Love in the Time of HIV: Testing as a Signal of Risk Laura Derksen∗

Joep van Oosterhout∗∗

May 9, 2017

Abstract Ending the AIDS epidemic in Africa is within reach due to lifesaving treatment which also prevents transmission. However, demand for HIV testing is low. We model HIV testing as a signal of infection. Test seekers are subject to discrimination from sexual partners. We test the theory by providing new information, randomized at the community level: antiretroviral drugs prevent HIV transmission, so a sexual partner who is diagnosed is relatively safe. This reduces discrimination and increases testing, with the strongest effects in communities where the information becomes common knowledge. These results suggest that HIV discrimination is partly explained by misinformation and can be mitigated. ∗ [email protected],

647-787-9333, 105 St. George St., Toronto, Ontario, M5S 3E6 Canada. ∗∗ Dignitas International, Malawi College of Medicine Laura Derksen is grateful to Oriana Bandiera and Greg Fischer for their guidance. This paper benefited from discussions with Marcella Alsan, Nava Ashraf, Victoria Baranov, Tim Besley, Patrick Blanchenay, Gharad Bryan, Jon De Quidt, Erika Deserranno, Kate Dovel, Thiemo Fetzer, Maitreesh Ghatak, Markus Goldstein, Willa Friedman, Marcos Vera Hernandez, Jason Kerwin, Alan Manning, Sam Marden, Steve Pischke, Munir Squires, Balazs Szentes, Susan Watkins, and Kelly Zhang, and seminar and conference participants at CSAE 2016, Dartmouth, Duke, IAS 2015, LSE, Mathematica, Melbourne, Ottawa, Oxford, PacDev 2016, PAA 2016, RES 2016, Sussex, Toronto, UBC, and UWO. We gratefully acknowledge assistance from Alfred Matengeni, Dignitas International and Invest in Knowledge Initiative, and financial support from STICERD and the Russell Sage Foundation. Errors remain those of the authors.

1

1

Introduction

One million Africans die of AIDS every year, despite the availability of free, effective treatment. Antiretroviral therapy (ART) prolongs life by decades and reverses the symptoms of AIDS. ART drugs also reduce HIV transmission by 96 percent (Cohen et al., 2011). This recent discovery has prompted a “Treatment for Prevention” strategy to bring an end to the AIDS epidemic. While the supply of free medication has increased dramatically, demand for HIV testing and treatment remain low. Low testing rates may be explained by discrimination towards people who seek HIV-related care. The underlying causes of such discrimination are not well understood. Many policy responses help clients cope with discrimination. For example, door-to-door testing and selftesting offer privacy. Monetary incentives obscure a client’s motivation for testing (Thornton, 2008). By focusing on the root cause, it might be possible to reduce discrimination, rather than simply helping people to cope. In this paper, we use an information experiment to lower one barrier to HIV testing. In particular, we investigate statistical discrimination by potential sexual partners. Those who seek an HIV test are more likely to be infected, and may face rejection from potential partners who fear contracting the virus. In fact, an HIV test should not be viewed as a bad signal, since a person who is diagnosed and treated for HIV is less contagious. Providing precise information, at the community level, on this public benefit of ART increases the HIV testing rate. We find a reduction in discrimination between sexual partners to be the likely mechanism. Our results suggest that discrimination is a barrier to HIV testing, and that providing accurate risk information can mitigate discrimination and its effects. We model HIV testing as a signalling game between potential sexual partners. Testing decisions are observed with some probability, which depends on the choice of testing location. A person with HIV has more to gain from testing, because it allows them to access ART drugs. A partially separating equilibrium emerges: some individuals reject potential 2

partners who’ve been tested. This is a form of statistical discrimination, based on a fear of HIV. Beliefs about the public benefit of ART will affect this particular form of discrimination and leave other types of stigma unaffected. That is, people are more likely to consent to matches who have been tested and treated if they know that treatment prevents HIV transmission. This reduction in discrimination should, in turn, increase HIV testing. We map the model to an information experiment, and test several predictions empirically. We randomly assigned 122 villages in Malawi to either partial or full intervention arms. In every village we organized an information campaign. We disseminated information about ART in public at community health meetings. In full intervention villages, we provided information on both the private and public benefits of ART. In particular, we informed attendees that ART reduces HIV transmission by 96 percent. In the partial intervention we provided information on only the private benefits and availability of ART. In general, the formation of beliefs depends on endogenous factors. A randomized experiment allows us to measure the causal effect of beliefs. We also view the information campaign as a policy intervention to increase HIV testing and treatment. Without an information campaign, incorrect beliefs might persist in equilibrium. First, ART is underadopted, which means there is limited scope for first-hand learning. Second, information sharing within networks may be strategic and lack credibility. For example, a potential sexual partner who claims that ART blocks HIV transmission may have ulterior motives. Finally, recent abstinence-focused public health messages have exaggerated the risk of HIV transmission. Providing new information on the public benefit of ART shifted beliefs and caused a decrease in reported discrimination between sexual partners. It also caused a 34 percent increase in HIV testing, as recorded over four months of administrative data. To our knowledge, this is the first evidence for an information intervention to increase voluntary HIV testing at clinics. Other empirical results are consistent with the theoretical predictions of the signalling model. Overall, one third of HIV tests take place far away from the HIV testing client’s home village. 3

A large shift in beliefs about the public benefit of ART is associated with an increase in the number of tests sought nearby. Common knowledge is important: perceived community beliefs predict HIV testing, while private beliefs do not. If altruism were driving the results, one would expect the opposite. Many public health messages exaggerate the risk of transmission. Policy makers are often reluctant to provide accurate information for fear of risk compensation. Our study highlights a negative consequence of such policies: fear based messages may increase discrimination, discourage testing, and thus increase the spread of HIV. To our knowledge, this paper provides the first evidence for an information intervention to increase voluntary HIV testing. Misinformation may explain the low adoption rates of many health measures. New, precise information on health risks impacts some types of behavior1 . In the context of HIV, most papers focus on risk taking2 (Bandiera et al., 2012, Dupas, 2011, Kerwin, 2014). The link between information and health seeking behavior is not well understood. Information on the protective benefits of circumcision appears to have no effect on demand for the procedure (Chinkhumba et al., 2012, Godlonton et al., 2011). The choice to seek medical treatment is often a signal of illness or infection. We show that providing information on the public benefit of medical treatment reduces discrimination towards those who seek care. The literature indicates that HIV testing responds to monetary incentives, and that this involves a social component. Thornton (2008) argues that such incentives conceal a person’s motivation for testing. Godlonton and Thornton (2012) and Ngatia (2011) investigate heterogeneous peer effects of offering incentives. These studies have a policy implication. However, using incentives to encourage ART takeup may prove infeasible. To obscure the signal, the entire community would have to visit the clinic at regular intervals. An intervention that reduces 1 See,

for example Bandiera et al. (2012), Dupas (2011), Jalan and Somanathan (2008), Kerwin (2014), and Madajewicz et al. (2007). 2 Interventions which provide new, precise information differ sharply from traditional anti-HIV messaging strategies which focus on behavior change directly. Duflo et al. (2015), for example, find that a behavior change program that pushes abstinence fails to reduce risk taking among adolescent girls in Kenya.

4

discrimination is attractive for several reasons. Such an intervention should increase both HIV testing and treatment, and generate persistent effects for a fixed cost. Discrimination between potential sexual partners imposes a high cost on those concerned about sexual prospects. This discourages HIV testing specifically among those with a high risk of infection. Therefore, a policy designed to reduce this type of discrimination may be particularly effective. This paper shows that a technology may be under-adopted if it benefits a group that is subject to statistical discrimination. We also show that new information can diminish discrimination. Arrow (1973)’s seminal work on statistical discrimination explains labor market discrimination by differences in average productivity. The effects of discrimination have been investigated in various contexts3 . Much of the economic literature on social stigma, both in the context of HIV and more generally, views stigma as an exogenous cost4 , and investigates the effects of such a cost. Benabou and Tirole (2011) and Besley and Coate (1992) investigate the theoretical basis and implications of stigma. In this paper we examine both the cause and effect of stigma, theoretically and empirically. We microfound stigma as statistical discrimination against HIV-infected sexual partners, and use an experiment to test the implications. The paper proceeds as follows. Section 2 describes the experimental context and design. Section 3 presents the model and its predictions. Section 4 provides a description of the data and results. Section 5 concludes.

