nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Is propensity score matching supported by JMP?

volker_kraft

Hi all,

do we have any update about doing propensity score matching (e.g. see Comparing apples and oranges) in JMP?

It has been reported as a very hot topic in biometrics by our users. SAS supports it and an R package with tutorial exists.

Thank you for any pointers,

Volker

Accepted answers

All comments

Peter_Bartell

Not sure but maybe this is something in JMP Clinical or JMP Genomics? Chris Kirchberg regularly surfs this space so maybe he can answer in a definitive fashion?

brady_brady

I have received a collection of scripts, ported over from Japanese by Yosuke Ono, the purpose of which is exactly this. I've yet to review it though; it's on "the list". I need to work with it a bit before it can be posted anywhere.

However, my question is, should we provide this capability, or should we instead point users to this paper, which posits that propensity scores should NOT be used for matching:

http://gking.harvard.edu/files/gking/files/psnot.pdf

(Grab coffee; discuss)

Chris_Kirchberg

Definitely not in JMP G or JMP C.

Would be nice to have functionality since I have heard this a couple of times and we have skeleton of an add-in. But not my call to say ya or nay

volker_kraft

Thank you all indeed!

What I heard about propensity score matching was

- that this method is very hip (from a prof in biometrics, Germany)

- that a paper has been rejected because proper propensity score matching was NOT done (from a researcher)

- that is is sometimes requested as a standard procedure (another researcher)

In case we want to look into this topic as a new feature the prof would be willing to support us.

My three cents, Volker

brady_brady

Just to be clear because I was not in my previous post: one of the overarching themes at JMP, at least in the time I've been here, is to help users to "do the right thing". Not necessarily the hip thing, or the thing they think they want, but the RIGHT thing. There is no question that we COULD provide this function if we chose to do so.

The December 2016 paper whose link I included above is titled "Why Propensity Scores Should Not Be Used for Matching." While this is a new paper (a working paper which has yet to be submitted for peer-review), given the pedigree of its authors (here's the lead's link: http://gking.harvard.edu/biocv), I would say that at the very least, it merits a careful reading by a development lead, before any of us should respond in anything but a noncommittal manner to customer requests for this feature.

JMP doesn't let people fit regression lines through categorical X and continuous Y, not because nobody wants this (I'm sure somebody does) but because it simply isn't the right thing to do. If we offer a tool to match based on propensity scores, it should be because we've recognized that this is the right thing to do--otherwise, why should customers and prospects trust our product, or our claims that one need not be a statistician to use it?

PS: I realize that much of what I've written here is not fact, but my opinion. Nevertheless, I would not have mentioned my opinions on this matter had I not felt so strongly about them. Please consider what I've written here.

Cheers,

Brady

volker_kraft

Hi Brady,

very much appreciated! I absolutely agree with your "opinion" what JMP should do. This was not my proposal but really the voice of a customer.

I will discuss with the customer when we will meet this week, and I will also ask Chris G if he has an "opinion" about this already.

Best,

Volker

Jordan_Hiller

Hello folks,

I just took a one-day short course in causal modeling with propensity scores.It's mind-bending stuff.

I learned you can use propensity scores in multiple ways for causal modeling of observational data: matching, weighting, and stratification. The course I took focused on the weighting approach, but I was left with the impression that all three are in common usage. I skimmed the paper Brady pointed to, but I don't have enough background to weigh in on the merits of matching.

Even if matching is falling out of favor, we will have customers requesting it for a long time to come. I'd love to see us offer something for this, even as an add-in.

jiancao

I have known PSM as a tool in observational studies for quasi-casual inferences for years. I would concur with Jordan that we offer it as an add-in. Remember "all models are wrong but some are useful"? PSM is such useful tool.

Thanks for sharing the paper.

volker_kraft

Thank you all for your comments, very helpful indeed! I will discuss next steps (add-in?) with Chris.

BTW, SAS has a training course about Propensity Score Matching and Propensity Score Adjustment.

Cheers, Volker