HIV Testing and Risky Sexual Behaviour
I thank Frederic Vermeulen, the editor, and three anonymous referees for their comments and suggestions that have greatly improved the manuscript. I also thank Ted Miguel for his insights and constant encouragement and Betty Sadoulet for her patient guidance. In addition, I thank Michael Anderson, Marshall Burke, Steve Buck, William Dow, Fred Finan, Paul Gertler, Susan Godlonton, Kelly Jones, Jeremy Magruder, John Maluccio, Rebecca Thornton and participants at the UC Berkeley Development Seminar, Northeast Universities Development Consortium, ARE Departmental Seminar, Pacific Development Conference, and Midwest International Development Conference for their advice and suggestions. Data for this study come from The Voluntary HIV1 Counselling and Testing study (which) was sponsored by UNAIDS/WHO, AIDSCAP/Family Health International and the Center for AIDS Prevention Studies, University of California, San Francisco. All errors remain my own.
Using a study that randomly assigns HIV testing in two sites in sub-Saharan Africa, I examine the effects of testing on sexual behaviour. Using sexually transmitted infections as markers of risky sex, I find behavioural responses to HIV tests when tests provide unexpected information. Individuals surprised by an HIV-positive (HIV-negative) test increase (decrease) their risky sexual behaviour. I simulate the effects of testing and find under certain conditions, new HIV infections increase when people are tested. The provision of anti-retrovirals for HIV-positive individuals immediately after testing mitigates these effects and leads to decreases in HIV infections in all cases.
HIV testing is regarded as the gateway to prevention and treatment (WHO, 2009). Learning your HIV status is believed to lead to safer sexual behaviour, while the provision of anti-retrovirals (ARVs) requires first identifying infected individuals. Under this premise, universal access to HIV testing has been a key policy response to the HIV/AIDS epidemic. In 19 countries in sub-Saharan Africa (SSA) with reliable data,1 the number of people tested for HIV increased from 4.6 million in 2007, to 8.3 million by 2008 – a yearly growth rate of 80% (WHO, 2009).2 Despite this emphasis, a major question remains: how does HIV testing affect risky sexual behaviour? Since testing serves two purposes (prevention and access to treatment), it can be a desirable policy intervention if at a minimum testing does not increase the number of HIV infections. However, if testing leads some people to undertake riskier sexual behaviour, it could counteract the effect that treatment has on the epidemic.
The two main challenges to empirical research on HIV testing are selection into testing and measuring risky sexual behaviour. Previous studies have relied on non-random variation in who is tested and used self-reported sexual behaviour, which is subject to bias; there is substantial evidence that people under-report their sexual behaviour to conform with social norms (Palen et al., 2008; Minnis et al., 2009).3 The notable exception is Thornton (2008), who uses random assignment of financial incentives for learning one's HIV status and improves on self-reported sexual behaviour by using observed condom purchases as the outcome of interest. Changes in condom purchases, however, may not fully capture changes in actual sexual behaviour.4 My article is the first to resolve both selection and measurement challenges simultaneously by using data from a study that randomly assigns offers of HIV testing and uses biological markers (gonorrhoea and chlamydia infections) as objective proxies of risky sexual behaviour.
Even when selection and measurement issues are resolved, it is not clear how people will respond to testing. Economic models predict asymmetric behavioural responses to HIV testing. Boozer and Philipson (2000) show theoretically that behavioural responses are greatest when HIV tests provide unexpected information. For example, if someone believed she was unlikely to be infected with HIV, an HIV-negative test result will have little effect on this person's behaviour. According to this framework, only people surprised by their test results will change their behaviour. Theoretical models, however, must assume the preferences of individuals. Individuals surprised by HIV-positive tests could reduce their risky sexual behaviour if they are altruistic (i.e. they do not want to infect others); on the other hand, they could increase their risky sexual behaviour if they feel they have ‘nothing to lose’. Ultimately, understanding the effects of HIV testing on risky sexual behaviour requires an empirical approach.
I use data from the Voluntary Counselling and Testing (VCT) Efficacy study conducted in Kenya and Tanzania, which randomly assigned people into HIV testing and followed up with them six months later (The Voluntary HIV-1 Counselling and Testing Efficacy Study Group, 2000). I construct a measure of people's beliefs about their HIV status before getting tested using questions on the baseline survey. To measure risky sexual behaviour, I use biological markers that are not susceptible to self-reporting bias. Data are collected on newly contracted infections of gonorrhoea and chlamydia (henceforward known as ‘sexually transmitted infection’ or ‘STI’) that occur during the study.5 An STI only results from unprotected sex with someone who has an STI and serves as an objective measure of risky sexual behaviour. The random assignment of testing enables me to identify the effect that HIV tests have on sexual behaviour conditioned on prior beliefs of HIV infection.
My findings suggest that HIV tests have the largest effects on risky sexual behaviour when test results provide unexpected information to an individual. I find that people surprised by an HIV-positive test (i.e. those who believed they were at low risk for HIV before testing and learn they are HIV-positive) have a 10.5 percentage point increase in their likelihood of contracting an STI compared to an HIV-positive control group who had similar beliefs of HIV risk but were untested at baseline.6 I interpret this increase in contracting an STI as an indication that those surprised by an HIV-positive test increased their risky sexual behaviour – an unintended consequence of testing. I estimate that these types on average increased their number of new partners by about 2.4 over a six-month time frame. People surprised by an HIV-negative test (i.e. those who believed they were at high risk for HIV before testing and learn they are HIV-negative) have a 5 percentage point decrease in the likelihood of contracting an STI compared to an HIV-negative control group with similar beliefs of HIV risk but were untested at baseline.7 This decrease in the likelihood of contracting an STI suggests that those surprised by HIV-negative tests decrease their risky sexual behaviour. Both of these results indicate that when people make decisions about risky sexual behaviour, self-interests dominate altruistic preferences. People who discover they are HIV-positive no longer have any incentive to practice safe sex (i.e. ‘nothing to lose’), while those who learn they are HIV-negative face greater incentives to avoid risky behaviour. Finally, when HIV test results agree with a person's beliefs of HIV status, the effects of testing on STI likelihood are not statistically different from zero. This is consistent with an economic model where the behavioural responses to HIV tests are greatest when they provide unexpected information.
I use the empirical results described above and combine them with a simple epidemiological model to simulate the short-run effect of rolling out HIV testing in an urban setting. While this exercise inherently requires a set of strong assumptions, and hence the results should be interpreted with caution, it does address an important policy question. I use the distribution of beliefs of HIV risk and actual HIV status from the Demographic Health Surveys in Kenya, Mozambique and Zambia – all three countries faced with a generalised HIV epidemic. I find that in the cases of Kenya and Zambia, testing leads to declines in new infections, while testing leads to an increase in infections in Mozambique. However, when ARVs are provided at an earlier stage in the infection, testing leads to large reductions in HIV infections in all three countries. Since ARVs greatly reduce the infectivity of HIV-positive individuals, the aggressive provision of ARVs can mitigate the risks posed by HIV-positive individuals who increase their risky sexual behaviour after testing.8
This study makes several contributions. It is the first work that provides empirical evidence that individuals who discover they are HIV-positive through testing increase their risky sexual behaviour. While some caution is warranted given the small number of HIV-positives in the study, this finding is at odds with conventional wisdom that those who learn they are HIV-positive will take steps to prevent infecting others (Potts et al., 2008; Gersovitz, 2010). My ability to simultaneously resolve the selection and measurement challenges is the key methodological contribution, an issue unresolved in the few existing sub-Saharan studies that have exogenous variation in who is tested (The Voluntary HIV-1 Counselling and Testing Efficacy Study Group, 2000; Thornton, 2008). As a result, these findings have important policy implications. The first is that it lends more support to a policy of greater access of ARVs to HIV-positive individuals. Current WHO guidelines recommend the provision of ARVs at a later course of the infection, when the immune system becomes compromised as measured by CD4+ cell counts that measure the strength of the immune system.9 Treating individuals at an earlier stage of the infection not only directly benefits the patient but can mitigate any adverse behavioural outcomes due to testing. Thus when testing is combined with greater access to ARVs, multiple benefits are accrued: testing leads those who overestimate their risk to reduce their risky sexual behaviour, and testing identifies HIV-positive individuals who can be led into ARV treatment that reduces both their own mortality rates and their risk of infecting others.10
This work also contributes to the emerging empirical literature on the important role that information and beliefs play on an individual's behaviour (Manski, 2004; Delavande et al., 2012). Dupas (2010) finds that providing teenage girls in Kenya with the relative risk of HIV infection by age leads to a decrease in unprotected sex with older men; the implicit assumption is that these girls did not know what these risks actually were. Both Nguyen (2008) and Jensen (2010) show that providing information on the returns to schooling leads to increased education. Both authors attribute this behavioural response to low perceived returns of schooling before information is provided. Goldstein et al. (2013) find that pregnant women with high expectations of being HIV-positive increase their uptake of neonatal services when tested for HIV. This article shows that HIV testing can have large effects on sexual behaviour if the test results are unexpected.
