Marieke van Son

143 PREDICTION MODELS FOR BF AFTER FOCAL SALVAGE HDR-BT Statistical analysis Baseline characteristics and survival Normally distributed determinants are presented as mean (± standard deviation [SD]). Skewed variables are presented as medians with interquartile ranges (IQR). Frequencies and percentages are used for categorical data. The Kaplan-Meier method was used to estimate biochemical disease-free survival (bDFS). For comparisons between groups, the log-rank test was used. Missing data handling Missing data was considered to be missing at random. Multiple imputation by chained equations was used to impute missing data, creating 20 imputation datasets. All pre- dictors listed above, additional patient and treatment characteristics listed in Supple- mentary File A, the outcome, and the cumulative baseline hazard, calculated with the Nelson-Aalen function, were included in the imputation procedure[22,23]. All subse- quent modelling steps were pooled over the 20 imputation datasets. Functional form of continuous predictors Before fitting the multivariable model, non-linear relationships between continuous predictors and the outcome were assessed visually by plotting the predictors against log-hazard using restricted cubic splines with three knots (10th, 50th, and 90th per- centile). In case of visible non-linearity, spline transformations were tested against linear modelling through univariable and multivariable Cox proportional hazards models (likelihood-ratio test). If model fit improved significantly, a spline-transformation was used. For pre-salvage PSA, a natural logarithm-transformation was used based on literature and model fit in our dataset[24]. Model development In case correlations between candidate variables were ≥0.75, the clinically most rel- evant variable was chosen for multivariable testing. MRI-based T-stage showed high correlation with seminal vesicle involvement (correlation coefficient 0.78). Based on clinical judgement, MRI-based T-stage was therefore excluded from multivariable re- gression analysis. A multivariable Cox proportional hazards regression model was fitted, providing hazard ratios (HR) with 95% confidence intervals (CI). Stepwise back- ward elimination was performed, using lowest Akaike’s Information Criteria (AIC) for selection[25]. No interactions were assessed due to the limited sample size. Model assumptions For both models the assumptions of the Cox proportional hazards model were checked. The proportionality assumption was assessed using Log-Log curves and Schoenfeld residuals for categorical and continuous variables, respectively. Linearity of continuous 8