Determining predictors of mortality in HIV positive people in South Africa, 2003 to 2009: a mixed methods approach incorporating unobserved variables

Abstract

Background The largest proportion of HIV-infected people resides in Southern Africa. In South Africa, the government has taken the lead in the provision of free HIV treatment with a high coverage rate. Provision of free antiretroviral treatment has led to a decline in mortality rates and an increase in life expectancy. However, a significant number of people with HIV continue to die despite the availability of free treatment. A large proportion of studies have concentrated on using quantitative methods of analysis. Very few have used mixed methods that combine quantitative time-to-event frailty models and qualitative methods in assessing risk factors for mortality in HIV-infected individuals. However, use of such mixed methods approach could provide insights that may lead to an improvement in patient care and management. Aim To determine mortality risk factors in HIV-infected people through incorporating unobserved variables using a mixed methods approach in which quantitative findings are explained by the qualitative. Methods To critically review statistical methods used for assessing risk factors for mortality in HIV-infected people between the years 2002 and 2011. We conducted a literature review on the design of studies, how data were analysed and whether suitable statistical methods were utilised in assessing mortality risk factors in HIV-infected people in the period 2002-2011. Only publications written in English and listed in Pubmed/Medline were considered. In this review, papers using time-to-event techniques were regarded as appropriate. Data were split into two equal periods allowing for the comparison of the statistical methods over time.

To compare the different time-to-event methods, we ran 1,000 simulations of parametric clustered data using parameters derived from an HIV study that was conducted in South Africa by the Perinatal HIV Research Unit (PHRU). Data for 5, 10 and 20 clusters of size 50 and 100 were simulated. Survival and censoring times were derived from a Weibull distribution. The minimum of survival and censoring times was taken as the study time. Using the simulated data, we compared the following time-to-event methods: Cox proportional hazards regression, shared Gamma frailty with Weibull and exponential baseline hazards (frequentist models), and the Bayesian integrated nested Laplace approximation (INLA) with Weibull baseline hazard. Parameter estimates, standard errors and their fit statistics were averaged over 1,000 simulations. Similarly, means and standard deviations from INLA were averaged (over the 1,000 simulations). Frequentist models were compared using the -2 loglikelihood fit statistics while all the four models were compared using the mean square error (MSE). Additionally, we simulated semiparametric clustered frailty models (using gamma and log-normal frailties) including INLA, h-likelihood, penalized likelihood and penalised partial likelihood estimations. Parameter estimates and their standard errors were presented graphically and compared using the MSE. To assess mortality risk factors in HIV-infected people in South Africa in different settings, factors associated with mortality in HIV-infected people were assessed by INLA survival frailty model using cohort data of HIV-infected people from South Africa. Two thirds were from Soweto (urban) and the rest from Mpumalanga (rural). Findings were evaluated by site. Mixed methods were used to evaluate risk factors for mortality by combining the best fitting model applied to retrospective data and qualitative analysis on prospective

data. In order to explain the unobserved frailty modelling results, we conducted a qualitative study that enrolled 20 participants who had confirmed knowing a person that had died as a result of HIV. Participants were recruited from the Zazi VCT in PHRU and were interviewed using a semi-structured interview guide. The aim of the qualitative study was to attempt to explain the unobserved factors influencing mortality in HIV-infected individuals using perceived reasons for death given by the participants. These were later used to complement the potential reasons for death as identified in the frailty modelling (quantitative) results. Results In the critical review, 189 studies met the inclusion criteria that included prospective (69%) and retrospective (30%) studies. Of the 189 studies, 91 were published in the period 2002-2006 and 98 in 2007-2011. Cox regression analysis with frailty was used in only 7 studies (~4%); of which 6 were published between the years 2007- 2011. The simulation study showed that the shared frailty models performed better than Cox-PH. Within the shared frailty models, the Gamma frailty model with a Weibull baseline performed better than the Gamma frailty model with an exponential baseline. The MSE showed that in general, the Bayesian INLA had better results. In the semiparametric simulations, results were similar but INLA had a slightly better fit with consistently lower MSE values relative to both gamma and log-normal frailty models. The random effects estimate for INLA, whose method is slightly different, had lower MSE values consistently relative to the other methods. In the HIV cohort study, 6,690 participants were enrolled with majority being female (78%) and most participants residing in an urban area (67%). Rural participants were older (36 years; IQR: 31-44) and with a higher mortality rate (11/100 person years).

Among those residing in rural areas, HAART treatment for between six and twelve months (HR: 0.2, 95% CI: 0.1-0.4) and more than 12 months (HR: 0.1, 95% CI: 0.1- 0.2) was protective relative to not being on treatment. Being on HAART treatment for greater than twelve months was protective in the urban participants (HR: 0.35, 95%CI: 0.27-0.46). Significant heterogeneity, assessed by frailty variance, was high in rural participants and lower in the urban. Since the frailty modelling results suggested that the unobserved variables had a significant effect on mortality in HIV-infected individuals, a qualitative study was conducted to explore the potential causes of death. In the qualitative study, participants perceived that mortality in HIV-infected individuals may have been influenced by engagement in risky sexual behaviour such as multiple sexual partnerships, negative attitude by healthcare workers towards HIV-infected people, believing in the healing power of religion, traditional medicine, food security and social support structure. Conclusions The study found that Cox proportional hazards regression with frailty is not commonly used in research on mortality in HIV-infected individuals as it is used in other fields of health research. Additionally, use of the more complex semiparametric frailty models was even lower in this population. From simulations, we found that frailty survival models provided a better fit in modelling mortality due to their ability to account for unobserved variables especially the Bayesian INLA. As the unobserved variables are complex to explain using only quantitative modelling techniques, qualitative analysis of perceived causes of death was explored. Unobserved variables affecting mortality were explored through qualitative analysis of perceived reasons provided by bereaved participants. This mixed methods approach optimised

data by using a quantitative approach followed by a qualitative one that complemented each other. Use of optimal methods in assessing morbidity and mortality in HIV-infected patients may improve patient care and management in South Africa and other countries.

Key words: HIV, Mortality, Rural, Urban, unmeasured variables, HAART, Frailty