Finally, this work contributes to the growing literature that examines the unintended effects of policies designed to improve health outcomes. In western Kenya, Duflo et al. (2012) examine the effects of two school interventions designed to improve educational and health outcomes:
- free school uniforms and
- an HIV/AIDS curriculum focused on abstinence until marriage.
They find that while providing school uniforms increased educational attainment and reduced early fertility, the additional provision of the HIV/AIDS curriculum mitigated both these positive effects. Kohler and Thornton (2012) find that conditioning cash transfers to men in Malawi for maintaining their HIV-negative status did not change HIV infection rates but did increase the men's risky sexual behaviour after receiving the cash rewards. This article shows that HIV testing which is intended to prevent the spread of HIV can lead to additional infections since those who discover they are HIV-positive optimise their individual behaviour (i.e. increase their number of partners) but do not take into account the negative externalities they are generating by this behaviour (i.e. infecting others).
The article is structured as follows. Section 'Conceptual Framework' outlines a simple model which shows that theoretically HIV testing has ambiguous effects on behaviour. Section 'Data' describes the features of the data. Section 'Empirical Analysis' provides the empirical strategy and presents the main results. Section 'Discussion' details how my findings contrast to those from the original paper published in the Lancet by The Voluntary HIV-1 Counselling and Testing Efficacy Study Group, (2000) which found that testing leads to reductions in unprotected sex. I discuss why my article and the original study arrive at different conclusions using the same data set. In addition, I propose ways that my findings can be reconciled with Thornton (2008), Delavande and Kohler (2012) and de Paula et al. (forthcoming), all of whom find that HIV-positive tests (or increases in beliefs of being HIV-positive) lead to decreases in risky sexual behaviour. Section 'Short-Term Effect of Testing on the Epidemic' does a simple simulation showing the effects of testing on new HIV infections, and Section 'Conclusion and Policy Implications' concludes.
1 Conceptual Framework
In this Section, I present a simple model to show:
- the role that beliefs of HIV infection play in determining risky sexual behaviour and
- the effects of HIV testing on behaviour are, a priori, ambiguous.
This model is influenced by Boozer and Philipson (2000) and similar to de Paula et al., (forthcoming). The main differences are that this model does not explicitly show how beliefs of HIV status are updated as de Paula et al. (forthcoming) do, and shows that testing has an ambiguous effect on individual sexual behaviour which differs from Boozer and Philipson (2000).
An individual chooses a level of risky sexual behaviour j which generates u(j) utility. While risky sex can take multiple forms, in this model j represents the number of sexual partners. The level of risky sexual behaviour is a function of beliefs of HIV infection, which I denote as π ∈ [0,1]. Those who believe they are at higher risk for HIV have higher values of π, where π = 1 for those who are certain they are HIV-positive and π = 0 for those who are certain they are HIV-negative. Each time an individual partners sexually with someone, they face a probability of HIV-infection λ(β,W) which is a function of β (HIV transmission rate) and W (prevalence of HIV).11, 12 Finally, c is the disutility that comes from knowing that you are HIV-positive. I assume u(j) is increasing in j and concave. Individuals then must choose j that maximises the following utility function:
The first-order condition equates the marginal benefit of risky sexual behaviour with the marginal cost:
where uj is the partial derivative of u(j) with respect to j. As beliefs of being HIV-positive increase, the marginal cost of risky sexual behaviour decreases, which leads individuals to choose higher levels of risky sex ( j ). From this model, it is clear that beliefs of HIV infection have an important role when an individual chooses a level of risky sexual behaviour.
I now introduce altruism to the model which takes the form of a discount to the utility one receives from risky sex:
where A(π) ∈ [0,1] is a function of beliefs of HIV infection and serves to discount the marginal benefit of risky sex. I assume that Aπ < 0 or that as beliefs increase, a greater discount is applied to the utility of risky sex.
How does risky sexual behaviour respond to HIV testing? We can think of HIV tests as shocks to beliefs (π), where someone surprised by an HIV-positive (HIV-negative) test has Δπ > 0 (Δπ < 0). When an HIV test confirms an individual's priors , beliefs are unchanged (Δπ = 0).
The comparative statics show how behaviour (j) responds to a change in beliefs (π):
Since by concavity, u′′(j) < 0, and given a non-zero HIV transmission rate (λ(β,W) > 0), the sign of ∂j/∂π depends on u′(j)Aπ + λ(B,W)c. When is large, or when the utility from risky sex is heavily discounted when beliefs increase (i.e. altruistic preferences), then u′(j)A′(π) + λ(B,W)c < 0 and risky sexual behaviour decreases as beliefs increase (∂j/∂π < 0). When is small, or when the utility from risky sex is not greatly discounted when beliefs increase, then u′(j)A′(π) + λ(B,W )c > 0 and people increase their risky sexual behaviour as their beliefs increase (∂j/∂π > 0). If altruistic preferences are not known before testing, then the ex ante effects of HIV testing on risky sexual behaviour are ambiguous. Individuals who receive HIV-positive test results and have strong altruistic preferences will decrease their risky sexual behaviour, while those who care only about their own interests will increase their risky sexual behaviour.
To summarise, the model shows the role that beliefs of HIV infection play when an individual chooses a level of risky sexual behaviour. HIV testing serves as a shock to these beliefs; an HIV-positive test increases these beliefs while an HIV-negative test decreases beliefs of HIV infection. Without altruism, an increase in the beliefs of HIV infection decreases the marginal cost of risky sex and increases risky sexual behaviour. When altruism is introduced, the effects of HIV testing on risky sexual behaviour are ambiguous.
The data are from the HIV Voluntary Counselling and Testing Efficacy study conducted in 1995–8 (The Voluntary HIV-1 Counselling and Testing Efficacy Study Group, 2000). The study was designed to assess whether HIV testing and counselling is effective at reducing risky sexual behaviour. My analysis uses data from the study sites in Nairobi, Kenya and Dar Es Salaam, Tanzania.13 In both places, a single study site was placed in/near a health centre. These sites enrolled, surveyed and tested participants. A combination of media (flyers, radio and TV advertisements) and recruiters were used to recruit study participants. The initial sample consists of approximately 2,900 people who were seeking HIV-related services, with 1/3 of them enrolling as a couple (see Kamenga et al., 2000 for an in-depth description of the study's design and methods).
A baseline survey was conducted and urine samples were taken of all individuals. These urine samples were frozen and used during the six-month follow-up survey. Study participants were then classified as either individuals or couples. They were then randomly assigned into either a treatment or control arm (see online Appendix A.1 for figure of study design). People assigned into the treatment arm were offered counselling and an HIV test, of which 93% accepted the test.14 Test results were available two weeks after testing; 78% of those in the treatment arm returned to the clinic to receive their HIV test results. Participants enrolled as a couple were strongly encouraged to share their HIV test results with each other. People in the control arm watched a 15-minute video that described ways to prevent HIV infection and had a question and answer session with a health information officer. As the treatment and controls arms differ not only due to HIV testing, but also due to different information interventions (counselling in the treatment arm and a video in the control arm), there may be differences between arms in what people learn about HIV. I compare changes in HIV/AIDS knowledge and awareness between the treatment and control arms during the study and find no differences (see online Appendix A.2).
