Bringing the state back in: Understanding and validating measures of governments' political commitment to HIV

Abstract

Analysis of the politics of HIV programme scale-up requires critical attention to the role of the state, since the state formulates HIV policies, provides resources for the HIV response and negotiates donor involvement in HIV programmes. However, conceptual and methodological approaches to analysing states' responses to HIV remain underdeveloped. Research suggests that differences in states' successes in HIV programme scale-up reflect their levels of ‘political commitment’ to responding to HIV. Few empirical measures of political commitment exist, and those that do, notably the AIDS Program Effort Index (API), employ ad hoc scoring approaches to combine information from different variables into an index of commitment. The indices are thus difficult to interpret and may not have empirically useful meaning. In this paper, we apply exploratory factor analysis to examine whether, and how, selected variables that comprise the API score reflect previously theorised dimensions of political commitment. We investigate how variables associated with each of the factors identified in the analyses correspond to these theorised dimensions as well as to API categories. Finally, we discuss potential uses – such as political benchmarking and accountability – and challenges of factor analysis as a means to identify and measure states' political commitment to respond to HIV.

Keywords:

• political commitment
• HIV/AIDS
• AIDS Program Effort Index
• exploratory factor analysis
• validation

Notes

1 In the data that we obtained from the 2005 round, the survey questions were the same as in the 2003 round. The difference in the two rounds was in the countries that were sampled. Data from the two years were pooled in order to get the maximum number of countries for analysis.

2 A potential constraint under exploratory factor analysis is that sample size can affect the precision and the stability of the solution: factors obtained from a relatively small sample may not be congruent with population factors and may vary in repeated samples (MacCallum, Widaman, Zhang, & Hong, 1999). However, researchers caution against relying solely on sample size in assessing the adequacy of the sample (Bandalos & Boehm-Kaufman, 2008; MacCallum, Widaman, Zhang, & Hong, 1999; Costello & Osborne, 2005). We discuss implications of our sample size in the ‘challenges and limitations’ section below.

3 In successive iterations of the factor analysis, variables were eliminated one at a time if (1) the variable had a loading of 0.3 or higher only on those factors that we would not consider significant to the analysis based on the scree test or (2) the variable did not have a loading of 0.3 or higher on any factor (Costello & Osborne, 2005). In all cases, we used prior knowledge and theory as a parallel guide to decide whether a variable should be left out (Costello & Osborne 2005). For example, if a variable loaded onto multiple factors, then we considered whether the cross-loadings could be plausibly interpreted based on the substantive content of the variable and the factors.

4 The scree plot graphs eigenvalues in descending order of their magnitude; the eigenvalue of a factor refers to the amount of variance explained by that factor (Preacher & MacCallum 2003). In the scree test, the inflection point of the scree plot indicates the point where further factors do not substantially increase the proportion of common variance explained and it is used as a cut-off for the number of factors. Whereas the scree test tells us where the addition of further factors does not increase the proportion of common variance explained relative to the explanatory power of the other factors, the eigenvalue test (i.e., retaining factors with eigenvalues > 1) assesses the percent of the common variance explained in absolute terms.

5 We applied a standard oblique (oblimin) rotation, allowing for correlation among the factors.

6 Three factor scores were computed for each country, one for each dimension of political commitment, using the regression scoring method.

7 Based on the alternative, arbitrary cut-off criterion of eigenvalues > 1, we would retain two factors, but we defer to the scree test and theoretical plausibility, which suggest three factors.

8 Factor loadings represent the weights or contribution of each variable to a given factor. Communalities refer to the proportion of a variable's total variance that is explained by common factors (Preacher & MacCallum 2003).