This article examines the effect of introducing a new HIV/AIDS service—prevention of mother-to-child transmission of HIV (PMTCT)—on overall quality of prenatal and postnatal care. My results suggest that local PMTCT introduction in Zambia may have actually increased all-cause child mortality in the short term. There is some evidence that vaccinations may have declined in the short term in association with local PMTCT introduction, suggesting that the new service may have partly crowded out existing pediatric health services.
There has been a dramatic expansion of publicly funded HIV/AIDS services in sub-Saharan Africa over the past decade. Hundreds of millions of people in this region of the world most heavily afflicted with HIV/AIDS have received fully subsidized access to public services such as voluntary counseling and testing (VCT), antiretroviral therapy (ART), and prevention of mother-to-child transmission of HIV (PMTCT). Made largely under the auspices of the U.S. President’s Emergency Plan for AIDS Relief (PEPFAR) and the Global Fund to Fight AIDS, Tuberculosis and Malaria, donor disbursements make up 88 % of spending on these services in low-income countries (UNAIDS 2010) and currently total approximately $8 billion USD per year (Kates et al. 2011).1 (For the remainder of the article, all dollar amounts are shown in U.S. currency.)
This large increase in mostly disease-specific public funding has come in the context of relatively weak health systems (De Cock et al. 2000). Low overall health spending, high caseloads for doctors and nurses, and insufficient infrastructure characterize health systems in much of sub-Saharan Africa (WHO 2006a; WHO 2013a, b).2 Health worker absence (Goldstein et al. 2013), counterfeit drugs (Bate et al. 2011; Bjorkman-Nyqvist et al. 2012), and weak incentives (Basinga et al. 2011) compound these problems. Some have argued that the large increase in HIV/AIDS spending may strengthen health systems in recipient countries (e.g., El-Sadr and Abrams 2007; Grepin 2012; Price et al. 2009; Rasschaert et al. 2011), whereas others have argued that the vertical (i.e., disease-specific) approach to HIV/AIDS programming may generate negative spillovers for non-HIV care (e.g., England 2007; Garrett 2007; Grepin 2011, 2012; Jaffe 2008; Rabkin et al. 2009; Shiffman 2008).3
This article examines the effects of the expansion of one of the main HIV/AIDS services—prevention of mother-to-child transmission of HIV (PMTCT)—on child mortality.4 Medical trials have indicated that PMTCT is highly effective at reducing vertical HIV transmission (Dabis and Ekpini 2002). Based on this evidence, many countries in sub-Saharan Africa have dramatically expanded public access to this service in the past decade. However, little evidence exists on the effect of PMTCT at scale, and the concerns about HIV/AIDS service expansion crowding out non-HIV care in international health suggest that the overall improvements in child health from PMTCT expansion may have been limited.
I constructed a geocoded monthly panel using a census of health facilities in Zambia that documents the expansion of PMTCT over roughly the first decade of scale-up. Several facts make Zambia an excellent context in which to examine this question. Approximately 14 % of pregnant women in Zambia are HIV-positive (Central Statistical Office et al. 2009), and nearly 7 % of infants are born HIV-positive or acquire HIV through breast-feeding (Torpey et al. 2010). Zambia is one of the PEPFAR 14 priority countries, and donor aid for HIV in Zambia has increased from $10 million in 2000 (Oomman et al. 2007) to $250 million in 2008 (Resch et al. 2008), or roughly $1 per capita to $25 per capita. In comparison, the mean annual total expenditure on health in Zambia over the period 2000–2008 was approximately $39 per capita (WHO 2013a). As I discuss in more detail in the following section, this increase in funding does not appear to have financed an increase in the number of health workers, highlighting the possibility that human resource or information constraints may have inhibited the benefits of PMTCT expansion.
Retrospective birth history modules from repeated national cross-sectional household surveys provide information on child mortality. Administrative records from these surveys allow me to identify the location of survey households and calculate their proximity to each health clinic. I use these data to measure the change in child mortality associated with local PMTCT introduction while controlling for time-invariant and time-varying factors associated with local PMTCT introduction.
My results suggest that PMTCT expansion increased all-cause child mortality, particularly in the short term. Four main regression results support this view. First, conditional on month and year of birth, the local introduction of PMTCT was associated with approximately a 2 to 3 percentage point increase in under-24 month mortality. Second, a semiparametric difference-in-differences analysis reveals no clear pre-introduction trend in locations receiving PMTCT. Third, the increase in child mortality appears to have been greater among households residing closer to PMTCT sites. Fourth, the increase in child mortality appears to have been concentrated among women who were less likely to be HIV-positive and hence less likely to directly benefit from PMTCT.
I find evidence that local PMTCT introduction may have reduced the quality of non-HIV child health services in the short term. Quality of non-HIV pediatric care appears to have rebounded around the event time that the deleterious effect of PMTCT introduction on all-cause child mortality dissipated. However, this evidence on a possible mechanism is less strong than the main result.
These results seemingly contrast with existing quasi-experimental evidence on the effects of antiretroviral provision at scale. Bendavid et al. (2012) and Lucas and Wilson (2013a, b) reported evidence suggesting that the expansion of ART for the treatment of HIV-positive adults improved mortality and morbidity outcomes for adults and for children at scale.5,6 However, the scope for non-HIV service crowded out due to HIV-specific funding is likely much greater for pediatric care than for adult care. In Zambia, HIV/AIDS is responsible for approximately 12 % of under-5 mortality (WHO 2010b), and other pediatric health conditions endangering life and requiring medical attention are relatively common.7 In contrast, HIV/AIDS is surely a much greater fraction of adult mortality in Zambia. In sub-Saharan Africa as whole, where 5 % of adults are HIV-positive (UNAIDS 2010), HIV/AIDS causes 29 % of mortality among adults aged 15–59 and less than 5 % of mortality among children aged 0–4 (WHO 2011a).
