Clinical trials: how to assess confounding and why so

Background: In large randomized controlled trials the risk of random imbalance of the covariates is mostly negligible. However, with smaller studies it may be substantial. In the latter situation assessment and adjustment for confounders is a requirement in order to reduce a biased assessment of the treatment comparison. Objective: In the current paper three methods for confounding assessment and adjustment are reviewed for a nonmathematical readership. Methods: First method, subclassification: the study population is divided into subclasses with the same subclass characteristic, then, treatment efficacy is assessed per subclass, and, finally, a weighted average is calculated. Second method, regression modeling: in a multivariable regression model with treatment efficacy as independent and treatment modality as dependent variable, the covariates at risk of confounding are added as additional dependent variables to the model. An analysis adjusted for confounders is obtained by removing the covariates that are not statistically significant. Third method, propensity scores: each patient is assigned several odds ratios (ORs), which are his/her probability, based on his/her covariate value of receiving a particular treatment modality. A propensity score per patient is calculated by multiplying all of the statistically significant ORs. These propensity scores are, then, applied for confounding adjustment using either subclassification or regression analysis. Conclusions: The advantages of the first method include that empty subclasses in the treatment comparison are readily visualized, and that subclassification does not rely on a linear or any other regression model. A disadvantage is, that it can only be applied for a single confounder at a time. The advantage of the second method is, that multiple variables can be included in the model. However, the number of covariates is limited by the sample size of the trial. An advantage of the third method is, that it is generally more reliable and powerful with multiple covariates than regression modeling. However, irrelevant covariates and very large / small ORs reduce power and reliability of the assessment. The above methods can not be used for the assessment of interaction in the data.