2

Background and Experimental Design

Antiretroviral therapy, or ART, is a combination of drugs that suppress the HIV virus and reverse the progression of AIDS. ART is now widely available for free in Africa, and mortality from AIDS has decreased by over 30 percent in the past decade (WHO, 2013). 3 See

for example, Altonji and Pierret (2001), and Coate and Loury (1993). the context of HIV, see Ngatia (2011) and Thornton (2008), and in the context of welfare payments, and Moffitt (1983). 4 In

5

A large randomized controlled trial5 demonstrated that regular ART use by HIV positive individuals reduces HIV transmission by 96 percent (Cohen et al., 2011). Universal testing and treatment could end the AIDS epidemic within 50 years, and is cost effective (Granich et al., 2009). However, in sub-Saharan Africa, many have never been tested for HIV, and only one third of those infected are taking ART (WHO, 2013). HIV is susceptible to many forms of discrimination and stigma6 . Because HIV is sexually transmitted, discrimination from potential sexual partners may be common. However, other types of discrimination exist; Hoffmann et al. (2014) find that some are willing to forgo payment to avoid coming into contact with an object handled by an HIV positive person. Discrimination might be statistical, taste-based, or a social equilibrium (Peski and Szentes, 2013, Benabou and Tirole, 2011). Our study took place in Zomba District, Malawi. HIV prevalence is approximately 14.5 percent (DHS, 2010). Both HIV testing and antiretroviral therapy (ART) are free7 . There are many accessible clinics. To obtain ART drugs, a person must first seek an HIV test. While the results of the test are private8 , the decision to test may be observable. clinics are public spaces, with a separate room dedicated to HIV testing. The availability of ART induces adverse selection among voluntary test seekers. In Malawi, HIV prevalence is 44 percent higher among those who report having been tested (DHS, 2010). This may generate statistical discrimination towards voluntary test seekers, which in turn 5 ART

drugs reduced HIV transmission by 96 percent, measured over a period of four years, among serodiscordant couples who took the drug at home. This reduction in transmission does not hinge on perfect adherence to the medication. It is accurate to say that ART drugs reduce HIV transmission by at least 96 percent relative to the absolute transmission rate over any time period up to ten years. The absolute transmission rate is low, so the the total reduction in transmission over repeated interactions is approximately linear. 6 See Mahajan et al. (2008) for a review of medical and sociological research on the stigmatization of HIV and AIDS. 7 At the time of the study, only those with symptoms or a low CD4 count (a measure of immune health) qualified for ART. However, most delay HIV testing until they are experiencing severe symptoms. In our data, 89 percent of patients who wish to initiate ART qualify. Many countries, including Malawi, have since adopted a policy of universal testing and treatment. 8 In Malawi, HIV tests are confidential, and a person who tests HIV negative does not receive any written proof of the result. The test is conducted twice with different test kits, so an incorrect diagnosis is unlikely.

6

discourages HIV testing. Our data shows an annual voluntary testing rate of only 9 percent. One third of test seekers travel further than necessary, possibly to avoid being seen. Discrimination from potential sexual partners may be costly, even for those in committed relationships. Concurrent sexual relationships are common in Malawi and predict HIV infection (Helleringer et al., 2009). Sexual partners are often a source of financial support. Cash transfers to young women in Malawi reduce both sexual behavior (Baird et al., 2010) and HIV (Baird et al., 2012). The Malawian Journals Project9 contain extensive qualitative data on HIV testing and the dating market. Many report seeking HIV testing far from home to avoid seeing potential sexual partners. The private benefits of ART are well understood, but the public benefits are not (Angotti et al., 2016) . Our study sample includes 122 villages in Zomba District. Villages are randomly assigned to either partial or full intervention arms. We stratified on population and nearest clinic. The sample appears balanced on observable village characteristics (see Table 1). Table 1: Balance on Village-Level Covariates (1) (2) Partial Full intervention intervention Km to clinic 4.449 5.066 Pre-intervention (2.5 months) % tested 1.315 0.885 % tested jointly 0.345 0.163 % tested nearby 0.817 0.454 % taking ART 2.546 2.080 Stratification variables Village population 378 402 Nearest clinic Observations 62 60

(3) p-value 0.147 0.126 0.103 0.056 0.177 0.574

p-values are for a regression of the covariate on village intervention status. HIV testing rates among target population: age 15-49, non-pregnant. A clinic is nearby if it no more than three kilometers further than the nearest clinic. 9 http://malawi.pop.upenn.edu/malawi-data-qualitative-journals

7

We conducted community health education meetings in both partial and full interventions. In this way, we isolate the effect of one fact: ART drugs provide a public benefit by preventing HIV transmission. In the partial intervention we provided information only on the private benefits of ART drugs. In the full intervention we provided information on both the private and public benefits of ART. In partial intervention villages, community educators started the meeting by asking for a show of hands: 75 percent of attendees believed that ART allowed HIV positive individuals to lead a long and healthy life. Educators then explained that ART increases life expectancy, hides the symptoms of AIDS, and is free at local clinics. Educators explained using an infographic that depicts a reduction in viral load (Figure 1). In full intervention villages, community educators provided the same basic information. Next, they asked whether participants believed that ART had an effect on HIV transmission. Only 5 percent of meeting attendees raised their hands. Educators explained that ART reduces the probability of HIV transmission by 96 percent10 . They used an infographic to explain that ART reduces a person’s viral load, which leads to a reduction in transmission risk (Figure 1). Educators emphasized the importance of correct adherence to ART.

(a) Partial Intervention

(b) Full Intervention

Figure 1: Infographics for Information Intervention Note: The infographics depict viral load with and without ART. The size of the arrow represents HIV transmission risk. Equivalent infographics with an HIV-positive woman were also used at each meeting.

A map of study villages and clinics, and detailed script of the information provided are available in the appendix. The meetings were 10 This

fact is based on ART as typically used and does not depend on perfect adherence (Cohen et al., 2011).

8

balanced in length (approximately 45 minutes) and community participation. We conducted the the intervention over three weeks. In order to avoid information spillovers, educators conducted the partial intervention before training for and implementing the full intervention. An average of 67 percent of the target population attended the meeting in each village. Meeting attendance was insignificantly higher in villages assigned to the partial intervention. The topic of the meeting was not announced ahead of time. Village chiefs advertised the meetings and took attendance, and in exchange received small personal gifts of soap and salt. We did not provide incentives for attendance. The educators knew they would receive incentives based on knowledge retention in both partial and full intervention villages11 .

3

Model of HIV Testing as a Signal

In this section we present a signalling model of the HIV testing decision, and illustrate a clean empirical test of the theory. This empirical test motivates the information experiment described in Section 2. Our theory is related to the sociological concept of stigma: a phenomenon whereby those suspected of having HIV are excluded from social interactions12 . This paper focuses on one potential source of stigma, and one type of social interaction. Those who are likely to be HIV positive are excluded from sexual interactions due to a rational fear of contagion. If HIV testing is a signal of underlying risk, then potential sexual partners discriminate against those who seek an HIV test. Other 11 The

maximum incentive was approximately 100 USD. These incentives were paid to community educators five months after the intervention in May, 2014, and the size of the incentive was based on the percent of community meeting attendees in full intervention villages who believed that ART prevents HIV transmisson, and the percent of community meeting attendees in partial intervention villages who believed that ART had private benefits, as recorded in the survey. The survey was conducted by a different set of enumerators who did not know the community educators, and were not aware of the incentive scheme 12 This is derived from the seminal definition by Goffman (1963), who defines stigma as “The phenomenon whereby an individual with an attribute which is deeply discredited by his/her society is rejected as a result of the attribute.”

9

types of stigma, and other barriers to HIV testing13 may exist, and are included in the model in reduced form14 . Comparative statics show that the information experiment reduces stigma if and only if it is due to a fear of HIV transmission. Other forms of stigma and costs of testing remain unaffected. We model the HIV testing decision as a one-sided signalling game between two players, over two periods. In the first period, the first player decides whether to seek an HIV test. He takes into account his private type, which corresponds to his risk of being HIV positive. In the second period, he is matched at random to a potential sexual partner, who might observe whether he has been tested for HIV. Based on this signal, the second player decides whether to consent to a sexual relationship. Consider a continuum of individuals A = [0, 1], and another continuum B = [0, 1] which represents potential sexual partners for A. A fraction h¯ of individuals a ∈ A are HIV positive. Every b ∈ B is HIV negative15 . We denote HIV status by h a ∈ {0, 1}. Individuals have private information about their own HIV risk, based on past behavior and symptoms. In the first period, each a ∈ A observes his private type θ a . This represents his probability of infection, which takes one of two values: θ a ∈ {θ L , θ H }

(1)

with 0 < θ L < θ H < 1. We refer to these two types as low-risk and high-risk types. a decides whether to seek an HIV test, and whether to test at a far f f away clinic t a ∈ {0, 1} or nearby tna ∈ {0, 1}, such that tna + t a ∈ {0, 1}. Testing benefits those who are infected, as it provides immediate access 13 HIV

testing has increased significantly over the past decade, in part due to increased supply of HIV testing services and ART. We do not wish to suggest that stigma is the only potential barrier to HIV testing. 14 For ease of exposition, we include only stigma from potential sexual partners. Allowing for stigma from the general population does not produce a substantially different model or predictions. 15 We shall refer to a as “he” and b as “she” for clarity, but the model is not meant to be gender specific. We use a one-sided signalling model for clarity of exposition; a two-sided model gives the same predictions.

10

to ART drugs16 . The total payoff to HIV testing is v if the test result is positive, and 0 otherwise. The parameter v includes both the benefit of ART drugs, and other benefits such as counselling, social programs, and easier access to other medication17 . The total cost of traveling to a nearby clinic is cn , while traveling to a far away clinic costs c f . These costs include the opportunity cost of time and psychic costs. We assume that for low-risk types, the cost is higher than the expected benefit of medication, as the test result is unlikely to be positive. Meanwhile, for high-risk types, the expected benefit is higher than the cost. θ L v < cn < θ H v (2) We also assume that the cost of traveling to a far away clinic is bounded in the following way. c f − cn < θ H v − cn 1−φ

(3)

If this does not hold, then individuals will never test far away. Individuals a ∈ A receive the following payoff at the end of the first period. f f u1a (tna , t a ) = tna (h a v − cn ) + t a (h a v − c f ) (4) In the second period, each a ∈ A is matched at random with a potential sexual partner b ∈ B. While b is aware of the population HIV ¯ she does not observe a’s risk type. prevalence h, b may observe a’s testing decision. His testing decision generates f a signal σa ∈ {0, 1} with P(σa = 1) = tna + φt a . If a tested nearby, he is observed with probability one, while if he tested far away, he is observed with probability φ ∈ (0, 1). b uses this signal to form beliefs about the probability that her partner a is HIV positive; we denote these beliefs by θˆb (σa ). 16 We

assume that every person who tests positive starts ART drugs. This is equivalent to an assumption that the substantial benefits of ART drugs are known; we provide this information in the partial intervention described in Section 2. 17 Many HIV positive patients require treatment for tuberculosis or sexually transmitted infections, as well as antibacterial medication which treats HIV-related infections.

11

b then decides whether to consent to a sexual relationship with her match, mb ∈ {0, 1}. If b consents, a and b obtain relationship benefits y a ∼ G and yb ∼ G respectively18 , where the distribution G has positive support. Consenting to the relationship comes with a cost to b. She risks becoming infected with HIV. The probability of contracting HIV from an infected partner depends on whether he has been diagnosed and treated. The relative reduction in HIV transmission risk is denoted by ρ, and the true value is approximately 0.96 (Cohen et al., 2011). τ is the absolute transmission rate without ART. The cost of contracting HIV is ch . We also allow for other forms of stigma. A relationship with an HIVpositive person might generate disutility for reasons contagion. This could be, for example, because of a propensity for infidelity, because of social norms, or because of a taste parameter. We capture these costs in reduced form, denoted by co . If b chooses to consent, the second-period payoff to a is the benefit of the relationship. u2a (mb ) = mb y a (5) The second-period payoff to b depends on both the benefit of the relationship and the risk of contracting HIV. h    i f u2b (mb ) = mb yb − h a ch τ 1 − ρ(tna + t a ) + co

(6)

An HIV-positive individual who has been tested is, on average, less f contagious, as captured by the term (1 − ρ(tna + t a )). We now make two additional assumptions. First, P(yb > h¯ (ch τ + co )) = 1;

(7)

the net payoff from a sexual relationship with the average, untreated

18 We

have assumed that a always obtains net benefit from the relationship. We do not model his consent decision.

12

member of the population is positive. Second, P(yb < ch τ + co ) = 1;

(8)

the net payoff from a sexual relationship with an untreated, HIV-positive match is negative. Individuals in B have heterogeneous beliefs about the extent to which ART drugs prevent the spread of HIV. Each b ∈ B has belief ρˆ b ∈ [0, 1] about the relative reduction in HIV transmission risk associated with ART19 . ρˆ = 0 corresponds to the belief that ART has no effect on transmission. ρˆ = 1 to the belief that it is impossible to contract HIV from a person taking ART. We denote by Fρˆ the distribution of beliefs in the population B . This links the model to our experimental design: our information intervention shifts20 the distribution Fρˆ . We solve for the pure-strategy Perfect Bayesian Equilibria of this game. In the first period, individuals in A choose an HIV-testing strategy {tn , t f } to maximize their total expected utility21 . E(u1a + u2a ),

(9)

They take the strategies of {b ∈ B}, as well as beliefs Fρˆ and {θˆb (σ ) : b ∈ B} as given. In the second period, each b ∈ B receives a signal σa about a’s testing decision. She forms beliefs about a’s type θˆb (σa ), which are consistent on the equilibrium path. She uses Bayes’ rule to calculate the probability that her match is infected, conditional on the signal. She then chooses whether to consent, maximizing her expected utility based on beliefs ρˆ b and θˆb . 19 For

simplicity, we assume that testing leads to immediate treatment, and the true value of rho is 0.96. Alternatively, we could define ρ as the average reduction in transmission associated with testing, taking into account the fact that not everyone will initiate treatment. The true value is then positive, but less than 0.96. The predictions of the model do not change. 20 The empirical distribution of beliefs, as well as the effect of the intervention are illustrated in Figure 4. 21 For simplicity and without loss of generality, we assume a discount rate of 1.

13

Proposition 1. There are two classes of pure-strategy Perfect Bayesian Equilibria in this game. The first is a partially separating equilibrium. A fraction S of individuals in B discriminate: they reject a match who has been observed seeking an HIV test. The remaining fraction 1 − S of individuals consent to any match. The second is a pooling equilibrium. A fraction P of individuals in B consent if and only if their match has been observed seeking an HIV test. We restrict attention to the partially separating equilibrium, in which discrimination occurs. The pooling equilibrium implies universal testing; our data reveals low levels of HIV testing. The pooling equilibrium is also unstable, in the sense that it does not satisfy the D1 criterion22 . We defer discussion of the pooling equilibrium, and a complete proof of Proposition 1 to the appendix. Let us characterize the partially separating equilibrium, which generates discrimination. In this equilibrium, every b ∈ B consents to a match who has not been observed testing. A proportion S of {b ∈ B} reject those who have been observed seeking an HIV test. Given this strategy by those in B , the decision to test for HIV is based on the expected impact on utility, which is equal to θ a v − cn − Sy a

(10)

if the test is sought at a nearby clinic, and θ a v − c f − φSy a

(11)

if the test is sought at a far away clinic. In both cases, there is an indirect cost of testing. a might face rejection from his sexual partner in the next period. This cost is higher if a chooses to test nearby. By assumption (2), low-risk types will never seek an HIV test in this

22 Intuitively,

an equilibrium satisfies the D1 criterion if off-equilibrium deviators are assumed to have whatever risk type has a stronger incentive to undertake the deviation. In our setting, in an equilibrium with universal testing, a person who does not test should be perceived as a low-risk type. These off-equilibrium beliefs would undermine the pooling equilibrium.

14

equilibrium. If θ a = θ H , a will select an action as follows.    (0, 0)         n f (t a , t a ) = (0, 1)          (1, 0)

if

Tf S

< ya

if

Tn S

< ya ≤

Tf S

Tn S

Tf S

if y a ≤

<

(12)

Where the thresholds are Tf =

θH v − c f φ

(13)

Tn =

c f − cn . (1 − φ )

(14)

and

In the absence of discrimination from sexual partners (S), high-risk types would like to seek an HIV test nearby. If S > 0, those who place high value on a future sexual relationship will not seek an HIV test. Those with intermediate values of y a will choose far away testing. Those who value the relationship least will test nearby, as they have less to lose by being observed. We now turn our attention to the consent decision. If b does not observe her match seeking an HIV test (σa = 0), she will consent to a sexual relationship. This stems from assumption (7). Low-risk types generate the signal σ = 0. Some high-risk types generate σ = 0 and others generate σ = 1. b’s match cannot be more likely to have HIV than the average member of the population. Formally, maximizing her expected payoff (6), b will consent if h     i f yb > E h a ch τ 1 − ρˆ b (tna + t a ) + co |σa = 0

(15)

where we have replaced ρ by the belief ρˆ b . The right-hand side of this inequality is always23 less than or equal to h¯ (ch τ + co ), so by assumption     f E h a ch τ (1 − ρˆ b (tna + t a ) + co |σa = 0) ≤ E(h a (ch τ + co )|σa = 0) = E(θ a |σa = 0)(ch τ + co ) ≤ E(θ a )(ch τ + co ) = h¯ (ch τ + co ) 23 Proof:

15

(7), b will consent. If, on the other hand, b observes that a has tested for HIV (σa = 1), she infers that he is a high-risk type. She will consent if24 

h

yb ≥ θ H c τ (1 − ρˆ b ) + c

o



.

(16)

She will only consent if her benefit from the relationship (yb ) is sufficiently high, or if she believes that ART is effective at preventing HIV transmission (ρˆ b is high). The total level of discrimination in the population, S, is the fraction of individuals in B who would reject a match who has been seen seeking an HIV test. ! Z Z o h h S=

∞

0

1+(c /c τ )−(y/θ H c τ )

0

f (ρˆ )dρˆ

g(y)dy

(17)

The functions f and g represent probability densities for ρˆ and y respectively. This paper evaluates an experiment designed to shift beliefs about the ART prevention parameter ρ. The experiment disseminates the information that ART reduces HIV transmission by 96 percent. In our model, we interpret this as a first-order stochastically dominant shift in the distribution Fρˆ . We now characterize the theoretical implications of such a shift in the distribution of beliefs. Proposition 2. If ch = 0, a change in the distribution of beliefs Fρˆ has no effect on the level of discrimination between sexual partners S. This comes from (16). If stigma is not based on the risk of HIV transmission, then beliefs about the risk of transmission will not affect the level of stigma. In what follows, we consider the case ch > 0, that is, there is some cost associated with contracting HIV. 24 Here,

we make use of the fact that she is certain her match is a f high-risk type, and has been tested for HIV: tna + t a =  1. In this case,    f h n o h o E[h a c τ 1 − ρˆ b (t a + t a ) + c |σa = 1] = θ H c τ (1 − ρˆ b ) + c .

16

Proposition 3. A first-order stochastically dominant shift in the distribution of beliefs Fρˆ weakly decreases the level of discrimination S. An upward shift in beliefs implies that each person increases their belief about the preventative effect of ART. Some might change their strategy from reject to consent, and none will change from consent to reject. S=

Z ∞ 0

 Fρˆ

y co − 1+ sτ θ H ch τ

 g(y)dy.

