Statistical Analyses

We summarized baseline respondent characteristics by race/ethnicity and gender. We used chi-square statistics to compare categorical variables and analysis of variance to compare continuous variables, as shown in Table 1. We fit random-effects multinomial regression models using the GLIMMIX procedure in SAS software (version 9.3; SAS Institute Inc., Cary, NC, USA) to compare the facility-specific proportions of participants who were “less than” or “somewhat” versus “very” satisfied (reference level) for each domain, specifying a random effect to account for clustering of Veterans within VAMCs.

Because preliminary analyses indicated a significant interaction between race/ethnicity and gender, we modeled gender-specific associations between race/ethnicity and satisfaction for each domain. Each model included fixed effects for gender, age (centered to the overall mean age [55] and scaled by 10 years), race/ethnicity, and the gender-by-race/ethnicity interaction, and a random effect for site. We adjusted for age in all models to account for differences in the age distributions by race/ethnicity. We assessed gender-specific differences between race/ethnicity subgroups using two-parameter 0.05-level Wald tests that simultaneously compared “less than satisfied” and “somewhat satisfied” to “very satisfied”; we also report pairwise Wald tests comparing these response categories. We constructed gender-specific linear contrasts of satisfaction for black versus white and Hispanic versus white Veterans, and race/ethnicity-specific contrasts from the same statistical model to assess gender differences within each race/ethnicity category for all domains except women’s health. P-values of < 0.05 were considered statistically significant, with no adjustment made for multiple comparisons.

In a sensitivity analysis, we assessed potential confounding in the age-adjusted models by considering each of the covariates shown in Table 1. We included variables significant at the 0.10 level in the domain of outpatient care in all domain-specific multivariable models. For each domain-specific model, we then used a backwards selection approach (removing variables with p > 0.10) to identify a parsimonious set of covariates. We assessed whether conclusions from the age-adjusted models changed based on this statistical adjustment for potential confounders.