Authors' Response * Population-average models and sexual network studies are complementary approaches to study HIV risk

We thank Helleringer and Reniers for their thoughtful response to our recent manuscript.1 They point out several important challenges of interpreting the average, population-level effect of an individual-level exposure (HIV testing and counselling, or HTC) on an individual-level outcome (condom use). Our paper was a secondary analysis of data originally collected to measure the effect of hormonal contraceptive use on women’s risk of HIV acquisition.2 We used these data to quantify the change in Ugandan and Zimbabwean women’s self-reported condom use both a short and longer time period after learning their HIV status. Helleringer and Reniers note that our analyses are limited by the lack of data on women’s sexual networks, a point with which we agree. Partner tracing is not common in Africa. The first randomized trial of partner tracing of HIV contacts in Africa has just been completed in Malawi.3 This kind of data was unfortunately not available for our analyses because it was not collected during the parent study. Of course, while valid and complete sexual network data can be revelatory in elucidating HIV risk, executing sexual network studies of any meaningful size is logistically difficult. Missing data can lead to substantial bias.4 The ethical challenges of soliciting identifying information about participants’ sexual partners are non-trivial.5 A study that is comprehensive enough to measure both prospective individual-level changes in behaviour following an intervention such HCT as well as dynamic sexual networks and mixing patterns across the study population would be extremely labour intensive and costly in settings which are already resource constrained. Dr Helleringer is involved with one of the first sexual network studies to be conducted in Africa (also in Malawi, the Likoma Network Study), the results of which are now being released, and which will provide new insights into the feasibility and utility of network studies in sub-Saharan Africa.6,7 Helleringer and Reniers also state that we erroneously ‘… treat behavioral change as a homogeneous process that is well described by statistical measures of central tendency (e.g. means and regression coefficients)’. We agree that epidemics are non-linear in ways that may be dependent on hidden heterogeneity. It is also true that, as epidemiologists, we model averages over populations (in this case, women aged 18–35 years recruited from family planning clinics in Zimbabwe and Uganda). We could have stratified or conditioned our analyses on variables for which the effect of HCT on condom use was meaningfully different. However, as Helleringer and Reniers suggest, hidden heterogeneity may persist. As noted in our manuscript, we assessed a number of different distributions before selecting a zero-inflated negative binomial regression model (ZINB, sometimes referred to as a ‘hurdle model’). Our outcome was the number of unprotected sex acts in a typical month. ZINB models, which are relatively uncommon in epidemiological research, combine two distributions to generate two measures of effect. The first compares (in our case) the odds that all sex acts are protected in a typical month after HIV diagnosis with the odds that all acts are protected in a typical month beforehand. The second estimates the change in the number of unprotected acts in a typical month after HIV diagnosis with the number reported beforehand. This method of analysis allowed for diversity in participant response, albeit in a limited way, through two mechanisms. First, rather than assuming that data follow a single distribution, as in most multivariable regression procedures, ZINB models permit the data to follow two different distributions (a logistic procedure and a negative binomial procedure). Second, we ran separate ZINB models to describe changes occurring over a shorter (2–6 months) and longer (12–16 months) period after HIV testing, which allowed us to capture and quantify time-dependent behavioural changes. Nevertheless, we could have more explicitly interpreted our selected method as showing the impact of HIV testing on unprotected sex at the population level. Lastly, Helleringer and Reniers point out that reduction in the number of unprotected acts (the outcome in our analyses) does not always translate to a reduced number of HIV transmissions. We agree, and they provide plausible scenarios where behavioural risk at the level of the population declines, but HIV transmission continues because of particularly risky sexual mixing patterns or behavioural disinhibition. (Our manuscript also described possible but unmeasured consequences of learning one’s HIV status, including relationship dissolution.) However, we disagree that the logical conclusion of these alternative scenarios is that reductions in the number of unprotected acts is a meaningless behavioural goal to measure or pursue. Reduction in the number of unprotected acts will lead to reductions in HIV transmission at least within serodiscordant partnerships. As we describe in our manuscript, 9–10% of HIV-positive women reported no sex at all after their diagnosis (one reason for the overall drop in number of unprotected acts), and 44% reported no unprotected acts. For these women (assuming truthful self-report about condom use), the probability of onward transmission during the observation period was zero. In sum, we agree with Helleringer and Reniers that population-level analyses often mask important subgroup heterogeneity, and that sexual network studies will contribute substantially to the overall understanding of sexually transmitted disease transmission. However, we disagree that population-level analyses ‘hardly shed light on the actual social process of behavioral change’. Taken to the extreme, this statement implies that the heterogeneity within the response would invalidate all population-level research, including randomized intervention trials. Interventions nearly always have an effect on many, but not all, participants, and yet we consider the (population-average) results of randomized trials the research gold standard. We firmly believe that epidemiological studies using population-average models and sexual network studies are complementary. Both are necessary to enhance our understanding of the HIV epidemic.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.