Analysis

Data from the 18 states for all 3 study waves were appended into a single dataset. Questions related to beliefs around Saleema, exposure to advertising, message receptivity, and FGMC intentions and social norms were compared descriptively. Each set of constructs was assessed using confirmatory factor analysis (CFA) to investigate scale variables that might represent each construct, using factor loading and Chronbach’s alpha (>0.6) tests to assess factor strength, following the widely cited Comrey & Lee (1992) criteria [26].

We then examined Saleema event participation by state. We summed the number of event participants for each state, and created per-capita indicators of community, government, and media event participation in order to reflect the scale of activity by state population. We examined the distribution of this indicator and noted a right-skewness in event participation across states. We log-transformed this variable to use as a linear predictor in regression modeling.

We conducted regression modeling, examining association with logged, per-capita event attendance and improved knowledge, attitudes and behaviors (KABs) around Saleema for wave 3. Intra-cluster coefficients (to account for the cluster randomized sampling design) were examined for each key outcome variable, and were determined to be minimal (F-test was insignificant) at the PAU-level. We then used state-level fixed effects regression modeling to compare respondents at waves 2 and 3 and examine the effect of increases or decreases in higher event participation, through use of an interaction term, in FGMC beliefs, intentions, attitudes, and norms. R software (version 3.3.2) was used in all analyses [27].