Development and validation of an exercise adherence prediction model 111 Model development In the backward elimination procedure (including bootstrapping), variables that remained in the model in more than half of the bootstrap samples (p < 0.05) were included in the final prediction model. Four predictors were remained and entered the model: intention, MRC-score, depression and alliance. Excluded were attitude, subjective norm, PBC, and exercise history, because they did not have a significant (p > 0.05) relation with adherence. The logistic regression analysis results of the four included variables are listed in Table 4. Table 4 Logistic regression analysis of predictors for adherence in patients with COPD Intercept and variables β coefficients p-value Odds Ratio (95% CI) Shrunken β coefficients* Intercept -4.996 -4.108** Intention 0.238 0.001 1.276 (1.109-1.469) 0.203 Depression -0.171 0.001 0.839 (0.738-0.952) -0.147 MRC -1.107 0.004 0.321 (0.148-0.695) -0.950 Alliance 0.064 0.008 1.068 (1.017-1.120) 0.055 * Result of bootstrapping technique used to correct for overoptimistic estimation of model. **After intercept adjustment; adding the calibration intercept (-1.015) to the model intercept (- 5.123). MRC: Medical Research Council dyspnea scale as a measure of disability in patients with chronic obstructive pulmonary disease; CI: Confidence Interval Model evaluation The AUC in the primary model was 0.79 (95% CI, 0.72-0.85); p = 0.00 (Figure 1A). After bootstrap internal validation, the optimism-corrected AUC was the same; 0.79 (95% CI, 0.72-0.85); p = 0.00, suggesting good discrimination (Figure 1B). The prevalence of adherence was 42.9% (84/196). The average estimated probability of adherence given by the prediction model was 41.9%, which indicates good estimations. The calibration slope and calibration-in-the-large were respectively 1.026 and -0.007 in the primary model (Figure 2A). After internal validation the probability of adherence given by the shrunken model was 24.4%, which indicates that there is a tendency to give underestimated scores for adherence. The new calibration slope was 1.198 and calibration-in-the-large was 1.015 (Figure 2B). So, after internal validation poorly calibrated predictions were found, with a calibration intercept far from 0. Intercept adjustment was performed by adding the calibration intercept (1.015) to the model intercept (-5.123). After this intercept update, the calibration curve of the intercept-adjusted model was close to the diagonal reference line of perfect moderate calibration (Figure 2C). The
RkJQdWJsaXNoZXIy MjY0ODMw