Six months after the baseline, a follow-up survey was given. Everyone who participated in the follow-up round was resurveyed, asked to give a urine sample and offered an HIV test. The urine sample was tested for two (STIs): gonorrhoea and chlamydia. For people who tested positive for an STI, their urine samples from baseline were unfrozen and tested for an STI. By doing this, we can determine STI incidence, which allows us to differentiate STIs contracted during the study from those that were pre-existing. Those in the control arm were offered HIV testing and counselling; 84% accepted an HIV test.15 Although the acceptance rate for HIV testing between the treatment (93%) and control arms (84%) is different, there do not appear to be any differences in observed characteristics between those accepting an HIV test in the treatment and control arms (see online Appendix A.3).
Baseline summary statistics for the treatment and control group are in Table 1. Demographic data are presented in rows 1–9 and relationship status is in rows 10–13; the average age is 28 and 39% of study participants are married. Under the HIV/AIDS section (rows 14–17), we see awareness of how HIV is transmitted is high (row 14),16 but few have been tested (row 16). Self-reported sexual activity during the two months prior to the baseline survey is reported in rows 18–26. Slightly over 20% of participants had two or more partners (row 18) and about 12% have engaged in commercial sex (row 20).17 It should be noted that the urine samples at baseline were not all tested for STIs; hence it is impossible to compare baseline STI prevalence between the treatment and control arms. As noted above, urine samples at baseline were frozen and only individuals who test positive for an STI at the follow-up had their urine samples at baseline tested.18 Overall, based on observed characteristics, the treatment and control groups appear balanced.
Table 1. Summary Statistics
| Demographics |
|8.||Tap water in home||0.54||0.54||0.96|
|9.||Electricity in home||0.44||0.45||0.49|
| Relationship status |
|10.||Enrolled as couple||0.33||0.32||0.90|
|13.||Number of living children||1.45||1.48||0.65|
| HIV/AIDS |
|14.||HIV/AIDS knowledge (out of 12)||9.73||9.76||0.75|
|17.||Baseline HIV+||0.20|| || |
| Sexual activity |
|18.||Two or more partners||0.22||0.21||0.70|
|19.||Unprotected sex with|| || || |
|23.||Episodes unprotected sex with|| || || |
| ||Sample size||1,477||1,465|| |
Baseline HIV tests for the treatment group (column 1, row 17) reveal HIV prevalence to be at 20%, this compares to HIV-prevalence of 16% in more recent Demographic and Health Surveys (DHS) of urban areas in Kenya, Mozambique and Zambia.19 Given the main intervention (treatment) of the VCT Efficacy study is to offer free HIV testing, the study population is sexually active individuals seeking HIV testing services. As the policy of universal access to HIV testing is focused on expanding the number of sites where HIV tests can be obtained, this is a relevant population.
Attrition in the study is high, where follow-up rates in the testing and control arms are 69% and 66% in each respective arm. Fortunately, there is no evidence of differential attrition between the testing and control arms; there are very few statistically significant differences across baseline characteristics between attriters in the testing and control arms (see online Appendix A.4 for further discussion). I now discuss three important aspects of how I use the data:
- measuring risky sexual behaviour;
- identifying people's HIV status; and
- measuring people's beliefs about HIV infection.
2.1 Measuring Sexual Behaviour
Sexual behaviour is difficult to measure because it is unobserved and, due to its sensitive nature, self-reports of sexual behaviour are subject to a high degree of social desirability bias. This may lead survey respondents to under-report risky sexual behaviour such as unprotected sex and multiple partnerships. The original study planners recognised this issue and designed the survey and interview ‘to limit reporting bias by adding preambles to sensitive questions shown to increase reports of risk behaviour’ (The Voluntary HIV-1 Counselling and Testing Efficacy Study Group, 2000). However, even when surveys and interviews are carefully designed, there is still evidence of significant under-reporting of sexually activity. In two well-cited studies in the health literature, Allen et al. (2003) and Minnis et al. (2009) used carefully designed surveys to collect self-reported sexual behaviour as well as biomarkers from samples in SSA.20 Biomarkers range from pregnancies, STIs and residual semen, which all act as objective proxies for risky sexual behaviour, as the likelihood of a biomarker is increasing in both acts of unprotected sex and number of partners. Both studies found significant under-reporting of sexual behaviour, with Allen et al. (2003) finding that 32% of pregnancies and HIV transmissions were detected among couples who self-report always using condoms, while Minnis et al. (2009) finding that 48% of those with biomarkers self-report having no unprotected sex.
In this article, the incidence of gonorrhoea and chlamydia infections is used as measures of risky sexual behaviour. The primary means of transmission for both infections is unprotected sexual contact and non-sexual transmission is extremely rare (Neinstein et al., 1984). Both infections are sensitive to risky sexual activity: transmission rates are between 20% and 80% per unprotected sexual act with an infected individual (Chen et al., 2008). Going forward, STIs will refer specifically to gonorrhoea and chlamydia infections (and not HIV).
2.2 HIV Status
The HIV status of everyone in the treatment arm that accepts an HIV test is known at baseline. However, the HIV status of those in the control group at baseline are unknown as they were not offered testing until the six-month follow-up. This is problematic, as I want to compare HIV-positive (negative) individuals in the treatment arm to those in the control arm. To create a counterfactual group for testing, I use the HIV test results from the six-month follow-up for the control group. For the control group, I assume that an individual's HIV test results at the six-month follow-up would have been the same result at baseline. Clearly, those who are HIV-negative at six months were also negative at baseline. For people who test HIV-positive at six months, I assume that all of these individuals were positive at baseline as well. This assumption relies on evidence which suggests that HIV is not easily transmitted, with estimated transmission rates of approximately 0.0015–0.0007 per coital act when your partner has an established HIV infection (Cohen and Pilcher, 2005).21, 22
How do new HIV infections that occur between baseline and the six-month follow-up in the control group affect the estimates of HIV testing on behaviour? Let Yi be risky sexual behaviour, Ti indicate random assignment into testing, HIVi be HIV status and subscript i denote an individual. The average effect of an HIV-negative test on risky sexual behaviour is:
As HIV status for the control group is not observed until six-month follow-up, I estimate:
where (HIV = 0)* is the HIV status at the six-month follow-up. If any individuals in the control group became HIV-positive during the course of the study, they would not be included in the HIV-negative control group, even though they were HIV-negative at baseline. Thus the average risky sexual behaviour of the true counterfactual group will be greater than the behaviour in the control arm:
which results in or that estimates of the effect of an HIV-negative test on risky sexual behaviour will be biased upwards.
What is the effect of using HIV-positive tests at the six-month follow-up to infer baseline status? The average effect of an HIV-positive test on behaviour is:
Again, using test results at the six-month follow-up generates this effect:
where (HIV = 1)* indicates an HIV-positive test result at the six-month follow-up. This group will consist of people who were HIV-positive at baseline and those who became infected during the course of the study due to risky sexual behaviour. The sexual behaviour for this control group then will be on average more risky than the behaviour for those who were HIV-positive at baseline:
which results in or that the estimated effect of a HIV-positive test will be biased downwards.
To conclude, my estimates for the effects of HIV-negative tests on risky sexual behaviour will be biased upwards and for HIV-positive tests the bias will be downwards.
As my main results show that those surprised by an HIV-positive test increase their risky sexual behaviour, this estimate becomes a lower bound for the true effect of HIV-positive tests on risky sexual behaviour. Correspondingly, my main results also show that those surprised by an HIV-negative test decrease their behaviour and, thus, this estimate serves as an upper bound to the effect of HIV-negative tests on risky sexual behaviour.