There are several reasons why disease-specific funding may harm overall health outcomes. In the context of HIV/AIDS, there are at least three main concerns. First, the vertical nature of HIV/AIDS programming in the developing world means that HIV/AIDS programs compete with the rest of the health sector for local health workers in an internal “brain drain” (England 2007; Garrett 2007; Grepin 2012; Jaffe 2008; Rabkin et al. 2009). Second, HIV/AIDS donor funding is often linked to achieving narrow target outcomes (e.g., number of individuals on ART, or number of pregnant women receiving antiretroviral prophylaxis for PMTCT) instead of broader measures of health outcomes (Garrett 2007). Third, donor funding for HIV/AIDS may crowd out donor funding for other health issues (England 2007; Shiffman 2008). Finally, although it should not harm overall health outcomes, if donor funding crowds out domestic funding, then the net benefit of donor funding is muted.8 The results of the current analysis are consistent with all these mechanisms.
HIV/AIDS and Child Health in Zambia
An HIV-positive woman may transmit the virus to her fetus or her newborn child through childbirth and breast-feeding. The cumulative transmission probability is as high as 45 % (Dabis and Ekpini 2002). In sub-Saharan Africa, mortality rates among HIV-positive infants appear to reach 50 % by 12 months of age (Brahmbhatt et al. 2001, 2006; Dabis et al. 2001; Newell et al. 2004a, b; Spira et al. 1999; Taha et al. 1999).9,10 PMTCT generally refers to a package of interventions including HIV testing, antiretrovirals (ARVs) for vertical prophylaxis, breast-feeding advice, and family planning. ARVs for vertical prophylaxis appear to be particularly effective.11
As part of a broader package of HIV/AIDS service expansion, donors began substantial efforts for facilitating PMTCT scale-up in sub-Saharan Africa in the early 2000s. Between 1996 and 2008, donor funding for HIV/AIDS in the developing world increased from $300 million to $7.6 billion (Kates et al. 2012). PMTCT spending appears to have followed a similar trend. Between 2004 and 2010, the proportion of HIV-positive pregnant women in sub-Saharan Africa receiving ARVs for PMTCT increased from 9 % to approximately 50 % (WHO 2010a, 2011b).
Throughout this process, there have been major concerns about health worker shortages (e.g., Kober and Van Damme 2004; WHO 2006b, 2007) and other possible spillovers to other services (e.g., England 2007; Garrett 2007; Grepin 2011, 2012; Jaffe 2008; Rabkin et al. 2009; Shiffman 2008). In response to health worker shortages, the World Health Organization (WHO) began promoting a “task shifting” system as part of its “Treat, Train, Retain” plan (WHO 2007). Under this system, health care workers (including those at antenatal clinics) are asked to take on tasks traditionally assigned to workers further up the delivery hierarchy (e.g., clinical officers conduct the initial consultation/clinical evaluation, a task traditionally assigned to doctors).
These concerns characterize PMTCT expansion in Zambia. Between 2000 and 2008, annual donor funding for HIV/AIDS in Zambia, a country of approximately 10 million people, increased from less than $10 million (Oomman et al. 2007) to roughly $250 million (Resch et al. 2008). These funds appear to have largely financed antiretroviral drugs. For example, of the nearly $190 million spent on HIV/AIDS in Zambia in 2006, nearly one-half was spent on ART for adults and care for treatment patients, and approximately 38 % of prevention spending was spent on PMTCT (UNAIDS 2008). Between 2001 and 2007, the number of health facilities offering PMTCT in Zambia grew from fewer than six to nearly 600, or approximately 40 % of facilities.12 During this period, few funds appear to have been allocated for health system strengthening (WHO 2009), and there does not seem to have been an expansion in the number of health facilities. Although international health aid as a percentage of Zambia GDP increased from 1.7 % in 2002 to 2.4 % in 2006, domestic government health funding as a percentage of Zambia GDP decreased from 3.0 % to 2.2 % (Ooms et al. 2010).
Evidence from a subset of health facilities in Zambia indicates little to no increase in the number of doctors, nurses, and other trained health workers during HIV/AIDS service expansion (Brugha et al. 2010a; Walsh et al. 2010), suggesting there may not have been a large increase system-wide and that PMTCT introduction may have increased health worker caseload. With approximately one physician per 10,000 population (WHO 2004), Zambia has fewer than one-third of the WHO–recommended physicians per person and a similar shortfall in the number of nurses (Schatz 2008).13 Salaries for public sector health workers were frozen from 2000–2005 (Makasa 2008) despite annual inflation rates ranging from 18 % to 27 % (International Monetary Fund 2006).14 Many clinics appear to have implemented a task shifting system (Morris et al. 2009). Although the WHO “Treat, Train, Retain” program was not launched until 2006 (WHO 2007), task-shifting in Zambia began no later 2004 (Morris et al. 2009). This shifting of workers up the delivery hierarchy appears to have been made implicitly toward HIV/AIDS service provision and to have come at the expense of labor previously allocated to non-HIV tasks (Walsh et al. 2010).
