A relationship of baseline risk to treatment effect size has been sugg
ested as a possible explanation of between-study heterogeneity in meta
-analyses. To address this question, we develop regression models to e
xamine the relationship between the logits (or other response measure)
in the intervention and control groups. A weighted least squares (WLS
) approach is described that allows for the heterogeneous sampling var
iation in the two groups, together with a correction of the coefficien
ts for sampling error. Two approximate maximum likelihood (ML) solutio
ns are also obtained, with or without an assumption of equal variances
between groups within studies. A closed form ML solution exists with
the assumption of equal variances. Both methods appear preferable to a
previously suggested regression model of the log odds ratio on the co
ntrol event rate; the methods proposed here use the same scale of meas
urement for both study groups, and eliminate an artifactual correlatio
n in the regression error structure. The ML. approach may be preferabl
e because of its symmetric treatment of study groups, but WLS is more
easily implemented with standard software. The methods are illustrated
with data from meta-analyses on pre-term delivery and on therapies to
lower serum cholesterol. (C) 1997 by John Wiley & Sons, Ltd.