iiinova — where your digital self keeps evolving

Bob Wilson @rwilson4.bsky.social · 07/09/2026

The method lends itself to sensitivity analysis, allowing us to incorporate uncertainty about unobserved confounders into p-values and confidence intervals, even leading to a kind of "power analysis for observational studies".

000

Bob Wilson @rwilson4.bsky.social · 07/09/2026

My new paper solves that integer program analytically, giving simple formulae for p-values and confidence intervals on Rosenbaum's attributable effects and Neyman's average treatment effect.

100

Bob Wilson @rwilson4.bsky.social · 07/09/2026

In 2015, Rigdon and Hudgens addressed both of these issues, offering a finite-sample exact analysis without the monotonicity assumption. Unfortunately, their analysis involves solving an integer program, infeasible in large sample sizes.

100

Bob Wilson @rwilson4.bsky.social · 07/09/2026

In many cases of practical interest, the assumption of monotonic treatment effects is quite strong.

100

Bob Wilson @rwilson4.bsky.social · 07/09/2026

Rosenbaum followed up with a 2002 paper analyzing attributable effects for matched-pair studies, but his argument was asymptotic, not finite-sample-exact, and more damningly, assumed treatment effects were monotonic (the treatment can be beneficial, or neutral, but never harmful).

100

Bob Wilson @rwilson4.bsky.social · 07/09/2026

In 2001, Paul Rosenbaum wrote a paper introducing attributable effects, which are closely related to Neyman's average treatment effects, but analyzed using Fisher's machinery. It's the best of both worlds: scientifically meaningful estimands analyzed using the elegant, finite-sample-exact machinery.

100

Bob Wilson @rwilson4.bsky.social · 07/09/2026

Jerzy Neyman developed an alternative approach that analyzed average treatment effects, arguing these to be of greater scientific interest. Fisher accused Neyman of calling his baby ugly, and they hated each other for the rest of their lives.

100

Bob Wilson @rwilson4.bsky.social · 07/09/2026

In the 1920s, statistician and eugenicist Ronald Fisher developed a way of testing "sharp null hypotheses". A sharp null hypothesis specifies the treatment effect for each individual in the study. Fisher's methodology is both elegant and exact in finite samples.

100

Bob Wilson @rwilson4.bsky.social · 07/09/2026

Causal Inference has generated many controversies over the years: DAGs vs potential outcomes, design- vs model-based, and average treatment effects vs sharp null hypotheses, to name a few. My paper lives in the world of that last one, but in a way that brings them together.

110

Bob Wilson @rwilson4.bsky.social · 07/09/2026

New Paper Alert! I have posted a pre-print to the arXiv: Randomization Inference for Matched Pairs with Binary Outcomes (link in comments). If you're doing propensity score matching, optimal matching, etc, and you have binary outcomes, give it a look!

121

Bob Wilson @rwilson4.bsky.social · 11/07/2026

This is bonkers to me, because clearly credentials are now meaningless. So why bother paying for them? Classic market for lemons.

010

Bob Wilson @rwilson4.bsky.social · 23/09/2025

I see. You'd prefer the appropriate terminology? In my view, it depends on the audience. I try to avoid jargon when talking to non-specialists.

100

Bob Wilson @rwilson4.bsky.social · 23/09/2025

Do you object to the implied *existence* of a genuine relationship, or the claim it has been identified?

100

Bob Wilson @rwilson4.bsky.social · 19/11/2024

Hi folks, new to Bluesky. I'm a data scientist at Meta where I work on marketing applications. I use a lot of causal inference, experiment design, and mathematical optimization in my work. I'm mostly a lurker on other platforms, and tbh don't have anything interesting to say that fits in these chara

130