We made use of pooled data from current Demographic and Health Surveys (DHS) conducted from January 1, 2010 and December 31, 2016 in 30 countries in SSA (see Figure 1). DHS is a nationwide survey collected every five-year period across low and middle-income countries. DHS focused on maternal and child health by interviewing women of reproductive age (15 – 49 years) and men between 15 and 64 years. DHS surveys followed the same standard procedures – sampling, questionnaire development, data collection, cleaning, coding and analysis which allow for cross – country comparison. The survey employed a stratified two stage sampling technique. The initial stage involves the selection of points or clusters (enumeration areas [EAs]) followed by a systematic sampling of households listed in each cluster or EA. For this study, the women file of the DHS data was used. All the participants were women in their reproductive age (15 – 49), who were usual members of the selected households and/or visitors who slept in the household on the night before the survey. In this study, only women who had information on all the variables of interest were included (N= 194,275).
Definition of variables
Outcome variable
The outcome variable was HIV testing. It was derived from the question “have you ever tested for HIV?” and the response was coded as “1=Yes and 0=No”.
Explanatory variables
Fourteen explanatory variables were considered in our study including the key explanatory variable (women decision-making capacity). Following a previous study [17], women decision-making capacity was derived from three questions “decision on personal health care”, “decision on large household purchase” and “decision on visits to family or relatives”. These response categories were recoded as “not alone = 0” and “alone = 1”. An index was created with all the “yes” and “no” answers with scores ranging from 0 to 3. The score 0 and 1 were labelled as “low capacity” and 2 and 3 were labelled as “high capacity”. A dummy variable was generated with ‘0’ score being women who were less capable and ‘1’ if women were more capable.
Besides, 13 additional variables were included in the study. These are; country, age, educational level, marital status, religion, wealth status, place of residence, parity, occupation, exposure to mass media (radio, television and newspaper) and whether healthy-looking person can have HIV or not. Apart from country of origin which was predetermined based on the geographical scope of the study, the selection of the rest of the variables were based on their association with HIV testing and counselling [6, 7, 8, 18, 19, 20, 21, 22]. Marriage was recoded into ‘never married (0)’, ‘married (1)’, ‘cohabiting (2)’, ‘widowed (3)’ and ‘divorced (4)’. Occupation was captured as ‘not working (0)’, ‘managerial (1)’, ‘clerical (2)’, ‘sales (3)’, ‘agricultural (4)’, ‘household (5)’, ‘services (6)’ and ‘manual (7)’. We recoded parity (birth order) as ‘one birth (1)’, ‘two births (2)’, ‘three births (3)’, and four or more births (4)’. Lastly, religion was recoded as ‘Christianity (1)’, ‘Islam (2)’, ‘Traditionalist (3)’, and ‘No religion (4)’.
Statistical analysis
The data was analysed with STATA version 14.2 for Mac OS. The analysis was done in three steps. The first step was the computation of the prevalence of HIV testing in SSA (see Figure 1). The second step was a bivariate analysis by which we calculated the prevalence and proportions of HIV testing across the socio-demographic characteristics with their significance levels (see Table 1). Afterwards, two stepwise logistic regression models were built in order to assess the predictors of HIV testing among women in SSA. Model I looked at a bivariate analysis between the key independent variable, women decision making capacity and HIV testing. Model II controlled for the effect of country and all the socio-demographic variables to build a multivariable logistic regression model (see Table 2). All frequency distributions were weighted while the survey command (svy) in STATA was used to adjust for the complex sampling structure of the data in the regression analyses. Multicollinearity was checked and there was no evidence of multicollinearity among the variables (Mean VIF=1.37). All results of the logistic regression analyses were presented as Crude Odds Ratios (CORs) and Adjusted Odds Ratios (AORs) at 95% confidence intervals (CIs).