2.3 Beliefs of HIV Status
Questions about an individual's HIV status are sensitive and respondents who believe they are HIV-positive face strong incentives not to reveal their true beliefs. There are numerous cases documenting that those who reveal they are HIV-positive are subject to employment discrimination, physical violence (including murder) and social stigma (Simbayi et al., 2007).23 Direct questions, such as, ‘What is the likelihood you are HIV positive?’ may lead to biased responses. The original study anticipated this and used a set of four questions designed to measure perceived HIV risk.24 The questions are as follows:
|A||What are the chances that you will get the AIDS virus?|
|B||What are the chances that you already have the AIDS virus?|
|C||How worried are you that you will get the AIDS virus?|
|D||How worried are you that you already have the AIDS virus?|
The responses for the questions use the following Likert scale:25
|Almost certainly will not happen||Not at all or hardly worried||1|
|It could happen||A little bit worried||2|
|It probably will happen||Quite a bit worried||3|
|It almost certainly will happen||Extremely worried||4|
Question B is the most direct means of measuring beliefs of HIV infection; however those who believe they are infected may bias their responses downwards.26 The use of questions A, C and D help resolve this problem. These additional questions are designed to measure perceived HIV risk (Smith and Watkins, 2004; Lauby et al., 2006), and slight changes in language may elicit more accurate responses. For example, someone who thinks they are HIV-positive might report having a low chance of being infected but may report having a high level of worry. To utilise the information from all four questions, I take the average response to questions A–D. The median of all the average responses is 2, which I use to divide the sample into a high and low-belief group (see online Appendix A.8 for the distribution of beliefs). Those with an average response of between 1 and 2 are classified as having low beliefs, while those with an average response greater than 2 are classified as having a high belief of HIV infection.27 I create a low and high-belief group using just the responses from question B as well.28 In online Appendix A.8, I present evidence that question B may be leading study participants to under-report their beliefs of HIV-status.
To test whether utilising data from all four questions provides a better belief measure than question B, I see how accurate each measure is with regard to actual HIV-status. If the belief measure is accurate, those with higher beliefs of being infected with HIV should have a higher probability of actually being HIV-positive. Regressing HIV-status (an indicator if HIV-positive) on both belief measures, I find that the one using all four questions has a much stronger correlation with actual HIV status compared to the measure only using question B (Table 2, columns 1–2). This suggests there is additional information about an individual's HIV status that is captured using all four questions that is not when using the more direct question about beliefs. Replacing HIV-status with self-reported sexual behaviour, I find that the belief measure using question B is negatively correlated with sexual activity, while the belief measuring using all four questions is strongly correlated with the past number of sexual partners (columns 3–6). This evidence again suggests that when individuals are directly asked about their HIV-status, there is a tendency to under-report their beliefs, while data using all four questions may provide a more accurate measure of underlying beliefs.
Table 2. Beliefs of HIV Infection
|High belief B||0.004||0.005||−0.054*||−0.024||−0.119||−0.063||−0.023||−0.013|
|High prior beliefs||0.041*||0.050**||0.023||0.037||0.404***||0.470***||0.059***||0.057***|
|(All 4 questions)||(0.023)||(0.025)||(0.023)||(0.023)||(0.135)||(0.170)||(0.016)||(0.018)|
Another test of the belief measures is to see which one best predicts future sexual behaviour.29 From the simple model of sexual behaviour, individuals take into account their beliefs of HIV-status when making decisions about their sexual behaviour (Section 'Data', equation (4)). The higher those beliefs, the lower the marginal cost of engaging in risky sexual behaviour. Belief measures that rely on biased responses may not be predictive of sexual behaviour. Using STI incidence as an objective measure of risky sexual behaviour, I regress STI incidence on both belief measures.30 The belief measure incorporating all four questions strongly predicts STI incidence (Table 2, column 7) and this relationship holds after including characteristics such as age, gender and education that may be correlated with beliefs (column 8). The belief measure that relies solely on question B does not significantly predict STI incidence.
Overall, it appears that the preferred belief measure (using all four questions) has a stronger correlation to actual HIV-status and past sexual behaviour as well as better predicting future sexual behaviour when compared to a belief measure relying on a single direct question about HIV-status.31 It should be stressed that the results in this Section should not be interpreted as causal. What this Section does is provide evidence that the preferred belief measure (using all four questions) is a valid measure of beliefs of HIV infection.
3 Empirical Analysis
3.1 Identification Strategy
This article has argued that risky sexual behaviour is a function of beliefs of HIV infection, and HIV tests may substantially change these beliefs depending on one's priors. Using measures of prior beliefs described in the previous Section, there are two groups where HIV tests should substantially change beliefs:
- low priors receiving HIV-positive tests and
- high priors receiving HIV-negative tests.
In these two groups, HIV tests should have large effects on risky sexual behaviour. Testing should have smaller changes in beliefs and behaviour in these other two groups:
- low priors receiving HIV-negative tests and
- high priors receiving HIV-positive tests.
The following Table presents the four groups and the predictions of the effects of testing in each group.
|Low prior beliefs||Tests have little effect on beliefs or behaviour||Tests increase beliefs ⇒ Change in behaviour|
|High prior beliefs||Tests decrease beliefs ⇒ Change in behaviour||Tests have little effect on beliefs or behaviour|
The goal is to identify the effect of HIV testing conditional on prior beliefs. Since testing is randomly assigned, I can compare STI outcomes between the control and testing arms in each of the four groups in the Table above to identify the effects of HIV testing. Before proceeding to the formal estimation, I present simple differences in STI incidence between the control and testing arms in Table 3. A simple comparison of STI incidence between the control and testing arms shows no statistically significant differences (panel (a)).32 Conditioning by HIV-status, there is some evidence that those receiving HIV-negative tests are reducing their risky sexual behaviour given the reduction in STI incidence (panel(b), row 2), while those receiving HIV-positive tests show slight increases in risky behaviour (panel (b), row 3), however, neither is significant. However, as discussed in the conceptual framework, when I condition on both HIV-status and prior beliefs, there appear to be significant differences in the groups where HIV-test results come as a surprise. Those surprised by an HIV-negative test (HIV-negative/High Priors) have significant decreases in STI incidence (panel (c), row 5), while those surprised by an HIV-positive test (HIV-positive/Low Priors) see significant increases in STI incidence (panel (c), row 6).33
Table 3. STI Incidence (Risky Sexual Behaviour) at Follow-Up
| Panel B: condition on HIV status |
| Panel C: conditioning on status and priors |
3.2 Estimating Equation and Results
The estimating equation is a linear probability model:
where STIij = 1 if individual i in country j contracts an STI during the study, Testi indicates assignment into the HIV testing arm, High Priorsi indicates if the individual has high prior beliefs, HIVi = 1 for those who are HIV-positive, and Couplei indicates if the individual enrolled in the study with his/her partner. The vector Ii includes all the interactions of Testi, High Priorsi, HIVi, Couplei that are not explicitly specified, is a vector of individual-level characteristics, and γj is a country fixed effect.
While assignment into the testing arm is randomly assigned, not everyone in the testing arm receives their test results (there is a two-week delay between testing and availability of results). I therefore employ intent to treat estimators (instrumenting for learning one's test result using the random assignment of testing generates similar results, see online Appendix A.7). The random assignment of testing implies that allowing the OLS estimate of β1 to be unbiased. As prior beliefs and HIV status were determined before testing occurred they are not affected by the intervention. Therefore, β5 estimates the causal impact of testing conditioned on high prior beliefs and β6 is the causal impact of testing conditioned on being HIV-positive.
Using the predictions from the previous Table, we should expect β1 = 0 (low priors receiving HIV- test), β1 + β6 ≠ 0 (low priors receiving HIV+ test), β1 + β5 ≠ 0 (high priors receiving HIV- test), and β1 + β5 + β6 + β7 = 0 (high priors receiving HIV+ test).
3.2.1 Biomarker results
Table 4 presents OLS estimates of (11).34 STI incidence across the whole sample is 3.91%. Column 1 includes each covariate of interest, while columns 2 and 3 include the full set of interactions. Column 3 also includes a set of controls such as gender, age, education, religion, marital status, number of children, assets, the language of the survey, and both interviewer and country fixed effects. The upper panel of Table 4 presents estimates for each of the individual covariates of interest. As the effect of HIV-testing depends on prior beliefs, the lower panel of Table 4 shows the pertinent linear combinations with the standard errors adjusted for the covariance between the variables.
