The case that she begins with:
But when Mr. Ryan, 22, was admitted to the trial in May, he was assigned by a computer lottery to what is known as the control arm. Instead of the pills, he was to get infusions of the chemotherapy drug that has been the notoriously ineffective recourse in treating melanoma for 30 years.
The question is:
With reasonable record keeping of existing outcomes for the standard treatments, there is no need to explicitly assign people to a control group with the standard treatment, as that approach is effectively explored with great certainty.
I see this as an argument for Observational (instead of experimental) research. The reason that experimental research is valuable is that you can rule out confounding and estimating quantities like the counter-factual outcomes is a lot easier (as counterfactuals have the nice properties of being missing completely at random).
If you only test the novel drug in the trial then you are assuming that there are not other changes (in population composition, in other forms of care) that do not explain some or all of the variability. You can use statistical adjustment to reduce differences but then things hinge on model specification and whether there are unmeasured differences.
I worry that we are some distance from trusting data where two analysts can get two different answers making different sets of assumptions (which years are the control years? who is included? which covariates are included?).
That being said, many key breakthroughs are entirely observational in nature and it would be good to see more use of this study design. But it is also clear that there are a lot of blind alleys that occur when non-randomized data is used (statins and cancer, anyone?).
[In a larger sense, I think that this is the danger of outside experts. Sometimes they point out that the Emperor has no clothes. However, they may not realize that is because it is currently bath-time . . . ]
Post a Comment