Statistical analysis

In the statistical analyses, we included all participants from the BRIDGE trial who completed the QIs at T₃. We analysed data in STATA IC, version 16, and set the statistical significance level to 0.05.

We performed descriptive analyses to report demographic data, quantify the quality of the received rehabilitation process, and describe the observed changes for each clinical outcome, calculated as the outcome score at T₄ minus the score for the same outcome at baseline. For the actual process indicators, there was no former established PR cut-off for high-quality care. Therefore, we used quartiles (0–25% = Q1, 25.1–50% = Q2, 50.1–75% = Q3, 75.1–100% = Q4) for the quality variable when we examined changes in outcomes by summary PR score for the process indicators.

As a preparatory analysis, we performed two regressions treating the summary PR for the quality variable as the response variable. In the first analysis, we regressed R on study centre alone. In the second regression, PR was regressed on the baseline predictors (age, sex, BMI, weekly training, comorbidity, paid employment, education level, civil status, and smoking), in addition to study centre.

We used a linear mixed model approach to assess the association between the process dimension of the quality of rehabilitation (the PR variable) and the study outcomes (goal attainment, physical function, and HRQoL, respectively). First, our primary independent variable was the summary PR for the process variables. For each outcome, its value at T₄ was treated as the response, and the fixed effects were its baseline value, the PR variable, and a variable capturing elapsed time since study start. To account for centre level clustering, we included centre as a random effect in the basic model. In the fully adjusted model, we included a wider range of baseline predictors: age, sex, BMI, weekly training, comorbidity, paid employment, education level, civil status, and smoking. In a separate analysis, the primary independent variable was replaced by the three summary PR values for the single indicators grouped into categories (Group A-C). We used the same basic and fully adjusted models as described above.

For each outcome, three models were fit: one without PR variable(s) (null model), one with the summary PR (to examine the quality variable as a sum score; alternative model I), and one with PRs for Groups A to C (to examine the quality variable as composed of the three PR variables; alternative model II). Subsequently, the association between the quality PR and the outcome was assessed by the likelihood ratio test, comparing each of the latter two models to the first. In other words, we used the likelihood ratio test to examine whether the alternative model (I or II, respectively) provided significant improvement (better fit) over the null model.

For the main outcome, we also performed mixed-logistic regression analyses in order to differentiate between those attaining minimal clinically important difference (MCID) for PSFS, and not. MCID for PSFS is 2 or more points [32] and therefore, we evaluated PSFS as a dichotomized outcome (change > = 2 yes/no). This was done first for PSFS, and next for PSFS-A1.