Table 4. Effect of HIV Testing on STI Incidence (Risky Sexual Behaviour) – Dependent Variable: STI Incidence (Mean = 0.039)
|2. High prior beliefs||0.023||0.056||0.053|
|4. High prior × HIV|| ||0.046||0.036|
|5. Test × high prior|| ||−0.051||−0.048|
|6. Test × HIV|| ||0.103||0.096|
|7. Test × high prior × HIV|| ||−0.096||−0.094|
| Linear combinations: effect of HIV tests by prior beliefs |
|HIV- test on low prior group|
|8. Test|| ||0.002||−0.001|
|HIV+ test on low prior group|
|9. Test + (Test × HIV+)|| ||0.105||0.095|
|HIV− test on high prior group|
|10. Test + (Test × High)|| ||−0.05||−0.049|
|HIV+ test on high prior group|
|11. Test + (Test × HIV+) + (Test × High) + (Test × High × HIV+)|| ||−0.043||−0.047|
I estimate the effects of HIV-positive and HIV-negative tests by each prior belief group. Individuals with low prior beliefs who receive HIV-negative tests have little change in STI incidence (row 8). The point estimate across both specifications is virtually zero, and standard errors are relatively small. This finding is consistent with a model where HIV-negative tests do not provide unexpected information to those with low prior beliefs which results in little change in behaviour.
To examine the effect of an HIV-positive test on individuals with low prior beliefs, I estimate the linear combination Test + (Test × HIV+) (row 9).35 The effect is large and statistically significant; those with low priors have about a 10.5 percentage point increase in STI incidence after receiving an HIV-positive test. Given that the STI incidence for the low prior/HIV-positive control group is 1.06%, this represents a more than ninefold increase in STI incidence after an HIV-positive test. While the magnitude of the effect of testing appears large, in Section 'Short-Term Effect of Testing on the Epidemic', I estimate that this effect could result if individuals added about two partners in a six-month time frame. Overall, this result is consistent with a model where people with low prior beliefs substantially increase them after receiving an HIV-positive test. The large increase in beliefs leads to greater risky sexual behaviour. This suggests that self-interests have a larger effect on sexual behaviour than altruism; as people revise their beliefs upwards the marginal cost of risky sex decreases and they face far less incentive to engage in safe sex.
Now I turn to the group with high prior beliefs of HIV infection. The effect of an HIV-negative test for individuals with high priors is the linear combination Test + (Test × High) (row 10). STI incidence decreases by 5 percentage points after an HIV-negative test. Given the mean STI rate of the high prior belief control group is 5.90%, it appears testing reduces STI incidence by 84%. Again, while the magnitude of this effect appears large, in a later Section I estimated that this effect could result if individuals decrease their partnerships by one in a six-month time frame. Those who substantially decrease their beliefs of HIV infection appear to be reducing their risky sexual behaviour. This is consistent with people having greater incentives to protect themselves when they learn they are uninfected. Finally, the effect of HIV-positive tests on high prior types is the linear combination Test + (Test × HIV) + (Test × High) + (Test × High × HIV) (row 11). There is no statistically significant effect on STI incidence but, given the wide confidence intervals, inference warrants caution.
As a robustness check, I vary how the low and high priors group are specified, and find that the main results presented here are consistent across several different specifications including gender (see online Appendices A.7, A.8, A.9).
3.2.2 Self-reported sexual behaviour
Before turning to the self-reported sexual behaviour, an important caveat is in order. Even if social desirability bias leads to the under-reporting of risky sexual behaviour, one might still compare average behaviour in the treatment and control arms assuming this bias is similar in both arms. Ex ante, as testing is randomly assigned, this may be valid. However, in this situation, the treatment itself, HIV testing, can affect the reporting of sexual behaviour. Those assigned into the testing arm have counselling sessions before and after testing. During these sessions, individuals review their sexual behaviour with a trained counsellor and a plan is developed to reduce risky sexual behaviour. At the end of these sessions, the counsellor elicits a commitment from the individual to reduce risky sexual behaviour (CAPS, 1995).36 In some sense, those in the testing arm are being told the ‘correct’ responses to follow-up surveys: namely that individuals after being tested should reduce behaviour such as unprotected sex and casual partnerships. This is especially true for those receiving HIV-positive tests; counsellors are trained to emphasise to HIV-positive individuals that they need to reduce their risky sexual behaviour to protect both their partners and themselves from reinfection (CAPS, 2003). Evidence of differential under-reporting in the control versus treatment arm can be found in the sample of individuals who contract an STI (‘STI Incidence’) during the study. This sample thus consists of individuals where we know sexual activity occurred for each of them. I compare the self-reports of being sexually active between the testing and control arms in Table 5. Of those who contracted an STI, 93% of those in the control arm reported being sexually active during the study compared to only 78% in the testing arm (row 4); in other words, individuals in the testing arm are more likely to under-report their sexual activity. This level of under-reporting is more severe with those receiving HIV-positive tests. Of the HIV-positives who contract an STI during the study, 92% of those in the control arm report being sexually active, while only 69% in the testing arm report any sexual activity (row 6).37
Table 5. Sexually Active During Study (Conditioned on Contracting an STI)
| Panel (a): sexually active at baseline |
| Panel (b): sexually active during study |
Given this evidence on differential under-reporting, I turn to the self-reported sexual behaviour cautiously. I look at three self-reported outcomes: being sexually active, number of partners and unprotected sex with a non-primary partner (NPP).38 I first look at the effects of testing itself on these outcomes (Table 6, panel (a)). Replicating the main result from the original study, testing leads to reductions in reporting unprotected sex (row 1; column 4), however, there is no accompanying change in STI incidence (column 1). Examining the sample by HIV-status (panel (b)), I find that those receiving an HIV-negative test also show reductions in reporting unprotected sex, which is accompanied by a marginally significant reduction in STI incidence (row 2), while those receiving HIV-positive tests report reductions in both sexual activity and unprotected sex but no statistically significant change in STI incidence (row 3). In fact, for those receiving an HIV-positive test, the point estimate for STI incidence is positive. These two results suggest that those receiving an HIV-positive test may be under-reporting risky sexual behaviour at a greater rate than those receiving an HIV-negative test.
Table 6. Effect of HIV Testing on Self-reported Sexual Behaviour
|Panel (a): overall sample (N = 1961)|
|Mean dependent variable||0.044||0.784||1.227||0.218|
| Panel (b): by HIV status |
|HIV− sample (N = 1589)|
|2. HIV− test||−0.022*||0.021||−0.121||−0.069***|
|Mean dependent variable||0.039||0.784||1.161||0.220|
|HIV+ sample (N = 372)|
|3. HIV+ test||0.031||−0.126**||−0.368||−0.129***|
|Mean dependent variable||0.067||0.783||1.511||0.211|
| Panel (c): by HIV status and priors |
|HIV+ and low prior sample (N = 188)|
|4. HIV+ test||0.131***||−0.155*||−0.352**||−0.172***|
|Mean dependent variable||0.011||0.787||1.106||0.213|
|HIV− and high prior sample (N = 687)|
|5. HIV− test||−0.043**||−0.029||0.035||−0.083**|
|Mean dependent variable||0.059||0.820||1.183||0.280|
Finally, I turn to the two groups where HIV tests come as a surprise: low priors/HIV+ and high priors/HIV- (panel (c)). These are the two groups where we find results using the STI outcomes. For those surprised by an HIV+ test, there are decreases in all three measures of self-reported sexual behaviour but large increases in STI incidence (row 4). The magnitude of the decreases in self-reported behaviour is similar to the entire HIV+ sample, and is consistent with the idea that those receiving HIV+ test results are more likely to under-report their sexual behaviour. This may explain the discrepancy between STI incidence and self-reported sexual behaviour for those surprised by HIV+ tests. Those surprised by an HIV-negative test (row 5) report decreases in unprotected sex, which accompanies decreases in STI incidence, and is similar to the overall HIV-sample. (Results for the two groups not surprised by their test results are presented in online Appendix A.10.)
