HIV and sexual behavior change: Why not Africa?

Abstract

Despite high rates of HIV in Sub-Saharan Africa, and the corresponding high mortality risk associated with risky sexual behavior, behavioral response has been limited. This paper explores three explanations for this: bias in OLS estimates, limited non-HIV life expectancy and limited knowledge. I find support for the first two. First, using a new instrumental variable strategy I find that OLS estimates of the relationship between risky sex and HIV are biased upwards, and IV estimates indicate reductions in risky behavior in response to the epidemic. Second, I find these reductions are larger for individuals who live in areas with higher life expectancy, suggesting high rates of non-HIV mortality suppress behavioral response; this is consistent with optimizing behavior. Using somewhat limited knowledge proxies, I find no evidence that areas with higher knowledge of the epidemic have greater behavior change.

Introduction

Five to ten percent of adults in Sub-Saharan Africa are infected with the human immunodeficiency virus (HIV) and the primary mode of transmission in the region is heterosexual sex. For this reason, sexual behavior change is a major focus of HIV prevention efforts and understanding changes in behavior is important both for predicting the future path of the epidemic and for developing policy.

Existing literature has shown mixed evidence of behavioral response in Africa. Caldwell et al. (1999) summarize a number of cases throughout Africa and generally suggest response to HIV has been quite limited. Oster (2005) shows evidence on lack of change in the share of women engaging in premarital sex in a number of countries in Africa through the 1990s. Stoneburner and Low-Beer (2004) argue that the 1990s saw limited changes in sexual behavior outside of Uganda.¹ This is not to say there has been no response. For example, Ng’weshemi et al. (1996) find reductions in risky behavior among men in Tanzania, and Bloom et al. (2000) find mixed evidence in Zambia, with reductions over some periods and not over others.²

However, even in papers which demonstrate behavioral response in Africa, this response is often quite small. A good example of this is Thornton (2008), who finds that when people learn they are HIV positive they increase their purchase of condoms, by only by about one condom. A statistically significant, but not economically significant, response. The initial cross-sectional analysis in this paper demonstrates a similar impact: if anything, areas with higher HIV rates appear to have more risky sexual behavior, even controlling extensively for demographics which we think might impact sexual activity. Limited behavior change is surprising in light of extensive behavioral responses among high risk groups – gay men in particular – in the United States (Winkelstein et al., 1987, McKusick et al., 1985, Francis, 2008). Most existing explanations for limited behavior change focus on Africa-specific cultural barriers to changing behavior – fatalism, low levels of female bargaining power and so on (Amuyunzu-Nyamongo et al., 1999, Caldwell et al., 1999, Lagarde et al., 1996a, Lagarde et al., 1996b, Philipson and Posner, 1995).

This paper explores three non-cultural explanations for the observation of limited behavioral response. First, I consider the possibility that the apparently limited response reflects bias in the existing estimates. Since HIV is a sexually transmitted infection, estimating the response of sexual behavior to HIV is plagued with problems of reverse causality which will naturally bias the estimates upward, making it more difficult to observe any response. Second, I explore whether behavioral response is limited by weak incentives to respond. Specifically, I consider whether low non-HIV life expectancy in Africa plays a role: individuals who expect to die young even without HIV have a weaker incentive to reduce risky sex in the presence of the epidemic. Finally, I consider a more standard explanation: individuals are prevented from changing behavior by lack of knowledge about the epidemic (i.e., Green, 2003).

Section 2 formalizes these explanations with a simple theory of the epidemic. The theory makes three basic predictions. First, individuals should decrease their risky sexual behavior in response to increases in HIV prevalence. Second, this decrease should be larger for individuals with high non-HIV life expectancy. The intuition behind this result is simple. Consider two men, one who expects to live for another eleven years, and a second who expects to live for another fifty years. In a world without HIV, the choice of sexual behavior need not depend on these future life expectancies. However, in a world with HIV, sexual behavior carries a risk of death, approximately 10 years after infection. Introducing HIV will affect the behavior of both men. However, for the first man HIV should not affect his behavior very much, since HIV infection only costs him one year of life. For the second man, HIV infection means losing forty years of life, so he should have a much larger response to the presence of HIV. Third, to the extent that knowledge of the epidemic is imperfect, behavior change should be more extensive for individuals with better information. Again, the intuition is straightforward: someone who knows that HIV is spread sexually should respond more than someone who does not know this.

Put in the context of this theory, existing literature has typically focused on the first of these predictions and found little support. I argue this result may be due to the reverse causality issues inherent in this estimation. I therefore begin by estimating this simple comparative static, but addressing the endogeneity. The problem is simple. HIV is a sexually transmitted infection. Areas with higher levels of risky sexual behavior will, on average, end up with higher HIV prevalence. Even if individuals respond to the epidemic by decreasing their risky behavior, this may be difficult to observe in the data. I address this concern using an instrumental variables strategy, instrumenting for HIV prevalence with distance to the origin of the virus in the Democratic Republic of the Congo. In principle, if the virus takes time to travel, moving from person to person, areas further from its origin should have lower prevalence on average.³

The analysis in this paper uses data from the Demographic and Health Surveys in a sample of 14 countries in Sub-Saharan Africa with surveys between 2001 and 2007. These surveys feature information on HIV prevalence, GIS location (for calculating distance) and sexual behavior. Using cluster-level data from these surveys, I show that there is a strong negative correlation between distance and HIV prevalence. This remains when controlling extensively for latitude and longitude, as well as demographic characteristics of clusters. This distance measure appears to be uncorrelated with the incidence of premarital sex in the period before HIV appeared, as well as uncorrelated with pre-epidemic education and income (as measured by durable good ownership). These latter facts should provide additional confidence that the instrument satisfies the exclusion restriction, even though it is obviously not random. The instrumentation strategy is discussed in more detail in Section 3.

