13.6.1  What is different when including non-randomized studies?

Review authors should expect greater heterogeneity in a systematic review of NRS than a systematic review of randomized trials. This is due to the increased potential for methodological diversity through variation between primary studies in their risk of selection bias, variation in the way in which confounding is considered in the analysis and greater risk of other biases through poor design and execution. There is no way of controlling for these biases in the analysis of primary studies and no established method for assessing how, or the extent to which, these biases affect primary studies (but see Chapter 8).


There is a body of opinion that it is appropriate to pool results of non-randomized studies when they have large effects, but the logic of this view can be questioned. NRS with large effects are as likely (perhaps more likely) to be biased and to be heterogeneous as NRS with small effects. Judgements about the risk of bias and heterogeneity should be based on critical appraisal of the characteristics and methods of included studies, not on their results.


When assessing similarity of studies prior to a meta-analysis, review authors should also keep in mind that some features of studies, for example assessment of outcome not masked to intervention allocation, may be relatively homogeneous across NRS but still leave all studies at risk of bias.


If authors judge that included NRS are both reasonably resistant to biases and relatively homogeneous in this respect, they may wish to combine data across studies using meta-analysis (Taggart 2001). Unlike for randomized trials, it will usually be appropriate to analyse adjusted, rather than unadjusted, effect estimates, i.e. analyses that attempt to ‘control for confounding’. This may require authors to choose between alternative adjusted estimates reported for one study. Meta-analysis of adjusted estimates can be performed as an inverse-variance weighted average, for example using the ‘Generic inverse-variance’ outcome type in RevMan (see Chapter 9, Section 9.4.3). In principle, any effect measure usedin meta-analysis of randomized trials can also be used in meta-analysis of non-randomized studies (see Chapter 9, Section 9.2), although the odds ratio will commonly be used as it is the only effect measure for dichotomous outcomes that can be estimated from case-control studies, and is estimated when logistic regression is used to adjust for confounders.


One danger is that a very large NRS of poor methodological quality (for example based on routinely collected data) may dominate the findings of other smaller studies at less risk of bias (perhaps carried out using customized data collection). Authors need to remember that the confidence intervals for effect estimates from larger NRS are less likely to represent the true uncertainty of the observed effect than are the confidence intervals for smaller NRS (see Section, although there is no way of estimating or correcting for this.