For meta-analyses that use summary data, meta-regression has been proposed as a means of exploring observed heterogeneity [55]. As in any regression analysis, metaregression tries to identify significant relationships between the outcome and covariates of interest. Clearly when IPD are not available, the unit of regression is restricted to the trial. The covariates maybe constant for the entire trial, e.g. the protocol dose of drug, or a summary measure of attributes describing the patient population, e.g. mean age or percentage of males. Several variables may be modelled together and provided that reasonable data are available, the technique maybe a useful exploratory tool. However, there are limitations. Not all publications will report on all the covariates of interest (and there could be potential bias associated with selective presentation of data that have shown a positive association within the trial). If a trial is missing a covariate it drops out of the regression, limiting the power and usefulness of the analysis, which is already likely to be based on relatively few data points. Furthermore, summary data may misrepresent individual patients [1]. This is known as the ecologic fallacy or regression to the mean. What is true of a trial with a median age of sixty may not necessarily be true for a 60-year-old patient. Potentially all the benefit could have been shown in the seventy year olds and none in the fifty year olds. It should always be borne in mind that a significant association does not prove causality and should rather be regarded as hypothesis generating.

0 0

Post a comment