Analysis of Adverse Events using Literature Data: a Simulation Study
F. Ezzet
Pharsight a Certara Company, St Louis, MO, USA
Objectives: To investigate the power of literature data in the analysis of Adverse Event (AE)
Methods: In meta analysis, a clinical outcome expressed as a proportion, such as %AE or %responders, may be treated as a continuous variable and modeled using standard mixed effects algorithms. Lack of fit will likely occur if proportions are close to the boundaries 0% or 100%. A Logit transformation (on proportions) in some instances helps but dose not resolve the problem completely. Instead, proportions are converted to a binary outcome and are modeled as such, using, say, glme in Splus or glmer in R. We explored the properties of the approach using simulated data that closely resemble outcome under clinical settings. The objectives were to determine 1) ability to estimate treatment effects accurately, 2) consequence of mixing studies of different sample size (N) and 3) usefulness of standard modeling diagnostics tools. Recently, this approach was implemented using literature data to investigate influence of treatment on AE's in cancer trials, and in another project on rates of AE's and dropout in osteoarthritis trials in patients treated with opioids [1]. The findings from patient data were compared with the simulations, but are not be reported here.
Results: Under different scenarios, model estimates using glme were in close agreement with the true values, including estimates of random effects and residual variances. Since the response is binomial, the contribution to the likelihood is inherently proportional to sample size, resulting in a proportional effect on model estimates, independent of random effects. Standard goodness of fit diagnostics based on fitted proportions can be misleading, especially when N is small. Instead, diagnostics should be based on residuals, e.g. Pearson residuals or adjusted deviance residuals. The findings of the meta analyses from the cancer trials and from the osteoarthritis trials where consistent with those of the simulation work.
Conclusion: Unlike meta analysis of mean data[2,3], no ad hoc weighting by sample size is necessary when outcome is a proportion. More importantly, expressing proportions as a binomial variable essentially re-constructs the observed outcome at the patient level. In meta analysis, the pooled trials thus retain original sample size and significantly increase statistical power. This is particularly useful when investigating rare events from a large number of small studies, as in early phase cancer trials.
References:
[1] F. Ezzet, K. Prins, M. Boucher. PAGE 19 (2010) Abstr 1771 [www.page-meeting.org/?abstract=1771]
[2] Ezzet, F., Ravva, P., Tensfeldt, T. Model-Based Literature Meta-Analysis: Virtues and Limitations, ACoP, Tucson, Arizona, March, 2008 (http://tucson2008.go-acop.org/pdfs/17-Ezzet_FINAL.pdf )
[3] Ezzet F. The Role of Literature-Based Disease Progression Models to
Support Knowledge Management and Decision-Making in Clinical Drug Development., AAPS National Biotechnology Conference, May 2008, Toronto, Canada
