American Economic Journal: Applied Economics 2013, 5(2): 58–85 92925 http://dx.doi.org/10.1257/app.5.2.58 The Effect of Absenteeism and Clinic Protocol on Health Outcomes: The Case of Mother-to-Child Transmission of HIV in Kenya† By Markus Goldstein, Joshua Graff Zivin, James Habyarimana, Cristian Pop-Eleches, and Harsha Thirumurthy* We show that pregnant women whose first clinic visit coincides with the nurse’s attendance are 58 percentage points more likely to test for HIV and 46 percent more likely to deliver in a hospital. Furthermore, women with high pretest expectations of being HIV positive, whose visit coincides with nurse attendance, are 25  and 7.4  percentage points more likely to deliver in a hospital and receive PMTCT medi- cation, and 9 percentage points less likely to breast-feed than women whose visit coincides with nurse absence. The shortcomings that pre- vent pregnant women from testing on a subsequent visit are common in sub-Saharan Africa. (JEL I12, J16, O15) S ervice provider absence in the health and education sectors in developing coun- tries is high and widespread. A recent multi-country survey of education and health providers reported rates of provider absence ranging from 20 percent for teachers to nearly 40 percent for health workers (Chaudhury et. al. 2006). Moreover, the effects of the absence can be compounded by shortcomings in the processes associated with public service delivery. While negative associations between * Goldstein: Africa Region Gender Practice and Development Research Group, The World Bank (e-mail: mgold- stein@worldbank.org); Graff-Zivin: University of California, San Diego, IRPS Dept, 9500 Gilman Dr. #0519, La Jolla, CA 92093-0519 (e-mail: jgraffzivin@ucsd.edu); Habyarimana: Georgetown Public Policy Institute, 37th and O Street NW, Old North Room 307, Washington, DC 20057 (e-mail: jph35@georgetown.edu); Pop-Eleches: Columbia University, SIPA, 1022 International Affairs Building, MC 3308, 420 West 118th Street, New York, NY 10027 (e-mail: cp2124@columbia.edu); Thirumurthy: Department of Health Policy and Management, Gillings School of Global Public Health, University of North Carolina at Chapel Hill, Campus Box 7411, Chapel Hill, NC 27599–7411 (e-mail: harsha@unc.edu). This project would not have been possible without the support of the Academic Model Providing Access to Healthcare (AMPATH) and members of the IU-Kenya partnership. We are grateful to David Cutler, Jishnu Das, Rajeev Dehejia, Bryan Graham, Duncan Thomas, and seminar participants at Harvard University and Georgetown University for valuable comments. Many individuals contributed to the implementation of the household survey under the direction of Duncan Ngare (deceased) and the authors. Anja Papenfuss and Leslie MacKeen assisted in managing the survey and the data collection was facilitated by the field supervision of June Ochanda. We also acknowledge the tremendous contributions of interviewer and data entry teams. Financial support for this project was received from the USAID, The World Bank, and Yale University’s Center for Interdisciplinary Research on AIDS (CIRA) through a grant from the National Institute of Mental Health to Michael Merson, M.D. (No. P30 MH 62294). Pop-Eleches acknowledges the generous support of NBER’s Fellowship in Aging and Health Care. The views expressed here do not necessarily reflect those of the World Bank or its member countries. All errors and opinions are our own. † To comment on this article in the online discussion forum, or to view additional materials, visit the article page at http://dx.doi.org/10.1257/app.5.2.58. 58 Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 59 income per capita and absence levels underscore the potential importance of this phenomenon, causal micro-evidence on the impacts of this absence on health and education outcomes is limited. A handful of studies have estimated the impacts of teacher absence on learning (Das et. al 2007; Duflo, Hanna and Ryan 2008; and indirectly, Kremer, Miguel and Thornton 2009). However, similar work has not been done in the area of health.1 This paper examines the role of absence in the context of maternal and child health services, where adequate service delivery can make the difference between life and death, and healthworker absence has the potential to be especially harmful. In the setting that we study in this paper, healthworker absence is not particu- larly high relative to rates reported elsewhere. But even low rates of absence can have significant impacts when processes at health facilities do not provide adequate alternatives for clients when healthworkers are absent. We show that the absence of a nurse who provides HIV testing and counseling services, in conjunction with the lack of an effective institutional response that enables pregnant women to test on a subsequent clinic visit, leads to significant health consequences for mothers and their children. By not receiving HIV testing and counseling during pregnancy, some women fail to receive services for the prevention of mother-to-child trans- mission of HIV (PMTCT). We provide evidence that the problem we describe at the study site is generalizable to the PMTCT outcomes in other countries in sub- Saharan Africa. Fundamentally, the absence of a service provider represents a shock to the supply of health care. There is a large literature on how individuals and house- holds respond to shocks (see, for example, Besley 1995 and Townsend 1995). These responses can be broadly classified into ex ante responses (for example, precautionary savings) and ex post responses (for example, mutual insurance). In the provider absence context, ex ante actions could include hiring additional staff or developing monitoring and incentive mechanisms that reduce unnecessary absence. Since ex ante plans are often both costly and imperfect, investments in ex post measures, such as the implementation of robust standard operating proce- dures that specify how clients who missed services can be identified and linked to care, will be necessary as well. While there is a literature on ex ante mechanisms to cope with absence and their impacts (see, for example, a review of a number of studies in Banerjee and Duflo 2006), little is known about the causal impacts of effective ex post coping mechanisms. In the rural antenatal care (ANC) clinic that we study, pregnant women are to be provided with the opportunity to test for HIV/AIDS. If found to be HIV positive, they are referred to an HIV clinic within the facility and provided with PMTCT services. Typically, the HIV test is provided on the first antenatal clinic visit on an opt-out basis and, theoretically, if women are not tested on their first visit and do not opt out, they should be offered an HIV test on subsequent ANC clinic visits. We show that despite low levels of provider absence, nurse absence 1  Bjorkman and Svensson (2009) find large health gains associated with a randomized intervention that improves service quality along several dimensions, including provider attendance. The impacts of provider absence alone cannot be separated from other program effects in their empirical framework. 60 American Economic Journal: applied economicsapril 2013 does indeed have significant negative health consequences. Women who fail to test during their first visit, when the sole nurse trained in HIV testing for pregnant mothers is absent, are more than 50 percentage points less likely to learn their HIV status during their pregnancy. Of course, it is not nurse absence alone, but also the process of HIV testing in antenatal clinics—specifically the lack of adequate ex post coping mechanisms for nurse absence—that accounts for these negative health consequences. To be fair, the complex nature of medical delivery systems, which involve multiple linkages (i.e., hand offs) between providers with specialized tasks, makes it particularly sus- ceptible to a breakdown at any point in this process (Wachter 2004). In fact, the mismanagement of patient handoffs between medical units and caregivers has been identified as one of the largest contributors to medical errors, even in sophisticated healthcare systems with state-of-the-art information and communication technolo- gies (Kohn, Corrigan, and Donaldson 1999). The response to these challenges in the health care systems of developed countries has been, in part, to redesign care protocols to include more process redundancies, such as the use of checklists, to ensure that critical care elements do not fall between the cracks (Grout 2003; Kumar and Steinebach 2008). In developing countries, where financial and human capi- tal resources are limited, the significant deployment of redundant processes and personnel are quite “costly,” if achievable at all. While the systematization of care through standard operating procedures (SOP) may be helpful, it is unclear how eas- ily care can be standardized without sacrificing specificity in actionable protocols. Indeed, the relevant existing government SOP for our empirical work—HIV testing in the ANC clinic—is clear in its goals, but quite vague in the mechanics of how to achieve them. Since the management of care in a medical clinic is both complex and highly varied over space and time, the “how to” in any protocol is essential. It is thus not surprising that, as we will show later in the paper, many countries appear to be operating in a manner that fails to significantly compensate for failing to test a pregnant woman for HIV on her first visit to the ANC clinic. Our results underscore that this failure, combined with nurse absence on the first visit, results in real health and human costs. The bulk of this paper examines the impact of health worker absence on vari- ous health outcomes among women attending an ANC clinic in an area of Kenya where HIV prevalence is high. We exploit across-time variation in the attendance of the sole nurse qualified to provide HIV testing to pregnant women as part of broader prevention of mother-to-child transmission (PMTCT) services—HIV counseling and testing and the provision of medications to women who test HIV positive—at the ANC clinic. Our empirical strategy is twofold. We begin by showing that the PMTCT nurse’s absence is uncorrelated with a wide range of observable characteristics of the pregnant women, as well as visit date informa- tion. We then present reduced-form estimates of the effects of the nurse’s absence on a range of health outcomes. The lone PMTCT nurse at the ANC clinic was absent on approximately 9 percent of clinic days during the study period. First time visitors to the clinic who arrived on a day when the PMTCT nurse was absent were nearly 60 percentage points less likely to receive HIV testing ser- vices over the entire course of their pregnancy. This impact of nurse absence is Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 61 large and robust to controlling for pretest beliefs about HIV status and date char- acteristics of the first ANC clinic visit. Women whose first visit coincided with the PMTCT nurse’s absence were also 13 percentage points less likely to deliver at a hospital or health center, where deliveries are safest. Since the benefits of a hospital birth and breast-feeding depend on the HIV sta- tus of the pregnant woman, in the second part of our empirical strategy we use an interacted specification to estimate separate effects for women with high and low self-reported pretest expectations of being HIV positive. Since we do not know the HIV status for nontesters in our sample, we proxy HIV status with self-reported pretest expectations of being HIV positive, a variable that we show is predictive of the actual HIV status of women who were tested at the clinic. We find large and significant effects of PMTCT nurse attendance that are consistent with underlying HIV status. Women with a high pretest expectation of being HIV positive, who had a nurse present during their first ANC visit, were nearly 27 percentage points more likely to give birth in a health center or hospital than similar high-risk women whose first visit coincided with nurse absence. More crucially for the long-run health of children, high HIV-positive expectations women whose first visit coin- cided with the nurse attendance are 7.5 percentage points more likely to receive PMTCT medications and 9 percentage points less likely to breast-feed their child than similar high HIV-positive expectations women whose first visit coincided with nurse absence. Given the efficacy of PMTCT medications and the importance of breast-feed- ing on the transmission of HIV from mother-to-child, these impacts indicate that health worker absence in our setting has far-reaching implications for the health outcomes of women and their children. Our estimates imply that the absence of the PMTCT nurse in our study translates into 3.7 additional HIV infections per 10,000 live births. Applying our estimates to the average multi-country study absence rate, and holding other features of the environment constant, implies a four-fold increase in the rate of new infections. To determine whether the findings in our paper (Goldstein et al. 2012) are generalizable to other settings, we also examine Demographic and Health Survey (DHS) data from Kenya, Ghana, Uganda, and Zambia on women’s experiences at antenatal clinics. When comparing the percentage of women who were offered an HIV test during their most recent pregnancy against the number of times that women visited ANC clinics during that pregnancy, we find that the likelihood of being offered an HIV test during pregnancy does not increase with the number of ANC visits at the rate that one would expect if women who did not test during a prior visit were always offered an HIV test during subsequent visits. Thus, it is likely that the process failure that we identify at the clinic studied in this paper is also prevalent at many other ANC clinics. In addition, we cite data on the uptake of PMTCT services, as well as official PMTCT guidelines, to show that the con- clusions from our empirical work in Kenya are consistent with other work on PMTCT services. The remainder of the paper is organized as follows. Section I provides back- ground information on counseling and HIV testing services during antenatal care, and also describes the data. Section II presents the empirical strategy. Section III 62 American Economic Journal: applied economicsapril 2013 presents reduced-form estimates of the effects of health worker absence, as well as the differential impacts by HIV status priors. Section IV focuses on the generaliz- ability of our results to other settings, and Section V concludes. I.  Background and Data The data used in this study were collected by the authors between July 2005 and February 2007. The first wave of data was collected as an in-clinic survey between July 2005 and February 2006. The second wave was a household-based survey implemented between May 2006 and February 2007. The study enrolled a sample pregnant women attending an antenatal clinic at a rural health center in western of ­ Kenya. The health center is located in Maseno Division, a region that has a popula- tion of over 60,000 individuals and lies within Kenya’s Nyanza Province. The health center serves a predominantly rural population, even though a number of patients from the peri-urban areas of Maseno Division use the clinic. The ethnic composi- tion of clinic users is predominantly Luo, although about 10 percent of the sample are Luhya. HIV prevalence in Nyanza Province is the highest of all the provinces in Kenya. Data from the 2007 Kenya AIDS Indicator Survey (KAIS) indicate that 17.2 percent of adult women in the province are HIV positive, compared to a national average of 8.4 percent (National AIDS and STI Control Program 2009).2 The health center offers outpatient, inpatient, and antenatal care services. It also includes an HIV care and treatment clinic that is managed by the US-Kenya academic medical partnership, USAID-Academic Model Providing Access to Healthcare (AMPATH) Partnership. AMPATH provides PMTCT medication for pregnant women who are HIV positive, as well as highly active anti-retroviral therapy (HAART) for patients who have developed AIDS, at no cost to the patient. Typically, women make three to four visits to the antenatal clinic during their pregnancy. In addition to receiving routine antenatal care, women are generally offered counseling and HIV testing services (CTS) at the first visit. If they decline these services during the first visit or if a PMTCT nurse counselor is not present, the women can obtain counseling and HIV testing during subsequent visits. All women are eligible for a pre- and post-HIV test counseling session. As part of the informa- tion provided to women in these sessions, women are encouraged to deliver at the health center or with a professional birth attendant. Women who test HIV positive are counseled on ways to prevent transmission of the virus to their partner and unborn children. For PMTCT, the women are typically referred to AMPATH’s HIV clinic, which is in the same health center. AMPATH provides a full course of HAART to these women during the period before and after delivery (as indicated above there is no charge for the treatment, and the administrative data from AMPATH allow us to establish whether the women in the study enroll in AMPATH). This is consistent with results from the 2003 Demographic and Health Survey (DHS) that 18.3 percent of adult 2  women in Nyanza Province were HIV positive, compared to a national average of just under 7 percent (Central Bureau of Statistics et al. 2004). Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 63 Enrollment into the study was limited to women visiting the ANC clinic for the first time for the observed pregnancy between July 2005 and February 2006. During enrollment we also obtained consent from the study participants to have access to their medical records (including the results of any HIV tests) and, in addition, a short intake questionnaire was administered prior to engaging with the staff at the ANC clinic (we refer to this as wave 1 of the study). Due to the space and time constraints at the clinic, the wave 1 questionnaire was kept fairly brief. This questionnaire obtained information on socioeconomic status, fertility pref- erences, HIV knowledge, and the subjective beliefs about a woman’s own HIV status as well as her partner’s. Data on the presence of the PMTCT nurse on any given day, whether the pregnant women consented to the HIV test, and the test result itself (with patient consent) were obtained from the administrative records of the antenatal clinic.3 Since patients who did not receive CTS during the first visit could do so on subsequent visits to the ANC clinic, administrative records were used to routinely update the CTS status of enrolled women. During the first wave, we also obtained consent from the women to visit them at their homes after delivery.4 Only a handful of first wave respondents did not consent to the home visit. Five hundred ninety-one women who were interviewed at the clinic during wave 1 were located in wave 2, and sample attrition between waves was under 10 percent.5 The second wave of the study was part of a large community-based study of maternal health. This wave of the study included a broader survey instru- ment that included a household roster, questions on education, health, consump- tion, marriage, sexual behavior, assets, income, and transfers. Interviews were also conducted with the husband or cohabiting partner of each woman (if he was present). The geographical coordinates of households and anthropometric data on women and children were also collected during the home visits. In order to ensure comparability of our data with nationally representative data, questions were worded similarly to those in the DHS. Care was taken to ensure that interviews were conducted with sufficient privacy. Wave 1 of the study lasted approximately 40 minutes, including the time taken for obtaining informed consent. Three experienced female enumerators conducted the inter- views in Kiswahili, Luo, or Luhya, depending on the language preferences of the subjects. Table 1 presents summary statistics of several key variables for the entire sample, as well as the subsamples, of women who report low and high priors that they are HIV positive. The average age of the women interviewed in both waves of the sur- vey is 24.7 years, and 59 percent of them report having completed primary school. 3  The PMTCT nurse was defined as absent if on a given day when the ANC clinic was open there was no entry in the PMTCT logbook. We also kept a direct-observation record of PMTCT nurse absenteeism in order to make sure that days on which all ANC visitors refused the test are not coded as days of PMTCT nurse absence. Such a coincidence did not occur during our sample period, and therefore our two approaches of measuring nurse absence produced identical outcomes. 4  Using the expected date of delivery from the administrative records, household visits for the intake respondents were scheduled for approximately two months after delivery. 5  In the majority of cases we could not complete the household interview because the respondent could not be located, despite considerable efforts to track down respondents as far as Nairobi. In analysis not shown here, we find no correlation between the probability of completing the second survey and the nurse absence or pre-test expecta- tions of testing HIV positive. 