Zach Shahn

Structural Nested Mean Models Under Parallel Trends with Interference Despite the common occurrence of interference in Difference-in-Differences (DiD) applications, standard DiD methods rely on an assumption that interference is absent, and comparatively little work has...

We (Audrey Renson and @pausalz.bsky.social) Just updated this paper on interference in time-varying DiD settings from a while ago: arxiv.org/abs/2405.11781. Still see papers coming out regularly about problems that I'm pretty sure this solves...

Is there any movement to get them to stop showing players’ postseason stats on tv? I wanna know how good they are, not how lucky they were the past few games

And I now see how this doesn’t depend on intervention. Eg conditioning on being on earth induces associations between variables related to falling explained mathematically by Newton. So I think I fully understand now, thanks for bearing with the stream of consciousness if you made it this far!

(Didn’t Google Boyle’s law, hope I got it right)

I guess it doesn’t need to be as magical as fate. You could imagine presetting a machine to maintain gas pressure at some level however it needs to. Then temperature and volume become counterfactually related. And P=TV is a mathematical non-mechanistic explanation like you discuss in the paper

Ok, but now I’m back to my original understanding that the distinguishing feature of pre-selection is that it occurs via intervention and induces association via “fate” as in your fairy godmother example

Ok, but C is in the past of E and D, the exposure and outcome of interest? I think I'm being dense, so I won't ask you to explain again!

In M-bias, you’re conditioning on C, which is something pre-baseline like an eligibility criterion for a study. But it doesn’t arise from actually intervening to set C. Is that what makes predecessor bias different?

Ha, I don’t think anything’s wrong, but I think M-bias from conditioning on pretreatment variables is one common case of what you’re describing

Yeah I was asking about your pinned post paper, read it with great interest

Does predecessor bias have to come from a 'fated collider' that seems to maybe be a uniquely quantum thing? Or do you use the term to refer to any setting where conditioning on a pre-intervention variable induces an association? We call the latter case 'M-bias' journals.lww.com/epidem/abstr....

Causal people are usually interested in causal effects and would want to adjust away what we would call that spurious association. So we would just call it confounding adjustment. Maybe more descriptive people like demographers have a word for adjusting for confounding when you don’t want to

Forgot to say, 'we' is me and my phd advisor David Madigan, who took a break from being a provost to do research again. Nice reunion!

Next, we ask 'when will the key assumptions be satisfied'? We argue that under a causal pie model the answer is 'basically never'! But we think violations should be small in practice and provide sensitivity analysis.

Then we develop Neyman orthogonal estimators for when S is only independent of Y(1) given Y(0) and covariates.

First, we simply point out that you don't really need multiple studies. You just need one study and a baseline covariate S that you can use to make your own 'substudies' that satisfy their assumptions. This gives you more control over whether the assumptions are approximately satisfied.

Identification and Estimation of Joint Potential Outcome Distributions from a Single Study Most causal inference methods focus on estimating marginal average treatment effects, but many important causal estimands depend on the joint distribution of potential outcomes, including the probabil...

I think this paper we just put out is kind of interesting (arxiv.org/abs/2509.20506)! Wu and Mao (arxiv.org/abs/2504.20470) cleverly showed that if you have multiple studies and 'study' is associated with Y(0) but not Y(1) given Y(0), you can identify the joint distribution of Y(0) and Y(1).

Dare I say that this post is ironically on the verge of starting a lively debate?

Ha, yeah, that was "unconstructive".

Guess you can usually tell who wants feedback and who's celebrating from the wording of their post. Definitely shouldn't rain on people's parade when they're clearly celebrating

I'm sure this a subtweet of something that I wouldn't defend, but I do post papers here in large part to hopefully get feedback...

Basic idea is: you want to know the effect of some exposure on longevity (via MR). You don't know how long the people in your sample will live. But you know how long their parents lived. So you use sample's genes as IVs, sample's exposure, parents' lifespan as outcomes. When/why would this work?

Constructed Measures and Causal Inference: Towards a New Model of Measurement for Psychosocial Constructs - PubMed Psychosocial constructs can only be assessed indirectly, and measures are typically formed by a combination of indicators that are thought to relate to the construct. Reflective and formative measurem...

Does this paper (pubmed.ncbi.nlm.nih.gov/34669630/) address the issues raised in that paper?

For anyone who's interested in how deck chairs should be arranged on the Titanic, we have a short note on the assumptions/rationale underlying the use of parental longevity as a proxy outcome (increasingly common) in Mendelian Randomization: arxiv.org/pdf/2508.03431

7 year old son and 3 year old daughter trade Beatles and frozen songs on car rides and he always chooses this one because it’s their longest. I’ve developed quite an appreciation for it

That, Baby You Can Drive My Car, and Paperback Writer are Paul's trifecta of genuinely dryly funny but still musically great songs. Can't think of other examples of that combo in pop/rock, but sure there are some...

Simple graphical rules for assessing selection bias in general-population and selected-sample treatment effects - PubMed When analyzing a selected sample from a general population, selection bias can arise relative to the causal average treatment effect (ATE) for the general population, and also relative to the ATE for ...

It can also sometimes be the contrast of interest, not “bias”: pubmed.ncbi.nlm.nih.gov/38904459/.

Simple graphical rules for assessing selection bias in general-population and selected-sample treatment effects - PubMed When analyzing a selected sample from a general population, selection bias can arise relative to the causal average treatment effect (ATE) for the general population, and also relative to the ATE for ...

This paper formalizes and gives identification conditions for contrasts where the exposure impacts selection into the sample: pubmed.ncbi.nlm.nih.gov/38904459/

I’m triggered

Also bootstrap requires even weaker assumptions. I teach my students that a good rule of thumb is when in doubt, use bootstrap