When a caliper is used, two patients with similar but not identical scores can be matched to one another. But if you require an *exact* match of the PS, that can almost only be achieved with patients that have the same covariate values used to compute the PS.
-
-
Replying to @ADAlthousePhD @MarkHoofnagle and
Unless two *covariates* in the PS model have exactly the same regression coefficient (in which case you could match a 67 year old smoker without diabetes to a 67 year old nonsmoker with diabetes - if smoking & DM had **exactly** the same regression coefficient in the PS model)...
1 reply 0 retweets 4 likes -
Replying to @ADAlthousePhD @MarkHoofnagle and
...but if each variable included in the PS model has a unique value of the regression coefficient, then the only “exact matches” on the PS will be patients that have exactly the same covariate values.
1 reply 0 retweets 3 likes -
Replying to @ADAlthousePhD @MarkHoofnagle and
This paper has plenty of things to critique, but I do not believe this is a copy-paste error nor is it fraudulent (at least, not based on this column of this table). In fact I would be more concerned if *exact matches* on the PS did *not* look like this.
1 reply 1 retweet 7 likes -
Replying to @ADAlthousePhD @MarkHoofnagle and
Andrew Althouse Retweeted Eric Weinhandl
As
@eric_weinhandl pointed out here: https://twitter.com/eric_weinhandl/status/1280630034894479360?s=21 …https://twitter.com/eric_weinhandl/status/1280630034894479360 …Andrew Althouse added,
Eric Weinhandl @eric_weinhandlReplying to @eric_weinhandl @SteveJoffeOTOH, maybe that is the explanation? The authors found only 96 (of 1985) HCQ-exposed patients with PS values that were exactly equal to HCQ-unexposed patients. These 192 patients occupied cells with perfectly equal covariate vectors—indeed, all covariates are binary.1 reply 0 retweets 4 likes -
Replying to @ADAlthousePhD @MarkHoofnagle and
I’m literally in the hospital with my wife and a newborn right now, otherwise I would make up a simulated dataset and R code to try and illustrate this. If you want to do it yourself, generate a dataset with a bunch of binary covariates, then...
16 replies 2 retweets 58 likes -
Replying to @ADAlthousePhD @MarkHoofnagle and
...run a propensity score matching algorithm with a caliper distance of 0, and examine the matched pairs you end up with. They’ll probably all be perfect matches on every covariate the included in the PS model.
1 reply 0 retweets 8 likes -
Replying to @ADAlthousePhD @MarkHoofnagle and
(The fact that the authors did this raises other red flags, most notably that no one really uses propensity score matching this way, just clarifying that the exact match of the two columns is oddly enough consistent with what they said they did)
2 replies 1 retweet 10 likes -
Replying to @ADAlthousePhD @MarkHoofnagle and
Agree. They used 21 dichotomized (!) characteristics for propensity score matching. Plus treatment, that yields 2^22 possible combinations. It's extremely unlikely to find 190 perfect matches in a finite real world sample, except when the variables are highly correlated.
2 replies 0 retweets 0 likes -
Replying to @jhan2qt @ADAlthousePhD and
But even with matching with a sample that size what is the probability you could get such a match over that many variables?
2 replies 0 retweets 0 likes
Nah, as Andrew said it's basically implied by their method. The question for me is the probability of 190 such individuals existing in a sample of just 2,500. I can do something similar with my database of 100,000 tests, but 2,500 isn't a lot of people
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.