Controlling for confounding

Imbalances in prognostic factors in NRS (e.g. ‘confounding by indication’ (Grobbee 1997)) must be accounted for in the statistical analysis. There are several methods to control for confounding. Matching, i.e. the generation of similar intervention groups with respect to important prognostic factors, can be used to lessen confounding at the study design stage. Stratification and regression modelling are statistical approaches to control for confounding, which result in an estimated intervention effect adjusted for imbalances in observed prognostic factors. Some analyses use propensity score methods as part of a two-stage analysis. The probability of an individual receiving the experimental intervention (the propensity score) is first estimated according to their characteristics using a logistic regression model. This single summary measure of case-mix is then used for matching, stratification or in a regression model.



The selection of patients with similar values for important prognostic factors results in more comparable groups. Therefore, matching can be seen as a type of confounder adjustment. Matching can be either at the level of individual patients (i.e. one or more control participants are selected who have a similar characteristics to an intervention participant) or at the level of participant strata (i.e. selecting participants so that there are roughly the same number of control participants in one stratum, for example 60 years or older, as in the intervention group). Where direct matching has been used, the paired nature of the data has to be considered in the statistical analysis of a single study in order to obtain appropriate confidence intervals for the estimated effect of the intervention. Matching on a single measure such as the propensity score is easier to achieve than matching individuals with a particular set of characteristics.



Stratification involves the division of participants into subgroups with respect to categorical (or categorized quantitative) prognostic factors, for example classifying age into decades, or weight into quartiles. The intervention effect is then estimated in each stratum and a pooled estimate is calculated across strata. This procedure can be interpreted as a meta-analysis at the level of an individual study. For dichotomous outcomes, the Mantel-Haenszel method is often used to estimate the overall intervention effect, with versions available for the odds ratio, the risk ratio and the risk difference as measures of intervention effect. Again, the propensity score may be used as the stratification variable.



In a modelling approach, information on intervention and prognostic factors is incorporated into a regression equation. Advantages of regression models include the possibility of incorporating quantitative factors without categorization and the possibility of modelling trends in confounders measured on an ordinal scale. For dichotomous outcomes, a logistic regression model is almost always used to estimate the adjusted intervention effect. Thus, the odds ratio is (implicitly) used as the measure of intervention effect. Regression models are also available for risk ratio and absolute risk reduction measures of effect but these models are rarely used in practice. A linear regression model is typically used for continuous outcomes (perhaps after transformation of one or more variables), and a proportional hazards regression (Cox regression) model is typically used for time-to-event data.  Regression models may also use the propensity score alone or in combination with other participant characteristics as explanatory variables.


Review authors should acknowledge that in any non-randomized study, even when experimental and control groups appear comparable at baseline, the effect size estimate is still at risk of bias due to residual confounding. This is because all methods to control for confounding are imperfect, for example for the following reasons.: