It’s been a while since I went down an epidemiology rabbit hole, but I just ran across this great discussion on the datamethods forum about chance imbalances in baseline characteristics among study arms in randomized trials, and the typical “Table 1” that reports summary statistics for each baseline characteristic stratified by study arm. This gets at something that has bugged me since I took a clinical trials class in grad school.
First of all, if you haven’t already closed this tab, you should read Frank Harrell post arguing for the elimination of the traditional “Table 1” from reporting clinical trial results. His alternative is really interesting, and he makes a compelling case.
But if you don’t have a traditional Table 1, how does a reader know that a chance imbalance of a critical baseline variable didn’t influence the results of the trial? Randomization of patients into study arms only guarantees balance of baseline characteristics on average if you randomize many times — but most clinical trials are just conducted once. If that one randomization happens to, for example, put all women in Group A and all men in Group B (unlikely but theoretically possible), it’s important to know this and account for it. This is something that’s always bothered me about clinical trials that rely on unadjusted comparisons among arms.
Frank’s suggestion is to:
“Pre-specify the most important prognostic factors” and adjust for them come hell or high water.
This seems like a clean solution, and it forces investigators to identify key baseline characteristics a priori, which is a good idea regardless of the analytical approach.
The CONSORT guidelines for clinical trials recommend reporting both unadjusted and adjusted analyses. But I’ve never understood what value the unadjusted analysis adds over the adjusted one (assuming the adjustment is for pre-specified baseline characteristics). If the baseline characteristics are balanaced among arms, then the adjusted analysis will show no effect for these covariates — in other words it will look essentially the same as the unadjusted analysis. If they are not balanced, then you want to adjust for them — and the unadjusted analysis isn’t meaningful.
I’m not a clinical trials analysis expert by any means, so I’m willing to be convinced that I’m wrong here. If there’s a compelling argument for reporting unadjusted analysis instead of (or along with) a pre-specified adjusted analysis, I’m all ears!
