Regression analysis

In this study, regression analyses were conducted to find the important predictor variables for treatment and access costs for PLHIV (as the direct costs), and productivity costs before coping strategies. Regression analyses for total direct costs to the household were not conducted because PLHIV who took an accompanying person obviously need to pay higher costs because of the additional costs required for the accompanying person/s. This might overshadow the impact of other important variables like the health status, and the income level of the PLHIV. Therefore, regression analyses for treatment and access costs in the place of total direct costs were used.

To conduct the regression analyses for treatment and access costs and productivity costs, the predictor variables which were significant to the relevant outcome variables in the descriptive analyses, were taken into consideration. Occupation, ethnicity, CD4 level, self-reported health status, PLHIV accompanied or not, household income and study district were found to be statistically significant variables for the direct costs in the descriptive statistics. Likewise, education, occupation, ethnicity, CD4 level, self-reported health status, PLHIV accompanied or not, and study district were found significant variables for productivity costs in descriptive statistics. As the health status was measured in two forms: self-reported health status and CD4 level, one regression with self-reported health status and another regression with CD4 level were conducted separately along with other significantly contributing variables. Likewise, household income was also presented in two forms: as a continuous variable, and in income quintiles; one regression with income (as a continuous variable) and another regression with income quintiles were conducted separately along with other significantly contributing variables. Thus, four different types of regression analyses for treatment and access costs and three regression analyses for productivity costs were conducted before conducting the tests (here test regression analysis means the analysis which was conducted to test the significant of other extra variables which did not show up as significant in the descriptive analyses for the outcome variable); and then, final regression analyses. After that, a number of separate regression analyses were conducted to check if the other variables which did not contribute significantly in the descriptive analyses did emerge significantly as synergistic, or interactive variables. After checking all results, a final list of significantly contributing predictor variables was prepared and used for the final, reduced form of regression analysis.

Before conducting the regression analyses, normality, linearity, multicolinearity, and outliers among the variables were checked. Normality and linearity were checked by using histogram and normal probability plot, multicolinearity was checked by using Pearson correlation, and outliers were checked by using Mahalanobis and Cook’s distance matrices. To correct the skewed data, log transformation (log10) was made for treatment and access costs, productivity costs (before coping strategies) and household income. All other variables which were either interval or nominal in nature, were rescored as binary variables. Examples of potential predictors which were changed into binary or dummy variables were: study districts, education of PLHIV, occupation of PLHIV, ethnicity of PLHIV, CD4 level of PLHIV, self-reported health status of PLHIV, and PLHIV accompanied with other.

As our outcome variables were log transferred and some of the predictor variables were also log transferred, the regression equation was:

Where,

Log Y = outcome variable (log transferred)

X _1, X _2, X _{3, and} X _k are predictor variables

α represents regression constant or intercept

b _1, b _2, b _3, and b _k are the unstandardized regression coefficients, where k represents the number of predictor variables, and e is the error.

In the above regression equation, predictor variable X ₂ is log transferred.