in the top paragraph which reads:
... taking at least 400 International Units of Vitamin E daily was associated with a reduced risk of coronary heart disease, even after adjusting for all known confounders. Of course, as noted at the beginning of this chapter, people who take Vitamin E are different from people who do not – for example, they might be more health conscious. But if that were the case, one might expect a favorable outcome in people taking a multivitamin pill or vitamin C as well, behaviors that are also associated with being health conscious. However, this was not observed. The lack of an association of the outcome with an alternative predictor variable that one would expect to suffer from the same confounding as the treatment of interest suggested causality strongly enough that Tom briefly took supplemental vitamin E. Unfortunately, as mentioned above, subsequent evidence from randomized trials suggests vitamin E is of no benefit[2] and may even be harmful.[16] Fortunately, he survived to tell the story.[1]
can you explain what you think went wrong in using the "measuring another predictor" approach in this example? it looks like they did a reasonable analysis approach, so why do you think they got a different answer than what was found in the RCT? was there a specific point we are supposed to be gleaning from this example? or just that research is hard. I am getting confused by this example because you are presenting it as a way to minimize confounding. Are you making a qualitative point about this particular method's limitations?