M.Sc. Module: Systematic Reviews

Meta-analysis: dealing with heterogeneity Dealing with heterogeneity We have already discussed the meaning and detection of heterogeneity in the previous lecture, ‘Meta-analysis: methods for quantitative data synthesis’. In this lecture we look at how to deal with it when we have it. There are a number of possibilities. First, we could decide not to pool the study estimates at all. Instead, we would carry out a narrative review. We do not get any numerical estimate. Second, we could ignore the heterogeneity and analyse the data as described in the previous lecture. We would use what is called a fixed effect model, assuming that the underlying effects are the same for all studies. As we shall see, this can result in a confidence interval which is too narrow, a pooled estimate which is difficult to interpret, and which may be biased. Third, we could explore the heterogeneity and try to explain it and remove it. We may be able to find a variable or variables which explains this heterogeneity and so give our meta-analysis estimate depending on this variable. Fourth, we could allow for the heterogeneity in our analysis and produce a much wider confidence interval, using what is called a random effects model. We shall look at all these options below. Measuring heterogeneity First, we ask how much heterogeneity is there? The chi-squared test provides a test of significance for heterogeneity, but it does not measure it. An index of heterogeneity can be defined as I 2 (Higgins and Thompson 2002), wher