A common problem in the analysis of multiple data sources, including individual participant data meta-analysis (IPD-MA), is the misclassification of binary variables. Misclassification may lead to biased estimators of model parameters, even when the misclassification is entirely random. We aimed to develop statistical methods that facilitate unbiased estimation of adjusted and unadjusted exposure-outcome associations and between-study heterogeneity in IPD-MA, where the extent and nature of exposure misclassification may vary across studies.

We present Bayesian methods that allow misclassification of binary exposure variables to depend on study- and participant-level characteristics. In an example of the differential diagnosis of dengue using two variables, where the gold standard measurement for the exposure variable was unavailable for some studies which only measured a surrogate prone to misclassification, our methods yielded more accurate estimates than analyses naive with regard to misclassification or based on gold standard measurements alone. In a simulation study, the evaluated misclassification model yielded valid estimates of the exposure-outcome association, and was more accurate than analyses restricted to gold standard measurements.

Our proposed framework can appropriately account for the presence of binary exposure misclassification in IPD-MA. It requires that some studies supply IPD for the surrogate and gold standard exposure, and allows misclassification to follow a random effects distribution across studies conditional on observed covariates (and outcome). The proposed methods are most beneficial when few large studies that measured the gold standard are available, and when misclassification is frequent.

Meta-analysis of randomized controlled trials is generally considered the most reliable source of estimates of relative treatment effects. However, in the last few years, there has been interest in using non-randomized studies to complement evidence from randomized controlled trials. Several meta-analytical models have been proposed to this end. Such models mainly focussed on estimating the average relative effects of interventions. In real-life clinical practice, when deciding on how to treat a patient, it might be of great interest to have personalized predictions of absolute outcomes under several available treatment options. This paper describes a general framework for developing models that combine individual patient data from randomized controlled trials and non-randomized study when aiming to predict outcomes for a set of competing medical interventions applied in real-world clinical settings. We also discuss methods for measuring the models' performance to identify the optimal model to use in each setting. We focus on the case of continuous outcomes and illustrate our methods using a data set from rheumatoid arthritis, comprising patient-level data from three randomized controlled trials and two registries from Switzerland and Britain.

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.

If individual participant data are available from multiple studies or clusters, then a prediction model can be externally validated multiple times. This allows the model's discrimination and calibration performance to be examined across different settings. Random-effects meta-analysis can then be used to quantify overall (average) performance and heterogeneity in performance. This typically assumes a normal distribution of 'true' performance across studies. We conducted a simulation study to examine this normality assumption for various performance measures relating to a logistic regression prediction model. We simulated data across multiple studies with varying degrees of variability in baseline risk or predictor effects and then evaluated the shape of the between-study distribution in the C-statistic, calibration slope, calibration-in-the-large, and E/O statistic, and possible transformations thereof. We found that a normal between-study distribution was usually reasonable for the calibration slope and calibration-in-the-large; however, the distributions of the C-statistic and E/O were often skewed across studies, particularly in settings with large variability in the predictor effects. Normality was vastly improved when using the logit transformation for the C-statistic and the log transformation for E/O, and therefore we recommend these scales to be used for meta-analysis. An illustrated example is given using a random-effects meta-analysis of the performance of QRISK2 across 25 general practices.