Overall, what are we to make of these self-reported behaviour? The results suggest that those receiving HIV-positive tests, especially those surprised by them, may have a greater tendency to under-report their risky sexual behaviour. I also examine whether these types forgo STI treatment, which might help explain this discrepancy but find no evidence that this is occurring (see online Appendix A.10). In addition, it does not appear that these types are more likely to match with other risky partners who might also be infected with HIV (see online Appendix A.10 for more details). The self-reported behaviour of those surprised by HIV-negative tests might be more accurate given the accompanying decrease in STI incidence. Given the conflicting results between STI outcomes and the self-reported sexual behaviour for those receiving HIV-positive tests, in Section 'Short-Term Effect of Testing on the Epidemic', I use a simple epidemiological model of STI & HIV transmission to estimate changes in risky sexual behaviour based on the STI results. These estimated changes in sexual behaviour will then be used to calculate the change in HIV infections as a result of testing.
4.1 Comparison to the Original Study
The original study by The Voluntary HIV-1 Counselling and Testing Efficacy Study Group, (2000) was published in the Lancet, a leading medical journal, and is an influential study as it was the first RCT of HIV testing in SSA.39 The original study's main finding is that those assigned to the testing arm self-report lower rates of unprotected sex with casual (non-primary) partners compared to the control arm but there is no corresponding decrease in STI incidence. These results contrast to the ones in this article which show testing does lead to significant changes in STI incidence, while the self-reported sexual behaviour suggest the opposite. The sharpest contrast between the two studies is that I find evidence that HIV-positive tests lead to increases in risky sexual behaviour. Given that both studies use the same data, what explains the different results? There are three important distinctions. The original study:
- relies on self-reported behaviour,
- does not use true counterfactuals for HIV-positive (negative) tests and
- does not take into account prior beliefs of HIV-status before testing.
The first distinction involves the use of self-reported sexual behaviour. As discussed previously, not only is there concern about social desirability bias when self-reporting sexual behaviour but there is also evidence of differential bias between the testing and control arms. Individuals assigned into the testing arm may be more likely to under-report risky sexual activity (see subsection 'Self-reported sexual behaviour'). This helps explains why the original study found that testing leads to less reported unprotected sex but no actual decrease in STI incidence. The discrepancy between self-reports and STI incidence is particularly striking when the original study examines the effects of HIV-positive tests. The original study finds that those receiving HIV-positive tests report lower rates of unprotected sex compared to those receiving HIV-negative tests. What is puzzling is that those receiving HIV-positive tests also have greater rates of STI incidence compared to the HIV-negative group. What explains this discrepancy? This leads to the second important distinction between the original study and this article.
The original study does not use true counterfactuals when examining the effects of HIV-status on behaviour. Comparing those receiving HIV-positive tests to those receiving HIV-negative tests is not a valid comparison. A contribution of this aticle is identifying a plausible counterfactual group: I compare those receiving an HIV-positive test to HIV-positives in the control arm who are not tested (and similarly with HIV-negative tests; see subsection 'HIV Status'). I thus can account for the possibility that HIV-positive types have compromised immune systems leading to greater risk of STI infection. For example, even if HIV-positives report safer sexual behaviour, their compromised immune systems might put them at higher risk of contracting STIs. However, even when accounting for one's HIV-positive status, I find that those in the testing arm still report large reductions in unprotected sex but no corresponding change in STI incidence (Table 6, row 3). This suggests that those receiving HIV-positive tests are also under-reporting their sexual behaviour.
Finally, the original study does not take into account one's prior beliefs of HIV-status before testing. The original study simply compares those in the testing arm to the control group and finds no significant effect on STI incidence. However, if there are asymmetric responses to HIV-testing, where those surprised by HIV-negative tests act differently to those surprised by HIV-positive test, then it is possible that the average effect of testing may be close to zero (Table 3, row 1). This article accounts for both prior beliefs and HIV-status and finds that simply comparing the testing to control arms masks important heterogeneity; there are large differential responses to testing that depend on both priors and the result of the test.
4.2 Are Results Consistent with Previous Work?
Two main findings of this article are:
- the behavioural response to HIV testing depends on prior beliefs of HIV status; and
- individuals surprised by HIV-positive tests increase their risky sexual behaviour.
The second result contrasts with the findings of Thornton (2008), de Paula et al. (forthcoming) and Delavande and Kohler (2012), all of whom find that when individuals learn they are at higher risk for HIV they reduce their risky sexual behaviour. Specifically, Thornton (2008) finds that individuals receiving HIV-positive tests buy more condoms in a follow-up survey, de Paula et al. (forthcoming) find that married men report fewer extramarital affairs when their beliefs of being HIV-positive increase and Delavande and Kohler (2012) show that individuals two years after receiving an HIV-positive test report increases in condom usage. All three studies sample from the Malawi Diffusion and Ideational Change Project (MDICP),40 which is a panel survey intended to be representative of the rural Malawi population. Two factors might explain the differences in results. The first is that two of the studies (Delavande and Kohler, 2012; de Paula et al., forthcoming) use self-reported sexual behaviour which may be subject to bias.41 The second is that the degree of altruism towards others determines the behavioural response to an HIV-positive test (see Section 'Conceptual Framework'). It may be that previous studies had a higher proportion of subjects with altruistic preferences. One measure of altruism might be marriage. If married individuals are more likely to be altruistic towards another person (their spouse) than unmarried types, then difference in marriage rates and altruism may reconcile the results. While de Paula et al. (forthcoming) focus exclusively on married men, both Thornton (2008) and Delavande and Kohler (2012) use a sample where over 70% are married. This contrasts to this article, where less than 40% of the sample is married. To test the hypothesis that married types act differently than singles following an HIV-positive test, I estimate the effects of testing separately on individuals who are single and married (Table 7). Both singles and married types have higher STI incidence when surprised by an HIV-positive test (row 2), although only the effect on singles is statistically significant at the 10% level. I cannot reject the null that the effect of being surprised by an HIV-positive test is the same between singles and married types (p-value = 0.81). A stronger measure of altruism may be whether if someone tests as an individual or couple (i.e. people who test with their sexual partner).42 Making the effort to test as a couple may signal a higher level of altruism towards your partner, and the results are consistent with this notion. The large increases in STI incidence following a surprising HIV-positive test appear concentrated with those who test individually (row 8; column 3). For those who test as a couple and are surprised by an HIV-positive test, the point estimate is negative (column 4; row 8), although not significant. I am able to reject the null that the effect of being surprised by an HIV-positive test is the same between individuals and couples (p-value = 0.005).43 It is important to note that couple testing was not randomly assigned and, thus, there are both selection effects (i..e more altruistic types) as well as the effect of testing with your partner that both affect sexual behaviour. However, these results are consistent with the notion that those who may have higher altruistic preferences respond very differently to HIV-positive tests.
Table 7. Effects of HIV Testing Conditioning on Relationship/Testing Status
| Linear combinations: effect of HIV tests by prior beliefs |
|HIV− test on low prior group|
|HIV+ test on low prior group|
|2. Test + (Test × HIV)||0.084||0.063||0.122||−0.032|
|HIV− test on high prior group|
|3. Test + (Test × High)||−0.051||−0.058||−0.044||−0.017|
|HIV+ test on high prior group|
|4. Test + (Test × HIV) + (Test × High) + (Test × High × HIV)||−0.028||−0.073||−0.075||0.009|
The finding in this article that the behavioural response to testing depends on a subject's prior beliefs of HIV status is consistent with previous work done in the US and SSA. Boozer and Philipson (2000), using non-experimental data, find that HIV testing elicits a differential response depending on a person's priors using a sample in San Francisco. Goldstein et al. (2013) find a differential response in health-related behaviour by pregnant women in Kenya to health worker absenteeism depending on their prior beliefs of HIV status. One of the empirical challenges to conditioning a test response on both prior beliefs and HIV status is that cell sizes can become quite small. The data in this article are unique: over 19% of the sample is HIV-positive which generates 465 individuals who are HIV-positive. This provides sufficient statistical power, even when conditioning on both prior beliefs and HIV status. In comparison, Thornton (2008) has 52 HIV-positive individuals in her sample and does not reject the null hypothesis that there is a differential response to an HIV-positive test depending on priors.