Data on the PMTCT cascade—the sequence of actions required to ensure that mother/infant pairs receive antiretroviral drugs to prevent MTCT— suggests that the expansion of access to PMTCT translated into increased use of PMTCT.15 Nonetheless, there appears to have been attrition at every step of the cascade (Stringer et al. 2005). National household survey data (i.e., the 2001 and 2007 Demographic Health Surveys (DHS) and the 2003 and 2005 Zambia Sexual Behavior Surveys (ZSBS)) have indicated that antenatal clinic (ANC) attendance among pregnant women in Zambia exceeds 90 %, even prior to local PMTCT availability (see Table 1). Among respondents in the 2003 ZSBS, conducted toward the beginning of PMTCT scale-up, 15 % of ANC attendees in all of Zambia were offered an HIV test, and 44 % of these accepted this offer (see Table 1). By the 2005 ZSBS, these figures had increased to 25 % and 63 %, respectively. The likelihood of completing these steps in the PMTCT cascade varied substantially across space, presumably because of variation in the availability of PMTCT services at ANC clinics. Clinical data from the same time period indicate that 82 % of ANC attendees in Lusaka were offered a HIV test, 71 % of these accepted the test offer, and 99 % of those tested received the result (Stringer et al. 2005). Since mid-2005, more than 90 % of ANC attendees at Lusaka ANCs have been tested for HIV (Stringer et al. 2008b). In Zambia as a whole, UNAIDS has estimated that 47 % of HIV-positive pregnant women in 2007 received ARVs for PMTCT (UNAIDS 2008).
The time series on PMTCT expansion and infant mortality suggests that this scale-up may have generated a substantial reduction in child mortality. Figure 1 presents the cumulative number of PMTCT sites, the individual-level PMTCT coverage rate, and the age 0–12 month child mortality rate in each year from 1997–2007. I describe these data in more detail in the Data section. The largest reduction in infant mortality appears to precede the period of most rapid PMTCT expansion in terms of cumulative number of PMTCT sites. However, PMTCT expansion occurred earlier and with greater intensity in urban areas, suggesting that the cumulative number of PMTCT sites understates mother–infant exposure to PMTCT. I calculate the individual-level PMTCT coverage rate as the proportion of adult females residing within 20km of a PMTCT site. Declining infant mortality tracks rising individual-level PMTCT coverage much more closely than it tracks the cumulative number of PMTCT sites. Notably, aggregate infant mortality appears to have been relatively flat prior to PMTCT expansion. However, as I discuss in the upcoming section on child mortality, after controlling for child date of birth, the association between local PMTCT availability and child mortality reverses sign, suggesting that the negative association in Fig. 1 largely reflects a secular decline in child mortality. Although not shown here, the child mortality time series for ages 0–24 months very closely tracks that for ages 0–12 months.
The Japanese International Cooperation Agency (JICA) 2006 Health Facilities Census (HFC) provides the exact latitude and longitude of each hospital and health clinic in Zambia. I augment these data with information on the month and year when each facility introduced PMTCT or whether the facility never introduced PMTCT. This information on timing was collected beginning in June 2008, so the data on PMTCT expansion in Zambia span the period from the first PMTCT introduction in Zambia through the middle of 2008.16
Data on child mortality come from the birth history modules in the 2001 and 2007 DHS. I use these birth history recall data to construct a child-level panel and measures of child death by 6, 12, 18, and 24 months. To help address concerns about possible recall bias, I limit my analysis of child mortality to children born January 1997 or later. Table 1 reports descriptive statistics on child mortality. For example, “died by 6 months” is an indicator variable equal to 1 if a child died by 6 months of age, and equal to 0 otherwise.
For the 2001 DHS, administrative records on primary sampling units allow me to identify the statistical enumeration area (SEA) of residence of each respondent. I record the GPS coordinates of the centroid of the SEA of residence as the respondent's location.17,18 For the 2007 DHS, I use the GPS data points provided as part of the survey.19 These are the GPS coordinates of the centroid of the SEA of residence with a randomly drawn vector of length 0km–10km added by survey management to ensure respondent confidentiality. In conjunction with the GPS information in the 2006 JICA HFC, these data allow me to calculate the distance from each household to each health facility. Information on the interview date in the 2001 and 2007 DHS allows me to exploit the monthly variation in PMTCT expansion documented in the health facilities data. The other household surveys that I use in the empirical analysis of the steps in the PMTCT cascade—the 2003 and 2005 ZSBS—also include information on the interview date for each respondent.
The DHS also include individual-level data on several of the steps in the PMTCT cascade, as do the 2003 and 2005 Zambia Sexual Behavior Surveys (ZSBS). In particular, these surveys provide information on the steps in the cascade leading to and including the respondent receiving the result of a HIV test administered during an ANC visit for her most recent pregnancy. The DHS and ZSBS do not include information on adherence to antiretroviral drugs.
Table 1 reports descriptive statistics on multiple steps in the PMTCT cascade. The vast majority of pregnant women visit an ANC at least once during their pregnancy, even in 2001 when very few women had access to PMTCT. In contrast, the proportion of women reporting being offered a HIV test at an ANC nearly tripled between 2001 and 2007, and the proportion accepting this offer increased by nearly 50 % between 2003 and 2005.20
Prenatal and Postnatal Care
I construct measures of the quality of prenatal and postnatal care using information on maternal and child health in the 2001 and 2007 DHS. These include an indicator variable for whether a blood sample was taken during a visit to an ANC during the most recent birth, as well as a count variable for the number of basic prenatal services received other than the blood sample. They also include indicator variables for a health worker visit in the two months following birth and for receiving a vitamin A dose in the two months following birth. For children, I also construct a measure of the number of vaccinations received and an indicator variable for complete vaccinations.21 Finally, for children under the age of 5 reporting a fever in the past month, I construct an indicator variable for whether the child received antimalarial drugs at a health clinic.