64 American Economic Journal: applied economicsapril 2013 Table 1—Summary Statistics All Low prob High prob Diffference: women HIV+ HIV+ low versus enrolled women women high Mean SD Mean SD Mean SD p-value Panel A. Selected ANC user characters Variables Age in years 24.69 6.36 24.33 6.25 25.51 6.53 0.04 Fraction completed primary school 0.59 0.49 0.60 0.49 0.58 0.50 0.58 Fraction married or living with partner 0.76 0.42 0.77 0.42 0.76 0.43 0.97 Freq. church attendance, past 4 weeks 3.34 2.53 3.43 2.57 3.13 2.44 0.20 Number of sexual partners 1.02 0.31 1.02 0.33 1.01 0.27 0.50 Fraction boils water 0.77 0.42 0.77 0.42 0.78 0.41 0.81 Number of livestock 2.09 3.48 2.07 2.93 2.14 4.52 0.83 Fraction iron roof 0.73 0.44 0.73 0.44 0.73 0.45 0.99 Fraction located with Maseno Division 0.74 0.44 0.75 0.43 0.74 0.44 0.75 Tested for HIV 0.77 0.42 0.75 0.44 0.83 0.38 0.02 Tested HIV positive 0.15 0.36 0.12 0.32 0.24 0.43 < 0.001 Nurse present at first ANC visit 0.91 0.29 0.89 0.31 0.94 0.24 0.10 Received counselling/testing—self-report 0.88 0.32 0.88 0.33 0.90 0.30 0.40 Delivered in the health center or hospital 0.39 0.49 0.41 0.49 0.34 0.48 0.15 Data from Waves 1 and 2 Subjective belief about HIV status (Scale 1–4 decreasing in risk) Wave 1 2.76 0.88 3.25 0.44 1.61 0.49 < 0.001 Wave 2 2.78 1.06 2.87 1.03 2.58 1.11 < 0.001 Data from Wave 2 only Received PMTCT medication at birth 0.06 0.24 0.06 0.24 0.07 0.25 0.76 Mother reports breast-feeding 0.95 0.22 0.96 0.19 0.92 0.27 0.05   newborn child Observations 587 409 178 Panel B. Distribution of subjective beliefs about HIV status Wave 1 Wave 2 Chances of being HIV positive N Percent N Percent Great 69 11.8 74 12.9 Moderate 109 18.6 177 30.9 Small 305 52.0 121 21.2 None 104 17.7 200 35.0 Total 587 100 572 100 Notes: SD is the standard deviation and N is the sample size. Sample of women enrolled during first ANC clinic visit (wave 1) and interviewed at home after delivery (wave 2). p-value reported in panel A corresponds to a null hypothesis of no difference in average characteristics between ANC users with high and low probability of being HIV positive. Seventy-six percent of the women report being married or living with their partner. Seventy-seven percent of women enrolled in our study and located in wave 2 were tested for HIV during one of their antenatal clinic visits. Among those tested, just over 15 percent were HIV positive. For 91 percent of the women, a PMTCT nurse was present on the day of their first ANC clinic visit. Several outcomes pertaining to the pregnancy and delivery are of interest. First, women’s self-reports during wave 2 on whether testing and counseling ser- vices were offered at the ANC clinic correspond well to the actual testing rate indi- cated by the PMTCT logbooks (the self-reported rates are in fact slightly higher). Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 65 While nearly half the women in our sample report that they delivered their child with the assistance of a traditional birth attendant and at home, 39 percent reported having delivered in a public or private health facility or hospital. Table 1 also summarizes the other key variables used in our empirical strategy. Subjective beliefs about one’s chances of being HIV positive were measured in each wave on a scale of 1–4 (with 1 indicating “great chance” and 4 indicating “no chance at all” of being HIV positive). The mean for this subjective measure of beliefs is 2.76 in wave 1 and the distribution of beliefs is shown in panel B. We define women with low priors as those who in the baseline survey believe that they have little or no chance of being HIV positive, and we also define women with high priors as those who believe that they have a moderate or great chance of being HIV positive. In the same table we also present all the summary statistics separately for women with low and high priors. For women who report high priors of being H ­ IV positive, the mean in wave 1 is clearly lower than the mean for low-prior women. On most dimensions of baseline household characteristics, women who report low priors of being HIV positive are similar to high-prior women. Age is the only covariate that is statisti- cally different between these two groups, as women with high priors are slightly older than low-prior women. Our ANC sample is very similar to the population of young or expecting moth- ers in this part of Kenya. Nearly three-quarters of the women in both our sample and the DHS Nyanza Province sample live in houses with a durable materials roof. Along the dimension of desired fertility, both samples report a similar aver- age desired number of children, four. Knowledge about HIV/AIDS is very high in both samples. Nearly 90 percent of women in both samples report knowing that an individual who appears healthy can have HIV and that HIV can be transmitted from a mother to a child. A similar proportion of women in both samples report knowing someone who has died of HIV/AIDS. Finally, HIV testing rates appear considerably higher in our sample. Women enrolled at the ANC clinic are three times more likely to have had an HIV test. This difference is likely driven by temporal differences in testing rates possibly related to the recent availability of anti-retroviral medications.6 II.  Empirical Strategy The first step of our analysis is to obtain the reduced-form effect of nurse absence on a range of health outcomes. We estimate regressions of the form (1)  ​Y​ β​  = ​ i​  + ​ o​ β​ ​​ 1​ ​Xi  + ​ β​ ​​ 2​ ​Wi β​  + ​ 3​ ​P​ + ​ i​ ε​  , i​ where Y is an outcome variable of interest; X is a set of individual characteristics, such as education, age, distance from the clinic, and marital status; W represents ­ ndicator visit date characteristics, such as the day of the week or month; and P is an i 6  There is also a sharp difference in mosquito net ownership. Nearly twice as many women in our sample report owning a mosquito net compared to the DHS sample. The difference likely arises from recent aggressive marketing and distribution of mosquito nets that has taken place in this area in the period between the surveys. 66 American Economic Journal: applied economicsapril 2013 for whether the counseling and testing nurse was present on the first visit to the ANC clinic. β ​​represents the reduced-form effect of nurse attendance on health ​3 outcomes, such as learning HIV status as a result of a test at the clinic, the choice of delivery location, receipt of PMTCT medication, and breast-feeding. In all our specifications, our standard errors are clustered on the visit date level. Our estimation strategy will not reproduce the reduced-form effect of nurse absence on outcomes (​β3 ) if nurse absence is correlated with unobserved patient ​​ characteristics that affect outcomes (Cov(​ P​  )  ≠  0). The identifying ­ , ​εi​ ​ i​ assumption underlying our analysis is that after controlling for observable household and ANC user characteristics, visit date characteristics, and priors about HIV status, the demand for the information and services provided by the PMTCT nurse for women who visit the clinic on days when the nurse is present is the same as on days when she is absent. Our empirical strategy would be invalid if, for example, a selected sub- sample of women with particular unobservable characteristics, who want to avoid CTS, come to the clinic for antenatal care on days when the PMTCT nurse is absent or more likely to be absent. In order to address this concern, we first start by showing that the presence or absence of a PMTCT nurse on the day of a woman’s first antenatal visit is uncor- related with observable characteristics of pregnant women, their beliefs about their perceived probability of having HIV/AIDS, and visit date characteristics. In Table 2, we report the results from a cross-sectional regression of an indicator of nurse presence on ANC user and first visit date characteristics. In column 1, we include a range of socioeconomic characteristics of the woman and her house- hold. The results in column 1 suggest that the likelihood that a nurse is absent on the woman’s first antenatal visit is uncorrelated with observable characteristics, such as the age, education, marital status, and other measures of household well- being.7 To address additional sources of bias we include, in column 2, the quarter in which the baby was conceived.8 The results suggest the lack of a systematic association between nurse attendance patterns and the timing of conception. In column 3, we include self-reported beliefs that the woman is HIV positive to con- trol for a wide variety of observable and unobservable determinants of the demand for counseling and testing. Holding observable characteristics constant, we find no systematic association between reported beliefs and the nurse’s likelihood to be absent. Finally, it is possible that women can use information about patterns of absence unknown to the researcher to select visit dates where the nurse is more or less likely to be absent. We examine this possibility by including controls for ­ olumn 4. In particular, we include indicator variables visit date characteristics in c for each day of the week and a quadratic in the day of the month. Compared to our base variable (Monday), most of our day of the week indicators are small and statistically insignificant. However, there is evidence that Fridays are associated with a 18 percentage point greater absence rate compared to Mondays. Of note is 7  Anecdotal evidence from the study area suggests that the reasons for absence include official reasons, such as collection of salaries and attendance at workshops, illness of self/members of the family, and funeral attendance. It is unlikely that information about these “shocks” to attendance would be available to any of the ANC users. 8  We use the quarter rather than the month of conception to deal with measurement error associated with premature birth as well as to conserve degrees of freedom. However, our results are robust to controlling for month of conception. Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 67 Table 2—Correlates of Nurse Attendance Nurse present at time of woman’s first visit (1) (2) (3) (4) Age in years −0.013 −0.013 −0.015 −0.012 (0.012) (0.013) (0.013) (0.013) Age in years, squared 0.000 0.000 0.000 0.000 (0.000) (0.000) (0.000) (0.000) Completed primary school −0.012 −0.011 −0.012 −0.010 (0.022) (0.022) (0.022) (0.024) Married 0.037 0.036 0.041 0.042 (0.027) (0.027) (0.028) (0.030) Frequency of church attendance in past four weeks −0.002 −0.002 −0.001 −0.000 (0.004) (0.004) (0.004) (0.004) Boils drinking water −0.026 −0.021 −0.018 −0.018 (0.027) (0.023) (0.024) (0.025) Number of livestock held at enrollment −0.000 −0.000 −0.000 −0.001 (0.004) (0.004) (0.004) (0.004) Lives in a non-thatched house −0.009 −0.007 −0.006 −0.005 (0.031) (0.030) (0.028) (0.029) Lives in Maseno division −0.005 −0.005 −0.006 −0.006 (0.026) (0.027) (0.026) (0.026) Log distance from clinic −0.004 −0.005 −0.005 −0.003 (0.021) (0.021) (0.020) (0.019) Quarter of conception == 2 0.017 0.017 0.013 (0.038) (0.038) (0.034) Quarter of conception == 3 −0.030 −0.035 −0.039 (0.057) (0.058) (0.054) Quarter of conception == 4 −0.045 −0.047 −0.066 (0.050) (0.052) (0.052) Moderate chance HIV + ve 0.017 0.012 (0.039) (0.040) Small chance HIV + ve −0.027 −0.029 (0.036) (0.037) No chance at all HIV + ve −0.069 −0.075 (0.054) (0.054) Day of week = Tuesday −0.028 (0.048) Day of week = Wednesday −0.045 (0.043 Day of week = Thursday −0.098 (0.068) Day of week = Friday −0.188 (0.093)* Day of the month 0.004 (0.011) Day of the month squared −0.000 (0.000) Constant 1.086 1.086 1.140 1.151 (0.217)*** (0.231)*** (0.252)*** (0.242)*** Observations 581 581 577 577 2 ​​ R ​ 0.01 0.01 0.02 0.07 Notes: The variables are defined in Table 1. “Nurse present at time of woman’s first visit” takes value one if the PMTCT nurse was present at the ANC clinic on the day of the first visit during this pregnancy, zero otherwise. Standard errors in brackets clustered at the visit date level. *** Significant at the 1 percent level.  ** Significant at the 5 percent level.   * Significant at the 10 percent level. 68 American Economic Journal: applied economicsapril 2013 the fact that we find no evidence of a systematic relationship between absence pat- terns and the day of the month. In sum, our evidence suggests that the composition of women whose first visit coincides with the nurse’s presence is not measurably different from those who visit when she is absent. A number of institutional details and additional robustness checks may help assuage any remaining doubts about this identification strategy. Firstly, for selection bias to be present, potential ANC visitors need to be able to observe and predict pat- terns of nurse absence. Based on our two year experience working with the clinic, this is unlikely to be the case since absences of the medical staff are rarely preannounced or advertised (consistent with the results in Banerjee, Deaton, and Duflo 2004). Moreover, since the majority of women travel significant distances to the clinic, it is unlikely that they could have access to such information at home, even if it were available. Secondly, the average rate of nurse absence (9 percent of days) and the variation by day of week, which ranges from 3.4 percent on Mondays to just under 22 percent on Fridays, is small enough that strategically choosing to visit on a day when the nurse is absent seems unlikely.9 This possibility is made all the more implausible by the fact that women can always opt out of CTS (20 percent of women who visit on a day when the nurse is present do indeed decline to be tested). Third, we have performed a number of additional robustness checks (not reported) and found no consistent relationship between the characteristics of women and the day of the week when they visit the ANC clinic. For example, there is no significant relationship between the distribution of pretest beliefs that women have about their HIV status and the day of the week of their first visit. In addition, there is no rela- tionship between the distribution of pretest beliefs about HIV status and the pres- ence of the PMTCT nurse. Next, we extend our main framework to capture the heterogenous treatment effects of absence by HIV status. Understanding these heterogenous responses are of particular interest in this setting because the benefits from contact with the PMTCT nurse are expected to be larger for women who are HIV positive, as well as for their infants. It is worth noting that since we do not observe in our data the HIV status of women who do not get testing and counseling services, we are unable to use actual HIV status as the variable that is interacted with nurse atten- dance. Instead, our specifications use the self-reported belief from the baseline survey as a proxy for actual HIV status. We thus estimate the following regression model: Y​ (2)  ​ β​  = ​ i​  + ​ o​ β​1​ ​X​ + ​ i​ β​ ​​ 2​ ​Wi  + ​ β​ ​​ 3​ ​Pi  + ​ β​ ​​ 4​  lo​wi + ​ β​5​  P ×   + ​ lo​wi​​ ε​ . i​ Most variables are defined in equation (1). lo​wi​​is a dummy taking value zero for women who in the baseline survey believe that they have a moderate or great chance of being HIV positive, and value one for women who believe that they have little or 9  It should be emphasized that the nurse’s rate of absence at the clinic is considerably lower than levels that have been documented in other developing country settings. Average levels of absence for nurses from a multi-country study are more than three times as large (see Chaudhury et al. 2006). Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 69 no chance of being HIV positive. The main coefficient of interest are ​ β3 β5 ​​and ​ ​​, as they indicate the impact of nurse attendance for women with high priors (​β3 ) and the ​​ impact for women with low priors (​β3  + ​ ​​ β​ ) of having HIV. 5​ Before proceeding, we discuss two identification challenges for our analysis of heterogenous treatment effects by HIV status. The first identification issue is that our measure of self-reported beliefs of HIV status may be correlated with a num- ber of observable and unobservable characteristics of the women. For example, one might worry that if age and self-reported beliefs about HIV are correlated, our coefficient of interest (​β​5​ ) might also pick up the differential effect of nurse absence by age. As a robustness check we will show specifications where we also add as a control the i ­nteraction of nurse absence with an index of socioeconomic status that includes age (quadratic), marital status, education, distance from the clinic, housing c ­ haracteristics, and livestock holdings. This principal components index ­captures potential earnings and/or wealth during the life cycle and con- serves degrees of freedom. The second issue is whether self-reported beliefs of HIV status is a good proxy for underlying HIV status. A priori, bias in self-reported beliefs about HIV status might arise from stigma-related concerns that prevent women from revealing their true beliefs to an enumerator or from poor survey comprehension. As mentioned earlier, not all women tested for HIV at the ANC clinic, but for the 77 percent who do get tested, we can confirm that the reported beliefs are good predictors of actual HIV status.10 Column 1 of Appendix Table A1 shows that compared to women who reported “no chance at all” of being HIV positive at the time of enrollment, women who reported a “moderate” or a “great” chance were approximately 17 and 27 ­ percentage points more likely, respectively, to test HIV positive (these differences are statistically significant). These results persist when we control for visit date characteristics as well as the timing of conception in column 2. Adding observable characteristics of women in column 3 (such as age, education, and wealth) reduces the predictive power of beliefs slightly, as indicated by the change in the p-value of the Chi-squared test of no predictive power of self-reported beliefs. Even then, we reject the null of no information in self-reported beliefs at the 5 percent level. It is noteworthy that conditional on HIV status priors, only age significantly predicts HIV status. In column 4, we show that sample selection driven by nonresponse on some control variables does not drive the results. In col- umns 5–8, we show that an indicator of whether the woman reports a moderate/ great chance of HIV increases the likelihood that she tests positive by 12 percent- age points. Overall, these results provide support for the strategy we implement to uncover heterogenous reduced-form effects of absence. 10  The regressions in Appendix Table A1 suffer from potential sample selection bias given that the choice to test is endogenous. A Heckman selection model (not reported) using the nurse absence as an instrument for selection into HIV testing corroborates the findings here that self-reported beliefs predict HIV status. 70 American Economic Journal: applied economicsapril 2013 Table 3—Effect of Nurse Absenteeism on Testing Dependent variable: Indicator for tested for HIV during pregnancy (1) (2) (3) (4) (5) Panel A PMTCT nurse present 0.587 0.568 0.558 0.557 0.587 (0.067)*** (0.066)*** (0.065)*** (0.065)*** (0.067)*** Constant 0.241 0.200 0.252 0.469 0.241 (0.065)*** (0.099)** (0.101)** (0.279)* (0.065)*** Visit date controls X X X HIV Priors X X Controls X Observations 588 588 584 577 577 ​R2 ​​ 0.16 0.19 0.19 0.22 0.17 Panel B PMTCT nurse present 0.596 0.575 0.586 0.585 0.575 (0.138)*** (0.138)*** (0.138)*** (0.145)*** (0.138)*** Low prior HIV +ve −0.040 −0.045 −0.035 −0.036 −0.045 (0.155) (0.157) (0.158) (0.164) (0.157) PMTCT nurse present × Low prior HIV +ve −0.018 −0.018 −0.034 −0.033 −0.018 (0.159) (0.160) (0.162) (0.168) (0.160) Constant 0.273 0.251 0.455 0.452 0.247 (0.135)** (0.152) (0.294) (0.300) (0.151) Visit Date Controls X X X X Controls X X SES Index × present interaction X Observations 584 584 577 577 577 ​R2 ​​ 0.17 0.19 0.21 0.21 0.19 Test: Presence no effect on low prior subjects 55.66 52.67 51.74 51.70 52.90 prob > F 0.00 0.00 0.00 0.00 0.00 Notes: Standard errors in brackets clustered at the visit date level. The dependent variables are defined in Table 1. “Tested for HIV” takes value one if a pregnant woman was given an HIV test during any visit at the ANC clinic dur- ing pregnancy, zero otherwise. PMTCT nurse present takes value one if the PMTCT nurse was present at the ANC clinic on the day of the first visit during a particular pregnancy, zero otherwise. Visit date controls include the day of the week, day of the month and day of the month squared. Controls include age, age squared, an indicator for primary school completion, married, church attendance, reports boiling water, has permanent roof, location in the district, number of initial livestock holdings, quarter of conception, and log distance to the clinic. *** Significant at the 1 percent level.  ** Significant at the 5 percent level.   * Significant at the 10 percent level. III. Results A. Impact of Nurse Presence on Uptake of HIV Testing In panel A of Table 3, we begin with the impact of the PMTCT nurse presence on the likelihood that women learn their HIV status during the observed preg- nancy. The dependent variable for the regressions is an indicator for whether or not a woman learns her HIV status during the course of this pregnancy. In column 1, we present the unconditional estimate and add visit date, self-reported beliefs, and ANC user and household characteristics in columns 2, 3, and 4, respectively. Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 71 In column 5, we show that sample selection due to nonresponse does not drive our results. Across all specifications we find a very large and statistically significant effect of nurse presence during the first ANC visit on the likelihood that women learn their HIV status. The point estimates from our different specifications range between 55 and 59 percentage points. The robustness of these results to the inclu- sion of different controls also alleviate the earlier concerns that the absence of the PMTCT nurse might be correlated with types of women who attend the clinic on such days. Despite the fact that women whose first visit coincides with the PMTCT nurse’s absence make additional visits to the clinic, only one out of four women learns their H ­ IV status during other ANC visits. In comparison, a woman whose first visit coincides with the nurse’s attendance is three times more likely to learn her HIV status. The very large effect of absence on the uptake of HIV ­ uggests that the referral system at this health center is broken. While testing s women whose first visit coincides with the nurse’s absence should in principle have about three more opportunities to learn their HIV status, poor records man- agement implies that three out of four such women are not identified as needing HIV counseling and testing. Overall, the estimates in panel A of Table 3 suggest that the presence of the PMTCT nurse is critical to important health outcomes. B. Impact of Nurse Presence on Delivery and PMTCT Outcomes The immediate impact of the absence of a PMTCT nurse is that it can affect the likelihood that women take up important services that influence child delivery outcomes. The principal reason for offering HIV testing and counseling during antenatal care is that it identifies HIV-positive women who can be given medica- tions for the prevention of mother-to-child transmission of HIV. To enhance the chances that PMTCT medications are taken at the time of delivery, it is typically advised that HIV-positive women deliver in a health center or at the very least use a professional birth attendant who can administer the PMTCT medications. More broadly, for all women who take advantage of HIV testing and counsel- ing, the PMTCT nurse reinforces the importance of delivering at a health center or using sufficiently trained birth attendants.11 Since pregnant women and their households may weigh the costs of delivery in a formal setting against the per- ceived benefits, information gained during pre- and post-test counseling sessions may alter the trade-offs toward safer delivery and greater take-up of PMTCT medications. The reduced-form impact of nurse presence on antenatal, delivery, and post- natal outcomes is reported in panels A and B of Table 4. In columns 1 and 2 of Table 4, we examine the impact of nurse presence on the likelihood that the women deliver in an environment where they can obtain relatively high-quality obstetric care. Columns 3 and 4 examine how nurse presence affects the number of sub- sequent ANC visits, while columns 5 and 6 focus on the self-reported uptake of 11  This evidence is based on an interview at the clinic with the PMTCT nurse. 72 American Economic Journal: applied economicsapril 2013 Table 4—Effect of Nurse Absenteeism on Health Outcomes Delivered at Number of times Given any medication hospital or visited clinic for this to prevent mother health center pregnancy to child HIV (1) (2) (3) (4) (5) (6) Panel A PMTCT nurse present 0.115 0.132 0.056 −0.031 0.045 0.037 (0.058)* (0.054)** (0.288) (0.276) (0.035) (0.033) Constant 0.281 0.304 3.339 1.658 0.025 −0.246 (0.098)*** (0.321) (0.434)*** (1.461) (0.064) (0.129)* Visit date controls X X X X X X HIV priors X X X X X X Controls X X X Observations 576 564 570 558 571 559 2 ​R​ ​ 0.03 0.12 0.03 0.08 0.02 0.06 Mean of 0.28 3.74 0.04   dependent variable | nurse absent Breastfed Enrolled in baby ampath treatment (7) (8) (9) (10) Panel B PMTCT nurse present −0.012 −0.008 0.035 0.037 (0.039) (0.039) (0.021) (0.023) Constant 0.995 1.330 −0.013 −0.372 (0.042)*** (0.119)*** (0.043) (0.144)** Visit date controls X X X X HIV priors X X X X Controls X X Observations 576 564 576 564 ​R2 ​​ 0.03 0.06 0.05 0.07 Mean of 0.94 0.05   dependent variable | nurse absent Notes: Standard errors in brackets clustered at the visit date level. The dependent variables are all indicators defined in Table 1. All specifications include controls for day of the week, date, and HIV status priors. PMTCT nurse pres- ent takes value one if the PMTCT nurse was present at the ANC clinic on the day of the first visit during a particular pregnancy, zero otherwise. Visit date controls include the day of the week, day of the month and day of the month squared. Controls include age, age squared, an indicator for primary school completion, married, church attendance, reports boiling water, has permanent roof, location in the district, number of initial livestock holdings, quarter of conception, and log distance to the clinic. *** Significant at the 1 percent level.  **  Significant at the 5 percent level.   *  Significant at the 10 percent level. ­ edication to prevent the vertical transmission of HIV. In panel B, columns 7 m and 8 look at the effects on whether mothers breast-feed, while columns 9 and 10 examine the effect on enrollment into the AIDS treatment program. We include controls for visit date characteristics and HIV-status priors in all specifications in the odd numbered columns and add ANC user socioeconomic characteristics in the even numbered columns. Our preferred estimates are drawn from specifica- tions 2, 4, 6, 8, and 10, which have the full set of controls. We find a large and significant effect of nurse attendance on the choice to deliver in a hospital or Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 73 health center. The estimate suggests that women whose first ANC visit coincides with the nurse’s attendance are 13 percentage points more likely to deliver in a hospital or health center than women whose first visit coincides with the nurse’s absence. This represents a large—nearly 50 percent—increase in the likelihood of delivering in a considerably safer environment. In columns 3 and 4, we find that nurse absence during the initial ANC visit does not appear to affect the number of subsequent ANC visits. We will discuss the interpretation of this result in the concluding section, but note that this result is suggestive with absence not having an impact on the future demand for health services caused by a loss of confidence in the medical system. We find no effects of nurse presence on the likelihood of reporting the use of medication to prevent the vertical transmission of HIV. While the point estimates on PMTCT uptake are economically large, they are imprecisely estimated. This finding could also be explained by the fact that in our reduced-form regressions, the sample includes a large fraction of HIV-negative women for whom the use of PMTCT medications is generally not recommended. ­ olumns 7 and 8, we find no effect of nurse absence on breast-feeding patterns In c and enrollment into the AIDS treatment program. These results are not surprising since during ­ ­ ounseling ­ prenatal c HIV-positive and HIV-negative women receive opposite advice regarding breast-feeding, and AIDS treatment programs are only appropriate for those testing positive. C. Do Impacts of Health Worker Presence Differ by HIV Status? Table 5 explores the differential impact of PMTCT nurse presence by HIV status for the same outcome variables used in Table 4. We estimate equations in which the variable for nurse absence on the day of the first PMTCT visit is inter- acted with an indicator of whether at baseline the pregnant woman believes she has a low probability of being HIV positive. As discussed above, we use the self- reported beliefs instead of the actual results from the HIV test since about 23 per- cent of women in our sample do not get tested at the clinic during their pregnancy. Nevertheless, since our data from the sample of testers indicates that baseline self-reported beliefs are good predictors for underlying status, it suggests that pretest beliefs can be used to understand the heterogenous reduced form impacts of nurse absence. The two key estimates are drawn from the main and interacted effects of nurse presence. The main effect measures the impact of nurse attendance for women who report a high likelihood of being HIV positive, while the sum of the main and interacted effects measure the impact of nurse presence on low-prior women. As in Table 4, we control for visit date and ANC user characteristics and our preferred estimates are drawn from columns 1, 3, 5, 7, and 9, which also include a set of background controls. We find considerable heterogeneity in the impact of health worker attendance on child delivery outcomes. High-prior women whose first visit coincides with the nurse’s attendance are 25 percentage points more likely to deliver in a health center or hospital than high-prior women whose first visit coincides with nurse absence. This effect is large and statistically significant at the 5 percent level. For 74 American Economic Journal: applied economicsapril 2013 Table 5—Effect of Nurse Absenteeism on Health Outcomes: Interactions with Beliefs about HIV Status Given any Delivered at Number of times medication to prevent hospital or visited clinic for this mother to child HIV health center pregnancy transmission (1) (2) (3) (4) (5) (6) Panel A PMTCT nurse present 0.246 0.234 −0.150 −0.158 0.074 0.076 (0.117)** (0.122)* (0.710) (0.755) (0.028)*** (0.031)** Low prior HIV +ve 0.216 0.204 −0.282 −0.289 0.051 0.052 (0.136) (0.138) (0.761) (0.792) (0.038) (0.032) PMTCT nurse present ·  low prior −0.148 −0.138 0.149 0.155 −0.049 −0.051   HIV +ve (0.142) (0.144) (0.788) (0.817) (0.046) (0.041) Constant 0.127 0.081 1.879 1.849 −0.292 −0.286 (0.334) (0.339) (1.518) (1.499) (0.127)** (0.125)** Visit date controls X X X X X X Controls X X X X X X SES index · present interaction X X X Observations 564 564 558 558 559 559 2 ​R​ ​ 0.12 0.12 0.08 0.08 0.06 0.06 Test: Presence no effect on low prior 2.08 1.98 0.00 0.00 0.36 0.35  subjects prob > F 0.15 0.16 1.00 0.99 0.55 0.56 Enrolled in Breastfed ampath baby treatment (7) (8) (9) (10) Panel B PMTCT nurse present −0.089 −0.102 0.098 0.111 (0.024)*** (0.027)*** (0.030)*** (0.033)*** Low prior HIV +ve −0.062 −0.074 0.011 0.023 (0.049) (0.054) (0.029) (0.033) PMTCT nurse present · low prior 0.103 0.115 −0.076 −0.088   HIV +ve (0.053)* (0.057)** (0.042)* (0.044)** Constant 1.355 1.303 −0.348 −0.297 (0.123)*** (0.117)*** (0.135)** (0.151)* Visit date controls X X X X Controls X X X X SES index · present interaction X X Observations 564 564 564 564 ​R​2​ 0.06 0.06 0.06 0.06 Test: presence no effect on low prior 0.09 0.07 0.51 0.72  subjects prob > F 0.76 0.78 0.47 0.40 Notes: Standard errors in brackets clustered at the visit date level. The dependent variables are defined in Table 1. All specifications include controls for day of the week and date. PMTCT nurse present takes value one if the PMTCT nurse was present at the ANC clinic on the day of the first visit during a particular pregnancy, zero oth- erwise. Visit date controls include the day of the week, day of the month and day of the month squared. Controls include age, age squared, an indicator for primary school completion, married, church attendance, reports boil- ing water, has permanent roof, location in the district, number of initial livestock holdings, quarter of conception, and log distance to the clinic. *** Significant at the 1 percent level.  ** Significant at the 5 percent level.   * Significant at the 10 percent level. Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 75 low-prior women, the effect size indicated by the sum of the main and interacted terms is considerably smaller relative to the impact on high-prior women. Low- prior women whose visit coincides with the nurse’s attendance are only 10 per- centage points more likely to deliver in a health center or hospital than low-prior women who arrive when the nurse is absent. The low-prior effect is also impre- cisely estimated with p-values ranging between 0.15 and 0.32 across the different specifications. In columns 3 and 4, we do not find heterogeneity in health worker presence on number of ANC visits. In columns 5 and 6, we document the heterogenous impact of health worker presence on the likelihood of receiving medication to prevent ver- tical transmission of HIV. High-prior women whose first visit is on a day when the PMTCT nurse is present are 7.4 percentage points more likely to report receiving PMTCT medication than high-prior women whose visit coincides with the nurse’s absence. As we would expect for this outcome, health worker absence has no sta- tistically significant effect on low-prior women. Similarly in columns 7 and 8, we estimate the differential effect of health worker presence on breast-feeding behav- ior for high- and low-prior ANC users. The impact of health worker presence on ­ high-prior women is to reduce the likelihood that they breast-feed by nearly percentage points. For low-prior women, we estimate a very small and statisti- 9 ­ cally insignificant impact of attendance on the likelihood of breast-feeding. The uptake of PMTCT medication and abstaining from breast-feeding are both strate- gies to reduce vertical transmission of HIV to children. Any impact of the health worker’s presence should only matter for those women most likely to be HIV positive. In particular, it suggests that information delivered in the pre- and par- ticularly the post-HIV test counseling sessions has large impacts on child health outcomes. Finally in columns 9 and 10, we examine the impact of nurse presence on enrollment in the free AIDS treatment program at the health center. Only 5 per- cent of our sample enrolls in this treatment program. Our preferred results in col- umn 9 suggest that for women most likely to test HIV positive, arriving on a day when the nurse is present increases the likelihood that you enroll in the treatment program by 10 ­ percentage points relative to when the nurse is absent. This point estimate suggests that nurse attendance has a three-fold effect on the likelihood of enrolling in a treatment program. Given recent evidence that AIDS treatment outcomes are considerably better when treatment starts earlier (Thompson et al. 2010), these results imply large long-term benefits to women likely to test positive and arriving on a day when the nurse is present. As with the breast-feeding and PMTCT result, the effect of nurse attendance on low-prior women is small and statistically insignificant.12 The results above are robust to including interactions between nurse presence and an index that summarizes age, education, marital status, distance, and wealth holdings of ANC users. In columns 2, 4, 6, 8, and 10, including an interaction of absence and this principal components index of social economic status does 12  In panel B of Table 3, we do not find evidence of a heterogenous impact for HIV testing for high- and ­low-prior women. 76 American Economic Journal: applied economicsapril 2013 not change the magnitude or significance of the coefficients reported above. The results suggest that over and above visiting an antenatal clinic, the PMTCT nurse’s presence has large effects on the behavior of pregnant women that translate into large gains in child and maternal health. In addition to the public resources leak- age associated with health provider absence, these results suggest considerable adverse effects on the health of the intended beneficiaries of HIV testing and their newborn children. IV.  Generalizability of Results One important issue that our analysis has not addressed so far is related to the external validity of our finding. Our analysis is based on data from one clinic in rural Kenya. In this section we provide two types of evidence that suggest our results appear applicable to a wider range of settings. First, we cite existing studies indicat- ing that failure to offer HIV testing is one important factor that has constrained the rapid scaling up of PMTCT interventions. Secondly, we use data from a number of DHS surveys implemented recently in African countries to show that in many countries providers of ANC care do not seem to offer enough additional HIV ­ testing opportunities to pregnant mothers who fail to get tested on their first visit to the health facility. Following the discovery of several antiretroviral interventions that were proven to reduce the risk of mother-to-child HIV transmission and the subse- quent development of clear guidelines to implement and use such medications during ­antenatal care and pregnancy, many policymakers believed that PMTCT would be a relatively simple matter of incorporating antenatal HIV testing and maternal-infant antiretroviral prophylaxis into routine pregnancy and newborn care. Intrapartum and ­neonatal single-dose nevirapine was considered to be a very simple intervention (Guay 1999). However, in practice, it has been extremely challenging to bring this and other PMTCT interventions to scale (World Health Organization 2009). Relatively few studies have rigorously identified the reasons for the failure of near-universal PMTCT coverage. Many programs, for example, do not identify the major programmatic bottlenecks to testing all women and providing prophy- laxis to those who need it. Hence, it has been unclear whether low PMTCT cover- age is the result of a failure to offer HIV testing as part of antenatal care, a failure of women prescribed single-dose nevirapine to self-administer the medication at labor onset, or various other factors. However, a landmark study conducted in four African countries (the PEARL study) sought to estimate the coverage of existing PMTCT services (Stringer et al. 2010) by measuring the population nevirapine coverage, defined as the proportion of HIV-exposed infants in the sample with both maternal nevirapine ingestion and infant nevirapine ingestion. The study, con- ducted in 43 randomly selected facilities providing delivery services in Cameroon, Côte d’Ivoire, South Africa, and Zambia, established a path (or cascade) of events that needed to take place in order to successfully prevent HIV transmission from mother to infant. Only 51 percent of HIV-exposed infants were determined to have received the minimal regimen of single-dose nevirapine. Failure to offer HIV Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 77 testing was one of several important reasons for infants not receiving PMTCT medications. Importantly, the study found that the problematic step in delivering PMTCT services varied across clinics. In some clinics the problem was that few women were offered HIV testing, whereas in others it was nondelivery of maternal or infant nevirapine and n ­ on-adherence to medication. Each clinic faced its own mix of challenges in maximizing service coverage—a finding that is relevant to our study, as it suggests that clinic-specific factors are important in explaining low PMTCT coverage. In addition to the PEARL study, a number of features in various countries’ PMTCT guidelines reveal why the process by which HIV testing is offered during antenatal care may be responsible for pregnant women never being tested for HIV. In several countries there is a disproportionate focus on offering HIV testing to new (i.e., first-time) ANC visitors. This makes sense as it is the first opportunity to offer HIV testing to pregnant women, but is problematic if women are not being screened for HIV testing during subsequent visits to the antenatal clinic. In fact, there is evidence from one program in South Africa that first-time visitors are actually segregated and told about mother-to-child transmission of HIV and offered an HIV test (DFID/SA checklist). In other countries, such as Zimbabwe, there is an empha- sis on group education about HIV and PMTCT for all women presenting for antena- tal care, which is easier to deliver to first-time visitors who can be easily identified on each clinic day (Chandisarewa et al. 2007). One way to determine the extent of gaps in exposure to HIV testing during pregnancy is to look at the relationship between the probability of being offered an HIV test during pregnancy and the number of antenatal visits made by women. We use DHS data from four African countries: Kenya 2008, Uganda 2006, Zambia 2007, and Ghana 2008. These data provide comprehensive socioeconomic and demographic information and also indicate the number of ANC visits during the most recent pregnancy, as well as whether or not a woman was offered an HIV test during any of her ANC visits.13 In our analysis, we are particularly interested in determining whether the probability of being offered an HIV test seems to differ for women who had only one ANC visit compared to those who had two, three, and four ANC visits. We present our results in graphical form in Figure 1 panel A to panel D. Each of the four country graphs presents the results sepa- rately for urban and rural settings. We plot how the conditional probability of being offered an HIV test during any visit varies with the number of self-reported ANC clinic visits. These results are based on a linear probability regression that includes a series of indicator variables for the number of ANC visits as well as a set of observable characteristics such as education, age, marital status, and wealth indicators. The height of the bars represents the cumulative probability of being offered an HIV-test for women during the most recent pregnancy. These actual fre- quencies are contrasted with results from a simulation that applies the probability of being offered an HIV test among one time ANC visitors to all the subsequent visits and calculates the projected cumulated probability of being offered an HIV 13  Unfortunately, the survey did not ask during which specific visits the HIV test was offered. 