4.3 Are the Results Still Relevant?
The original study was conducted in the mid-1990s, and clearly the HIV/AIDS environment has changed dramatically since then. Two of the biggest changes in SSA regarding HIV/AIDS have been the increase in access to both HIV testing and ARVs. It should be stressed that the main result in this article is not the effects of making HIV testing available, it is that testing can affect behaviour if it substantially changes beliefs. The idea that testing can change an individual's beliefs is not specific to any time period. In addition, a vast majority of individuals throughout SSA have still not been tested for HIV; for example, over 70% of individuals sampled in Rwanda, Tanzania and Uganda have never had an HIV test (Mishra et al., 2009). The increased accessibility of ARVs in SSA over the past decade has also changed the environment; those who become infected with HIV now have hope and that learning their HIV-positive status is no longer a death sentence. However, using a simple model (see Section 'Conceptual Framework'), learning that you are HIV-positive still reduces the marginal cost of engaging in risky sex. In other words, there is still a cost of becoming infected with HIV even when ARVs are widely available. In Section 'Short-Term Effect of Testing on the Epidemic', I discuss further the implications of ARV availability to my results. Overall, while the environment with respect to HIV/AIDS in SSA has changed since the original study was conducted, given that this article is focused on how testing can alter beliefs and how these changes in beliefs also change the marginal costs of risky sex, the findings in this article still carry policy relevance.
5 Short-Term Effect of Testing on the Epidemic
What are the effects of testing on new HIV infections? As previously shown, the effects of testing depend on both the distribution of beliefs and HIV status in the population. To estimate these joint distributions, I use the latest round of DHS that measure both subjective beliefs and actual HIV status. I limit the analysis to countries in SSA with high HIV prevalence and where surveys were conducted over the past four years. I use the urban samples from Kenya 2008, Mozambique 2009 and Zambia 2007 DHS and predict the effects of testing on new HIV infections.44 In this DHS urban sample, HIV prevalence is 16% compared to 19% in this study. This suggests that the level of risky sexual behaviour is fairly similar between the DHS sample and the one in this article. For this analysis, I make the strong assumption that the behavioural response to testing is the same between the sample in this article and the population represented by the DHS. In addition, assumptions on behaviour and epidemiological parameters are made to make this model tractable. Overall, we should interpret the following results with caution.
The AVERT epidemiological model is used to estimate the change in the number of HIV infections due to testing (Rehle et al., 1998). Using the estimated effects of testing on STI incidence, I estimate the underlying change in sexual behaviour. For example, if STI incidence increased by 9 percentage points, I estimate the additional number of sexual partners necessary to generate this increase. Using these estimates of sexual behaviour, I then predict the resulting change in HIV-infections. Full details of the model can be found in online Appendix A.14.
Table 8 presents the predicted changes in HIV infections due to testing. I find that in the case of Kenya, testing leads to a 19% reduction in new infections, while in Mozambique and Zambia, testing leads to increases of 11% and 1%, respectively. These differences are due to the variation in the joint distributions of beliefs and HIV status across these three countries. For example, in Kenya, only 3% of the urban population is surprised by HIV-positive tests, but this is over 8% in Mozambique and Zambia (see online Appendix A.14). There are also a much higher percentage surprised by HIV-negative tests in Kenya (53%) compared to Mozambique (23%) and Zambia (25%). This exercise demonstrates that the effects of testing on the epidemic depend crucially on the distribution of both beliefs and actual HIV status.
Table 8. Effect of Testing on HIV Infections
|1. Change in HIV infections||−15||10||1|
|2. Percentage change CD4 thresholds||−19%||11%||1%|
|6. Test and treat||−86%||−73%||−76%|
A major policy question currently underway is, ‘when should ARVs be provided to HIV-positive individuals?’ Current WHO guidelines recommend that HIV-positive individuals who have CD4+ counts of 350 or lower should be put on treatment. However, there is a growing body of evidence that shows that when ARVs are administered at an earlier stage of the infection, HIV-positive individuals benefit from fewer opportunistic infections and lower mortality rates (Thompson et al., 2012). In addition, ARVs reduce the viral load in HIV-infected individuals, dramatically reducing the probability that they transmit the infection to an uninfected partner (Cohen et al., 2011). If the CD4 threshold is raised to cell counts of 500/μL, increasing the number of HIV-positive individuals eligible for treatment, the number of HIV-infections decreases considerably in all three countries (Table 8, row 5). If ARVs are immediately provided to all HIV-infected individuals regardless of CD4 counts, a policy known as ‘test and treat’, we see massive reductions in HIV infections as testing is rolled out (row 6).
It should be noted that there may be behavioural responses to the increased availability of ARVs, which this simple model does not take into account. There is evidence in the US and SSA that individuals increase their risky sexual behaviour when ARVs are more widely available (de Walque et al., 2010; Friedman, 2012).
Combining a simple epidemiological model with well-identified estimates on the effects of HIV testing on sexual behaviour, I show that under certain conditions, HIV testing in the short term can lead to an increase in HIV infections. This result is driven by the increase in risky sex by those surprised by an HIV-positive test. However, the provision of ARVs at earlier stages of the infection (higher CD4+ counts) completely mitigates this response and leads to an overall decrease in HIV infections.
A few final caveats are in order. First, in the long run, as testing increases the risky behaviour of those surprised by an HIV-positive test, the pool of potential sexual partnerships becomes riskier. HIV-negative types may respond to this by decreasing their risky sexual behaviour (Kremer, 1996; Mechoulan, 2004). I am therefore unable to say how steady-state HIV prevalence would be changed by HIV testing. Second, the population of interest in this study are sexually active urban individuals. The effect of testing maybe different on a rural population that is less sexually active. This remains a topic for further research.
6 Conclusion and Policy Implications
This study shows that the effects of HIV testing depend crucially on both the distribution of HIV status and prior beliefs in the population. Empirically, I find evidence that groups surprised by HIV-positive tests increase their risky sexual behaviour while those surprised by HIV-negative tests decrease their risky sexual behaviour. There is evidence that the former result is not limited to the context of this study. In rural Malawi, Kaler (2003), using observational journals,45 finds that men who have high beliefs of being HIV-positive continue to engage in multiple partnerships and unprotected sex as ‘this behaviour is no longer dangerous if one has already contracted the virus’. The results from this study raise questions about the implicit assumption in HIV testing policies that those who receive HIV-positive tests will behave altruistically and take steps to prevent infecting others.
As noted previously, the main result that those surprised by HIV-positive tests is based on inference from a sample of 188 HIV-positive individuals with low priors at baseline. Given this, some caution is warranted when interpreting the main results of this article. Further work is necessary to test whether these findings hold in other settings and time periods.
From a policy perspective, it should be stressed that this article does not in any way suggest that HIV testing be curtailed. Testing is necessary to identify HIV-positive individuals to link them into treatment. The results from this study do suggest that the provision of ARVs at earlier stages of the infection (i.e. higher CD4+ counts) not only would improve health outcomes for those infected but it could mitigate any adverse behavioural responses to HIV-positive tests. In addition, there is an active discussion on whether to ‘test and treat’ or immediately provide ARVs to all those who test HIV-positive, regardless of their CD4+ counts. Epidemiological models predict that it has the potential to reduce HIV prevalence dramatically over the long term (Granich et al., 2009). An additional benefit of a ‘test and treat’ policy is that the likelihood that HIV-positives who believe they have ‘nothing to lose’ after testing will pose much less risk to infecting others. Finally, this study suggests that targeting of HIV testing might be both feasible and desirable. Using population-based surveys, such as the DHS, we may be able to identify populations that overestimate their HIV risk (high priors/HIV-negative). Focusing testing on these groups may lead them to engage in safer sexual behaviour.