Descriptive statistics for these variables are in Table 1. For most outcomes, there were improvements between 2001 and 2007. Two notable exceptions are the likelihood of receiving no vaccinations and the likelihood of receiving drugs for a fever. The increased likelihood of receiving no vaccinations is consistent with the WHO assessment of a general downward trend in the incidence of several vaccinations in Zambia (WHO 2013b). For example, diphtheria tetanus toxoid and pertussis (DTP3) immunization among 1-year-olds in Zambia declined from 85 % in 2001 to 80 % in 2007 (WHO 2013b). During this period, polio immunization rates also fell from 86 to 77 % WHO 2013b). Similarly, the large reduction in the likelihood of receiving drugs conditional on having a fever is consistent with the trend in the 2006 and 2008 National Malaria Indicator Surveys (MIS) (WHO 2013b).
There are several sources of measurement error in these data. Information on PMTCT expansion may overstate PMTCT availability if clinics that received PMTCT subsequently experienced service interruptions. Similarly, the fact that the location of each survey respondent is measured with error means that exact distance to PMTCT locations is measured with error. Information on the PMTCT cascade, child mortality, and reproductive health inputs are provided using birth recalls and does not include women who have died. Instead of exact HIV status, the available HIV data for multiple DHS rounds are likely HIV status. To the extent that these sources of measurement error do not change specifically in PMTCT locations simultaneously with local PMTCT introduction, these measurement errors should only attenuate any estimated effects of local PMTCT introduction toward zero.
My primary empirical specification treats a respondent as being near a health clinic if the respondent lives within 20km of the nearest health clinic.26 Alternative specifications relax restrictions that the local introduction of PMTCT had the same effect on child mortality invariant of distance conditional on distance being less than or greater than 20km.
I cluster the standard errors at the SEA level. This is the geographic unit at which local PMTCT varies according to my spatial measure of PMTCT availability. There are more than 300 SEAs in the 2001 DHS and more than 3,000 SEAs in the 2007 DHS, so standard asymptotic tests are appropriate (Cameron et al. 2008). The other household surveys that I use in the empirical analysis of the steps in the PMTCT cascade (i.e., the 2003 and 2005 ZSBS) contain fewer SEAs but still include more than 100 SEAs.
Before turning to the analysis of the effects of local PMTCT availability on child mortality, I examine the effects of local PMTCT availability on several of the steps in the PMTCT cascade. The 2001 and 2007 DHS and the 2003 and 2005 ZSBS include information from respondents on the following steps in the PMTCT cascade: (1) whether the respondent visited an ANC at least once during her pregnancy, (2) whether the ANC offered an HIV test, (3) whether the respondent accepted the offer, and (4) whether a blood sample was taken during the ANC visit. Because the vast majority of pregnant women in Zambia report visiting an ANC multiple times (see Table 1), there is little scope for local PMTCT introduction to increase ANC attendance. Thus, I focus on constructing indicator variables for completing each of the steps in taking a HIV test at an ANC.27 I regress each of these indicator variables on the full set of controls indicated in Eq. (1).
Table 2 presents the estimates of the effect of local PMTCT availability on these steps in the PMTCT cascade. All specifications include an indicator variable equal to 1 if the respondent resides within 20km of a clinic that ever offered PMTCT. In addition, all specifications include the full set of controls as indicated in Eq. (1). Standard errors are clustered by SEA of residence.28
The results of this analysis suggest that local PMTCT introduction increased the likelihood the respondent reported having a blood sample taken during an ANC visit. The point estimate in column 2 indicates that local PMTCT introduction was associated with a more than 5 percentage point increase (statistically significant at the 5 % level) in the likelihood of having a blood sample taken. Although the point estimate on local PMTCT availability in the “offered test” regression is positive, it is not statistically significant. Two possible explanations for this are: (1) the sample size in column 1 is much smaller than in column 2, and (2) respondents may misreport answers to the “offer” question with greater frequency than to a question about an actual procedure (i.e., having blood drawn).
The main regression results suggest that the local introduction of PMTCT may have increased child mortality in the short term. Estimates of the effect of local PMTCT availability on child mortality rates appear in Table 3. These regressions use the child-level panel constructed using the birth history modules from the 2001 and 2007 DHS. All specifications include an indicator variable equal to 1 if the respondent resides within 20km of a clinic that ever offered PMTCT.
Panel A examines the effect of local PMTCT availability on the likelihood of child death by 6 months of age. Column 1 presents the results of a simple regression that controls only for whether a clinic within 20km of a respondent ever offered PMTCT. The point estimate is negative (albeit relatively small and statistically insignificant), suggesting that local PMTCT introduction may have reduced 0–6 month mortality but that the reduction (if any) was not large. In column 2, where I control for childbirth month × birth year, the estimated effect of local PMTCT availability reverses sign and becomes statistically significant at the 10 % level. Columns 3–6 include additional time, geographic, and individual-level controls. Throughout, the point estimate on local PMTCT availability remains positive and statistically significant at (at least) the 10 % level.
