I 8. ‘ z

Don’t worry I’m searching for a Joe Rogan of the “toddlers shouldn’t fly planes” party, and then this problem will be fixed

9/9 So any test showing baseline imbalance in covariates is by definition a Type I error.

8/9 fact that there is not a difference in covariates between conditions in the population. Randomization ensures that the theoretical populations of those assigned to Treatment A and those assigned to Treatment B are identical.

7/9 purpose of inferential statistics is to measure a subset of a population (a sample) and make an inference about the larger population. When it comes to random assignment to condition, absent some fatal flaw in which patients were not actually randomly assigned, we know for a

6/9 Additionally, even if we accept the premise that we want to balance covariates, tests for baseline imbalance, which report p-values to support statements that a covariate is balanced or not, are inherently invalid. p-values are an output of inferential statistical tests. The

5/9 reading of tea leaves, like trying to figure out if a nonsignificant result is due to low power or no effect. The answer is unknowable.

4/9 difference when there isn’t one, and a subset of those cases will be because impactful covariates that are the true causes ended up imbalanced. We accept that possibility with our chosen alpha. Trying to figure out if a result is due to baseline imbalance seems like just more

3/9 matter if covariates are “balanced”. It is true that large imbalances in covariates across conditions at baselines can lead treatments to look different even if they aren’t. But this possibility is already baked into the chosen alpha. Some studies by chance will find a

2/9 it breaks the link between X (treatment condition) and potential confounders, so that the observed effect of X can be attributed to the difference between treatments and not some other variable (hence why randomization is said to prioritize internal validity). It doesn’t