The 19 countries include: Benin, Botswana, Cape Verde, Central Africa Republic, Democratic Republic of Congo, Eritrea, Ethiopia, Gambia, Ghana, Guinea-Bissau, Lesotho, Mauritania, Niger, Sao Tome & Principe, Senegal, Sierra Leone, Somalia, Swaziland and Uganda.
The number tested in 2008 represents just 5.9% of the 142 million people who live in these countries.
See Weinhard et al. (1999) and Denison et al. (2008) for comprehensive reviews of the HIV testing literature.
Thornton (2008, p. ••) notes that ‘condom purchases may not reflect the true demand for safe sex. If knowledge of HIV status increases abstinence, the demand for condoms could fall in response to obtaining test results’.
HIV is also a STI. However, in this article, an STI will refer specifically to either a gonorrhoea or chlamydia infection.
The mean STI rate for the control group (not tested at baseline) who believed they were at low risk for HIV at baseline but are actually HIV-positive is 1.06%.
The mean STI infection rate for the control group (not tested at baseline) who believed they were at high risk for HIV at baseline but are actually HIV-negative is 5.90%.
See Granich et al. (2009) and Cohen et al. (2011) for how ARVs reduce the viral load in HIV-positive individuals thereby reducing their transmission risk to uninfected individuals.
CD4+ cells are ‘T-cells’ that make up an important component of the human immune system. Current WHO guidelines are for ARVs to be provided when CD4+ cell counts are 350 cell/μL or lower.
HIV-positive individuals who go on ARV treatment at an earlier stage of the infection (i.e. higher CD4+ counts) have lower rates of both morbidity and mortality (Thompson et al., 2012).
Following Kremer (1996), this model assumes that λ(B,W) and j are relatively small (i.e. fewer than 10 partners). Under this condition, the probability of infection is approximately linear in partnerships. In the sample I use in this study, the self-reported number of recent partnerships is 1.30. In situations where j is very large (i.e. 100 partnerships), the probability of infection is no longer linear in partnerships.
Since the focus of this study is the effects of HIV-tests conditioned on beliefs, beliefs of HIV infection (π) are assumed to be exogenous in this model. One can model beliefs as a function of previous sexual behaviour and information about HIV risk. See de Paula et al. (forthcoming) for examples of more dynamic modelling of beliefs.
Port of Spain, Trinidad, was the third study site. It was excluded from the analysis since the focus of this article is on the effects of HIV testing in SSA.
Of the 1,477 in the treatment arm, 1,385 opted to take an HIV test.
Of the 1,223 in the control arm who returned for the six-month follow-up survey round, 1,022 accepted an HIV test.
The HIV/AIDS knowledge test asks participants 12 questions about how HIV is transmitted. Examples of questions include: ‘Can a person get AIDS or the AIDS virus from: working near someone, eating food cooked by someone who has the AIDS virus, using public toilets, having sexual intercourse without a condom with someone who has the AIDS virus?’
Commercial sex partners are defined as when money is exchanged for sexual activity.
The reason for not testing all urine samples for STIs at baseline at the time of the baseline survey was that if the results for these were known, it would be unethical to withhold this information from the study participants. The original study team wanted to compare the effects of HIV testing to a control group only receiving health information and not the additional information of an STI test result. In the follow-up round, the study did not unfreeze all baseline urine samples for STI testing due to the cost.
Estimated HIV prevalence in urban areas are 7.2% (Kenya 2008), 15.5% (Mozambique 2009) and 19.5% (Zambia 2007).
Allen et al. (2003) use a sexual diary to collect data on sexual behaviour and interview study participants on a three-month basis, while Minnis et al. (2009) use both audio computer-assisted self-interviewing (ACASI) surveys and face-to-face interviews.
Of the 750 individuals who tested HIV-negative at baseline and retested at six months, only 12 became infected, an infection rate of 1.6%.
Gonorrhoea and chlamydia increase HIV transmission rates increase by a factor of 10 and 5 respectively (Chesson and Pinkerton, 2000).
By extension, those who reveal that they believe they are likely to be infected with HIV face similar costs.
All four questions are on the baseline survey but removed from the six-month follow-up survey, because ‘Interviewers needed to be blinded to the baseline serostatus of participants during the follow-up interview;’ (Grinstead et al., 2001).
Given that the responses use a Likert scale and are not subjective probabilities, interpersonal comparisons warrant some caution. Two people may have identical beliefs about being HIV infected, but one may respond as ‘not at all or hardly worried’ (1) while the other person may respond as ‘a little bit worried’ (2).
Given the evidence that people misreport their sexual behaviour (see subsection 'Self-reported sexual behaviour') due to social desirability bias, it follows that people may also misreport their beliefs of HIV infection.
The results in this article are not sensitive to this cut point for dividing the sample into low and high-belief groups; the main results are robust when this cut point is varied (see online Appendix A.9).
Those who respond to question B with a 1 or 2 are placed in the low-belief group, and those that respond with a 3 or 4 are placed in the high-belief group.
Examining whether beliefs measured in one period predict future behaviour as a test for the validity of beliefs has precedent. Jensen (2010) measures beliefs on the returns to education for students in the Dominican Republic. He finds that those who have higher beliefs on the returns to education go on to complete more years of school.
I restrict this analysis to the control group since the HIV tests in the treatment arm would change beliefs of HIV infection.
It is worth noting that the preferred measure assigns an equal weight to each of the four belief questions. This was done to be fully transparent about the construction of the belief measure. An alternative is to weight each question by how predictive it is of HIV status (see online Appendix A.8).
The effect of assignment into the testing arm conditioned on if tested as an individual or as a couple and by gender also shows no statistically significant effect on STI incidence. Results are presented in online Appendix A.6.
Sample size for both the HIV-positive group (N = 372) and the HIV-positive/Low Priors group (N = 188) compares favourably to other studies on HIV-testing, where Thornton (2008) uses 52 HIV-positive individuals and Delavande and Kohler (2012) use 97 HIV positive individuals to estimate the behavioural effects of receiving an HIV-positive test. I also find behavioural effects of testing in the HIV-negative/High Priors group where sample sizes are larger (N = 687).
Intent-to-treat estimates of assignment to the testing arm can be found in Table 3.
I exclude the HIV indicator because I compare HIV-positive individuals with low prior beliefs who get tested versus HIV-positives with low prior beliefs who are not tested.
This commitment itself does not appear to change actual sexual behaviour as there are no statistically significant differences in STI incidence between the treatment and control arms. Only when HIV-status and priors are taken into account that we observe changes in STI incidence due to testing.
This under-reporting does not seem to be an artefact of pre-existing propensities to under-report sexual activity. Panel (a) compares the control and testing arms of the same sample at baseline (pre-assignment) and if anything, it appears that those who will be assigned into the testing arm are a bit more forthcoming about their sexual activity (rows 1–3).
A non-primary partner is someone who is not a spouse or boyfriend/girlfriend. It is analogous to a casual partner.
Full details of the study design are described in Kamenga et al. (2000).
The MDICP consists of a long-term panel data set that begins in 1998. More information can be found at http://www.Malawi.pop.upenn.edu.
Thornton's (2008) use of condom purchases may reflect demand for safe sex but not actual sexual behaviour.
Not everyone who tests as a couple is married to their partner – 27% who test as a couple are not married to their sexual partner.
Baseline characteristics between testing and control arms for both individuals and couples segmented by HIV status and priors are presented in online Appendix A.12.
Comparing the DHS sample to the one in this study, I find similarities along age, sexual activity and, most importantly, HIV prevalence. The average age in the DHS urban sample is 29.3 and over 85% have had sex. This compares to an average age of 28.3 and over 80% have been sexually active in the two months prior to the survey. Comparisons of self-reported sexual activity are difficult because of major differences in the types of questions asked between the two surveys.
Observational journals are a record of all conversations (either direct or overheard) about AIDS that were kept by local researchers in Malawi. They are a standard methodological tool used by anthropologists.