Should you adjust for baseline characteristics within randomized controlled trials?

At first glance, one may think the answer is no. Randomization should insure that baseline characteristics are balanced across trial arms. In practice, however, sometimes baseline characteristics due differ somewhat by trial arm simply by chance, especially in smaller randomized controlled trials (RCTs). A JAMA Guide to Statistics and Methods by Holmberg et al. 2022 provides justification for why you would want to adjust for baseline characteristics in many cases.

…adjusting for baseline characteristics in the analysis of RCTs is advised by both the European Medicines Agency and the US Food and Drug Administration because it may improve statistical efficiency, enhancing the ability to draw a reliable conclusion from the available dat

It is most important to adjust for baseline characteristics if there are differences in characteristics across treatment arms and if the characteristic(s) of interest are known prognostic factors or effect modifiers. Ideally, these characteristics would be pre-specified in the trial protocol. If a researcher does decide to make these adjustments, they must determine which variables to adjust for, the statistical method to use, how to handle missing data, and how to report the unadjusted and adjusted results.

Common statistical adjustment approaches include regression model (linear for continuous variables, negative binomial for count variables, logistic for binomial variables), but other methods are possible (eg, inverse probability of treatment weighting).

What do you gain (statistically) by adjusting for baseline characteristics?

By accounting for factors influencing the outcome other than the randomly assigned intervention, adjustment leads to increased statistical power (ie, the ability to detect a treatment effect when present) and may increase precision in the estimation of the treatment effect, depending on the type of outcome. Analyses of hypothetical trials have suggested that the relative increase in effective sample size may be up to 20% by adjusting for baseline characteristics, although the actual increase is strongly dependent on the prognostic value of the baseline characteristics included in the model.

Note, however, that there are some limitations and concerns to keep in mind when considering adjusting for baseline characteristics.

Adjustment for nonprognostic variables will not lead to an increase in statistical efficiency and could potentially decrease precision in the estimation of treatment effects (ie, widen confidence intervals) or decrease the statistical power even compared with an unadjusted analysis. Post hoc selection of variables for adjustment (eg, based on the magnitude of the observed imbalance between treatment groups or on an evaluation of the effect of adjusting for different variables on the results of the analysis) can lead to bias in the estimates of the treatment effects.

Additionally, in cases where there are stratified trials (e.g., strata based on trial sites), adjustment may be problematic if there are per observations per strata as it will be difficult to disentangle differences due to baseline characteristics and those due to being in a specific strata.