Every effect size has its place: a commentary on the avoidance of pre–post effect sizes

Along with the increasing popularity of meta-analysis, their methods have very much improved in the last decades, leading to highly sophisticated methods, for example multiple treatment comparisons and advanced multiple regression models (Sutton & Higgins, 2008). However, these improvements of meta-analytic methods should not obstruct one's view of the basic decisions and assumptions, because these may have a significant impact on result and interpretation of a meta-analysis, as was nicely illustrated by Cuijpers et al. (2016) in this journal for the choice of the effect size.

In their paper, the authors illustrate that pre–post effect sizes do not reflect a specific treatment effect only, but will also be affected by various processes, for example spontaneous remission. I think this cannot be questioned, but I am less sure if I am in support of a general statement that suggests avoiding pre–post effect sizes in meta-analysis. Thus, I will play the devil's advocate in defence of the pre–post effect sizes and will give some thoughts to their usefulness.

As already mentioned by the authors, pre–post effect sizes can for example be informative in a comparison of efficacy and effectiveness trials (van der Lem et al. 2012). Most effectiveness trials use single group designs, and therefore it is impossible to calculate comparative effect sizes. Consequently, excluding these trials by focusing on controlled effect sizes only may miss a substantial part of the evidence. This may be less important for trials of single interventions, but it is of major importance when summarising evidence in areas where randomised controlled trials are rarely conducted, for example in research on complex interventions (Girlanda et al. 2017). Besides this rather pragmatic view, pre–post effects may be more interesting for clinicians than the comparative effects, because I suppose many clinicians are interested in an overall improvement of their patient, irrespective of whether the effect is a specific or unspecific effect of the medication. The pre–post effects may therefore be more interesting for clinicians comparing their treatment outcomes to the literature. This may be particularly true for data derived from placebo controlled trials, which are testing against a treatment option, which is not available for clinicians in daily practice. Due to the drug approval requirement by the drug agencies (Barbui & Bighelli, 2013), the basis for the body of evidence for drug treatments is derived from these trials, and one may argue, that the focus on reporting of relative drug placebo differences may contribute to the observation of the efficacy-effectiveness gap.

In their paper, Cuijpers et al. (2016) also highlight the methodological problems connected to the pre–post effect sizes, and I very much appreciate that the authors address the frequently ignored statistical dependency of the pre–post measurements and illustrate the use of common assumptions with a real data example. As illustrated by their example, the pre–post effect size estimates can be misleading if the assumption of the pre–post correlation is incorrect. However, for a wide range of assumed correlation (e.g., r = 0.0 to r = 0.5 compared a true correlation 0.18 for Hamilton Depression Rating Scale) the effect size estimations were quite robust and may be within an acceptable error margin. Furthermore, the authors point out that heterogeneity can be very high in meta-analyses using pre–post effects. Although very high heterogeneity of effects can also be found in comparative effectiveness research (Koslowski et al. 2016), it is plausible to assume that meta-analyses using pre–post effect sizes show higher heterogeneity of the treatment effects. These analyses will more often include non-controlled studies and may have less restrictive inclusion criteria. However, the heterogeneity itself contains some important information, thus the exploration of the heterogeneity may be valuable in understanding moderators of the treatment effects.

As shown by Morris & DeShon (2002), every effect size can be biased and the sources of bias differ between effect sizes. Cuijpers et al. (2016) mention that the pre–post effect size is affected, for example by time effects. However, the comparative effect size may be affected by group-related effects, which may not be directly related to the treatments. In a double-blind placebo controlled trial, adverse events of the medication may debunk group assignment, and this may influence the expectation and ratings of clinicians and patients (Basoglu et al. 1997; Moncrieff et al. 1998). Although this bias related to the study design and conduct, it will affect a comparative effect size and may result in an overestimation of the group difference. A pre–post effect size is also prone to biased ratings, but because the effect sizes can be derived at group level, they may be more helpful in evaluating such bias. Pre–post effect sizes could be used to identify unusual low effect size in the placebo group, which is hard to see in analysis of comparative effect sizes.

In conclusion, pre–post effect sizes may not be a good choice for summarising experimental research, and it is difficult to use them properly. However, every effect size has its place, and the decision in favour of a specific effect size measure should depend primarily on the research question.