30 Chapter 2 background (the Netherlands, western countries, and non-western countries); a two-digit pseudonymised residential address postcode to account for variations in urbanisation and provider availability; primary diagnosis identified by the main practitioner and coded in 97 categories at the third level of the DSM-IV classification (Table A4); and binary indicators for having concurrent diagnoses of other mental disorders (two diagnoses and three or more diagnoses). In the analysis of treatment minutes, functional improvement, and additional treatment record, we also controlled for disease severity at baseline. Furthermore, in the analysis of functional improvement and additional treatment record, we controlled for the natural logarithm of treatment minutes. Statistical analysis We used ordinary least squares regression models to estimate the association between household income quintiles and the outcomes. For binary outcomes, ordinary least squares regressions are linear probability models, which in comparison with logistic regression allow for the direct interpretation of the coefficients as changes in likelihood [26]. We estimated models adjusted only for demographic characteristics and models adjusted for covariates measuring need and treatment intensity as well. Furthermore, we studied income gradients for both functional improvement and additional treatment, stratifying by five reasons of record closure: on the patient side (e.g. non-attendance or changing place of residence); on the provider side (e.g. referral to other provider or setting); by mutual agreement between patient and professional (e.g. joint decision on treatment completion); treatment records consisting only of intake, diagnosis, or a crisis episode; and administrative reasons (the most common of which was reaching the 1-year maximum record length). Finally, we investigated heterogeneity in our findings by gender, using a stratified analysis, and by type of diagnosis, using interaction terms between the income quintiles and the diagnostic classes for depressive disorders, anxiety disorders, and all other disorders. To test the robustness of our findings we estimated models adjusting for different sets of covariates, primary diagnosis aggregated at the second level of DSM-IV, and additionally controlling for a proxy of addiction (any diagnosis of substance use disorder). We also controlled the analyses of treatment outcomes for three binary indicators identifying each type of therapy used at any point during the record (pharmacotherapy, psychological therapy, and other therapies). We used alternative definitions of the functional improvement outcome (a continuous difference between final and baseline GAF, a binary indicator for being in the normal functioning range of 70–100, a binary indicator if baseline and final GAF categories were the same, and a binary indicator if final GAF category was lower than baseline). We also estimated our main outcomes of functional improvement and additional treatment stratifying for the initial level of GAF and using logistic regression models. We estimated robust standard errors to account for heteroskedasticity and performed two-sided t-tests for 95% CIs. All analyses were performed using Stata version 17.
RkJQdWJsaXNoZXIy MjY0ODMw