Harmonization of survey datasets
First, we reviewed survey design documents to describe the survey design and weighting procedures used for all three surveys. We compared sampling design and survey weighting procedures across surveys to identify differences that could potentially influence comparisons. We developed an analysis strategy to both facilitate comparisons and minimize the influence of differences in survey design or weighting procedures on comparisons between survey estimates. Once we chose a weighting approach, we developed a list of variables to extract and harmonize across surveys based on perceived importance, availability, and consistency of definitions across surveys. Once extracted, the weighted estimates of these variables were assessed for consistency across surveys. Finally, we used selected variables to identify and describe appropriate statistical methods for comparisons and trend analysis.
Data extraction and manipulation
We reviewed data dictionaries and other survey documentation to identify relevant survey design and analysis variables pertaining to HIV biomarkers and behavioral and demographic variables across the three surveys for inclusion in the analysis.
These surveys were originally designed to provide data used by various stakeholders to monitor Kenya’s population and HIV-related health outcomes. This section briefly summarized the survey design and weighing approaches used in the surveys. All three surveys utilized two-stage stratified, cluster sampling designs based on the National Sample Survey and Evaluation Programme (NASSEP) household-based sample frames created by the Kenya National Bureau of Statistics and revised after each decennial population census.
KAIS 2007 was the first AIDS Indicator Survey conducted in Kenya to monitor progress on key indicators in the national HIV prevention, care, and treatment programs . The survey was designed to obtain a nationally representative sample of persons aged 15–64 years and to provide estimates of HIV-related outcomes stratified by urban/rural residence and the 8 provinces. The first stage included a selection of 415 clusters (70% rural and 30% urban) from the NASSEP IV (based on the 1999 census); the second stage included selecting a sample of 25 households within each cluster.
KAIS 2012 selected 372 clusters from NASSEP V (based on the 2009 census) using a systematic random sampling method. KAIS 2012 sampled 9,300 households within 9 of the 10 National AIDS and STI Control Programme (NASCOP) programmatic regions: Nairobi, Central, Coast, Eastern North, Eastern South, Nyanza, Upper Rift, Lower Rift, and Western regions, designated as either urban or rural. The sampling frame was not available for the North-Eastern region at the time of the survey, and this region (and hence seven NASCOP regions) was excluded from the survey. The target population was persons aged 18 months–64 years. Half of the households were targeted for children aged 18 months–14 years. The survey was designed to provide estimates of HIV-related outcomes for adults aged 15–64 years stratified by urban/rural area and the nine included NASCOP regions.
Like KAIS 2012, KENPHIA 2018 also was based on NASSEP V. KENPHIA was a cross-sectional, household-based survey conducted among persons aged 0–64 years in 800 clusters from 96 urban/rural county strata covering the entire household population of Kenya. In 2012, following the promulgation of the 2010 Constitution of Kenya, these counties became the geographical units of devolved government in place of districts. Survey data collection was conducted from June 2018 to February 2019. Of the 34,610 persons targeted by the survey, 27,897 were adults aged 15–64 years, and 6,713 were children aged 0–14 years. One in three households were targeted for the inclusion of children. The survey was designed to provide estimates for adults aged 15–64 years for all 47 counties in Kenya.
Each of these studies were carried out in accordance with the Helsinki Declaration.
Table 1 presents detailed summaries of the three surveys.
The KAIS 2007 design was stratified by district and residency (urban/rural). Urban areas were further stratified by socioeconomic status. Both KAIS 2012 and KENPHIA designs were stratified by county and residency. Household nonresponse adjustments in KAIS 2007 were computed by province and residency, whereas in KAIS 2012, they were computed by NASCOP region and residency, resulting in the following nineteen design strata: Nairobi (Urban), Central (Urban/Rural), Nyanza (Urban/Rural), North Rift (Urban/Rural), South Rift (Urban/Rural), Eastern North (Urban/Rural), Eastern South (Urban/Rural), Western (Urban/Rural), and Coast (Urban/Rural). In KENPHIA, household nonresponse adjustments were computed by county.
