Alejandro Schuler

A Non-Parametric Sensitivity Analysis for Bounding Bias in Hybrid Control Trials In the digital era, it is easier than ever to collect and exploit rich covariate information in trials. Recent work explores how to use this information to integrate external controls, including the u...

Using external data in trials can up power, but also bias. Some of my previous papers have tackled this, zeroing bias at the cost of power. In new work, we take the opposite fork: gaining max power and taking the bias hit, but with a rigorous way to protect from that bias.

arxiv.org/abs/2507.18876

Mark your calendars: CTML graduate student researchers Wenxin Zhang, Kaiwen Hou, and Alissa Gordon will be presenting their work next week at the 2025 Joint Statistical Meetings (JSM), held August 2–7 in Nashville, Tennessee. Innovations in causal inference, adaptive study design, and statistical methodology will be highlighted. If you’re attending JSM, don’t miss their sessions—come show your support!

CTML GSRs Wenxin Zhang, Kaiwen Hou, and Alissa Gordon will be presenting their work next week at the 2025 Joint Statistical Meetings (JSM). Innovations in causal inference, adaptive study design, and statistical methodology will be highlighted.
If you’re attending JSM—come show your support!

How hazard ratios can mislead and why it matters in practice - European Journal of Epidemiology Hazard ratios are routinely reported as effect measures in clinical trials and observational studies. However, many methodological works have raised concerns about the interpretation of hazard ratios ...

A nice recent article on why you should abandon hazard ratios.

#statssky #episky

I complain about the weather a lot, and honestly it's super important to me. I come alive when it's hot and sunny out.

Robert F. Kennedy Jr will require all new vaccines to undergo placebo testing, WaPo reports U.S. Health and Human Services Secretary Robert F. Kennedy Jr intends to shift the way vaccines are tested, The Washington Post reported on Wednesday, citing a spokesperson from the HHS.

Just to be clear, this is a violation of the fundamental ethical principle in clinical trial design called "equipoise."

Equipoise means that no-one is exposed to excess risk just to test the therapy. That's specifically *why* vaccine trials don't use placebo arms.

Ethical violation as POLICY.

what the fuck are you talking about how dare you talk shit about A Moon Shaped Pool

As always, reality is complex. In the distant past you will find as much and brutality as there is today, but also evidence for the same kinds of selflessness, beauty, and tolerance. See eg work of David Graeber. We have always been exactly who we are: ruthless, kind, intentional, thoughtless...

The problem with this study is not confounding. The problem is that it is a meaningless 'risk factor association study'.

The results cannot be interpreted. The study has zero scientific or practical value. Such studies need to be outlawed for undermining faith in science.

A Mother and Father Were Deported Under Trump. But What Happened to Their Daughter? (Gift Article) The Trump administration sent the mother of a 2-year-old to Venezuela and the father to a Salvadoran prison. Their daughter remains somewhere in the United States.

The Trump administration continues to ignore the courts.

278 men are still in CECOT prison without a trial, sentence, or any form of due process.

Here's one of their stories. Dad is in CECOT. Mom, deported to Venezuela. Their beautiful toddler, in foster care in the US, whereabouts unknown.

"if I was a billionaire I would not be cooking my brain in social media" if you were a billionaire you too would be desperate for what feels like earned regard of other apes

Multiple articles have been written about Bluesky being a dangerous echo chamber while the world’s richest and influential men are having their brains cooked in increasingly extremist group chats

the shift from “I never thought leopards would eat MY face” to “eat my face harder leopard daddy” must be studied

In theory you have more room to overfit the pscore, but no need to since the outcome regression is already exploiting the prognostic vars. So should suffer a little in small samples for over-specifying? Might be easiest to see in sims for a simple rct attempting to fit pscore v using known 0.5.

I feel like I’m losing my mind with the lack of wall to wall coverage on disappearing *children with cancer*

i keep saying this, but it’s not 5% of the uproar that came from Biden’s debate.

“kids peeing in litter boxes”, a thing that never happened, got orders of magnitude more coverage.

yeah this is IPW, not AIPW, right?

However, if the outcome model is misspecified in the DR estimator then including those prognostic covariates in the pscore could definitely help

for cross-fitting the emp proc remainder is n^-1/2 * |est_EIF - EIF| and this will also be slower I guess, so more finite-sample variance as well

so I'd expect that in finite samples w/ a DR estimator, you actually pay for putting extra stuff into the pscore model b/c you'll get slower convergence rates (more bias from 2nd order remainder).

hmm... interesting. you can definitely improve asymptotic efficiency over IPW w/ known pscore by "learning" the pscore, but can't boost efficiency over an efficient estimator (by definition) where nothing is misspecified and you have the right rates.

Trump Has Now Deported Multiple U.S. Citizen Children With Cancer Two families who had lived in the United States for years, including a child with cancer and a pregnant mother, were deported by ICE on Friday.

BREAKING: @aclu.org says Trump deported a child (U.S. citizen) with metastatic cancer without medication or consultation with physicians. This is the second child with cancer he's deported. www.rollingstone.com/politics/pol... @rollingstone.com

Dunno about feeds but I see good epi content from @statsepi.bsky.social @pwgtennant.bsky.social and @epiellie.bsky.social

Exactly- then it would make sense as positive motivation. But in that case it would still be misleading to omit that, by the same coin, we might expect such architectures to fail to exploit outcome explainers and suffer in efficiency

Why does anyone pay any attention to this man

... yeah, no shade to authors but I think it's fair to say it's "wrong" as written:

"...we expect conditioning on these parts to hurt finite-sample performance..."

I guess if "finite sample performance" = pure bias then this is ok but I don't think that's how most people would read it

yes, it's clearly sufficient to condition on pscore but the way its written makes it seem like that's all you would ever want to do.

at least according to this mechanism, barring any useful transfer learning (eg both functions are smooth/not smooth in similar regions)

and this would imply that these kinds of architechtures are better than separate models when there are strong instruments, and worse when there are strong "explainers" (outcome-only predictors).

it seems they should have actually made the opposite point: that having parts of X relevant only to the *treatment mechanism* playing a role in the *outcome model* is useless. I.e. that you want to filter out *instruments*

am I insane or is this motivation in the 2019 Dragonnet paper completely wrong? If this were true, then it would be pointless to ever adjust in an RCT (where the part of "X" relevant to predict treatment is none of it)

Riders on a BART platform

🚨News: BART’s efforts to put rider safety first are paying off with one of the largest drops in crime in the more than 50-year history of the agency.

For the first three months of the year crime on BART fell by 50% compared to last year.

Thread 🧵⬇️