Section 4.1 uses this instrument, and the Demographic and Health Survey data, to estimate response to HIV. Throughout, I use three measures of sexual behavior: whether the individual has multiple sexual partners, whether he or she has multiple partners with no condom use and the number of non-marital partners. I estimate responses separately for married women, unmarried women, married men and unmarried men and, in addition, for all married and all unmarried individuals. As expected, the OLS relationship between sexual behavior and HIV prevalence is positive: more risky sex in areas with more HIV. The IV estimates, however, are largely negative and, for married individuals, are significant. Focusing on married individuals, I find that a doubling of HIV prevalence leads to a 20% drop in the chance of having multiple sexual partners and a 30% drop in multiple partners with no condom use. There is no significant evidence of changes for unmarried individuals, and no significant evidence of changes in the number of partners for either type.

Section 4.2 turns to the effects of non-HIV life expectancy on behavior change. Based on the theory, I expect greater behavioral response for people with lower mortality. I test this by estimating the response of sexual behavior to HIV interacted with measures of non-HIV mortality. I begin with a simple measure: child mortality. Child mortality is highly correlated with adult mortality in non-HIV settings (this can be seen, for example, from the life tables in Coale et al., 1983), likely because young children and older people often die from similar causes. Actual older adult life expectancy in these areas is difficult to measure, due to limited data, and is likely to be correlated with HIV prevalence; child mortality is, therefore, a useful proxy. Of course, children also die from HIV, transmitted from their mothers. To avoid this confound as much as possible, I look at deaths for children over two and under six, in which range a large share of HIV-infected children have already died.

Using this measure, I find strong evidence that behavioral response varies with life expectancy. The interaction between prevalence and child mortality is positive and significant: those who live in areas with higher child mortality (i.e., lower life expectancy) change their behavior less. However, even if one accepts that child mortality is a good measure of non-HIV adult mortality, this analysis still faces the issue that mortality is likely to be correlated with a variety of other variables, which may well drive our results. So while this may be suggestive, it is certainly not conclusive.

To address this identification issue, I also consider how responsiveness varies with two explicit mortality shifters: malaria prevalence and, for young women only, maternal mortality.⁴ I argue that these analyses are less contaminated by the omitted variable bias. I calculate malaria levels based on climate factors alone – using the malaria model from Tanser et al. (2003), along with temperature and precipitation data – so the measure is not driven by, for example, bednet usage or other behaviors. In the case of maternal mortality, although the level of mortality is unlikely to be exogenous, I take advantage of the fact that it contributes more to mortality risk for young women. By comparing the response of young women to older women, and then that difference in response to the difference between younger and older men, I am able to employ a difference-in-difference identification strategy which avoids issues created by the correlation in levels.

With both measures I again find strong evidence that the response to HIV is greater in areas with lower non-HIV mortality. In the case of malaria, there are sizable reductions in risky behavior in areas with no malaria, and risky behavior is actually estimated to increase in areas with high levels of malaria. The estimates based on maternal mortality show similar evidence: young women in areas with high risk of maternal mortality (measured by childbirth deaths among siblings of respondents) change their behavior less than older women in these areas, relative to the difference in changes across age groups for men.

Although neither of these measures is perfect – non-HIV mortality is not randomized – I argue that, taken together, the data make a credible case that individuals with lower non-HIV life expectancy change their behavior less. In addition to the relevance to the specific case of HIV, these results may also be interesting as a test of the economic theory of competing mortality risks, as outlined by Dow et al. (1999).

Finally, in Section 4.3 I turn to the third possible explanation for limited behavioral response: lack of knowledge. Echoing the methodology for life expectancy, I explore whether the response to HIV is greater in areas with higher levels of knowledge about the epidemic. Knowledge is measured based on the share of individuals who know that one can reduce the chance of infection by using a condom or by limiting oneself to a single partner. By focusing on average knowledge in the area I hope to avoid the fact that high risk individuals are likely to know more about the epidemic because they are high risk. There remain two issues with this measure. One is that these questions may not capture knowledge well, and the answers may be noisy. Second, even if well-measured, the level of knowledge may not be exogenous. If knowledge campaigns are more extensive in areas with otherwise higher or lower behavioral response, this may bias our estimates. With these caveats in mind, I estimate effects of the interaction between knowledge and HIV prevalence. I find no evidence that response is more substantial in areas with more knowledge. If anything, the coefficients have the opposite sign.

Taken together, these results have a number of lessons for policy makers. They provide first some encouraging news on behavior change: despite what has been seen in the literature thus far, there does seem to be some evidence of behavioral response to the epidemic. More important, perhaps, are the results on variations across individuals. The fact that low life expectancy impedes behavior change suggests that increasing life expectancy through treating other illnesses – by, for example, improving maternal care or eradicating malaria – could have positive spillovers to HIV prevention. On the flip side, however, although the results are more speculative, knowledge about the epidemic does not seem to impact behavioral response. This may well be due to the already high levels of knowledge – in our data, 65% of individuals correctly respond that condoms prevent HIV – but, regardless of the mechanism, this argues against extensive continued spending on educational campaigns (Green, 2003).

In addition, these results may suggest a more limited role for some of the more traditional explanations for limited behavioral response – fatalism, bargaining power, etc. Certainly the results here do not rule out a role for these variables. However, the results do suggest that standard economic theory may provide significant insight and explanatory power, without having to rely on cultural or taste-based differences across areas.

The rest of this paper is organized as follows. Section 2 outlines a simple theory and 3 discusses the data and instrumental variables strategy. Section 4 presents results, and Section 5 concludes.