All the variables were initially analyzed descriptively by univariate statistics, including examining frequencies, percentages, means and standard deviations. The data were examined for the presence of missing values, influential outliers, skewness and kurtosis for key continuous predictor variables. Missing information was imputed using multivariate imputation by chained equation method .
A breakdown of the demographic characteristics of the GROs based on the intervention and control cities (see Fig. 1) at baseline is shown in Table 1. The socio-economic backgrounds of the GROs were generally similar among the four cities, with some slight variations. Iloilo GROs were slightly older, with an average age of 24.7 years, compared to GROs from Legaspi (t = 3.2, P < .05), Cagayan de Oro (t = 3.1, P < .05), and Cebu City (t = 3.6, P < .001). Iloilo GROs also earned slightly more relative to the other cities (1,480 Pesos, t = 3.6, P < .05; At the time of the study, $1 = 24.50 Pesos). Cagayan de Oros’ GROs were slightly more educated (9.63 years of education) compared to GROs from Legaspi (t = 2.8, P < .05), Cebu (t = 2.5, P < .05), and Iloi-ilo (t = 2.6, P < .05). A significantly higher percentage of Cagayan de Oros’ GROs were married (16.8%, χ2 = 30.5, P < .05), compared to the other cities. Legaspi City had the lowest percentage of married GROs (2.4%, χ2 = 21.3, P < .05).
Preliminary analysis of the data consists of examining the change in knowledge, attitude and condom use behavior from pretest to post-test. As illustrated in Table 2, there were significant increases in knowledge on HIV/AIDS transmission from baseline to post-test for the intervention groups, and no significant change in knowledge for the control group (t = 1.1, P = .92). Baseline to post-test change in attitude on condom use also increased significantly for all the intervention groups but did not change significantly for the control group. There was also a significant increase in condom use for the intervention groups and not for the control group. These preliminary results show the overall effects of the intervention at the unadjusted aggregate level without controlling for individual, establishment and community variables.
The next phase in the analysis consists of a more detailed analysis, using the three-level hierarchical models to analyze the data at the different levels. The condom use outcome was normally distributed, and we estimated the three-level hierarchical linear model using restricted maximum likelihood (REML) that assumes multivariate normality for the model residuals. In all the models, continuous predictors were mean-centered, including age, years in education, knowledge index, condom use attitude and self-efficacy measures.
Based on the findings of the study, it was found that the GROs’ knowledge on HIV/AIDS had a significant effect on condom use (β = .064, P < .05). GROs that had higher knowledge concerning the transmissibility of HIV/AIDS tend to use condoms more often. There was also a significant effect of attitude on condom use—the more positive the attitude towards condom use, the higher the level of condom use (β = .062, P < .01).
Among all the establishments surveyed, 46.6% of the establishments had a policy of mandatory condom use when engaging in sex with customers. The results of the establishment level analysis in Table 3 shows that mandatory condom use was significant (β = .705, P < 0.01). GROs that worked in an establishment with a mandatory condom use rule had a higher level of condom use.
Among the different intervention groups, the combined peer and managerial intervention (Cebu City) had the largest main effect on condom use (β = 1.304, P < 0.01) relative to the usual care group. The managerial intervention was also significant with the magnitude of the main effect at a slightly lower level than the combined peer and managerial intervention (β = 0.628, P < 0.01). The peer intervention group also had a significant main effect on condom use compared to the usual care group (β = 0.25, P < 0.05).
In examining the cross-level interaction effects, there was a significant cross-level interaction effect between peer intervention × establishment rule (β = 0.421, P < .01). Similarly, we found significant interaction effects between managerial intervention × establishment rule (β = 0.563, P < .01) and the combined intervention × establishment rule (β = 0.840, P < .01). These interaction effects indicate that participants who had the peer, managerial or combined interventions and at the same time worked in an establishment with the condom use rule had significantly higher condom use compared with intervention participants who worked in establishments with no condom use rule.
To illustrate this interaction effect, we estimated the expected values from the hierarchical linear models, and computed separately three scenarios assuming a hypothetical increase of 10% in (1) knowledge, (2) attitudes and (3) self-efficacy in Table 4. The results show an increase in condom use ranging in magnitude from 15.8 to 19.7% for the intervention groups when we compare establishments with condom use policy vs establishments without condom use policy. The difference in condom use for the control group in establishments with and without condom use policy were lower, and ranges from 10.3 to 10.8%.