Hmsic::find.matches()

1

Hi Professor Harrell

It may be a bit late to say that I hope you had a nice holiday season.

I would like to use this function to perform matching. I find it more intuitive than certain specialised packages. I suppose that you have written some auxiliary function to depict the result and to help in judging whether the matching is well done. Perhaps I do I myself?

I am completely new to matching techniques. My understanding is that I can use the matchCases() afterwords. I got no idea how do I use the result.

I may be unfair to ask the second question. I have seen people discussed propensity score matching with the RHC data you bundled from Connors et al. (1996). They discussed using caliper=0.2 to do the matching. From what I have seen from Connor et al.'s (1996) paper is that they did two rounds of matching. One is to use the PS to do the matching. Then they perform another round of matching with the absolute values of PS differences from the first round of matching. It requires some data structuring skills, but it is doable. However, I have not seen anyone did it. I am convinced that I misunderstood something. May I have your teaching?

Thank you very much for your advice.

Best wishes

Sandro

2

Hi Sandro - I don’t like any matching methods so unfortunately I’m not motivated enough to work on this. I do recommend you look at the several R packages for matching on CRAN rather than rolling your own. Sorry!

3

Dear Professor Harrell

Now, it makes to me. I am a big fan of your work. I was thinking… It is not logical.

May I learn the reasons why you do not like matching? I know Professor Gary King is very explicit that he does not like matching too.

It would be nice to learn from you.

Best wishes

Sandro

4

I lay this out in Chapter 10 of BBR. Matching is inefficient, leads to non-reproducible research due to arbitrariness in matching algorithms, and requires specialized analytical methods that respect the matching, which few practitioners bother to do. They mainly use methods design for ordinary sample that are not very valid when you’d done matching. There are other problems. Matching should be avoided unless it’s a cost-savings part of a prospective sampling strategy where the matching is planned before the money is spent.

1 Like