Setting
Data were taken from the Manicaland General-Population Cohort Study (Manicaland Study), an open-cohort study of a representative sample from three districts in Manicaland, Zimbabwe. Manicaland is a province in eastern Zimbabwe where 85% of the population live in rural areas. (36) Manicaland is characterised by an above-average proportion of individuals living under the national poverty line (37) and one of the lowest values of the human development index and life expectancy in the country (38). In Manicaland, HIV prevalence stabilised at current levels of about 11% after a peak of over 25% in the late 1990s. (39) This is the lowest prevalence of any province in Zimbabwe, but the number of PLHIV in Manicaland is one of the highest in the country due to the large population size. (40) Behaviour change has been documented to have contributed to declines in HIV prevalence in Manicaland. (41,42) Despite significant decreases, HIV incidence remains high at just under 1% for females and 0.5% for males in the general population. (43) Sexual relations between young women and older men, characterised by limited condom use, have been identified as a driver of new HIV infections. (44) Uptake of VMMC has been slow (45), and PrEP is only available in small-scale research and pilot projects (46). Manicaland is a priority in the Zimbabwe National HIV and AIDS Strategic Plan, with the objective of reducing HIV incidence by half by 2020 compared to 2013. (47)
Data and measures
The Manicaland Study completed six surveys between 1998 and 2013. Data were taken from the four most recent surveys (2003-13). Study participants were selected from a household census in 12 sites (eight in the 2012/13 survey). These sites represent four different socio-economic strata of the population in Manicaland: Small towns, subsistence farming areas, agricultural estates, and roadside business centres. The study was introduced to the communities in each study site through meetings involving the public and community leaders. Members of the community acted as guides to support the implementation of the study by identifying households and members of the community eligible for the study. Between 8000 and 15000 adults aged 15-54 years participated in each survey, with participation rates ranging from 73.0% to 79.5%. HIV status was objectively determined for each participant on a dried blood spot sample. Other information was collected in a face-to-face interview, conducted by an interviewer of the same sex and in the local language (Shona), covering socio-demographic characteristics, sexual behaviour, perceptions and HIV-specific beliefs. To reduce social desirability bias in the reporting of sensitive information, including sexual behaviour, informal confidential interview techniques were used. (48,49) Time between surveys was about three years and, among those not lost to follow-up due to out-migration or death, follow-up ranged between 77.0% and 96.4%. The Imperial College London Ethics Committee and Medical Research Council of Zimbabwe provided ethical approval for the Manicaland Study. Written informed consent was obtained from each study participant for each survey round. The results of the study were disseminated through workshops and meetings with community members, exhibitions, and booklets in the local language. A comprehensive profile of the Manicaland Study has been published elsewhere (43) and more information is available online (50).
Sexually active participants who participated in at least two consecutive surveys and remained HIV-negative were included in this analysis. Participants who participated in non-consecutive surveys were not included given long intervals between surveys. Only participants who remained HIV-negative between survey rounds were included as HIV infection is likely to impact both risk perception and condom use. Moreover, only participants who were already sexually active at first observation were included as sexual debut influences risk perception and data on condom use are only available for those sexually active. Only data from survey three of the Manicaland Study onwards (2003-2005) were included as survey measures on condom use and risk perception changed after survey two. Risk perception was measured with one survey question (“If you are not infected, do you think you are in danger of getting infected now or in the future?”), allowing for ‘yes’, ‘no’, and ‘don’t know’ responses. ‘Don’t know’ responses (9.17% of observations) were excluded from main analyses. As discussed in Additional file 1 (section 3), these excluded participants represent a diverse set of individuals that could not be easily grouped together with those perceiving or not perceiving a risk for HIV infection. Considering ‘don’t know’ as a separate category was also not meaningful as the sample was small (see Additional file 1, section 3).
Increased risk perception was defined as reporting risk perception in one survey but not in the preceding one. Decreased risk perception was not reporting risk perception in one survey but reporting risk perception in the preceding one. Condom use referred to reporting condom use during the last sexual intercourse. An increase in condom use occurred when the participant reported condom use in one survey but not in the preceding one. Further information on data and measures is provided in Additional file 1 (sections 1-2).
Analysis
The primary hypothesis of this study was that an increase in risk perception causes an increase in condom use. However, the relationship between risk perception and condom use is bi-directional, so we further hypothesised that an increase in condom use leads to a decrease in risk perception as the protective behaviour is implemented. These hypotheses are described in Table 1. While the methods of this study only test for associations between two changes (changes in risk perception and changes in condom use) and the exact temporal relationship between these changes cannot be established (which of these changes came first), it can be tested whether the directions of associations are in theoretically expected directions (Table 1).
Generalised estimating equation (GEE) with a logit link function for a binomial response distribution and exchangeable correlation structure was used to model changes in condom use dependent on changes in risk perception. (51) The main unit of analysis of the regressions, therefore, was a pair of two survey responses contributed by an individual. Modelling changes ( ) over a time period ( ) removes cross-sectional interpretations of coefficients, thus making results more straightforward to interpret, and removes confounding from time-invariant unobservable factors. To model a binary outcome, the sample was restricted to those not reporting condom use at the beginning of ( ), so the outcome of all regressions was increase in condom use against continuing not using condoms (no change). Decreasing condom use was also considered but sample sizes were small; results are presented in Additional file 1 (section 7). GEE account for non-independence of survey responses from the same individual over time. More details on methods are provided in Additional file 1 (section 4).
Hypotheses outlined in Table 1 would only be supported by associations in expected directions in the absence of confounding by changes in other factors, although modelling changes removes the impact of time-invariant factors. For example, change in marital status is likely to impact both condom use and risk perception. It is therefore vital to account for changes in other variables associated with both risk perception and condom use. Several time-variant socio-demographic and behavioural factors were identified as potential confounders in preliminary analyses (Additional file 1, section 5). These include age, marital status, school enrolment, education, socio-economic status, having symptoms of sexually transmitted diseases (STDs), HIV testing, sexual risk behaviour, and having a partner who has other partners. For each of these, change between two surveys was modelled as described in Additional file 1 (section 6).
All GEE models, with the outcome of increase in condom use (vs. no change), included an independent variable for change in risk perception (increase/decrease vs. no change [reference category]). The change in risk perception variable can be seen to separately represent those reporting risk perception at (who may decrease risk perception) and those not reporting risk perception a (who may increase risk perception). Models were estimated including change in age as an additional variable (model 1) and change variables for all potential confounders (model 2). Models were estimated separately by sex. Secondary analyses estimated models by age group (15-24 vs. 25+ years) and marital status (not married vs. currently married) (Additional file 1, section 9). Additionally, these models were estimated with change in risk perception as no change against increase or decrease in risk perception broken down by reason for risk perception. These reasons refer to the reason for perceiving a risk after not reporting risk perception at or previous reason before decreasing risk perception. Reasons for risk perception included having multiple partners, having a partner who has other partners, marrying someone who may be HIV-positive, or ‘other’.
Population attributable fractions (PAFs) were estimated for proportions of ‘cases’ of increase in condom use attributable to increase and decrease in risk perception, respectively, as described elsewhere. (52-54) Model 2 of the regression estimates were used for this.