,
14 tweets,
4 min read
Read on Twitter
As a member of #epitwitter that (currently) works mostly on RCT's, I can't resist saying a few words about this, even though @statsepi already did.
1) I do not like the way articles like this tend to divide us into two camps, that either a) WE TOTALLY NEED AN RCT FOR EVERYTHING or b) WE DON'T ACTUALLY NEED ANY RCT AT ALL.
2) it is clear that there are questions which will never be answered in an RCT. For those, the latest and greatest causal inference methods are clearly necessary, and we hope that they are able to accurately estimate/predict whatever it is we are trying to estimate/predict.
3) there are some questions which clearly *can* be put to the test in an RCT. The real debate here is how many of them should, and how much more effectively an RCT answers those questions than an observational study, even one with super-duper causal inference methods.
4) that is likely to be situation-specific, but in most studies evaluating a medical intervention of some sort, an RCT is (probably) the best way to accurately answer the question "does this treatment work?"
5) this is unscientific, but I have worked on a lot of comparative-effectiveness-type stuff using "real world" data, and in most cases I do not believe that the "measured confounders" we have available would have been sufficient to address the reality that there are differences..
5b) in "who gets what" - this is an unquantifiable thing (pause for laugh at the statistician relying on the anecdotal feelz), but in speaking to physicians, it's clear that a list of "age, gender, race, BMI, smoking history, comorbidities..." does not fully capture the...
5c) "we decided to treat this patient surgically" versus "this patient chose a more conservative medically managed approach" decision. Variables like physical function and frailty are rarely assessed. As @ProfDFrancis put it, there's the "end-of-the-bed-o-gram"
6) So, overall (while acknowledging my "intellectual COI" that I work on RCT's) I have reservations that retrospectively collected observational data is (IMO!) able provide a reliable answer to "does *this treatment* work better than *that treatment* for this disease?" questions
7) In conclusion, the scorching hot "Can't we all just get along" take is basically: RCT's are the most efficient tool for doing some things. We should apply them where appropriate.
8) In the *legion* of other questions where it is impossible, impractical, or unethical to do an RCT, we should still try our best to answer those questions, which requires careful thinking about "What data source" / "What statistical approach" and so on.
9) I do not like pieces that say RCT's are the only way to answer a question; nor do I like pieces that say RCT's are no longer necessary. Both are inflammatory & unrealistic viewpoints.
10) One last thing - I just re-read the piece. I understand that it's a clickbait headline, but I don't think the actual contents of the piece merit some of the semi-outraged I've seen responses from the #epitwitter community. The last line is probably what upsets most people.
Also, finding some irony in this recent Tweet:
