Analyses

Analyses were weighted to account for SHEP sampling design and non-response to the SHEP survey. The average response rate over the study period was 51% (55% in 2009, 53% in 2010, 51% in 2011, 48% in 2012). As in previous analyses,^{24 , 26} survey weights were truncated at 500 to limit the impact of any one respondent. Unweighted and weighted percentages were used to characterize the study sample.

To determine whether increases in documented alcohol counseling resulted from increased counseling or only from changes in documentation practices, the overall prevalence of patient-reported and provider-documented alcohol counseling for each of the 4 fiscal years was estimated using logistic regression. Two logistic models were estimated: one for patient-reported and one for provider-documented alcohol counseling. Although weighted for SHEP sampling and non-response, models did not adjust for patient or clinical information; each model included the 4 fiscal years as dummy variables. The estimated prevalence and 95% confidence interval (CI) were calculated for patient-reported and provider-documented counseling in each fiscal year. For each counseling measure, the absolute percentage difference in prevalence between 2009 and 2012 was calculated. We calculated the p-value associated with a Wald statistic,^{27 , 28} calculating standard errors with the robust sandwich estimator²⁹ to account for the survey weights, to assess for significant differences across the 4 years.

To further characterize the mechanisms underlying changes in rates of documented alcohol counseling, patients were categorized into four mutually exclusive groups based on the presence or absence of provider documentation and patient report of alcohol counseling. For each year, the proportion of patients in each of the four groups was estimated using weighted, unadjusted logistic regression, with fiscal years included as a dummy variable. Results are presented as the estimated prevalence of each group, with 95% CIs, for each fiscal year. The absolute percentage difference in prevalence between 2009 and 2012 is reported with p-values calculated as described above. All analyses were performed using STATA software (version 13.1; StataCorp LP, College Station, TX, USA).