ABSTRACT
In many systematic reviews and meta-analyses statistical heterogeneity occurs. This means that the effect measure differs between trials more than it could be expected under homogeneity of effect. However, the question might be not if there is heterogeneity but more how its amount could be quantified. As Higgins and Thompson (2002) put it:
A small amount of heterogeneity in a MAIPD might lead to a valid analysis similar to an analysis under complete homogeneity, whereas a large amount of heterogeneity might require larger efforts in coping with it including a more substantial way of modeling it, as has been suggested in previous chapters. One historic form of quantifying heterogeneity is the Q-statistic usually attributed to Cochran (1954) (see also Whitehead and Whitehead (1991) or Normand (1999)). It is defined in our setting as
Q = k∑ i=1
wi(φˆi − φ¯)2 (9.1)
where wi = 1/V ar(φˆi) for φi = log θi and
φ¯ = ∑k
.
