OBJECTIVES: Observational studies evaluating the association between pharmacy choice and influenza vaccination can account for confounding in various manners. Our objective was to compare three statistical approaches that account for confounding variables in different ways and examine the association between the type of pharmacy a patient uses and influenza vaccination likelihood.

METHODS: Within 2018 Medical Expenditure Panel Survey data, US adults who used community pharmacy (CP) or mail order pharmacy (MOP) were included. Associations between patient-reported pharmacy type and vaccination status were evaluated using unadjusted weighted logistic regression (WLR), multivariable WLR, and a propensity score (PS)-matched sample using WLR. Multivariable analyses adjusted for sociodemographic, health status, and healthcare utilization confounders. PS of confounders were used to create a 3:1 community:mail matched sample. All analyses were stratified by age (<65 and ≥65 years).

RESULTS: The final sample had 8,074 (95,930,349 weighted) respondents <65 and 4,037 (40,163,276 weighted) respondents ≥65. Unadjusted odds of vaccination for CP users compared to MOP users were OR:0.55(0.47,0.64) <65 and OR:0.76(0.64,0.90) ≥65. Adjusted odds were significantly lower for CP users than MOP users among those <65 (AOR:0.71[0.58,0.87]), but not those ≥65 (AOR:0.87[0.69,1.09]). The PS-matched analysis included 2,962 (37,555,393 weighted) respondents <65 and 3,183 (32,681,992 weighted) respondents ≥65, and showed similar odds to multivariable WLR (<65 OR:0.77[0.65,0.91]); ≥65 OR:0.88[0.74,1.05]).

CONCLUSIONS: Multivariable adjustment and PS-matched analyses produced results similar to each other, but significantly different from unadjusted results. While PS matching facilitates comparability of the exposure groups, multivariable WLR quantifies the effect of confounders on the outcome and can be a useful tool to evaluate relationships between variables with high potential for multiple confounders. However, it is unclear whether either technique has adequately accounted for bias in this relationship. The comparison of these fairly well-known techniques within the context of a weighted sample provides further complexity that warrants additional discussion.