Methods for conducting trends analysis: roadmap for comparing outcomes from three national HIV Population-based household surveys in Kenya (2007, 2012, and 2018)

In 2017, despite the rapid increase in antiretroviral therapy (ART) use over the previous two decades and the corresponding decline in mortality, approximately one-third of people living with HIV in East and Southern Africa and less than half of those living with HIV in West and Central Africa were not receiving any life-saving treatment [1, 2]. By 2017, HIV/AIDS was a major cause of death in sub-Saharan Africa (SSA), where 71% of all people living with HIV resided. Globally, 75% of HIV-related deaths and 65% of all new HIV infections occurred in SSA [3‐5].

Against this background, it is important to assess whether interventions over the last two to three decades have decreased HIV incidence and to identify geographic regions and sociodemographic groups with high HIV prevalence [3, 6]. HIV data obtained from national population-based surveys play an important role in monitoring the HIV epidemic and response in the general population. These surveys estimate incidence, prevalence, and various parameters related to the HIV pandemic in high-HIV-prevalence countries. These surveys were designed to monitor progress toward ending the AIDS epidemic [6‐8]. Additionally, they were designed to monitor the UNAIDS 90–90-90 targets by the year 2020: 90% of all HIV-positive people know their HIV status; of these, 90% are receiving sustained ART; and of these, 90% have achieved viral load suppression [9‐11]. These surveys have also been used to describe associations between high-risk behavior and HIV status and to assess HIV prevention, care, and treatment services.

Unlike in high-income countries where longitudinal studies provide nationally representative trend estimates for health outcomes, for example, the National Health and Nutrition Examination Survey [12, 13], HIV surveys in low-income countries and high-prevalence settings are generally cross-sectional and are independently implemented approximately once every 5 years. Therefore, it is important to develop methods that can be used to assess trends across independent surveys for countries interested in employing similar techniques. We used Kenya to showcase this approach as there had been several HIV population-based surveys conducted, with varying sampling and survey weighting considerations, in the past two decades. Such methods must account for differences in survey design, weighting, coverage, and indicator definitions. Over the past two decades, five national population-based surveys [14‐18] have included HIV testing and HIV modules in their algorithms in Kenya.

We present methods that can be used to assess temporal trends in outcome variables of interest as a means to answer such questions as: “Has HIV risk behavior significantly declined over time in Kenya, and if so, in which demographic groups or regions?” and “Has access to HIV testing services increased over time in Kenya?” Our tools also can help HIV programs appropriately analyze trends in recent population-based HIV surveys in Kenya and provide guidance regarding appropriate statistical comparisons between surveys, including tests for trends. These suggestions may also serve as a roadmap for other cross-survey comparison analyses applicable to other countries or indicators. The methods presented here are being utilized to examine trends in specific indicators of interest in other KENPHIA-focused studies. Therefore, the programmatic implications of selected trends comparison presented in this study are not discussed.

We developed an approach for assessing and harmonizing independent population-based surveys to assess trends in HIV-related indicators. After describing the methods used to design and weight each survey, we harmonized stratification, demographic variables, and survey weights to ensure comparability before proceeding with a trend analysis. In this analysis, the survey weights for the latest survey (KENPHIA 2018) were revised to ensure comparability with the previous two surveys. It is important to note that we developed these methods strictly to allow for comparisons between surveys. The methods are not meant to provide revised or improved estimates for the most recent survey analyzed (KENPHIA). The original weights for KENPHIA are optimal and should be used to analyze and present the results of the KENPHIA survey. Similar approaches to making comparisons between surveys are documented elsewhere in the literature [23‐30]. For reproducibility, we also provide the analysis codes that demonstrate how the analysis was carried out and how the comparison was done.

The weighted distributions of demographic variables were consistent across surveys with some exceptions. There was an increasing proportion of the sample that resided in urban areas, as expected given broad development trends in Kenya. The age structure showed spikes in subsequent age groups across surveys, consistent with a cohort effect from reductions in fertility 15–20 years before the KAIS 2007 survey, given that the surveys were spaced at approximately 5-year intervals, consistent with the historical fertility reductions observed in the recent Demographic and Health survey [31] and census [13] in Kenya. Other differences are difficult to explain. For example, the sex distribution seemed skewed in 2007, with 42% of the survey population being men, compared to higher proportions of in the other two surveys (48%–49%), perhaps indicating coverage issues among men in that survey.

We used two outcome variables expected to change over time (HIV testing and age at sexual debut) to demonstrate a methodology to carry out trend comparisons. For HIV testing in the last 12 months, we highlighted two approaches that can be used to assess trends in dichotomous outcomes. We first computed survey-weighted proportions and plotted the resulting trends over time by selected covariates. We then fitted logistic regression models with the year as a covariate, adjusting for age, sex, residence, marital status, and wealth index. This approach has been employed in several other previous studies. Trends in HIV-testing rates have also been discussed extensively in the literature [23‐25]. The use of chi-square tests of trends and logistic regression were extensively used in the literature.

For age at sexual debut, we show two approaches for assessing trends for a continuous variable. Several studies have also used survey data from low-income countries to assess trends in the HIV-related outcomes considered in our analysis. Several studies have treated age at sexual debut as a time-to-event outcome, assessing this outcome variable's trends among different cohorts observed [32, 33]. These studies have used survival analysis-based approaches to assess trends in the outcome variables of interest. In our analysis, we used two approaches where the first ignored censoring in the age at sexual debut and presented a summary and regression-based results as an example of how trends in continuous outcome variables could be assessed [34]. We then used the survival approach and found a decrease in the risk of early sexual debut over time.

Our analysis is subject to several limitations. The trend comparison was based on using three time points (2007, 2012, and 2018), so we were only able to make relatively short-term assessments of the trajectory of the indicators considered. Previous studies used Cochran Armitage chi-square tests or Z-tests to assess the significance of trends [35‐38]. Survey weighted versions of this statistic were not implemented in our analysis due to limitations in the software we used. Another challenge encountered was the change in the definition of certain variables and indicators over time, adding uncertainty in interpreting the meaning of observed trends. Furthermore, not all questions asked across the three surveys were the same, making it difficult to analyze some of the outcomes across the three surveys. Our study did not address all relevant issues for every conceivable trend analysis that could be conducted with these surveys. For example, changes in HIV-testing algorithms may affect estimates. Great care is needed in interpreting results with potential underlying methodological differences. Finally, there are alternatives to population-based surveys for measuring trends in health conditions. For example, in 2018, Kenya established an HIV Case-Based Surveillance system to measure progress along the HIV care cascade to provide high-quality, timely, and reliable HIV data by population characteristics. Despite current limitations and challenges, this system will provide an opportunity for future assessment of trends based on a census of events rather than population-based sampling, as presented here.

The pooling of data from multi-year surveys aimed at assessing trends in key public health performance indicators is an important part of interrogating the impact of programs and interventions. However, this pooling of data from large complex surveys leads to data sets with large sample sizes which inadvertently increase statistical power [39]. This increased power leads to a tendency of finding statistically significant differences however small they are. Therefore, researchers need to distinguish between statistical difference and scientific difference. The statistically significant difference that arises from larger sample sizes may not be scientifically meaningful.

We have provided approaches and considerations that can be used to support trend comparisons for various outcome variables in HIV surveys in low-income settings. Our approach has demonstrated that independent national household surveys conducted over time can be assessed for comparability, adjusted as appropriate, and used to estimate trends in key indicators. Analyzing trends over time can not only provide insights into Kenya’s progress toward HIV epidemic control but also identify gaps in key HIV indicators.