(18)

Consider a first-order stochastically dominant shift in the distribution Fρˆ . For each value of y, the integrand will decrease. Therefore, the level of discrimination S will also decrease. If other forms of stigma co are large relative to the cost of contracting HIV, ch , a shift in beliefs about ρ will have no effect, as the integrand in (18) will always equal 1. By shifting community-level beliefs, the information experiment reduces discrimination between sexual partners. This reduction in discrimination in turn affects HIV testing. Proposition 4. A decrease in discrimination between sexual partners leads to an increase in the total number of HIV tests. It also leads to an increase in the number of tests at nearby clinics. These two statements follow from the fact that a cumulative distribution function G (y) always increases in y, and limy→∞ G (y) = 1. Proposition 5. A moderate decrease in S has an ambiguous effect on the number of HIV tests at far away clinic. But, as S → 0, the number of far away tests approaches zero. The fraction of { a ∈ A} who choose far away testing equals G

θH v − c f φS

!

−G

c f − cn (1 − φ ) S

!

which, depending on the distribution G, may be locally increasing or decreasing in S. Figure 2 depicts the relationship between discrimination and HIV testing location. G is assumed to be normally distributed and the 17

near y High discrimination

far

y Moderate discrimination

y Low discrimination

Figure 2: Discrimination and HIV Testing Location Note: This figure illustrates the proportion of individuals who seek an HIV test. Each graph depicts the probability density function for the benefit of a relationship, y a . To the left of the first threshold, individuals seek a nearby HIV test. Between the first and second thresholds, individuals seek a far away test. To the right of the second threshold individuals do not test. These thresholds shift up as discrimination S decreases.

HIV testing rate is initially low. In this case, a moderate increase in beliefs increases far away testing. As stated in Proposition 5, this is not a general result. The effect depends on the distribution G and on initial testing levels. However, a large increase in beliefs should always increase nearby testing. In this model, HIV testing behavior depends on the beliefs of potential sexual partners. This leads to another specific empirical implication. Proposition 6. The decision to test for HIV depends on the beliefs of the community Fρˆ , and not on one’s own beliefs. The effect of a shift in beliefs on sexual behavior is ambiguous. A decrease in discrimination corresponds to a direct increase in sexual behavior. However, the rate of HIV testing increases, so the discrimination that remains will affect a larger number of people. Further theoretical and empirical discussion of sexual behavior is deferred to the appendix.

4

Empirical Strategy and Results

In this section, we use results from the experiment described in Section 2 to test assumptions and predictions of the model. We investigate the causal impact of a community meeting on survey measures of beliefs 18

and discrimination. We then estimate the impact of the intervention on our main outcome of interest: the village-level rate of HIV testing as recorded in administrative data. We also test additional predictions of the model. We combine survey data with administrative data to examine the relationship between beliefs and HIV testing. Finally, we attempt to rule out alternative mechanisms.

4.1

Beliefs and Discrimination

We conducted a survey four months after the intervention. We administered the survey to a subset of 1,358 community health meeting attendees, as identified by the chief’s attendance sheet25 . Out of 122 study villages, 119 were successfully surveyed26 . We investigate the effect of the full information intervention on the beliefs of meeting attendees. We regress each survey belief measure on the intervention arm of the village. Belie f ij = α + βFj + δ0 χij + eij

(19)

Belie f ij is a belief measure elicited from respondent i in village j. Fj ∈ {0, 1} is the indicator for the full intervention arm. χij is a set of individual and village-level covariates27 . Standard errors are clustered at the village level. Because Fj is randomly assigned, we expect E(eij | Fj ) = 0, so the OLS estimate βˆ is unbiased. The full intervention had a large and significant effect on beliefs 25 The

interviewers selected respondents by conducting a random walk within each village. Two interviewers were assigned to each village; one began the random walk at the center and the other at an outer edge of the village. The interviewers were hired from a pool of candidates who were not socially connected to the community educators employed for the intervention. Interviewers were not aware of the intervention. This results in selection bias among those interviewed relative to a random sample of the village as a whole, but this selection should be the same in both partial and full intervention villages. 26 In one full intervention village and two partial intervention villages, village authorities denied permission for a survey to take place due to the recent death of a village leader. 27 Individual-level controls include age, gender, whether the person is married, employed, primary school educated, secondary school educated, has livestock, and has a brick house. Village-level controls consist of covariates listed in Table 1 and distance to each clinic.

19

about the public benefit of ART, according to five different measures (Table 2). Respondents in the full intervention arm were much more likely to believe that ART prevents transmission (Columns 1, 2, 4 and 5). Most also understood an important implication: that a person taking ART is a safer sexual partner than a person who has never been tested (Column 3). To elicit beliefs about the transmission rate we used the bottle caps measure advocated by Delevande and Kohler (2009). Each of ten caps represents a serodiscordant couple: an untreated HIV-positive person and their HIV-negative spouse. Respondents were asked to remove one cap for each case of new infection within one year. The process was repeated under the alternative assumption of ART use. These measures allow us to calculate beliefs about the relative reduction in risk associated with ART use (Column 4 of Table 2). On average, those in the full intervention arm believed that the relative reduction in risk was 56 percent. This is higher than in the partial intervention arm (9 percent) but still below the true value ρ = 0.96. We also captured beliefs about the ART prevention parameter ρ using an infographic (Figure 3). We elicited this measure last, as it might remind respondents of the intervention. The distribution of beliefs about ρ is uniformly higher in the full intervention group (Figure 4, Column 5 of Table 2).

Figure 3: Infographic to Elicit Beliefs Note: Respondents stated their beliefs by selecting one of eight options. The top left corresponds to the belief that an infected person taking ART drugs is not contagious. The bottom right corresponds to the belief that ART drugs have no effect on contagion.

20

70

60

60

50

50

40

40

percent

percent

70

30

30

20

20

10

10

0

0 0

0.1

0.25

0.5

ρˆ

0.75

0.9

0.95

0

1

(a) Partial Intervention

0.1

0.25

0.5

ρˆ

0.75

0.9

0.95

1

(b) Full Intervention

Figure 4: Beliefs ρˆ (ART Prevention Parameter) Note: Beliefs were elicited from meeting attendees 4 months after the intervention, using the infographic in Figure 3. The true value is ρ = 0.96.

The full intervention reduced discrimination towards sexual partners taking ART. Respondents in full intervention villages are more likely to prefer a partner taking ART to one who has never been tested, and more likely to believe that a person taking ART will find a new sexual partner (Columns 1 and 2 of Table 3). Every village received information about the private benefits of ART drugs. Column 3 of Table 3 shows that respondents’ beliefs do not differ between the two intervention arms28 . Neither beliefs about the absolute transmission rate29 , nor beliefs about HIV prevalence were affected (Columns 4 and 5 of Table 3).

28 Respondents

were asked to agree or disagree, on a five-point Likert scale, with the statement An HIV-positive person can live a long and healthy life if he takes ART. 29 Beliefs about the absolute probability of HIV transmission are much higher than the true value, which, according to the Malawi National AIDS Commission, is approximately 10 percent per year. These beliefs are consistent with the overestimates of HIV transmission rates measured by Kerwin (2012) in Malawi, and may be explained by a health education policy which purposely overstates the risk of contracting HIV in an effort to discourage risk taking behavior.

21

Table 2: Beliefs Post-Intervention (4 Months): Does ART Reduce HIV Transmission? (1) (2) (3) Selected Likert Partner ART scale taking from (rescaled ART is list 0-1) less risky

(4) ρˆ (calculated)

(5) ρˆ (infographic)

Full intervention (F)

0.581*** (0.040)

0.364*** (0.023)

0.277*** (0.048)

0.443*** (0.028)

0.490*** (0.032)

Mean dep var in (P) Individual controls Village-level controls Obs (Individuals)

0.19 Yes Yes 1343

0.44 Yes Yes 1340

0.43 Yes Yes 1341

0.09 Yes Yes 1310

0.18 Yes Yes 1340

Note: (1) Selected ART from a list of possible HIV prevention methods: faithfulness, abstinence, ART, circumcision, condoms, and mosquito nets. (2) If an HIV-positive person takes ART it will reduce the chance that he transmits HIV to a partner. Likert scale: 5=strongly agree, 1=strongly disagree. Rescale: response is divided by 5. (3) The respondent believes that a person on ART drugs is less likely to transmit HIV than a person who has never been tested. (4) Used ten bottle caps to show beliefs about (a) absolute transmission probability: Ten couples are serodiscordant (one HIV positive and the other negative). Suppose they do not use condoms or ART drugs. After one year, how many will transmit HIV to their partner? (b) transmission probability with ART: Suppose instead that they are all taking ART. How many will transmit HIV to their partner?. These were used to calculate the relative reduction in risk. (5) The respondent’s selection from Figure 3, converted into a measure of ρ by rescaling. ρ = 0.96 relative reduction in HIV transmission associated with antiretroviral drugs. Survey to meeting attendees 4 months post-intervention. OLS, at the individual level, with a constant and controls: age, gender, married, employed, primary school educated, secondary school educated, has livestock, has a brick house, variables listed in Table 1, village population and distance to each health facility. Standard errors are clustered at the village level, in parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

22

Table 3: Attitudes and Beliefs Post-Intervention (4 Months ) Attitudes

Other beliefs

(1) Prefers untested vs. taking ART

(2) Thinks ART user won’t find new partner

(3) ART leads to a long and healthy life

(4) Absolute transmission (one year)