In panels B–D of Table 3, I examine the effect of local PMTCT introduction on child death by 12, 18, and 24 months, respectively. These results follow a pattern similar to that in panel A. After controlling very flexibly for secular birth time effects, the causal effect of local PMTCT introduction on child mortality appears to be an increase in mortality. Moreover, at older ages (i.e., 18 and 24 months), the point estimate is statistically significant across most of the regression specifications and usually at (at least) the 10 % level. As a whole, the results in Table 3 suggest that the decline in child mortality in Fig. 1 may have been a secular change and that the causal effect of local PMTCT introduction may have actually been an increase in child mortality.
The spatial nature of these data provides a useful test of whether we should take a causal interpretation of the baseline results. Namely, the effect of local PMTCT introduction should be greater for respondents residing closer to the clinic where PMTCT is introduced. In Table 4, I allow the effect of local PMTCT introduction on under-2 mortality to vary semiparametrically by the distance at which the respondent resides from the clinic where PMTCT is locally introduced. Notably, the three PMTCT availability measures are not mutually exclusive. For example, individuals residing within 10km of a PMTCT site also reside within 20km and 30km of a PMTCT site. Therefore, the estimated effect of local PMTCT for individuals residing within 10km is the sum of the coefficient estimates on “PMTCT within 10km,” “PMTCT within 20km,” and “PMTCT within 30km.” For individuals residing within 11km–20km, the estimated effect is the sum of the coefficient estimates on “PMTCT within 10km” and “PMTCT within 20km.” For individuals residing 21km–30km, the estimated effect is simply the coefficient estimate on “PMTCT within 30km.”
The results presented in Table 4 suggest the effect of local PMTCT introduction on child mortality may have been concentrated among respondents living closest to the health facility where PMTCT was locally introduced. To fix ideas, consider the estimates from the specification with the full set of controls (i.e., column 6). Among individuals residing within 10km of the PMTCT site, under-2 mortality increased by approximately 3.4 percentage points. In contrast, under-2 mortality increased by only 1.1 percentage points among individuals residing 10km–20km from the PMTCT site, and the estimated effect is not statistically significant at conventional levels. Likewise, the point estimate for PMTCT within 30km suggests neither a large nor a statistically significant effect on under-24 mortality for respondents residing very far from the PMTCT site.
The coefficient estimates plotted in Fig. 2 are consistent with a causal interpretation of the baseline child mortality results. Conditional on the birth month × birth year fixed effects, there is little evidence of a consistent prelocal introduction trend in child mortality in locations ultimately receiving PMTCT.30 Furthermore, there is evidence of an upward trend in the point estimates beginning immediately with the local introduction of PMTCT.31
Fig. 2 also reveals that the deleterious effect of PMTCT on child mortality appears to have dissipated around 48 months after local PMTCT introduction. To further investigate this reversal, I estimate a version of Eq. (1) allowing for an additional effect of local PMTCT availability in locations where PMTCT had been available at least 48 months. Panel A in Table 5 presents these results. After controlling for birth month × birth year fixed effects, the results suggest that local PMTCT increased under-2 mortality by 3 to 4 percentage points in the short term and had a little effect on under-2 mortality for children born 48 months or more after local PMTCT introduction.
The results in panel A could be driven by unobserved heterogeneity between clinics that received PMTCT earlier and those that received PMTCT later. To investigate this possibility, panel B in Table 5 reports the results of allowing the effect of local PMTCT to vary simultaneously by whether local PMTCT introduction had been available at least 48 months and by whether the location was an early PMTCT recipient.32 These estimates suggest that the deleterious effect of PMTCT introduction may have been concentrated in locations that were early PMTCT recipients.
Heterogeneity by Likely HIV Status
This section tests whether the estimated effect of PMTCT on child mortality varies by the likelihood that the respondent was HIV-positive. The 2007 DHS includes an HIV testing module with results that are linked to the rest of the individual-level information in the survey. Approximately 75 % of 2007 DHS respondents consented to providing a blood sample for the HIV test, and there is evidence from other settings of selection into the DHS HIV testing module by HIV status (Reniers and Eaton 2009). I use these data to construct a measure of HIV prevalence in a respondent’s demographic group defined by the interaction of five-year age group and province of residence. In these data, there is substantial variation across province of residence and across the age profile in the likelihood of being infected with HIV. The fact that HIV prevalence has remained relatively constant in Zambia over the period 2001–2007 suggests that aside from possible downward bias generated by selection into the HIV testing module, this approach may yield a reasonable (albeit noisy) measure of the likelihood that a respondent was HIV-positive.
Table 6 reports the results of allowing the effect of local PMTCT availability to vary by this continuous measure of likely HIV status. Interpreting the results of this exercise in the regression specifications that do not include individual-level controls or province fixed effects requires substantial caution because the measure of likely HIV status is highly correlated by construction with five-year age group and province of residence. Thus, although the point estimate on the PMTCT availability term interacted with likely HIV status is positive in columns 1–4, this is likely a spurious correlation driven by underlying heterogeneity across individual characteristics and province of residence. In columns 5–6, where I control for these important omitted variables, the point estimates suggest that the increase in child mortality associated with local PMTCT introduction was concentrated among respondents who were relatively unlikely to be HIV-positive.
Other Pediatric Health Services
Presumably local PMTCT introduction should have decreased MTCT. In a context where roughly 7 % of infants are born HIV-positive or acquire HIV through breast-feeding (Torpey et al. 2010), increasing access to PMTCT among a population with ANC attendance rates above 90 % should have reduced child mortality. Why does the evidence presented thus far largely suggest the opposite?
