69 Inclusion of both patient level and study-level covariates in a meta-analysis

Abstract

69 INCLUSION OF BOTH PATIENT LEVEL AND STUDY-LEVEL COVARIATES IN A META-ANALYSIS Julian P. T. Higgins and Anne Whitehead Institute of Child Health University College London London, United Kingdom A me&analysis using individual patient data allows a detailed investigation of the effect of covariates on the efficacy of a treatment. This paper presents a method for including both patient-level covariates, such as age, sex and baseline severity of disease, and study-level covariates, such as dose and other aspects of design, in such an investigation. A treatment effect and interactions between treatment and covariates are included in the model. The analysis, complicated by the presence of different within-trial variances, may be performed using the Gibbs sampler, a popular technique in Bayesian statistics. The implications are discussed of imposing the standard random treatment effects between studies to account for other sources of heterogeneity. The method is general to most types of outcome variable. As an example, we analyse individual patient data from twelve trials of tacrine for treating Alzheimer’s disease (AD). These were acquired for a systematic review by the Cholinergic Trialists’ Collaboration under the auspices of the Cochrane Dementia and Cognitive Impairment Group. It has been hypothesized that tacrine should work better in moderately severe cases of AD than in mild or severe cases, that the age and sex of the patient should have little impact on the treatment and that there is a positive dose response relationship. The first three hypotheses are investigated using individual patient data, and the last using average dose in each study, with change in cognitive function (as measured on an ordinal scale) as the outcome variable. 70 CHOOSING AN APPROPRIATE DICHOTOMOUS EFFECT MEASURE FOR META-ANALYSIS: EMPIRICAL EVIDENCE OF THE APPROPRIATENESS OF THE ODDS RATIO AND RELATIVE RISK J.J. Decks, D.G. Altman, G. Dooley and D.L.S. Sackett Centre for Statistics in Medicine Oxford, United Kingdom Systematic reviews of randomized controlled trials have traditionally used odds ratios (ORs) to analyze and express relative treatment effects for event-based outcomes, typically estimating the common odds ratio using the Mantel-Haenszel method for combining odds ratios from stratified data. A similar method is available for combining relative risks (RRs). Little guidance is available on choosing between the OR and RR models. The debate reflects issues of statistical ideology (concerns about bounding of RR and behavior of the metric at extremes) and of interpretational simplicity (the OR not being easily interpreted or applied at the bedside and having a potential for gross overinterpretations of the treatment effects when event rates are