The KAIS 2007 and KENPHIA surveys covered the entire national territory, but KAIS 2012 excluded one geographic region, North Eastern. Therefore, to ensure that differences in coverage did not bias trend analyses, this region was omitted from the analysis, thereby stratifying by 17 remaining NASCOP region/residency strata across all three surveys.
To compensate for over- or under- sampling of cases or for disproportionate stratification along with the non-response, studies often include several types of survey weights in the datasets that are made available after the survey. Individual, child, and HIV-testing (blood) weights ensure that adults aged 15–64 years, children aged 0–14 years, and individuals selected for HIV testing, respectively, are representative of the population sampled. The survey design and nonresponse weighting approach for KAIS 2007 and KAIS 2012 were similar, and so no adjustments were made to the weights used in these studies. The KENPHIA 2018 survey design weights differed from the KAIS design weights in that no household-level post-stratification adjustments were done, and nonresponse weights were developed using a least absolute shrinkage and selection operator regression and chi-square automatic interaction detection methodology rather than the simpler inverse proportional weighting done by sex and geographic area variables. Furthermore, post-stratification weights were developed to age and sex control totals from the national population projections for 2019 for KENPHIA. Therefore, to remove potential biases in comparisons resulting from the differing nonresponse and post-stratification weighting approaches, KENPHIA was reweighted to increase comparability between weighted estimates across the surveys.
Revised KENPHIA weights
A primary sampling unit (PSU) or enumeration area (EA) base weight was computed as the inverse of the probability of selection of the EA. No PSU nonresponse adjustment was made, apart from two ineligible EAs whose weights were set to 0. A household’s initial weight was then computed as a product of the PSU base weight and the inverse of the probability of selection of the household within the EA. An unknown eligibility household nonresponse adjustment was computed as a product of the household initial weight and the inverse of the probability of the household having unknown eligibility. The household weight was further adjusted for the eligible household member nonresponse rate.
Adult person-level weights were assumed equal to the household weight since all adults (aged ≥15 years) were eligible in a household. In the case of children (aged 0–14 years), only children in every third household were included in KENPHIA 2018. The child weight was then computed as three times the household weight. For adults, nonresponse adjustments cells were created by NASCOP region, urban-rural residence, and sex, whereas nonresponse-weighting classes for children were not stratified by sex. The post-stratification cells are produced by NASCOP region and sex. The child weights were not post stratified.
A similar approach was used to compute the HIV-testing (blood) weights included in the study.
Data manipulation and merging
Using the three individual survey datasets, we created a dataset that included survey year, the design variables (weights, strata, and cluster), demographic characteristics, and HIV-specific indicators. The stratification variable in the combined dataset consisted of the 17 NASCOP regions. The cluster was uniquely characterized by the survey year and the cluster identifier in each survey. The weights in the combined dataset were normalized such that the normalized weights summed to the total number of respondents in each survey. The SAS program that combines the three datasets and renames and recodes variables to facilitate comparative analyses is available in Supplementary File 1.
To create the combined data file, we combined 2007, 2012, and 2018 files so that the number of respondents in the combined data file was the sum of the respondents from the three individual files. We then ensured that the analysis variables had the same names and values or categories in all three data files. Table 2 illustrates how the variables used in this analysis were redefined. Secondly, the approach to creating the new set of statistical weights is provided in Supplementary File 2.
The study investigators did not interact with human subjects or have access to identifiable data or specimens. This was a secondary data analysis using anonymized data from each of the surveys that were included.
Figure 1 describes our suggested approach for harmonization of variables and datasets to perform trend analysis.
Assessing comparability of reweighted surveys across key population characteristics
Ideally, a set of unchanging population characteristics could be used to assess the comparability of the original and re-weighted datasets before proceeding with trend analyses. In the absence of such ideal variables, several demographic characteristics such as age, sex, marital status, residency, wealth index, and education, which have predictable trends and have been measured in other surveys over time, can be assessed for trends. In this analysis, we assessed the weighted distribution of each of these variables and used survey-weighted logistic regression to assess changes in the selected characteristics over time (Table 3). We found that there was no significant difference (trend) in key demographic variables selected for comparative assessment of original and re-weighted KENPHIA 2018 datasets.