2.8 Statistical analyses

Data were analyzed by intention-to-treat [50]. We used generalized linear mixed models that accounted for repeated measurements [51] to assess the impact of adding the DA to patient education. All models included a group variable (education + DA versus education alone), a time variable (two-month follow-up versus baseline), and a group-by-time interaction term. With the repeated measurements, we assumed a specified form of covariance structure among the two visits in which estimates and standard errors were based on a restricted likelihood function given the observed data (REML). Using an unstructured covariance matrix, this specification permitted to handle missing values at the follow-up visit [52]. Models estimated means or prevalence with corresponding 95% confidence intervals, as well as within and between-group differences in means and prevalence ratios. The estimate of the group-by-time interaction term was of primary interest [51]. To fulfill model assumptions, decisional conflict scores underwent a natural log transformations [53]. To facilitate interpretation, means and 95% confidence intervals were back-transformed on their original scale [53]. Determinants of decisional conflict (knowledge [54]), appropriate use of pharmacotherapy (knowledge [42], age [55]), and asthma control (knowledge, body mass index, age, allergy and respiratory tract infections [56]) identified a priori were not included in statistical models since they did not result in a >10% change in the mean differences or prevalence ratios [57]. SAS 9.4 (SAS Institute Inc., Cary, North Carolina, USA) was used to perform all statistical analyses. A two-tailed P-value <0.05 was considered indicative of statistical significance.