OBJECTIVES: Network meta-analysis (NMA) is a powerful tool for comparing multiple interventions simultaneously by combining direct and indirect evidence. However, ensuring the accuracy and reliability of NMA models is challenging due to potential biases and inconsistencies in the data. Predictive cross-validation is a method that can enhance model performance by evaluating how well a model predicts new data, thus helping to identify and mitigate sources of error. This study aims to enhance the robustness of NMA models by implementing predictive cross-validation techniques.

METHODS: A network meta-analysis was conducted using data from multiple randomized controlled trials (RCTs) comparing various oncology treatments. The model incorporated initial assumptions about trial conditions and used random effects to account for treatment diversity. Each trial's contribution to the overall analysis was evaluated by estimating treatment effects relative to a control, considering trial-specific characteristics and variability. To improve model accuracy and reliability, predictive cross-validation was employed. This involved systematically excluding each study, refitting the NMA model, and predicting the outcomes of the excluded study. Bayesian methods compared the predicted outcomes with actual results, providing a robust framework for assessing model performance.

RESULTS: Using a "leave one out" approach, we conducted cross-validation to evaluate trial 16 as an "outlier," resulting in a p-value of 0.004. This indicates that achieving a result as extreme as trial 16 is unlikely based on our model for the remaining data. Convergence occurred after 60,000 burn-in iterations, with results derived from 180,000 samples across three independent chains.

CONCLUSIONS: Implementing predictive cross-validation in network meta-analysis enhances the assessment of model performance and aids in the detection of influential outliers. This approach improves the reliability and robustness of NMA models, leading to more accurate and credible conclusions in comparative effectiveness research. Future research should prioritize predictive distribution checks to refine inferences and uphold NMA integrity.