The impact of individual patient data in a network meta-analysis: An investigation into parameter estimation and model selection

The use of individual patient data (IPD) in network meta-analysis (NMA) is becoming increasingly popular. However, as most studies do not report IPD, most NMAs are performed using aggregate data for at least some, if not all, of the studies. We investigate the benefits of including varying proportions of IPD studies in an NMA. Several models have previously been developed for including both aggregate data and IPD in the same NMA. We performed a simulation study based on these models to examine the impact of additional IPD studies on the accuracy and precision of the estimates of both the treatment effect and the covariate effect. We also compared the deviance information criterion (DIC) between models to assess model fit. An increased proportion of IPD resulted in more accurate and precise estimates for most models and datasets. However, the coverage probability sometimes decreased when the model was misspecified. The use of IPD leads to greater differences in DIC, which allows us choose the correct model more often. We analysed a Hepatitis C network consisting of 3 IPD observational studies. The ranking of treatments remained the same for all models and datasets. We observed similar results to the simulation study: The use of IPD leads to differences in DIC and more precise estimates for the covariate effect. However, IPD sometimes increased the posterior SD of the treatment effect estimate, which may indicate between study heterogeneity. We recommend that IPD should be used where possible, especially for assessing model fit.