Stata release 13

We used means and standard deviations, or numbers and proportions, for descriptive statistics. We calculated means and 95 % confidence intervals of pre- and post-assessments for each of the four domains, and then compared them, using linear regression with robust standard errors, so we could account for the possibility some patients were included in both phases of the assessment. We then created multivariable models to control for the potential influence of confounders by simultaneously including the following variables in the model: age, sex, GP assignment, and, duration of assignment. We used STATA release 13.1 for all analyses (Stata Corp, College Station, TX, USA). A p-value of <0.05 was considered statistical significant.

We compared baseline characteristics (age, sex) using Chi-square or Fisher's exact test for categorical data, and differences in estimation of stroke risk using t-test for continuous data. We analysed GP data separately, and stratified by confidence in risk estimates, experience with TIA, and age of the physician. We dichotomized co-variables and omitted underestimates of risk so we could understand the possible confounding mechanism in overestimated risk assessments. To compare the third risk estimate (carotid endarterectomy) with the other two, we collapsed both overestimated risk reductions (40% and 50%) into one category. We calculated p-value using Chi-square test and Spearman's rho for risk estimation as the independent variable. Each of the other co-variables (experience, age, self-confidence) was used as the dependent variable to detect correlation. Finally, we performed a logistic regression analysis to assess predictors of immediate emergency referral by GPs in cases of suspected TIA. We considered a p-value of 0.05 to be statistically significant. All analysis was done with STATA release 13.1 (Stata Corp, College Station, TX, USA).

Data were managed and analyzed with the statistical package Stata release 13.1 (StataCorp, College Station, TX). We used descriptive statistics (mean, standard deviation of the mean [SD], proportions) to summarize the data, and Pearson's chi-square statistic (χ²) to compare the distribution of various measured factors according to participants' depression status. We used Cronbach's alpha to evaluate the internal consistency (reliability) of the KICA-dep, and two by two tables to estimate sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) associated with different cut-points of the scale. We generated a receiver operating characteristic (ROC) curve using the ‘roctab’ command of Stata, which allowed us to estimate the area under the ROC curve. We estimated the point-prevalence of depression by dividing the number of people with a depressive disorder by the total number of people in the sample, and calculated the 95% confidence interval (95%CI) of the prevalence estimate using generalized linear modeling (glm). The odds ratios (OR) of depression and respective 95%CI associated with relevant exposures were calculated using logistic regression. Alpha was set at 5% and all probability tests reported were two-tailed.