Details of the cohort have been published previously (Shafer et al. 2008). Briefly, the study is located in a subcounty in south-western Uganda, in which household surveys of socio-demographic and behavioural characteristics, as well as HIV serostatus of all consenting residents aged 13 years and above have been conducted annually since 1989. The cohort was initially comprised of about 10 000 people residing in 15 neighbouring villages (the old villages) but was expanded at the 1999/2000 survey to approximately 18 000 people in 25 villages. The average annual participation rate is about 60–65%. Community sensitisation activities precede each survey round, including local council briefings and village meetings. All households are visited by, in turn, the mapping, census and survey teams. Consenting residents are interviewed at home in the local language by trained survey staff and provide a blood sample for HIV testing. Each annual survey begins in November and ends in October the following year. This study used data and stored blood samples collected between 1990 and 2007 from participants aged 13–59 years. For simplicity, we denote survey rounds by the year in which the survey ends (e.g. data collected from November 1989 to October 1990 are denoted 1990).
HIV infection at each visit was determined using two parallel ELISA tests (Wellcozyme HIV-1 recombinant VK 56/57, Murex Biotech, Dartford, UK; and Recombigen HIV-1/2, Trinity Biotech, Galway, Ireland) confirmed with Western blot in case of first-time positives or discordant ELISA (Van der Paal et al. 2007). HSV-2 was determined using the Kalon ELISA (Kalon Biological, UK) with cut-off optical density for a positive result at 1.5 (Biraro et al. 2011). To minimise unnecessary testing, all samples were tested for HSV-2 at key survey rounds (1990, 1994, 2000, 2007). Samples from participants testing positive at these rounds were back-tested in the previous rounds until a negative result was detected.
Analyses were restricted to the 15 villages with data from the whole follow-up to allow estimation of changes over time. An a priori decision was taken to stratify analyses by gender because incidence rates and risk factors for HIV infection differ by sex. Annual HIV incidence rates were estimated from 1990 to 2007. Participants were censored at the earliest of the last HIV-negative test or estimated date of seroconversion (estimated as the mid-point of the last negative and first positive result). Both observed and smoothed incidence rates were plotted for easier visual interpretation of trends. Smoothing was carried out using the locally weighted scatterplot smoothing (LOWESS) procedure (Cleveland & Devlin 1988).
HSV-2 test results were available from 48.9% of the eligible resident population overall, and from 48.2%, 41.9%, 50.6% and 53.9% in 1990, 1994, 2000 and 2007, respectively. To minimise bias due to missing data, HSV-2 prevalence was estimated using multiple imputation (Sterne et al. 2009). Data were assumed to be ‘missing at random’ as HSV-2 data were missing mainly because participants chose not to provide a blood sample due to study fatigue, due to absence from home, or lack of availability of samples (Wagner et al. 1994; Kamali et al. 1999). Variables included in the multiple imputation were sex, age, marital status, HIV status and observed HSV-2 status results. These were selected because they were the main independent predictor variables for HSV-2 infection that were available in the data set. Annual HSV-2 incidence rates were calculated using an approach similar to that for HIV incidence described above, on the basis of actual HSV-2 test result, rather than longitudinal or multiple imputation. However, unduly long intervals between the last negative and first positive results will result in imprecise estimated dates of seroconversion and may cause an artificially high estimate of incidence in the periods corresponding to the mid-points. To examine this potential bias, sensitivity analyses were conducted by comparing incidence estimates from analyses including all participants irrespective of the length of the interval between last negative and first positive test with estimates from analyses limited to participants with intervals less than 4, 3 and 2 years, respectively. Estimates for annual HSV-2 incidence were similar, so all participants with interviews of up to 4 years retained in the analysis.
To study the effect of HSV-2 infection on HIV incidence, crude and adjusted incidence rate ratios (IRR) and 95%CI were estimated using Poisson regression. A hierarchical modelling approach was used (Victora et al. 1997). An a priori decision was made to adjust all analyses for time period and age group, as risk of HIV incidence is likely to be associated with these factors. Four time periods were defined. These were two 5-year periods in the 1990s when HIV prevalence/incidence was known to have declined in this population and two periods in the 2000s when declines in HIV prevalence/incidence were no longer observed (Shafer et al. 2008). The time periods are 1990–1994, 1995–1999, 2000–2004 and 2005–2007. P-values for association were estimated using the likelihood ratio test. As the primary exposure of interest, HSV-2 was retained in the model irrespective of the level of association with HIV incidence. Initially, the association between HIV and socio-demographic factors was analysed, and factors with P < 0.15 were fitted in a core socio-demographic model and retained if they remained independently associated with HIV (P < 0.15). Subsequently, each behavioural factor was adjusted for this core socio-demographic model. Behavioural factors in the model were then removed one at a time and retained if significant at P < 0.15. Finally, as the primary exposure of interest, HSV-2 was added into the model and other factors retained if independently associated with HIV (P < 0.15). Factors that remained with P < 0.05 in the final model were considered to be independently associated with incident HIV infection. Demographic factors included in the analysis were period in time, age, level of education, marital status and religion; whereas behavioural factors were age at first sex, number of lifetime sexual partners, number of sexual partners in the last year. As risky behaviour is known to increase risk of HIV incidence, risk factor analysis for behavioural variables was limited to sexually active participants to compare risk for different levels of risky behaviour, but all participants were included in the final model.
To estimate the impact of HSV-2 on HIV incidence at the population level, population attributable fractions (PAFs) were estimated using adjusted incidence rate ratios (aIRR) for HSV-2 from Poisson regression models for HIV incidence adjusting for the effect of other factors that were found to be independently associated with HIV incidence at the individual level (Brady 1998). These factors included increasing age, being an internal migrant, not being Muslim, being presently or previously married, and being HSV-2 infected, as well as time period (men only) and belonging to tribe other than Muganda (women only). PAFs were estimated separately by sex and age group. The age group cut-off was defined as the median age at HIV seroconversion to maximise statistical power. The median age at HIV seroconversion was found to be 28.7 years overall and 31 in men and 26 in women. As both HIV and HSV-2 are sexually transmitted infections and may be acquired simultaneously, PAFs were estimated for incident HSV-2 (i.e. seroconversion to HSV-2 in the same time period as HIV seroconversion) and prevalent HSV-2 (HSV-2 infection prior to HIV seroconversion), separately.
Analysis for risk of acquisition examined the effect of HSV-2 prevalence/incidence on HIV incidence in the same gender, whereas analysis for risk of transmission examined effect HSV-2 prevalence/incidence from one gender on HIV incidence in the opposite gender. The association of population-level annual HIV incidence with annual HSV-2 incidence/prevalence was analysed by calculating crude and adjusted regression coefficients and 95% confidence intervals using linear regression. The effect of HSV-2 on HIV incidence was adjusted for population-level measures of the other factors that were found to be independently associated with HIV incidence at the individual level. In this population, almost all HIV transmission occurs through heterosexual contact (White et al. 2007). Therefore, the effect of HSV-2 on HIV acquisition in the population was investigated by examining how HIV incidence in a given gender is related to HSV-2 in the same gender. The effect of HSV-2 on HIV transmission was investigated by examining how HIV incidence in one gender is influenced by HSV-2 in the opposite gender. As HSV-2 is more transmissible than HIV, it was hypothesised that at a population level, changes in HSV-2 incidence are likely to precede changes in HIV incidence. For this reason, analyses examined both the correlation between annual HIV incidence and HSV-2 infection in the same year, and correlations between HIV incidence and HSV-2 infection 1 year and 3 years previously, respectively.