number of Country-Regions

0-6 Mos

7-12 Mos

13-24 Mos

-0.043***

-0.021***

-0.044**

(0.009)

(0.006)

(0.016)

0.004

-0.013

0.019

(0.011)

(0.009)

(0.012)

51,663

40,917

24,113

Yes Yes 20 185

Yes Yes 20 185

Yes Yes 20 185

+ p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001 NOTE: Dependent variable is an indicator equal to 1 if the child survived over the specified interval. Uganda is omitted due to lack of age-at-death data. All standard errors are robust and clustered at the region level. The results are sample weighted. ’Household Controls’ are linear and quadratic controls for the mother’s age at the child’s birth, mother’s education, categorical variables for the child’s parity among siblings, household wealth quartile, and urban residence.

TABLE 6: Di↵erence-in-Di↵erences for Child Survival, within HIV+ Comparison 0-6 Mos Previous HIV Test Post * Previously Tested for HIV N Birth Year and Country-Region FEs? Individual Controls? Number of Countries Number of Country-Regions

7-12 Mos

13-24 Mos

-0.005

0.013

-0.024

(0.020)

(0.015)

(0.019)

0.022

-0.007

0.014

(0.025)

(0.024)

(0.021)

4,096

3,203

2,078

Yes Yes 20 158

Yes Yes 20 157

Yes Yes 20 144

+ p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001 NOTE: The sample is limited to children of HIV+ mothers only. The dependent variable is an indicator equal to 1 if the child survived over the specified interval. Uganda is omitted due to lack of age-at-death data. All standard errors are robust and clustered at the region level. The results are sample weighted. ’Household Controls’ are linear and quadratic controls for the mother’s age at the child’s birth, mother’s education, categorical variables for the child’s parity among siblings, household wealth quartile, and urban residence.

TABLE 7: Child Survival by Early Weaning Panel A 7-12 Months HIV+ Household has Piped Water

1

2

3

4

-0.028***

-0.029***

-0.022***

-0.023***

(0.006)

(0.006)

(0.006)

(0.006)

0.009***

0.007**

0.007***

0.005**

(0.002)

(0.002)

(0.002)

(0.002)

-0.055**

-0.054**

Weaned by 6 Mos HIV+ * Weaned by 6 Mos Weaned by 6 Mos * Piped Water N Birth Year and Country-Region FEs? Individual Controls? Number of Countries Number of Country-Regions

(0.016)

(0.017)

-0.044

-0.044

(0.034)

(0.035)

0.043*

0.046*

(0.021)

(0.020)

40,594

40,594

40,594

40,594

No Yes 20 190

Yes Yes 20 190

No Yes 20 190

Yes Yes 20 190

+ p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001 NOTE: The dependent variable is an indicator equal to 1 if the child survived between 7 and 12 months. Uganda is omitted due to lack of age-at-death data. All standard errors are robust and clustered at the region level. The results are sample weighted. ’Household Controls’ are linear and quadratic controls for the mother’s age at the child’s birth, mother’s education, categorical variables for the child’s parity among siblings, household wealth quartile, and urban residence.

TABLE 7: Child Survival by Early Weaning Panel B 13-24 Months HIV+ Household has Piped Water

1

2

3

4

-0.037**

-0.037**

-0.035***

-0.034***

(0.012)

(0.012)

(0.010)

(0.010)

0.003

0.002

0.002

0.001

(0.006)

(0.006)

(0.006)

(0.006)

-0.021

-0.021

Weaned by 6 Mos HIV+ * Weaned by 6 Mos Weaned by 6 Mos * Piped Water N Birth Year and Country-Region FEs? Individual Controls? Number of Countries Number of Country-Regions

(0.014)

(0.014)

-0.024

-0.022

(0.034)

(0.034)

0.019

0.020

(0.020)

(0.020)

23,928

23,928

23,928

23,928

No Yes 20 190

Yes Yes 20 190

No Yes 20 190

Yes Yes 20 190

+ p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001 NOTE: The dependent variable is an indicator equal to 1 if the child survived between 13 and 24 months. Uganda is omitted due to lack of age-at-death data. All standard errors are robust and clustered at the region level. The results are sample weighted. ’Household Controls’ are linear and quadratic controls for the mother’s age at the child’s birth, mother’s education, categorical variables for the child’s parity among siblings, household wealth quartile, and urban residence.