Background: Modelling non-linear associations between an outcome and continuous patient characteristics, whilst investigating heterogeneous treatment effects, is one of the opportunities offered by individual participant data meta-analysis (IPD-MA). Splines offer great flexibility, but guidance is lacking.

Objective: To introduce modelling of nonlinear associations using restricted cubic splines (RCS), natural B-splines, P-splines, and smoothing splines in IPD-MA to estimate absolute treatment effects.

Methods: We describe the pooling of spline-based models using pointwise and multivariate meta-analysis (two-stage methods) and one-stage generalised additive mixed effects models (GAMMs). We illustrate their performance on three IPD-MA scenarios of five studies each: one where only the associations differ across studies, one where only the ranges of the effect modifier differ and one where both differ. We also evaluated the approaches in an empirical example, modelling the risk of fever and/or ear pain in children with acute otitis media conditional on age.

Results: In the first scenario, all pooling methods showed similar results. In the second and third scenario, pointwise meta-analysis was flexible but showed non-smooth results and wide confidence intervals; multivariate meta-analysis failed to converge with RCS, but was efficient with natural B-splines. GAMMs produced smooth pooled regression curves in all settings. In the empirical example, results were similar to the second and third scenario, except for multivariate meta-analysis with RCS, which now converged.

Conclusion: We provide guidance on the use of splines in IPD-MA, to capture heterogeneous treatment effects in presence of non-linear associations, thereby facilitating estimation of absolute treatment effects to enhance personalized healthcare.

Context: While urinary incontinence (UI) commonly occurs after radical prostatectomy (RP), it is unclear what factors increase the risk of UI development.

Objective: To perform a systematic review of patient- and tumour-related prognostic factors for post-RP UI. The primary outcome was UI within 3 mo after RP. Secondary outcomes included UI at 3-12 mo and ≥12 mo after RP.

Evidence acquisition: Databases including Medline, EMBASE, and CENTRAL were searched between January 1990 and May 2020. All studies reporting patient- and tumour-related prognostic factors in univariable or multivariable analyses were included. Surgical factors were excluded. Risk of bias (RoB) and confounding assessments were performed using the Quality In Prognosis Studies (QUIPS) tool. Random-effects meta-analyses were performed for all prognostic factor, where possible.

Evidence synthesis: A total of 119 studies (5 randomised controlled trials, 24 prospective, 88 retrospective, and 2 case-control studies) with 131 379 patients were included. RoB was high for study participation and confounding; moderate to high for statistical analysis, study attrition, and prognostic factor measurement; and low for outcome measurements. Significant prognostic factors for postoperative UI within 3 mo after RP were age (odds ratio [OR] per yearly increase 1.04, 95% confidence interval [CI] 1.03-1.05), membranous urethral length (MUL; OR per 1-mm increase 0.81, 95% CI 0.74-0.88), prostate volume (PV; OR per 1-ml increase 1.005, 95% CI 1.000-1.011), and Charlson comorbidity index (CCI; OR 1.28, 95% CI 1.09-1.50).

Conclusions: Increasing age, shorter MUL, greater PV, and higher CCI are independent prognostic factors for UI within 3 mo after RP, with all except CCI remaining prognostic at 3-12 mo.

Patient summary: We reviewed the literature to identify patient and disease factors associated with urinary incontinence after surgery for prostate cancer. We found increasing age, larger prostate volume, shorter length of a section of the urethra (membranous urethra), and lower fitness were associated with worse urinary incontinence for the first 3 mo after surgery, with all except lower fitness remaining predictive at 3-12 mo.

Objective: To illustrate how to evaluate the need of complex strategies for developing generalizable prediction models in large clustered datasets.

Study Design and Setting: We developed eight Cox regression models to estimate the risk of heart failure using a large population-level dataset. These models differed in the number of predictors, the functional form of the predictor effects (non-linear effects and interaction) and the estimation method (maximum likelihood and penalization). Internal-external cross-validation was used to evaluate the models' generalizability across the included general practices.

Results: Among 871,687 individuals from 225 general practices, 43,987 (5.5%) developed heart failure during a median follow-up time of 5.8 years. For discrimination, the simplest prediction model yielded a good concordance statistic, which was not much improved by adopting complex strategies. Between-practice heterogeneity in discrimination was similar in all models. For calibration, the simplest model performed satisfactorily. Although accounting for non-linear effects and interaction slightly improved the calibration slope, it also led to more heterogeneity in the observed/expected ratio. Similar results were found in a second case study involving patients with stroke.

Conclusion: In large clustered datasets, prediction model studies may adopt internal-external cross-validation to evaluate the generalizability of competing models, and to identify promising modelling strategies.