(5) Preva -lence

Full intervention (F)

-0.155*** (0.034)

-0.112** (0.043)

0.009 (0.008)

-0.011 (0.009)

0.013 (0.013)

Mean dep var in (P) Individual controls Village-level controls Obs (Individuals)

0.46 Yes Yes 1224

0.68 Yes Yes 1276

0.95 Yes Yes 1340

0.96 Yes Yes 1330

0.54 Yes Yes 1224

Note: (1) Would prefer a partner who has never been tested for HIV to one who is taking ART. (2) Believes that a person taking ART will definitely not find a new sexual partner. (3) An HIV-positive person who takes ART can live a long, healthy life. Likert scale: 5=strongly agree, 1=strongly disagree. Rescale: response is divided by 5. (4) Used ten bottle caps to show one-year probability of HIV transmission for a serodiscordant couple who are not using condoms or taking ART. (5) Used ten bottle caps to show HIV prevalence in the village. Survey to meeting attendees 4 months post-intervention. OLS, at the individual level, with a constant and controls: age, gender, married, employed, primary school educated, secondary school educated, has livestock, has a brick house, variables listed in Table 1, village population and distance to each health facility. Standard errors are clustered at the village level, in parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

4.2

HIV Testing

The main outcome of interest is HIV testing. We digitized data from handwritten patient registers at all 18 free clinics in the study area. We obtained HIV testing data for 2.5 months before the start of the intervention and 4 months after the intervention ended30 , to coincide with the survey. The HIV testing register lists the client’s sex, age, and

30 We

ignore all HIV tests that took place during the intervention period, which lasted 22 days.

23

address or home village31 . This allows us to link an HIV test to the patient’s village, and to an intervention arm. We use GPS coordinates to calculate the distances between villages and clinics. During the study, clinics did not experience a shortage of HIV testing kits, counsellors, or antiretroviral drugs. Our intervention is designed to increase voluntary HIV testing among a sexually active population. We therefore define our target population as non-pregnant individuals aged 15 to 4932 . Pregnant women are excluded, as they must undergo HIV testing to receive antenatal health care. We must construct this outcome at the village level, as our data is deidentified. We interpret results as the effect of the community meeting on the village-level testing rate, regardless of meeting attendance. To identify the causal effect of the information intervention on the HIV testing rate, we perform an OLS regression at the village level, weighted by village population. We control for pre-intervention levels of the outcome variable and other village level covariates33 . This intervention may be subject to spillovers. 60 percent of the villages in the study are within 1 kilometer of another study village. We use a specification that controls for nearby study villages and nearby full intervention villages. Because of random assignment, including these regressors does not introduce endogeneity, and removes the bias caused by spillovers. Our regression specification is as follows: Percent HIV tested j = α + β 1 Fj + γ0 Nj + γ1 NFj + δ0 χ j + e j .

(20)

Fj is an indicator for the full intervention arm, χ j is a vector of villagelevel covariates including the pre-intervention testing level (see Table 1 31 HIV

testing registers also indicate whether the patient was pregnant, whether or not the test was a joint test, the time since the client’s last HIV test, and the result of the test. 32 Census data provides the approximate size of the target population in each village. 33 This estimator (ANCOVA) is unbiased in a randomized study, and has lower variance than either a simple regression of post-intervention testing levels on treatment, or a difference-in-difference estimator (McKenzie, 2012). Weighted OLS accounts for the heteroskedasticity in village-level averages, increasing efficiency while remaining unbiased.

24

for a complete list). Nj is the total number of study villages within one kilometer, and NFj is the number of full intervention villages within one kilometer. Standard errors are robust. Because Fj is randomly assigned, E(e j | Fj ) = 0, and OLS estimate βˆ 1 is unbiased. We report weighted OLS estimates and logistic odds ratios with robust standard errors. These results do not differ greatly from estimates obtained using an unweighted OLS or probit specification (see the appendix). Table 9 shows that in the partial intervention group, the four-month HIV testing rate is low. 2.64 percent of the target population sought an HIV test (approximately 5 tests per village). The full information intervention caused a significant increase in the total number of HIV tests, as shown in Table 9. Controlling for spillovers, the estimated effect size is 0.91 percentage points. This is a relative increase of 34 percent, or approximately 2 additional HIV tests per village. The results of a logistic regression are similar (Columns 3 and 4 of Table 9): the adjusted odds ratio is 1.42 (95% confidence interval [1.12, 1.80]). Spillovers from the full intervention arm are positive and significant. Such spillovers may be due to the spread of information, or a perceived change in attitudes towards those who seek an HIV test. The robust standard errors reported in the main specification (Column 2 of Table 9) are more conservative than standard errors adjusted for spatial correlation (Conley and Molinari, 2007) (p=0.011), or that rely on the wild cluster bootstrap method proposed by Cameron et al. (2008) for a small number of clusters34 (p=0.004). Granich et al. (2009) use an epidemiological model to link low HIV testing rates to the spread of HIV. According to the epidemiological model, our intervention averts approximately two new HIV infections per primary infection. This model abstracts from some behavioral considerations. But, it is worth noting that a small increase in HIV testing may have large epidemiological implications.

34 We

can cluster at the clinic, with only 11 clusters.

25

Table 4: Voluntary HIV Tests Post-Intervention (4 Months) % tested (pp) Weighted OLS Full intervention (F)

Odds ratio Logit

(1)

(2)

(3)

(4)

0.738** (0.352)

0.906** (0.354)

1.329** (0.163)

1.420*** (0.171)

# P or F villages < 1km

-0.236 (0.311)

0.941 (0.105)

# F villages < 1km

0.806** (0.373)

1.401** (0.193)

Mean of dep var in (P) Village-level controls Proportional increase Obs (Villages)

2.64 Yes 28% 122

2.64 Yes 34% 122

Yes

Yes

122

122

Note: Administrative data from 18 clinics. Dependent variable: percent (/100) of village target population tested for HIV post-intervention. 15-49, non-pregnant. The post-intervention period is 4 months. All regressions are at the village level and include a constant. Village-level controls: variables listed in Table 1, village population and distance to each health facility. Robust standard errors in parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

We have demonstrated a causal effect of the full intervention on beliefs, discrimination, and HIV testing. Guided by our theoretical predictions, we now examine the correlation between beliefs and HIV testing behavior. Before the intervention, one third of those seeking an HIV test traveled further than necessary. This reduces the likelihood of being seen. HIV test seekers may also travel to far away clinics for other reasons, such as clinic quality. In this case beliefs about the public benefit of ART should play no role. We say that an HIV test takes place far away if a client travels at least three kilometers further than the nearest clinic to their home village35 . We again run a weighted OLS regression with specification (20). 35 For

many villages, two clinics are roughly equidistant, and the walking distance might differ from the GPS distance by a few kilometers. We define far away with a three-kilometer buffer to exclude such situations.

26

Columns 1 and 2 of Table 5 show an increase in both nearby and far away testing, though coefficients are imprecisely estimated. The model makes a particular prediction about the link between testing location and the distribution of beliefs about the ART prevention parameter ρ. Moderate beliefs correspond to a moderate reduction in discrimination. This might increase both the number of people who test nearby and the number who test far away. High beliefs correspond to a large reduction in discrimination. This should only increase the number of nearby tests. We define high beliefs as beliefs ρˆ ≥ 0.95. This threshold approximately36 matches the true value of ρ. From a policy perspective, it is useful to know whether shifting beliefs to the true value of ρ is effective. As the majority of individuals in partial intervention villages hold beliefs ρˆ = 0, we define moderate beliefs to be 0 < ρˆ < 0.95. The full intervention shifted beliefs from low beliefs ρˆ = 0 to either moderate or high beliefs, as shown in Figure 4. The full variation in beliefs is explained partly by the intervention, and partly by unobserved factors. We use the following OLS specification. y j = α + β 1 (moderate beliefs) j + β 2 (high beliefs) j + δ0 χ j + e j

(21)

y j is either the percent of the target population in village j that seeks an HIV test nearby (far away) from their home village. The regressors (moderate beliefs) j and (high beliefs) j represent the proportion of the village that has moderate or high beliefs, as estimated from survey responses. We interpret βˆ 1 and βˆ 2 as correlations between community beliefs and the number of HIV tests either nearby or far away. χ j is a vector of village-level covariates. Standard errors are robust. Results are consistent with the predictions of the model. High beliefs about the public benefit of ART predict a significant increase in the number of nearby tests (Columns 3 and 4 of Table 5). The link between beliefs and test location follows the pattern predicted by Figure 2. 36 These

beliefs are elicited in the survey using the infographic in Figure 3. Note that this beliefs measure is discrete, and the selection ρˆ = 0.95 is closest to the true value of ρ = 0.96.