One possible explanation is that introducing PMTCT at a clinic may have affected the quality of care associated with other critical pediatric health services (Both and van Roosmalen 2010). As discussed in the earlier section, HIV/AIDS and Child Health in Zambia, a shortage of health workers may have forced health facilities to support PMTCT activities with labor formerly allocated to non-HIV tasks. Table 7 explores this broad hypothesis by regressing measures of prenatal and postnatal care on the baseline measure of local PMTCT availability. The point estimates suggest that local PMTCT introduction may have reduced the quality of non-HIV prenatal and postnatal care. Although the coefficient on local PMTCT availability is statistically significant only in the vaccination regressions, it is always negative and often large in absolute value. Moreover, the estimated effect of local PMTCT on receiving complete vaccinations is broadly consistent with the estimated effect of local PMTC on child mortality. The point estimate in column 2 of Table 6 suggests that local PMTCT introduction was associated with approximately a 4 percentage point reduction in the likelihood of receiving a complete set of vaccinations. If one-quarter to one-half of unvaccinated children in this setting die of vaccine-preventable diseases, then this channel alone would explain one-half to all the roughly 2 percentage point increase in child mortality associated with local PMTCT introduction.
Panel A of Table 8 examines the dynamic effects of local PMTCT introduction on basic prenatal and postnatal care. The results suggest that the effect on the likelihood of receiving a blood test more than doubled in the medium term compared with the short term.
The dynamic effects on measures of the quality of non-HIV prenatal and postnatal care in panel A of Table 8 provide additional support for the crowd-out hypothesis. For each of the measures, aside from the postnatal health worker visit, the deleterious effects of local PMTCT introduction appear to have been partly to more than fully mitigated after 48 months. This is consistent with the elasticity of service integration being greater in the long term than in the short term.
Panel B of Table 8 explores whether this seemingly dynamic effect of local PMTCT may have actually been heterogeneity in the effect by the timing of local introduction. On the whole, these results do not provide strong support for the hypothesis that the seemingly short-term deleterious effects of local PMTCT on basic child health service indicators was simply driven by being an early adopter.
An additional piece of evidence that is consistent with the hypothesis of non-HIV service crowd-out is the heterogeneous effect of local PMTCT introduction on child mortality by the likelihood the respondent was HIV-positive. As shown in Table 6, for women who were relatively unlikely to be HIV-positive and hence unlikely to directly benefit from PMTCT, local PMTCT introduction appears to have increased child mortality. For women who were very likely to be HIV-positive and hence likely to benefit from PMTCT, local PMTCT introduction appears to have not affected child mortality or even reduced it, consistent with the competing direct and indirect effects of PMTCT. The differential effects by the likelihood the respondent was HIV-positive also help rule out an increase in breast-feeding as the mechanism by which PMTCT increased child mortality. This mechanism would increase child mortality among HIV-positive women, if among any women, compared with HIV-negative women.
This article examines the effect of prevention of mother-to-child transmission of HIV (PMTCT) expansion on all-cause child mortality in Zambia. I use a geocoded census of health facilities in Zambia to construct a monthly panel documenting the nationwide expansion of PMTCT. Panel data on child mortality come from the birth history modules in the 2001 and 2007 DHS.
I find that local PMTCT introduction appears to have increased all-cause child mortality in the short term. This adverse mortality effect appears to have diminished with event time, and non-HIV pediatric health service quality appears to have followed an inverse pattern after local PMTCT introduction. These results suggest that disease-specific funding may distort health provider incentives toward the delivery of particular services and away from health outcomes. A high child mortality rate due to competing risks and a shortage of health workers may make this dynamic particularly acute in the context of PMTCT expansion in poor countries. This analysis highlights the value of additional research on alternative funding and performance evaluation mechanisms for health providers in the developing world (e.g., Miller and Babiarz 2014; Basinga et al. 2011; Leonard and Masatu 2010; Miller et al. 2012), with a particular focus on the possibility of heterogeneous effects across pediatric and adult care.
I would like to thank Eran Bendavid, Marianne Bitler, Alfredo Burlando, Raluca Buzdugan, Christopher Carpenter, Joseph Cummins, Carlos Dobkin, William Dow, Pascaline Dupas, Frederico Finan, Gunther Fink, Emmanuela Gakidou, Paul Gertler, Erick Gong, Laura Guay, Gianmarco Leon, Ethan Ligon, Jeremy Magruder, Zoe McClaren, Sandra McCoy, Edward Miguel, Christopher Murray, Nancy Padian, Bei Qin, Jonathan Robinson, David Sahn, Lucie Schmidt, T. Paul Schultz, Manisha Shah, Lara Shore-Sheppard, Jeffrey Stringer, Anand Swamy, Harsha Thirumurthy, Waly Wane, Tara Watson, David Weil; two anonymous referees; and seminar participants at the NBER Africa Project Zanzibar Conference; the 6th Annual PopPov Conference on Population, Reproductive Health, and Economic Development in Accra, Ghana; the University of California, Berkeley; the University of California, Irvine; the University of California, Santa Cruz; the University of Oregon; the University of Washington; and Williams College for many excellent comments. Madeleine Watson and Wentao Xiong provided superb research assistance. The NBER Africa Project provided generous financial and institutional support. This research would not be possible without the assistance of Kunyima Banda and the Network of Zambian People Living with HIV/AIDS (NZP+). All errors are my own. The findings, interpretations, and conclusions expressed in this article are those of the author and do not necessarily represent the views of the aforementioned individuals or the agencies that employ them.
