Prediction interval

In Meta-Essentials, the larger, green, interval around the combined effect size on the bottom row of the forest plot is the prediction interval (Figure 1). (See Borenstein et al., 2009: 129-131, for how a prediction interval is estimated.) The 95% prediction interval gives the range in which the point estimate of 95% of future studies will fall, assuming that true effect sizes are normally distributed through the domain. Because the prediction interval is estimated based on the effect sizes observed in the studies that are meta-analysed, the prediction interval corresponds more or less (depending on whether sampling variation in the meta-analysed studies is large or small) with the range of effect sizes that are meta-analysed and that are represented in the forest plot. This implies that the prediction interval can only “predict” with some accuracy if no relevant selection bias exists in the set of populations that have been studied (i.e., if the populations of which the effect size estimate is included in the meta-analysis are “representative” for the domain). If selection bias exists, then effect sizes observed in future studies might occur beyond the limits of the prediction interval. Because selection bias is more likely than not, it is recommended to interpret the prediction interval as a description of the range of observed effect sizes rather than as a prediction of the range of effect sizes that will be observed in future studies (despite its name “prediction” interval).