27

Table 5: Distribution of Beliefs and HIV Testing Location % of population tested for HIV (3) Nearby

(4) Far

Proportion of village with high beliefs (ρˆ ≥ 0.95)

1.714*** (0.591)

-0.219 (0.353)

Proportion of village with moderate beliefs (0 < ρˆ < 0.95)

-0.028 (0.534)

0.491 (0.446)

Yes 119

Yes 119

Full intervention (F)

28

Mean of dep var in (P) Village-level controls Obs (Villages)

(1) Nearby

(2) Far

0.247 (0.309)

0.659* (0.341)

1.64 Yes 122

1 Yes 122

Note: HIV testing rate and location are based on administrative data from 18 health clinics. Dependent variable: percent (/100) of village target population tested for HIV post-intervention far/nearby. Far clinics defined as >3km further than nearest free clinic. (1)-(2): Reference category is partial intervention. Controls for spillovers. (4)-(6): ρ = the relative reduction in HIV transmission associated with antiretroviral drugs. True value: ρ = 0.96. Community beliefs about ρ are approximated by the village-level average of beliefs, as obtained from the survey using the infographic in Figure 3. Omitted reference category: proportion of village with low beliefs ρ = 0. 15-49, non-pregnant. The post-intervention period is 4 months. All regressions are weighted OLS at the village level and include a constant. Village-level controls: variables listed in Table 1, village population and distance to each clinic. Robust standard errors in parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

In the model, HIV testing decisions depend on common knowledge. A person is more likely to test if they know that the community is aware of the public benefit of ART. This differs sharply from a model involving social preferences, in which one’s own beliefs matter most. It also differs from a household bargaining model, in which one’s spouse’s beliefs are important. To test this prediction, we measured perceived beliefs of the community. After using Figure 3 to indicate their own beliefs, each respondent was asked about others’ beliefs. They indicated the option their spouse would likely choose, and the option most members of the community would choose. We regress self-reported HIV testing37 on this measure of perceived beliefs, with specification HIV testij = α + β 1 (respondent has high beliefs)i + β 2 (spouse has high beliefs)i + β 3 (community has high beliefs)i + δ0 χij + eij (22) where the dependent variable is an indicator for respondent i in village j reporting an HIV test post-intervention. The regressors of interest are the respondent’s own beliefs, her perception of her spouse’s beliefs, and her perception of the community’s beliefs. These three variables are indicators for high beliefs about the public benefit of ART, ρˆ > 0.95. The set of covariates χij includes individual and village-level controls. Standard errors are clustered at the village level. As predicted, respondents are more likely to report an HIV test if they believe that the community knows about the public benefit of ART (Table 6). The respondent’s own beliefs and perceived spouse’s beliefs do not play a significant role.

37 We

cannot match de-identified administrative data to survey beliefs at the individual level, and therefore use self-reported testing as an outcome in this regression. While self-reports may be an imperfect measure of HIV tests, even an increased willingness to report an HIV test would provide evidence for our theory.

29

Table 6: Common Knowledge of ρ and HIV Testing

Believes ρ ≥ 0.95

HIV test: post

pre

(1)

(2)

(3)

-0.050 (0.039)

-0.018 (0.059)

0.053 (0.054)

-0.100 (0.075)

-0.024 (0.061)

0.110** (0.044)

0.151*** (0.050)

-0.081* (0.046)

Yes Yes 1330

Yes Yes 1079

Yes Yes 1079

Thinks spouse believes ρ ≥ 0.95 Thinks community believes ρ ≥ 0.95 Village controls Individual controls Obs (Individuals)

Note: Outcome is self-reported HIV test post-intervention or 4 months preintervention. ρ = the relative reduction in HIV transmission associated with antiretroviral drugs. True value: ρ = 0.96. Beliefs about ρ are obtained from the survey using the infographic in Figure 3. (2)-(3): sample restricted to married respondents. Survey to meeting attendees 4 months post-intervention. OLS, at the individual level, with a constant and controls: age, gender, married, employed, primary school educated, secondary school educated, has livestock, has a brick house, variables listed in Table 1, village population and distance to each health facility. Standard errors are clustered at the village level, in parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

4.3

Other Mechanisms

We claim that the increase in HIV testing we observe is due to a reduction in discrimination towards those who seek an HIV test. However, other models may generate similar predictions. Both social preferences and household bargaining would cause an increase in HIV testing in response to new information about the public benefit of ART drugs. We will attempt to rule these out as first order explanations for our results. Altruism is an obvious alternative mechanism. People may derive utility from protecting their sexual partners from HIV. If this were the primary mechanism, HIV testing would depend on one’s own beliefs about the public benefit of ART, as opposed to others’ beliefs. As shown in Table 6, this is not the case. Additionally, while altruism could cause 30

either nearby or far away testing to increase, it does not explain the pattern observed in Table 5. High beliefs about the public benefit of ART are linked to nearby testing. Altruism may nevertheless play a role in HIV testing. A person who is altruistic towards his partner may be less likely to be HIV positive. In addition, an altruistic person may choose to test even without knowledge of the public benefit of ART. For example, they might want to make informed choices to limit transmission risk. Household bargaining can play a role in family planning decisions (Ashraf et al., 2013). HIV testing decisions may also be subject to household bargaining. A person who is aware of the fact that ART reduces HIV transmission may put greater pressure on her spouse to get tested and treated for HIV. In this case, we would expect the testing decision to be strongly linked to one’s spouse’s beliefs about the benefits of ART. Table 6 shows that this is not the case. A model of household bargaining would also predict an increase in joint testing. A person has little to gain by putting pressure on her partner to seek an HIV test if he is not intrinsically motivated to seek treatment. The results of the test are private, and he can claim a negative test result. However, if a couple is tested jointly, the results are seen by both. Each member of the couple can exert pressure on the other to seek ART if necessary. Column 1 of Table 7 shows that the effect of the full intervention on joint testing is small and insignificant. Survey results do not indicate an increase in pressure to test, a preference for joint testing, or willingness to pay for joint testing (Columns 2, 3, and 4 of Table 7). Household bargaining does not seem to explain the increase in HIV testing we have seen as a result of the information campaign. It may nevertheless play some role in HIV testing decisions. Most respondents report a strong preference for joint testing. Approximately 20 percent of HIV tests in the clinic data are joint tests. Those with enough bargaining power might pressure their partners to test even if they do not know that ART blocks transmission. Joint testing allows couples to make informed decisions about condom use and family planning.

31

Table 7: Pressure to Test for HIV Joint tests

Full intervention (F)

32

Mean of dep var in (P) Individual controls Village-level controls Obs (Villages) Obs (Individuals)

Survey measures

(1) % joint tested

(2) =1 if tested due to pressure

(3) =1 if prefers joint testing

(4) WTP joint vs. private

0.131 (0.097)

0.006 (0.007)

-0.017 (0.017)

-6.119 (15.704)

.34 No Yes 122

0.01 Yes Yes

0.94 Yes Yes

105.40 Yes Yes

1010

1010

1006

Note: (1) Administrative data from 18 clinics. Dependent variable: percent of village target population joint tested for HIV in 4 months post-intervention.Target population: age 15-49, non-pregnant, calculated from the Malawian National Statistics Office census. Regression is weighted OLS at the village level, with village-level controls, controls for spillovers and a constant. (2)-(4): Survey to meeting attendees in intervention villages. Dependent variables are individual responses to survey questions. Sample restricted to married respondents. The post-intervention period is approximately 4 months. All regressions are OLS, at the individual level, and with controls and a constant. Individual-level controls: age, gender, employed, primary school educated, secondary school educated, has livestock, has a brick house. (2) Respondent sought HIV test due to pressure. (3) Respondent would prefer a joint test over private test. Sample: married respondents. (4) Respondent’s willingness to pay for door-to-door joint testing campaign vs. private testing (Malawi Kwacha: 400MK = 1USD). Village-level controls: Table 1, village population and distance to each clinic. Robust standard errors are given in the parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

While we have focused our attention on discrimination, other barriers to HIV testing and treatment surely exist. For example, there may be psychic costs of testing. Oster et al. (2013) proposes an optimal expectations model to explain low uptake of a test for Huntington’s Disease. A desire to ignore one’s HIV status may partly explain low HIV testing rates. However, there is near universal demand for private door-to-door testing in Malawi (Angotti et al., 2009). Supplies of HIV testing kits, counsellors and ART are consistently available in Zomba District. However, travel costs may explain a large part of the gap in HIV testing. Indeed, testing is significantly negatively correlated with distance to the nearest clinic. A reduction in perceived transmission risk may lead to risk compensation. However, this is not necessarily the case. A reduction in discrimination increases HIV testing, and the net effect on consent to risky sex is ambiguous (proof in the appendix). We collected self-reported measures of sexual risk taking. We found a small and insignificant effect on risk taking (results in the appendix), consistent with the limited behavioral response to the HIV epidemic documented by (Oster, 2012) and limited HIV risk compensation observed by Friedman (2012) and Kerwin (2014). ART is a very effective HIV prevention method. Among ART users, there would need to be a 25-fold increase in unprotected sexual activity to offset the 96 percent reduction in HIV transmission38 .

5

Conclusion

This study indicates that statistical discrimination from sexual partners is a barrier to HIV testing. HIV testing disproportionately benefits those infected with the virus, and is a signal of underlying risk. Community members often observe testing decisions, and reject potential partners who’ve been tested. 38 Risk

compensation may be of concern for other reasons, for example, the spread of other sexually-transmitted infections which ART does not prevent. Gong (2014) and Baird et al. (2014) find an increased risk of sexually-transmitted infection after a person tests positive for HIV.

