Objective: To focus on methodology of metaregression and demonstrate to clinicians, through 2 published examples, some strengths and limitations. EXAMPLE 1
Methods: Metaanalysis of data from 20 years of randomized trials of lidocaine prophylaxis in preventing primary ventricular fibrillation (VF) in myocardial infarction used separate data for control and active-treatment groups to model the risk of VF as a function of year of publication and other study characteristics. EXAMPLE 1
Results: Collinearity between pairs of predictor variables can lead to difficulty in interpreting logistic regression models. EXAMPLE 2
Methods: Metaanalysis of data from 7 trials (323 patients) measured treatment effect of immunosuppressive therapy for acute Crohn's disease by response-rate difference (RD, experimental minus control group). EXAMPLE 2
Results: Weighted least-squares regression models of the RD suggested an association between RD and various study characteristics, but collinearity again led to difficulty in interpretation of multiple regression results.
Discussion: Warning signs of colinearity include: large pairwise correlations between predictor variables, large changes in coefficients caused by the addition or deletion of other variables, and extremely large SEs for coefficients. Suggestions for coping with collinearity include: removing redundant variables from the model, reducing reliance on interpretation of coefficients for confounding variables, forming one or more summary variables, centering the data, collecting more data, and using more sophisticated regression methods.
Conclusions: Metaanalysis can explore variations in as well as summarize results of randomized trials. Although metaregression has advantages, study characteristics are often strongly associated with each other, leading to collinearity.