agree in principal. however, for rare diseases where to have RCT is not feasible, we have to compare with historical control or data from observational studies. Propensity score matching is no worse than cherry picking match
Since the confidence goes down as you add potentially spurious data from additional variables, at a given confidence the measured effect also goes down.
If you think it is in theory possible to create a model which accurately predicts everyone who smokes marijuana, you simultaneously deny the notion of free will and the nondeterminism and randomness.
Disappointing. Nobody is suggesting that they are anywhere as good as RCTs. At worse Propensity Scores is a flawed system to test for correlation for which doing an RCT is unconscionable (imagine assigning half of your non-pot-smoking participants to smoke pot), but in the case of such an unconscionable RCT, a propensity score is all we're left with to use (at least for doing weighting for example, note that it is inferior to other methods for matching datapoints). In the absence of being able to do an RCT it's better than nothing. Sadly in the last dozen years Dr. Wilson's opinion here has become so commonplace and popular ... like a clap trap really, that the medical community has become almost giddy with the prospect of disparaging any study that isn't an RCT. This prejudice against Propensity Scores where RCTs are not suitable is really shameful, and the prejudice has been doing a lot of damage in the field of medical research. Tremendous progress has already been held back as a result of doctors insisting on RCT's before they make a change when in fact an RCT is impractical or imprudent. It almost seems scripted since it leaves all decision making only to those funded enough to run RCT's, ie. pharmaceutical companies, who will only do RCTs on their drugs.