33

This statistical discrimination is based on misinformation. Most are unaware of the fact that an HIV-positive person who is tested and treated for HIV is much less contagious. Providing information about the public benefit of ART reduces discrimination. This in turn increases HIV testing by 34 percent. A large shift in beliefs predicts a large increase in the number of HIV tests sought at nearby clinics. The testing decision appears to be based on common knowledge of the community’s beliefs about ART, and not one’s own beliefs. These results suggest that providing new information can reduce statistical discrimination. If the increase in HIV testing we observed was due to a reduction in discrimination, the effect should persist. The theory suggests that the intervention targeted those who care most about the number of sexual prospects. Individuals who prefer to have many potential sexual partners are also most at risk of contracting and spreading HIV. Our information intervention increased the annual testing rate from 9 percent to 12 percent. While this suggests that discrimination is an important barrier to HIV testing, it is not the only one. Travel costs, psychic costs, time inconsistency, and other misconceptions are important areas for future research. This work may be relevant to contexts in which a technology is under-adopted because it signals an underlying type. For example, a person who suffers from mental illness may be reluctant to seek psychiatric care, for fear his friends or employer might find out. If an employer believes that the mentally ill are less productive workers, an employee might fear statistical discrimination. This leads to a bad equilibrium. Mental illness goes untreated, and an employee is less productive than he would be if treated. Of course, one solution is to ensure privacy. We propose that providing employers with correct information about the productivity benefits of treatment might also be effective. A community-level information campaign is an inexpensive way to increase HIV testing. A policy that reduces discrimination is in some ways more attractive than one that conceals HIV tests or their motivation. For example, monetary incentives increase HIV testing. However, to be effective in the face of discrimination, they must be paid repeat34

edly to the entire community. This might make monetary incentives for treatment infeasible. A reduction in the level of discrimination should affect all health seeking behavior, and have permanent effects. Many countries, including Malawi, have adopted policies of universal testing and treatment. Any such policy should involve community outreach and the provision of accurate information. A scaled policy intervention should have two important features. First, information on the public benefit of ART should be provided at the community level, to ensure common knowledge. Second, the information should be provided in a credible way, for example, by trusted health authorities. These aims are achievable at low cost by employing community health workers, or by using billboards and radio messages. In the presence of discrimination, learning and information sharing are hampered. Under-adoption of ART drugs means that most do not learn of their effects first hand. Discrimination makes people reluctant to discuss these effects with social contacts. Credibility is also a challenge: those who are infected stand to benefit from a shift in beliefs about ART drugs. In the absence of an information campaign, discrimination allows incorrect beliefs to persist. While our experiment took place in Malawi, there is reason to believe its policy lessons apply elsewhere. Southern Malawi is representative of sub-Saharan Africa in many ways. HIV prevalence is high, and ART drugs are widely available. The population is concentrated in rural villages, in which health seeking behavior is often observable. Our model of discrimination between sexual partners is based on rational behavior, and might apply in communities outside of our study area. People who know that ART reduces HIV transmission are less likely to discriminate against people with HIV. Many public health campaigns inflate perceptions of risk in order to reduce risk taking. Such policies may have unintended consequences. They exacerbate discrimination and inhibit health seeking behavior. Information that increases HIV testing improves the lives of those who are infected by providing access to treatment, and serves the community as a whole by reducing the spread of the virus.

35

References Altonji, J. G. and Pierret, C. R. (2001). Employer learning and statistical discrimination. The Quarterly Journal of Economics, 116(1):313–350. Angotti, N., Bula, A., Gaydosh, L., Kimchi, E. Z., Thornton, R. L., and Yeatman, S. E. (2009). Increasing the acceptability of hiv counseling and testing with three c’s: convenience, confidentiality and credibility. Social Science & Medicine, 68(12):2263–2270. Angotti, N., Kaler, A., and Ramaiya, A. (2016). You cannot believe your eyes: Gendered dimensions of art anxiety in rural malawi. Presented at Population Association of America Conference, Washington, DC. Arrow, K. (1973). The theory of discrimination. Discrimination in labor markets, 3(10):3–33. Ashraf, N., Field, E., and Leight, J. (2013). Contraceptive access and fertility: The impact of supply-side interventions. Working Paper. ¨ Baird, S., Chirwa, E., McIntosh, C., and Ozler, B. (2010). The shortterm impacts of a schooling conditional cash transfer program on the sexual behavior of young women. Health economics, 19(S1):55–68. Baird, S., Gong, E., McIntosh, C., and Ozler, B. (2014). The heterogeneous effects of hiv testing. Journal of Health Economics, 37:98–112. ¨ Baird, S. J., Garfein, R. S., McIntosh, C. T., and Ozler, B. (2012). Effect of a cash transfer programme for schooling on prevalence of hiv and herpes simplex type 2 in malawi: a cluster randomised trial. The Lancet, 379(9823):1320–1329. Bandiera, O., Buehren, N., Burgess, R., Goldstein, M., Gulesci, S., Rasul, I., and Sulaiman, M. (2012). Empowering adolescent girls: Evidence from a randomized control trial in Uganda. Working Paper. Benabou, R. and Tirole, J. (2011). Laws and norms. Working Paper. Besley, T. and Coate, S. (1992). Understanding welfare stigma: taxpayer resentment and statistical discrimination. Journal of Public Economics, 48(2):165–183. 36

Cameron, A. C., Gelbach, J. B., and Miller, D. L. (2008). Bootstrapbased improvements for inference with clustered errors. The Review of Economics and Statistics, 90(3):414–427. Chinkhumba, J., Godlonton, S., and Thornton, R. (2012). The demand for medical male circumcision. American Economic Journal: Applied Economics, 6(2):152–177. Coate, S. and Loury, G. C. (1993). Will affirmative-action policies eliminate negative stereotypes? The American Economic Review, pages 1220– 1240. Cohen, M. S., Chen, Y. Q., McCauley, M., Gamble, T., Hosseinipour, M. C., Kumarasamy, N., ..., and Fleming, T. R. (2011). Prevention of HIV-1 infection with early antiretroviral therapy. New England Journal of Medicine, 365:493–505. Conley, T. G. and Molinari, F. (2007). Spatial correlation robust inference with errors in location or distance. Journal of Econometrics, 140(1):76– 96. Delevande, A. and Kohler, H.-P. (2009). Subjective expectations in the context of HIV/AIDS in Malawi. Demographic Research, 20(31):817– 875. DHS (2010). Malawi DHS, 2010 - Final report. Duflo, E., Dupas, P., and Kremer, M. (2015). Education, hiv, and early fertility: Experimental evidence from kenya. The American Economic Review, 105(9):2757. Dupas, P. (2011). Do teenagers respond to HIV risk information? Evidence from a field experiment in Kenya. American Economic Journal: Applied Economics, pages 1–34. Friedman, W. (2012). Antiretroviral drug access and behavior change. Working Paper.

37

Godlonton, S., Munthali, A., and Thornton, R. (2011). Behavioral response to information? Circumcision, information, and HIV prevention. Bureau for Research and Economic Analysis of Development Working Paper. Godlonton, S. and Thornton, R. (2012). Peer effects in learning hiv results. Journal of Development Economics, 97(1):118–129. Goffman, E. (1963). Stigma: Notes on the management of spoiled identity. Prentice-Hall. Gong, E. (2014). HIV testing and risky sexual behaviour. Working Paper. Granich, R. M., Gilks, C. F., Dye, C., Cock, K. M. D., and Williams, B. G. (2009). Universal voluntary HIV testing with immediate antiretroviral therapy as a strategy for elimination of HIV transmission: a mathematical model. The Lancet, 373(9657):48–57. Helleringer, S., Kohler, H.-P., and Kalilani-Phiri, L. (2009). The association of hiv serodiscordance and partnership concurrency in likoma island (malawi). AIDS (London, England), 23(10):1285. Hoffmann, V., Fooks, J. R., and D. Messer, K. (2014). Measuring and mitigating hiv stigma: A framed field experiment. Economic Development and Cultural Change, 62(4):701–726. Jalan, J. and Somanathan, E. (2008). The importance of being informed: Experimental evidence on demand for environmental quality. Journal of Development Economics, 87(1):14–28. Kerwin, J. (2012). Rational fatalism: non-monotonic choices in response to risks. Working Paper. Kerwin, J. T. (2014). The effect of HIV infection risk beliefs on risky sexual behavior: Scared straight or scared to death? Working Paper. Madajewicz, M., Pfaff, A., Van Geen, A., Graziano, J., Hussein, I., Momotaj, H., Sylvi, R., and Ahsan, H. (2007). Can information alone change behavior? Response to arsenic contamination of groundwater in bangladesh. Journal of Development Economics, 84(2):731–754. 38

Mahajan, A. P., Sayles, J. N., Patel, V. A., Remien, R. H., Ortiz, D., Szekeres, G., and Coates, T. J. (2008). Stigma in the HIV/AIDS epidemic: A review of the literature and recommendations for the way forward. AIDS, 22(2):S67. McKenzie, D. (2012). Beyond baseline and follow-up: The case for more T in experiments. Journal of Development Economics, 99(2):210–221. Moffitt, R. (1983). An economic model of welfare stigma. The American Economic Review, pages 1023–1035. Ngatia, M. (2011). Social interactions and individual reproductive decisions. Technical report. Oster, E. (2012). Hiv and sexual behavior change: Why not africa? Journal of Health Economics, 31(1):35–49. Oster, E., Shoulson, I., and Dorsey, E. R. (2013). Optimal expectations and limited medical testing: Evidence from Huntington disease. American Economic Review, 103(2):804–30. Peski, M. and Szentes, B. (2013). Spontaneous discrimination. The American Economic Review, 103(6):2412–2436. Thornton, R. L. (2008). The demand for, and impact of, learning HIV status. The American Economic Review, 98(5):1829. WHO (2013). Report on the global AIDS epidemic.

39

A A.1

Appendix Intervention Details Table 8: Intervention

Topic Initial beliefs (private benefits) Private benefits Mechanism

Infographic P Initial beliefs (public benefit)

Public benefit of ART

Infographic F Other information about ART Prevention Availability Questions

Script Raise your hand if you believe that person with HIV can live a long and healthy life with ART. A person who has HIV can live a long and healthy life, as long as he or she takes ART. A person taking ART will still have HIV, but they will have a reduced viral load and few or no symptoms. Figure 1 Imagine a couple. One person is HIV positive and the other is HIV negative. If the HIV-positive person takes ART, does that reduce the chance that the virus is passed to his or her partner? Raise your hand if you think the answer is yes. If a person with HIV takes ART, it will greatly reduce the chance of spreading HIV. Imagine an area where no one takes ART, where 100 people contracted HIV last year. If all of their partners had been taking ART, only 4 people would have contracted HIV. An HIV-positive person taking ART is 96 percent less contagious. This is true for both men and women. ART reduces the amount of virus in the body, which reduces the chance that the virus will be transmitted from one person to another. Figure 1 Only HIV-positive people should take ART, and should adhere properly. If he or she forgets to take the pills, his or her viral load will increase. For maximum protection, you should practice faithfulness and use condoms. Health clinics offer free HIV testing and ART. For other questions, ask at the health clinic.

40

Partial Full X

X

X

X

X

X X

X

X

X X

X

X

X X

X X

Figure 5: Study Area Note: The study included 122 villages in Zomba District, Malawi, represented by blue dots. Administrative data was obtained from 18 clinics, represented by red squares.

A.2

Proofs

We now characterize the pure-strategy Perfect Bayesian Equilibria of the model in Section 3. We analyzed the partially separating equilibrium in Section 3. Proposition 1 claims that there is only one other type of equilibrium: a pooling equilibrium. For some parameter values, our game has a class of pure-strategy Perfect Bayesian Equilibria in which a fraction P of the population consents only to matches who have been tested. The remaining fraction 1 − P consent to all matches. P must be large enough to induce universal HIV testing. This is because the equilibrium relies on extreme off-equilibrium beliefs about those who do not test. In this sense, the pooling equilibrium is not stable (it does not satisfy the D1 criterion). Intuitively, high-risk individuals have more to gain from testing. So, it is not natural to believe that a person who has not tested is riskier than someone who has. Fix P, and consider the strategy of each a ∈ A. His benefit to testing nearby is access to ART drugs plus the guarantee that his potential sexual partner will consent, θ a v + Py a , and the cost is cn . There is no additional benefit to testing far away, and he will not do so. He will test

41

nearby if θ a v + Py a > cn .

(23)

High-risk types θ a = θ H will always test, because we assumed θ H v > cn . The proportion of low-risk types who test is equal to P (θ L v + Py a > cn ) .

(24)

If this measure is less than one, then this is not an equilibrium. To see why, consider the beliefs of b ∈ B . If her match has not been tested, she concludes that his type is θ = θ L . But in this case, she will consent (assumption (7)), and therefore P = 0. This implies that for parameters such that P (θ L v + y a > cn ) < 1 there is no pooling equilibrium. Even if P = 1, that is, everyone rejects those who have not been tested, universal testing will not be achieved. Some low-risk types are not sufficiently motivated by the prospect of a sexual relationship to bother seeking an HIV test. Conversely, a pooling equilibrium exists for any 0 < P ≤ 1 such that P (θ L v + Py a > cn ) = 1.

(25)

All a ∈ A best respond to the strategies of those in B by testing nearby. In this case, beliefs about a match who has not been tested (σa = 0) can take any value, as this action is off the equilibrium path. To prove the existence of the equilibrium, we set these beliefs equal to θˆb (0) = 1 for all b ∈ [0, P] and θˆb (0) = 0 for all b ∈ ( P, 1]. That is, a fraction P of the population believe that a person who has not been tested is HIV positive with probability one. The remaining fraction 1 − P believe that a person who has not been tested is HIV negative. Because it is never worth consenting to a match who is infected and untreated, P(yb < ch τ + co ) = 1 (assumption (8)), every b ∈ [0, P] will reject a partner who has not been tested. Meanwhile, assumption (7)

42

implies P(yb > 0) ≥ P(yb > h¯ (ch τ + co )) = 1.

(26)

In equilibrium, HIV testing is universal. We now describe and rule out other potential types of equilibria. In the case of either universal testing or universal abstention from testing, consistent beliefs on the equilibrium path will equal the average ¯ and by assumption (7), b will HIV prevalence in the population: θˆb = h, consent, because yb > h¯ (ch τ + co ) > h¯ (ch τ (1 − ρˆ b ) + co ).

(27)

Consider the case in which some test and others do not. Highrisk types have more to gain. This implies adverse selection among those who test. Those who do not test are therefore lower-risk than the average member of the population. So, consistent beliefs about those ¯ and by assumption (7), b will consent to who do not test are θˆb (0) < h, those who do not seek an HIV test. We can now restrict attention to equilibria in which each b ∈ B adopts a strategy of consenting to some or all matches. The partially separating and pooling equilibria cover cases in which some always consent, and others all adopt the same strategy. The only other potential class of equilibrium is one in which different b ∈ B adopt opposite strategies towards those who seek an HIV test. That is, S > 0 and P > 0. Suppose S > P > 0. In this case, all low-risk types will not test, and some high-risk types may not either. Among those who do not test, the average risk is less than that the average risk in the population. ¯ By assumption (7), all b ∈ B will Consistent beliefs are θˆb (0) < h. consent to those who do not test, so P = 0, which is a contradiction. Suppose instead that P > S > 0. In this case, all high-risk types will test, as the net cost of discrimination is negative (S − P < 0). We consider two subcases. First, suppose this does not induce universal testing. That is, P (θ L v + ( P − S)y a > cn ) < 1. (28) 43

Any person who does not test is a low-risk type. So, rational beliefs are θˆb (0) = θ L . By assumption (7), all b ∈ B will consent to those who do not test, so P = 0, which again leads to a contradiction. Instead, suppose that this does induce universal testing. So, P (θ L v + ( P − S)y a > c) = 1.

(29)

Now, rational beliefs about those who do seek an HIV test are θˆb (1) = ¯ so by assumption (7) everyone will consent. This implies S = 0. h,

A.3

HIV Testing: Alternative Specifications Table 9: Voluntary HIV Tests Post-Intervention (4 Months) % tested for HIV (pp) Unweighted OLS

Full intervention (F)

Binomial Probit

(1)

(2)

(3)

(4)

0.655* (0.398)

0.834** (0.385)

0.121** (0.052)

0.150*** (0.051)

# P or F villages < 1km

-0.134 (0.339)

-0.029 (0.047)

# F villages < 1km

0.810* (0.456)

0.144** (0.057)

Mean of dep var in (P) Proportional increase Village-level controls Obs (Villages)

2.64 25% Yes 122

2.64 32% Yes 122

Yes 122

Yes 122

Note: Administrative data from 18 clinics. Dependent variable: (1)-(2) percent (/100) of village target population tested for HIV post-intervention. (3)-(4) binomial aggregated by village, estimates are coefficients of the probit link function. 15-49, nonpregnant. The post-intervention period is 4 months. All regressions are at the village level and include a constant. Village-level controls: variables listed in Table 1, village population and distance to each health facility. Robust standard errors in parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

44

A.4

Risk Compensation

Proposition 7. The effect of the level of discrimination, S, on the number of high-risk individuals who obtain consent from their matches is ambiguous. All low-risk individuals (θ a = θ L ) obtain consent. The fraction of high-risk types who obtain consent is equal to
M (θ H ) = 1 − φSG T f /S − (1 − φ)SG ( Tn /S) .

(30)

If we take the first order condition with respect to the level of discrimination S, we obtain           Tf Tf Tf Tn Tn dM (θ H ) Tn + (1 − φ ) g . =φ g −G −G dS S S S S S S (31) The sign of this derivative depends on the distribution of preferences G. For example, if G is the Uniform distribution, the net effect of a decrease in discrimination on sexual behavior is zero. During the survey, we collected self-reported measures of sexual risk taking. We found a small and insignificant effect on risk taking (Table 10), consistent with the limited behavioral response to the HIV epidemic documented by (Oster, 2012) and limited HIV risk compensation observed by Friedman (2012) and Kerwin (2014).

45

Table 10: Risk Compensation

Full intervention (F) Reference category Mean of dep var in ref cat Individual controls Village-level controls Obs (Individuals)

(1) Sex acts

(2) Condoms used

0.462 (0.302)

-0.054 (0.115)

P 1.59 Yes Yes 1330

P 0.34 Yes Yes 1325

Note: (1) Number of sex acts, recalled over past 7 days. (2) Number of condoms used, recalled over past 7 days. Survey to meeting attendees 4 months post-intervention. OLS, at the individual level, with a constant and controls: age, gender, married, employed, primary school educated, secondary school educated, has livestock, has a brick house, variables listed in Table 1, village population and distance to each health facility. Standard errors are clustered at the village level, in parentheses. Stars indicate *** p < 0.01, ** p < 0.05, * p < 0.1.

46