78 American Economic Journal: applied economicsapril 2013 Panel A. Kenya 2008 DHS Panel B. Ghana 2008 DHS 1 1 0.8 0.8 Fraction Fraction 0.6 0.6 0.4 0.4 0.2 0.2 0 0 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 Urban Rural Urban Rural Sample: All pregnancies in the last 24 months Sample: All pregnancies in the last 24 months Kenya 2008 DHS Ghana 2008 DHS Panel C. Zambia 2007 DHS Panel D. Uganda 2006 DHS 1 1 0.8 0.8 Fraction Fraction 0.6 0.6 0.4 0.4 0.2 0.2 0 0 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 Urban Rural Urban Rural Sample: All pregnancies in the last 24 months Sample: All pregnancies in the last 24 months Zambia 2007 DHS Uganda 2006 DHS Panel E: Authors’ data 1 0.8 Fraction 0.6 0.4 0.2 0 1 2 3 4 5 6 Sample: Most recent pregnancy Authors’ data Actual Simulated Figure 1. Fraction of Women Offered HIV Test versus Number of ANC Visits Notes: Figure 1, panels A-D present the results separately for urban and rural settings based on DHS data from Kenya, Uganda, Zambia and Ghana. The graphs plot the conditional probability of being offered an HIV test during any pregnancy versus the number of self-reported ANC clinic visits. These results are based on a lin- ear probability regression that includes a series of indicator variables for the number of ANC visits as well as a set of observable characteristics such as education, age, marital status, and wealth indicators. The height of the bars represent the cumulative probability of being offered an HIV-test for women during the most recent preg- nancy. These actual frequencies are contrasted with results from a simulation that applies the probability of being offered an HIV test among one time ANC visitors to all the subsequent visits, and calculates the projected cumu- lated probability of being offered an HIV test by the number of ANC visits. Panel E presents similar data from our own survey sample. Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 79 test by the number of ANC visits. The graphical results provide evidence that is consistent with the fact that women are much more likely to be offered an HIV test on the first visit compared to subsequent visits. As an example for rural Kenya (Figure 1, panel A), the actual probability of being offered a test increases from 64 percent for one-time visitors to 80 percent for two-time visitors, and 83 ­percent for three-time visitors. The simulated probabilities for one-, two-, and three-time visitors are 64 percent, 87 percent, and 95 percent, respectively. The results for rural Uganda (Figure 1, panel D) are even more striking. For women who visit the ANC clinic three times during a recent pregnancy (the median number of visits in the sample), the difference between the actual probability of being offered a test (31 percent) and the projected probability (60 percent) is even larger. The results for the remaining two countries presented in Figure 1, panels B and C, as well as for all urban regions, are also broadly consistent. These patterns are also consis- tent with similar graphs presented in Figure 1, panel E, using data from our own survey based on information on actual testing information and the number of ANC visits during the latest pregnancy. The results presented above should be treated with care in light of the fact that, apart from using observable characteristics as controls, we cannot rule out potential ­ umber selection effects that drive both the reporting of being offered a test and the n of ANC visits. Nevertheless, we believe that our results using the DHS data com- bined with the existing evidence on PMTCT coverage are indicative that the find- ings in our paper apply more generally. They suggest that health facilities do not seem to provide adequate testing opportunities to pregnant women beyond the first ANC clinic visit. V.  A Discussion and Conclusion Using a panel dataset of pregnant women who sought antenatal care in a high HIV prevalence region of Kenya, we assess the impact of healthcare provider absence on a number of health outcomes. Our results show that in the study area, health worker absence is one of the important determinants of uptake of HIV ­ testing and counseling services, and that it also influences the probability that pregnant women give birth in a hospital or health center. We test for differential impacts of nurse attendance using pretest beliefs that predict HIV status for those who test. For those women who are more likely to be HIV positive, we find that the presence of the PMTCT nurse increases the probability of receiving PMTCT medications at the time of delivery, decreases the probability of breast-feeding, and increases the prob- ability of enrollment in an HIV treatment program. While our analysis has focused on the reduced-form impacts of PMTCT nurse presence on health outcomes, at least two plausible and possibly overlapping mechanisms could underpin this relationship and merit a brief discussion. First, the presence of the PMTCT nurse is required for being tested for HIV and for the pro- vision of HIV and pregnancy counseling. Learning one’s HIV status and ­ receiving counseling are the main channels for helping women learn about the risks and ben- ­ efits of breast feeding for HIV-positive mothers and the benefits of delivery in a safe setting. Nonetheless, an alternative mechanism may also be at play here. If women 80 American Economic Journal: applied economicsapril 2013 who arrive at the clinic on a day that the nurse is absent lose confidence in the medical system, then they may similarly be less likely to demand downstream health services, independent of their knowledge regarding the potential benefits of those services. While the absence in our setting does not preclude patients from accessing all forms of antenatal care apart from PMTCT counseling and testing during that visit, we nonetheless cannot rule out this discouragement explanation as at least a partial driver of our results. That nurse absence during the initial ANC visit does not appear to affect the number of subsequent ANC visits (columns 4–6 of Table 4) provides at least suggestive evidence that this is not the primary mechanism through which these absence effects operate. Given the pervasiveness of health worker absence across the developing world, it is instructive to translate these impacts into an estimate of the num- ber of new HIV cases averted (see Appendix A2 for details on calculations). The lone PMTCT nurse in our setting is absent 9 percent of the time and this absence results in a 58 percentage point reduction in the likelihood that patients test at any point during their pregnancy. Combining this with data on patient flow at the antenatal clinic and the effectiveness of medications in reducing mother-to- child transmission yields the result that PMTCT nurse absence contributes to an additional 3.7 mother-­­to-child infections per 10,000 live births. If we apply these estimates to the 35 percent absence rate documented in some other developing country settings (Chaudhury et al. 2006) and assume a similar population and quality of health facility, then nurse absence contributes to about 14.6 additional infections per 10,000 live births. This number appears staggeringly large when compared to the seemingly small expenditure that would be required to provide substitute nurse coverage in the clinic. In addition, improvements in the referral system such as the deployment of well designed electronic medical records sys- tems could mitigate the effects of absence in this setting (Siika et al. 2005). Of course, implementing effective and long lasting reductions in absence or interven- tions meant to reduce the effects of absence may be hard when the system is not conducive to change (Banerjee, Duflo, and Glennerster 2008). National and global policy makers need to take the costs and benefits associated with these effects into account when deciding on priority investments for health. Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 81 Appendix A1. Do Reported Priors about Being HIV Positive Predict HIV Status? Table A1—Subjective Beliefs before HIV Test and Actual Test Results Dependent variable indicator tested positive (1) (2) (3) (4) (5) (6) (7) (8) Chance of having HIV- great 0.272 0.247 0.208 0.263 (0.097)*** (0.098)** (0.099)** (0.097)*** Chance of having 0.171 0.156 0.111 0.166   HIV- moderate (0.081)** (0.082)* (0.079) (0.082)** Chance of having HIV- small 0.077 0.060 0.021 0.069 (0.059) (0.060) (0.061) (0.060) Chance of having HIV- great 0.126 0.125 0.120 0.129   or moderate (0.043)*** (0.043)*** (0.043)*** (0.043)*** Day of week = Tuesday 0.004 −0.001 0.004 −0.000 (0.057) (0.056) (0.057) (0.056) Day of week = Wednesday −0.005 0.004 −0.002 0.005 (0.058) (0.059) (0.058) (0.059) Day of week = Thursday −0.019 −0.041 −0.011 −0.038 (0.052) (0.050) (0.053) (0.050) Day of week = Friday 0.083 0.073 0.089 0.077 (0.072) (0.071) (0.072) (0.072) Day of the month 0.006 0.001 0.005 0.001 (0.009) (0.009) (0.009) (0.009) Day of the month squared −0.000 −0.000 −0.000 −0.000 (0.000) (0.000) (0.000) (0.000) Quarter of conception == 2 0.006 0.018 0.004 0.017 (0.047) (0.048) (0.047) (0.048) Quarter of conception == 3 0.069 0.089 0.073 0.092 (0.057) (0.060) (0.057) (0.060) Quarter of conception == 4 −0.120 −0.115 −0.126 −0.118 (0.054)** (0.054)** (0.052)** (0.052)** Age in years 0.087 0.087 (0.027)*** (0.026)*** Age in years, squared −0.001 −0.001 (0.000)*** (0.000)*** Completed primary school −0.049 −0.045 (0.041) (0.041) Married −0.073 −0.079 (0.061) (0.061) Frequency of church attendance 0.007 0.007   in past four weeks (0.008) (0.008) Boils drinking water 0.007 0.005 (0.045) (0.046) Number of livestock held at −0.005 −0.005  enrollment (0.005) (0.006) Lives in a non-thatched house −0.007 −0.009 (0.043) (0.043) Lives in Maseno division 0.009 0.009 (0.044) (0.044) Log distance from clinic −0.011 −0.011 (0.031) (0.031) Observations 453 452 446 446 453 452 446 446 F-stat: test no effect of priors 12.55 11.41 9.98 12.29   on actual status prob > ​ χ​  2​ 0.01 0.01 0.02 0.01 Notes: The variables are defined in Table 1. Table reports marginal probit estimates. Tested positive takes value 1 if the subject was tested positive during the pregnancy and 0 if HIV negative. Standard errors in brackets clustered at the visit date level. *** Significant at the 1 percent level.  ** Significant at the 5 percent level.   * Significant at the 10 percent level. 82 American Economic Journal: applied economicsapril 2013 A2. Number of New Cases of HIV Resulting from Absence and Poor Clinic Protocols Below we provide a more detailed explanation for the imputation of the number of HIV infections that could be averted by the elimination of nurse absences. First, we provide an estimate of the prevalence rate of eventual nontesters whose first ANC visit happened on a day when the nurse is absent. Second, we combine these estimates with information from the medical literature on the relationship between PMTCT medication and reductions in HIV transmission at birth. Third, we calculate the impact of absence on the number of transmissions in a given year for the absence level at our clinic, as well as for typical absence rates in the health sector in develop- ing countries more generally. Based on a number of plausible assumptions, we generate five distinct estimates of the prevalence rate of pregnant women who did not test due to nurse absence on the first ANC visit:   • We assume that the prevalence rate of nontesters is equal to the prevalence rate of testers (19.7 percent).   • We assume that the prevalence rate of nontesters is equal to the adult preva- lence rate in the 2003 Kenyan DHS for the Nyanza region (18.3 percent).   • We assume that the prevalence rate of women who turn up for their first ANC visit on days when the nurse is absent (group 1) is the same as on days when she is present (group 2). Among eventual testers for these two groups, the prevalence rate is 19.9 percent (group 2) and 15.4 percent (group 1). The testing rates for these groups are 82.5 percent (group 2) and 24.1 percent (group 1). The resulting prevalence rate for nontesters who would have tested if the nurse was present is 21.8 percent.   • We use the background characteristics of the women who test to predict in a regression framework the prevalence of all nontesters (20.7 percent).   • We use the background characteristics of the women who test to predict the prevalence of all nontesters whose first visit is on a day when the nurse is absent (19.1 percent). Across each of the five different assumptions, the calculated prevalence rate for the group of interest is roughly 20 percent and varies between 18.3 percent and 21.8 percent. Next, we turn to estimates of the efficacy of PMTCT interventions. Using the estimates reported in UNAIDS (2005), rates of mother-to-child transmis- sion and the impact of different PMTCT regimens are as follows:   • Default mother to child transmission rate without any intervention: 32 percent   • No intervention, long breast-feeding (18–24 months): 35 percent Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 83   • No intervention, short breast-feeding (6 months): 30 percent   • No intervention, replacement feeding: 20 percent   • Single-dose NVP14 (mothers and infants), combined with short (6 months) breast-feeding (6 months): 16 percent   • Single-dose NVP (mothers and infants), combined with replacement feed- ing: 11 percent   • AZT15 long (from 28 weeks) and single-dose NVP (mothers and infants), combined with short breast-feeding (6 months): 10 percent   • AZT long (from 28 weeks) and single-dose NVP (mothers and infants), combined with replacement feeding: 2 percent According to the treatment regimen in place at the time of the survey, the most common treatment was AZT long with single-dose NVP combined with short breast-feeding, which has an estimated transmission rate of 10 percent. Therefore, the treatment with PMTCT in our setting reduces the transmission rate at birth among HIV-positive women by approximately 22 percentage points (32 percent to 10 percent). On a typical day, a PMTCT nurse conducts testing and counseling to an average of 4.1 pregnant women. When she is absent, about 58 percent of first time ANC visitors do not test during the pregnancy. Since the prevalence rate is estimated to be around 20 percent for this group and testing increases the chance of receiving medication to prevent MTCT for those who are positive by 18 percentage points, this means that a one day absence results in roughly 0.09 (= 4.1 × 0.58 × 0.2 × 0.18) positive women who do not receive PMTCT. This translates into an increase in the HIV transmission from the mother to the child of 0.019 (0.09 × 0.22) cases. If we apply this estimate to the typical absence rate in our clinic (9 percent), then nurse absence contributes to an additional 0.42 infections per year (assuming 250 working days in a year). If we apply these estimates to the much larger absence rates found in the literature (35 percent), then nurse absence contributes to about 1.65 infections per year per nurse. Taking into account the fraction of women that visit ANC clinics (88 percent) and neonatal mortality (33 per 1,000 live births), these numbers trans- late into 0.37 infections per 1,000 live births (9 percent absence) and 1.46 infections per 1,000 live births (35 percent absence rates). 14  NVP- nevirapine 15  AZT-azidothymidine 84 American Economic Journal: applied economicsapril 2013 References Banerjee, Abhijit, Angus Deaton, and Esther Duflo. 2004. “Wealth, Health, and Health Services in Rural Rajasthan.” American Economic Review 94 (2): 326–30. Banerjee, Abhijit, and Esther Duflo. 2006. “Addressing Absence.” Journal of Economic Perspectives 20 (1): 117–32. Banerjee, Abhijit V., Esther Duflo, and Rachel Glennerster. 2008. “Putting a Band-Aid on a Corpse: Incentives for Nurses in the Indian Public Health System.” Journal of European Economic Asso- ciation 6 (2/3): 487–500. Besley, Timothy. 1995. “Nonmarket Institutions for Credit and Risk Sharing in Low-Income Coun- tries.” Journal of Economic Perspectives 9 (3): 115–27. Bjorkman, Martina, and Jakob Svensson. 2009. “Power to the People: Evidence from a Randomized Field Experiment on Community-Based Monitoring in Uganda.” Quarterly Journal of Economics 124 (2): 735–69. Central Bureau of Statistics, Ministry of Health, Kenya Medical Research Institute, National Council for Population and Development, ORC Macro, and Centers for Disease Control and Prevention. 2004. Kenya Demographic and Health Survey 2003. Central Bureau of Statistics (CBS). Calver- ton, MD, July. Chandisarewa, W., L. Stranix-Chibanda, E. Chirapa, A. Miller, M. Simoyi, A. Mahomva, Y. Maldo- nado, and AK Shetty. 2007. “Routine offer of antenatal HIV testing (“opt-out” approach) to prevent mother-to-child transmission of HIV in urban Zimbabwe.” Bulletin of the World Health Organiza- tion 85 (11): 843–50. Chaudhury, Nazmul, Jeffrey Hammer, Michael Kremer, Karthik Muralidharan, and F. Halsey Rog- ers. 2006. “Missing in Action: Teacher and Health Worker Absence in Developing Countries.” Jour- nal of Economic Perspectives 20 (1): 91–116. Das, Jishnu, Stefan Dercon, James Habyarimana, and Pramila Krishnan. 2007. “Teacher Shocks and Student Learning: Evidence from Zambia.” Journal of Human Resources 42 (4): 820–62. Das, Jishnu, Jeffery Hammer, and Kenneth Leonard. 2008. “The Quality of Medical Advice in Low- Income Countries.” Journal of Economic Perspectives 22 (2): 93–114. Duflo, Esther, Rema Hanna, and Stephen Ryan. 2008. “Monitoring Works: Getting Teachers to Come to School.” Center for Economic and Policy Research (CEPR) Discussion Paper 6682. Goldstein, Markus, Joshua Graff Zvin, James Habyarimana, Cristian Pop-Eleches, and Harsha Thirumurthy. 2013. “The Effect of Absenteeism and Clinic Protocol on Health Outcomes: The Case of Mother-to-Child Transmission of HIV in Kenya: Dataset.” American Economic Journal Applied Economics. http://dxdoi.org/10.1257/app.5.2.58. Grout, JR. 2003. “Preventing medical errors by designing benign failures.” Joint Commission Journal on Quality and Patient Safety 29 (7): 354–62. Guay, Laura A., Philippa Musoke, Thomas Fleming, Danstan Bagenda, Melissa Allen, Clemensia Nakabiito, Joseph Sherman et al. 1999. “Intrapartum and neonatal single-dose nevirapine com- ­ pared with zidovudine for prevention of mother-to-child transmission of HIV-1 in Kampala, Uganda: HIVNET 012 randomised trial.” Lancet 354 (9181): 795–802. Kohn, Linda T., Janet M. Corrigan, and Molla S. Donaldson, eds. 1999. To Err is Human: Building A Safer Health System. Washington, DC: National Academies Press. Kremer, Michael, Edward Miguel, and Rebecca Thornton. 2009. “Incentives to Learn.” Review of Eco- nomics and Statistics 91 (3): 437–56. Kumar, S., and M. Steinebach. 2008. “Eliminating US hospital medical errors.” International Journal of Health Care Quality Assurance 21 (5): 444–71. National AIDS, and STI Control Program. 2009. “Kenya AIDS Indicator Survey (KAIS) 2007 Data Sheet.” Nairobi: National AIDS and STI Control Program. http://www.prb.org/pdf09/kaiskenya- datasheet.pdf (accessed March 15, 2012). Siika, Abraham M., Joseph K. Rotich, Chrispinus J. Simiyu, Erica M. Kigotho, Faye E. Smith, John E. Sidle, Kara Wools-Kaloustian, et al. 2005. “An electronic medical record system for ambula- tory care of HIV-infected patients in Kenya.” International Journal of Medical Informatics 74 (5): 345–55. Stringer, Elizabeth M., Didier K. Ekouevi, David Coetzee, Pius M. Tin, Tracy L. Creek, Kathryn Stinson, Mark J. Giganti, et al. 2010. “Coverage of Nevirapine-Based Services to Prevent Mother- ­ to-Child HIV Transmission in 4 African Countries.” Journal of the American Medical Association 304 (3): 293–302. Vol. 5 No. 2 goldstein et. al : Supply-side factors and health outcomes 85 Thompson, Melanie A., Judith A. Aberg, Pedro Cahn, Julio S.G. Montaner, Guilano Rizzardini, Amalio Telenti, José M. Gatell, et al. 2010. “Antiretroviral Treatment of Adult HIV Infection: ­ ­ ecommendations of the International AIDS Society—USA Panel.” Journal of the American 2010 R ­Medical Association 304 (3): 321–33. Townsend, Robert M. 1995. “Consumption Insurance: An Evaluation of Risk-Bearing Systems in Low-Income Economies.” Journal of Economic Perspectives 9 (3): 83–102. UNAIDS. 2005. “Rates of Mother-to-Child Transmission and the Impact of Different PMTCT Regi- mens: Report of a Consultation Organized by the UNAIDS Reference Group for Estimates, Mod- eling, and Projections.” http://www.epidem.org/Publications/PMTCT%20report.pdf (accessed February 10, 2013). Wachter, Robert M. 2004. “The End Of The Beginning: Patient Safety Five Years After ‘To Err Is Human.’” Health Affairs 23 (Supplement Web Exclusives): W4-534–45. World Health Organization. 2009. Towards Universal Access. Scaling up priority HIV/AIDS interven- tions in the health sector. Progress Report 2009. Geneva: World Health Organization. http://www. who.int/hiv/pub/tuapr_2009_en.pdf (accessed December 24, 2011).