iiinova — where your digital self keeps evolving

Alec McClean

@alecmcclean.bsky.social

Postdoc @ NYU Grossman; stats / ML + causal inference alecmcclean.github.io

Alec McClean @alecmcclean.bsky.social · 16/07/2025

Although estimator is complex, some nice properties arise from the construction: in particular, we can examine distribution of cumulative weights across subjects, like in single-timepoint weighting

Distribution of cumulative weights, with one for each subject.

000

Alec McClean @alecmcclean.bsky.social · 16/07/2025

Cross-world"ness" --> nuances in identification and estimation - ID: Need strong seq. rand., but still possible w/out positivity - Est: new EIF for doubly robust estimator involves additional term w/ covariate density ratio across the target regimes

Covariate density ratio across two target regimes

Second half of EIF, w/ additional novel term involving covariate density ratio across regimes

ID. No positivity needed. Just need weights to behave well, which is possible by construction (eg, overlap, trimming)

100

Alec McClean @alecmcclean.bsky.social · 16/07/2025

These fx are - "Cross-world" - "Mechanism-relevant" (they target mean diff in POs we care about) - **Not** "policy-relevant" (they're not implementable) This tradeoff arises elsewhere (mediation, censoring by death). Ours is another example: What you want to know != what you can implement

100

Alec McClean @alecmcclean.bsky.social · 16/07/2025

New paper 📜 We construct longitudinal effects tailored to isolated mean diff in two POs while adapting to positivity violations under both regimes. Some notes vv

170

Alec McClean @alecmcclean.bsky.social · 08/04/2025

Excited to present on Thursday @eurocim.bsky.social on new work with @idiaz.bsky.social on (smooth) trimming with longitudinal data! "Longitudinal trimming and smooth trimming with flip and S-flip interventions" Prelim draft: alecmcclean.github.io/files/LSTTEs...

053

Alec McClean @alecmcclean.bsky.social · 28/12/2024

We analyze the effect of mothers’ smoking on infant birthweight, and see that accounting for uncertainty in estimating M alters CIs for ATE. This was fun work with Edward and Zach Branson (sites.google.com/site/zjbrans...) and was a great project to finish my PhD! 9/9

010

Alec McClean @alecmcclean.bsky.social · 28/12/2024

We incorporate M into our “calibrated” sensitivity models. Generically: U <= GM where G is sensitivity parameter. We outline many choices for U and M and develop three specific models. Then identify bounds on ATE and give estimators that account for uncertainty in estimating M. 8/9

100

Alec McClean @alecmcclean.bsky.social · 28/12/2024

#2 Calibrated sensitivity models arxiv.org/abs/2405.08738 Sensitivity analyses look at how unmeasured confounders (U) alter causal effect estimates (when, eg, trtment not random). To understand U, we can calibrate by estimating analogous ~measured~ conf. (M) by leaving out variables from data 6/9

110

Alec McClean @alecmcclean.bsky.social · 28/12/2024

3. A slower-than-root-n CLT w/ non-smooth nfs (some “fun” technical results deep in appendix :D) 4. Simulations illustrating 1-3. Manufacturing Holder smooth fns was an interesting challenge! 4/9

110

Alec McClean @alecmcclean.bsky.social · 28/12/2024

We focus on estimating expected cond. cov. Four main contributions: 1. Structure-agnostic linear expansion for DCDR est. Nuis func est. bias more important than var. 2. Rates with local lin smoothers for nfs under holder smoothness. Semiparametric efficiency and minimax optimality possible 3/9

100

Alec McClean @alecmcclean.bsky.social · 28/12/2024

#1 arxiv.org/abs/2403.15175 The DCDR estimator is quite new (2018, arxiv.org/abs/1801.09138). It splits training data and trains nuisance fns on independent folds It can get faster conv rates than usual DR estimator, which trains nuis funcs on same sample. We analyze the DCDR est. in detail! 2/9

100