Statistical challenges in individual patient data meta analysis
In a typical meta analysis, estimates of the parameter of interest (e.g. odds ratio) are extracted from the literature or by contacting researchers and pooled together. In contrast, for IPD-MA line-by-line patient data are obtained from each study.
IPD-MA data permit researchers to define exposures and outcomes consistently across studies, and to analyze the association of interest consistently (e.g. adjusting for the same confounders), which may minimize heterogeneity. IPD-MA are of particular value for synthesis of observational studies. In traditional MAs based on published data, it is difficult to investigate differences between study results, to adjust for differences in populations across studies and to pool effects that have been adjusted for other variables.
IPD-MA have higher power than meta-regression to detect covariate-treatment interactions, and are preferable when the aim is to estimate interactions with patient-level covariates. IPD-MA are not prone to ecological bias, because patient-level data are not aggregated.
While in some cases IPD-MA and aggregate data MA may give similar results, this is unlikely when evaluating treatment-covariate interactions, incorporating nonlinear relationships, when trials are small, and there is heterogeneity across trials, and particularly for pooling of non-randomized studies that may need to adjust for several confounders.
For these reasons, IPD-MA are considered the gold-standard of MA, despite the complexity and cost of collecting the data, and are published with increasing frequency.
For IPD-MA, two broad analytic strategies (one- and two-step approaches) are possible; both preserve the clustering of subjects within studies, comparability of study arms, and may be either fixed or random. A fixed effects analysis assumes that the estimated effect is the same across all studies; a random effects analysis assumes that the estimated effect varies across studies due to differences in patient populations, study procedures, etc.
A one-step approach offers more flexibility to explore the differences that may exist between patients in the same study as well as across studies. Overall, most statisticians and meta-analyst methodologists agree that a one-step approach is better and more flexible than a two step approach.
While the biggest drawback of an IPD-MA is the time/expense to assemble the data, our experience demonstrates that despite many advantages, the wide range of methods used and lack of a standardized data analytics plan is also a serious problem. Compared to conventional MA, methods for IPD-MA are described as “more complex and not well-known”,14 maybe due to several open questions about the analytic plan. Next, we address these.
IPD-MA are the gold standard of MA, offering many opportunities for the sophisticated modelling of dose-response curves, effect of and adjustment for patient-level covariates as well as treatment interactions when a one-step approach is used. However, several methodologic challenges remain. This project aims to investigate these methodologic challenges and propose the best data analytic strategies. This project will introduce students to research in Biostatistics – that is at the intersection of Statistics and Epidemiology.
