A Meta-Analysis Method Based on Simulated Individual Patient Data
Yusuke Yamaguchi1, Wataru Sakamoto1, Shingo Sirahata1, Masashi Goto2
1Division of Mathematical Science, Graduate School of Engineering Science, Osaka University, Toyonaka, Osaka, Japan; 2Biostatistical Research Association, Nonprofit Organization, Toyonaka, Osaka, Japan

In clinical evaluation processes, meta-analysis is carried out to synthesis results of several trials. However, most of meta-analysis methods are based on summary statistics reported from each trial, so that a pooled effect size is estimated with ignoring scheme of sampling individual patient data (IPD), which should have been measured in each trial. Above all, meta-regression models, which are techniques for modeling the relationship between an effect size and trial-level covariates with intending to search characteristic factors, are often criticized (Berlin et al., 2002). In identifying characteristics of patients by meta-regression models, statistics on background factors of patients summarized for each trial, such as a mean age and a proportion of male patients, are included as covariates in the models. This means to evaluate characteristics of patients expediently with unavailable patient-specific covariates replaced by trial-specific covariates. It often involves a technical issue that is referred to as ecological bias, and lead to limitation in interpretation. Some meta-analysis methods based on IPD models have been discussed among many researchers as alternative solutions to this kind of problems; however, they cannot always be applied to all kind of situations due to difficulties in obtaining IPD.

A meta-analysis method based on simulated IPD is suggested, that is, we reconstruct quasi-IPD by statistical simulation using available summary statistics. Thus more flexible and more comprehensive statistical methods can be applied to this reconstructed data, and difficulties in IPD collection would be cleared out. Moreover going back to the data on individual patient profiles enables us to make inference on characteristics of actual interests. The power to detect interaction effects between treatment and covariates is especially focused on in the case of comparison between two groups. It has been well-known that the detection of this interaction effects in the meta-regression model tends to give lower power due to the ecological bias (Simmonds & Higgins, 2007), while the IPD model, which takes variability in covariates into account, gives higher power. A simulation study found that the model applied to simulated IPD improved the power to detect interaction effects over that in the meta-regression model, under various situations which arises the ecological bias.


Berlin J.A., Santanna J., Scmid C.H., Szczech L.A. & Feldman H.I. (2002). Individual patient versus group-level data meta-regressions for the investigation of treatment effect modifiers: Ecological bias rears its ugly head. Statistics in Medicine, 21, 371-87.

Simmonds M.C. & Higgins J.P.T. (2007). Covariate heterogeneity in meta-analysis: criteria for deciding between meta-regression and individual patient data. Statistics in Medicine, 26(15), 2982-2999.

Keywords: meta-analysis; statistical simulation; individual patient data; treatment-covariate interaction

Biography: 2010. 3. Passed through the master course of the division of Mathematical Science, Graduate School of Engineering Science, Osaka University.