From 1990 to 2007, the fraction of development assistance for health (DAH) disbursed through targeted programs such as the Global Fund increased substantially, whereas the fraction of DAH disbursed through development banks and the United Nations decreased (Ravishankar et al. 2009).
An early report on antiretroviral expansion in four Southern African countries (Kober and Van Damme 2004:103) concluded, “[T]he lack of human resources for health is deplored as the single most serious obstacle for implementing national treatment plans.”
Although we do not know the counterfactual trajectory of non-HIV health funding, the United States, the single largest source of HIV/AIDS donor funding (Schneider and Garrett 2009), increased annual spending on non-HIV international health from $1.3 billion to $1.7 billion between 2001 and 2008 (Government Accountability Office (GAO) 2010). In contrast, U.S. spending on HIV/AIDS in international health increased from $204 million to $3.3 billion over this period (GAO 2010).
By measuring effects on child mortality rather than rather than HIV transmission, the current analysis provides evidence on the net effect of PMTCT on a broad measure of child health. The Bellagio HIV/Health Systems Working Group has identified this as a priority research question.
Brugha et al. (2010b) provided a largely descriptive analysis of service quality in 39 health facilities in three districts in Zambia in the context of HIV/AIDS service scale-up and found mixed evidence of a link between HIV/AIDS service introduction and non-HIV quality of care.
A previous version of the current article presented results indicating that child mortality may have fallen in response to local PMTCT introduction. However, all those results came from empirical specifications that failed to properly control for childbirth time fixed effects. As shown in column 1 of Table 3, when I do not control for childbirth date fixed effects, the data suggest a negative association between local PMTCT introduction and child mortality. However, this reflects a strong downward secular trend in child mortality during this period rather than the causal effect of local PMTCT introduction on child mortality. Unfortunately, the time fixed effects implemented in the earlier version of this article were defined with respect to the survey date rather than the childbirth date.
The 2006 National Malaria Indicator Survey (MIS) indicated a malaria prevalence of 22 % among children age 0–59 months (Ministry of Health et al. 2009). The 2007 DHS indicated that 15.5 % of children age 0–59 months experienced diarrhea in the two weeks preceding the survey, and 5.2 % of children age 0–59 months had symptoms of an acute respiratory infection (ARI) in the two weeks preceding the survey (Central Statistical Office et al. 2009). WHO reported that 15 % of under-5 mortality in Zambia is due to malaria, another 15 % is due to diarrhea, and another 15 % is due to pneumonia (WHO 2010b).
The scope for domestic health expenditure crowd-out may be large. In many sub-Saharan African countries, donor funding for HIV/AIDS exceeds not only recipient countries’ domestic contributions to their HIV/AIDS budgets but also the domestically financed health budgets in recipient countries (Shiffman 2008). Lu et al. (2010) suggested that during the period 1996–2006, every one dollar in development assistance for health to developing countries reduced domestic expenditure on health in developing countries by forty-three cents.
Even in the absence of acquiring HIV/AIDS, child mortality in Zambia is high. Overall, the neonatal mortality rate in Zambia is 36 per 1,000 live births, the under-1 infant mortality rate is 92 per 1,000 live births, and the under-5 child mortality rate is 148 per 1,000 live births. HIV/AIDS accounts for 12 % of under-5 mortality in Zambia (WHO 2010b).
Mother-to-child transmission of HIV (MTCT) is not the only way in which having a HIV-positive mother increases the probability of child death. Maternal mortality due to HIV/AIDS further reduces the resources available to poor households in poor countries. Consistent with this burden, in a longitudinal study in Rakai, Uganda, Brahmbhatt et al. (2006) found under-2 child mortality rates of 166 per 1,000 for HIV-negative children of HIV-positive mothers, compared with 128 per 1,000 for HIV-negative children of HIV-negative mothers. However, MTCT appears to be the largest effect, with an under-2 child mortality rate of 547 per 1,000 for HIV-positive children of HIV-positive mothers.
Administering single-dose nevirapine (NVP) at the three stages at which MTCT may occur reduces MTCT by 10–25 percentage points (Guay et al. 1999; Jackson et al. 2003). Combination therapy (i.e., zidovudine (ZDV) and nevirapine (NVP)) may virtually eliminate MTCT (Dabis and Ekpini 2002).
I discuss these data in more detail in the section, PMTCT Expansion.
As of 2007, there were approximately 50 doctors in Zambia who had been trained there, despite the fact that approximately 600 doctors had been trained in Zambia over the preceding 40 years (Garrett 2007).
Salaries for health workers in Zambia are relatively low, and wages are often in arrears. The average annual salary for a doctor in Zambia is $17,208 and $4,128 for a nurse (McCoy et al. 2008). A 2005 health worker survey indicated that four of five respondents received late salary payments, nearly one of six respondents were not paid what they were owed, and one of 10 respondents paid “expediter’s fees” to receive their salaries (McCoy et al. 2008).
After the introduction of PMTCT drugs at a clinic, there are steps in the PMTCT cascade. The PMTCT cascade refers to the sequence of actions required to ensure that mother/infant pairs receive antiretroviral drugs to prevent MTCT. Generally speaking, the cascade consists of the following steps: (1) a pregnant woman visits an ANC, (2) the ANC offers voluntary counseling and testing to the woman, (3) the woman accepts the offer, (4) the woman receives the result of the test, (5) the woman agrees to antiretroviral prophylaxis, with (6) adherence to the maternal dose, and (7) adherence to subsequent maternal and infant doses (Stringer et al. 2005, 2008a). Although in practice, the PMTCT cascade requires that clinically eligible women attend an ANC during their pregnancy, PMTCT in Zambia during this period did not require delivery at a health facility (Stringer et al. 2003).
There does not appear to be an authoritative official account of aggregate PMTCT expansion in Zambia. The Zambia DHS Service Provision Assessment (SPA) contains information on health services in Zambia but only in a single cross-section (i.e., 2005). The National AIDS Council (2006) reported that by mid-2005, there were 270 PMTCT sites in Zambia. PEPFAR reported supporting PMTCT services in Zambia at 95 clinics in fiscal year 2004 (PEPFAR 2005), 200 clinics in Zambia in fiscal year 2005 (PEPFAR 2006), and 284 clinics in fiscal year 2006 (PEPFAR 2007). The Ministry of Health and the National AIDS Council (2008:26) stated, “Overall, PMTCT services have been rolled out to all the 72 districts of Zambia, representing an increase from 67 in 2005, 307 in 2006, and 678 as of September 2007. This scaling up of PMTCT services resulted in an increase in pregnant women who completed prophylaxis from 14,071 in 2005 to 25,578 in 2006, and by September 2007 this figure had reached 35,314.” Bweupe (2006) reported other data from the Ministry of Health suggesting that the number of PMTCT sites in Zambia increased from 6 in 2001, to 64 in 2003, to 265 sites in 2005. Using data from the UNICEF Zambia office, Ngashi et al. (2006) reported that in a single year, the number of PMTCT sites increased from 80 to 254.
To identify the centroids of the SEA of residence of respondents in the 2001 DHS, I use a digitized census map provided by the Zambia Central Statistical Office. This map is missing approximately 7 % of the more than 15,000 SEAs in Zambia. Hence, I am unable to identify the precise location of survey respondents in these missing SEAs and exclude them from the empirical analysis.
For the other household surveys that I use in the empirical analysis of the steps in the PMTCT cascade (i.e., the 2003 and 2005 ZSBS), administrative records on the primary sampling units allow me to identify the SEA of residence, and I follow the same centroid procedure.
The 2001 Zambia DHS does not include DHS-provided GPS coordinates. Across all DHS countries, DHS-provided GPS coordinates are much more common in recent survey years than in earlier survey years. See http://dhsprogram.com/data/available-datasets.cfm.
Surprisingly, the proportion of pregnant women offered a HIV test during an ANC visit fell from 22 % in 2001 to 15 % in 2003.
The 2001 and 2007 Zambia DHS record the following vaccinations: BCG, DPT (three doses), polio (three doses, alternatively one dose), and measles (one dose).
For example, Case and Paxson (2011) suggested that non-HIV related health services declined in regions of sub-Saharan Africa with high HIV prevalence relative to those with low HIV prevalence coincident with the rise of the HIV/AIDS pandemic. Because PMTCT expansion occurred with greater intensity in urban (i.e., high HIV prevalence) areas, my main empirical specification allows for time-invariant differences and differences in time trends between PMTCT and non-PMTCT locations.
The evidence in Ashraf et al. (2010) suggested that the Zambia national malaria control campaign, initiated in 2003, reduced all-cause under-5 mortality, and that household bed net ownership (rather than indoor residual spraying) was the main reason for this decline. Evidence from across the developing world has suggested that access to piped water is a major determinant of child health (Fewtrell et al. 2005) and that access to piped water in Zambia may have increased in parallel to PMTCT expansion as well. There was a large expansion in access to ART in Zambia during this period, and ART for adults may affect child health by improving adult health and household ability to improve child health (Lucas and Wilson 2013b). Few infants or children received ART in Zambia during the period that I examine in this analysis.
I also control semiparametrically for PMTCT expansion date by including indicator variables for year of local PMTCT introduction.
In a difference-in-differences interpretation of this regression specification, PMTCTeverij is the “treatment” indicator variable, the birth month × birth year fixed effects correspond to the standard “post” variable, and PMTCTij(mt – 9) is “treatment” interacted with “post.”
Among female respondents in the 2007 DHS, distance is cited as being a primary barrier to seeking health care (Central Statistical Office et al. 2009). Steckelenburg et al. (2004) found that maternal health care usage declines substantially at distances greater than a two-hour walk.
I omit whether the respondent accepted the offer from the regression analysis because of the extremely small sample. There are two reasons for the small sample size in this regression. First, the question is asked only in the 2003 and 2005 ZSBS. Second, it is defined conditional only on being pregnant, visiting an antenatal clinic, and being offered an HIV test.
Standard errors are clustered by SEA of residence in all regressions throughout the analysis.
As in a standard semiparametric difference-in-differences specification, I include untreated respondents (i.e., individuals residing further than 20km from an eventual PMTCT site) in these regressions.
For example, the point estimate on the indicator variable for 48 months before local introduction indicates that locations receiving PMTCT had under-24 month child mortality rates approximately 1 percentage point higher than locations not receiving PMTCT at that time, whereas the point estimate on the indicator variable for 60 months before local introduction indicates the exact opposite. None of the prelocal introduction quasi-event study parameters are statistically significant at conventional levels.
In Fig. 2, 0 on the x-axis denotes 9 to 20 months after local PMTCT introduction and corresponds to the first period that I include in the main measure of PMTCT availability in main regression specification (i.e., Eq. (1)).
I define a location as an early PMTCT recipient if it received PMTCT prior to January, 2003.