iiinova — where your digital self keeps evolving

Arman Oganisian @stablemarkets.bsky.social · 08/10/2026

I think we’re saying the same but when i say manipulation i mean it’s hard to conceive of a trial in which i enroll patients from a target populations, flip a coin, then assign them to be 55 versus 45 or something

010

Arman Oganisian @stablemarkets.bsky.social · 08/10/2026

If this worked then why wouldn’t we just spawn an AI agent t stand in for each non-responder in the survey. All done. Missing-Not-At-Random solved.

030

Arman Oganisian @stablemarkets.bsky.social · 08/10/2026

I think my skepticism comes more from lack of manipulability of calendar age than race. But inclusion of race in this way is also bizarre to me since it’s not well-defined - as the paper you linked seems to discuss. Eg I was literally born in the caucasus yet many would not see me as “caucasian.”

100

Arman Oganisian @stablemarkets.bsky.social · 08/10/2026

I like to stay pretty open-minded about causal discovery results but when someone shows me a DAG with an arrow from age to race it’s really hard to take it seriously.

120

Arman Oganisian @stablemarkets.bsky.social · 07/10/2026

I’m pretty good now at intuiting where I need to go and breaking it down into sub-tasks. So LLMs are great for me! But I got there by doing things manually and messing up and trying again. I’m worried that over-reliant students are cheating themselves of the friction they need to grow.

120

Arman Oganisian @stablemarkets.bsky.social · 07/10/2026

What makes a productive vs unproductive user of LLMs in biostatistics research? Imo an ability to think in an organized & modular way about getting to a working solution. One-shotting it usually gives me a complete mess. Modular usage keeps things tight and understandable + keeps my brain engaged

110

Arman Oganisian @stablemarkets.bsky.social · 29/09/2026

I have thoughts on what it means for a study to be “descriptive,” but simply calling a study descriptive doesn’t immunize it from such errors.

000

Arman Oganisian @stablemarkets.bsky.social · 29/09/2026

Some dislike explicit causal/target trial emulation methods because they feel it provides a permission structure to assume away unmeasured confounding. But I frequently see the opposite too: responding to issues of confounding, measurement, missingness as “oh this study is just descriptive”

110

Arman Oganisian @stablemarkets.bsky.social · 18/09/2026

Pairs well with bayesian bootstrap too!

010

Arman Oganisian @stablemarkets.bsky.social · 13/08/2026

Good use if AI: LaTeX all those old seminal papers so that they’re readable.

030

Reposted by Arman Oganisian

Sameer Deshpande @skdeshpande91.bsky.social · 02/08/2026

As the great American philosopher Russell Stringer Bell said, I want you to put the word out there that we back up. Stop by the SBSS+ ISBA table in the exhibition hall at #JSM2026 and find out about all the exciting Bayesian sessions taking place (and get some candy) @isba-bayesian.bsky.social

161

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

This is because there is a common denominator across interventions of B+C+E+F

000

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

An example of post-baseline conditioning that leads to what many would say is a valid causal contrasts are principal strata estimands. Eg contrasts of proportion of subjects who would have died between times 2 and 3 among the subset who would have survived until time 2 under *both* intervention.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

One may think this is simply an issue of conditioning on post-baseline event - specifically, survival up to time 2. That’s kind of right but incomplete. it’s not necessarily post-baseline conditioning that is the problem - it’s that the conditioning set differs across interventions.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

It may instead be that the subset of patients who would have survived up to time 2 under treatment 1 is, say, better of in unobservable ways relative to the subset that would have survived up to time 2 under treatment 0. It’s a consequence of h1/h0 not being a valid causal contrast.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

The “built-in selection bias” of hazard ratios (Hernan, 2010) refers to the fact that if h1/ h0 =\= 1 we cannot tell if this is due to treatment’s effect “at” time 2.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

For example the units in region H are included in the denominator of haz0, but not in haz1 because they would have survived until time 2 under treatment 0, but not under treatment 1. Thus, such contrasts of hazards in general fail condition B.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

But contrasts of h1 and h0 do not compare event counts under the two interventions within the same set of units - unlike with survival rates, s1 and s0, the denominators differ across the two terms h1 and h0.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

This is haz1 = (D+E+F) / (A+B+C+D+E+F) Under intervention, it’s 0 is haz0 = (B+E+H) / (B+C+E+F+H+I).

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

𝐂𝐨𝐦𝐩𝐚𝐫𝐢𝐬𝐨𝐧 𝐨𝐟 𝐇𝐚𝐳𝐚𝐫𝐝 𝐑𝐚𝐭𝐞𝐬 Now consider the proportion of subjects that would have died between time 2 and 3 among those who would have survived at least up to time 2 under intervention 1 - i.e. a discrete-bin hazard at time 2 (to avoid limits).

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

Contrasts of these two proportions s1 and s0 (differences, ratios) is a contrast of potential event counts under two different interventions within the same set of n=5,000 units. Thus, condition B is satisfied and this is a valid causal contrast.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

𝐂𝐨𝐦𝐩𝐚𝐫𝐢𝐬𝐨𝐧 𝐨𝐟 𝐒𝐮𝐫𝐯𝐢𝐯𝐚𝐥 𝐏𝐫𝐨𝐩𝐨𝐫𝐭𝐢𝐨𝐧𝐬 The proportion of these 5,000 units that would have survived past two years under intervention 1 is s1 = (A+B+C+D+E+F) / n. The proportion that would have survived past two years under intervention 0 is s0 = (B+C+E+F+H+I) / n.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

Suppose I want to compute causal quantities within this set of n=5000 units. I have both POs for all subjects. Thus, there is no inferential or modeling task here so we can focus purely on issues of the estimand

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

The plot is divided into regions A-I. In a minor abuse of notation, I also use A-I to denote the number of points in the respective region. E.g. The total number of points is n = A + B + ... + I = 5,000.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

Consider this plot. Each point is a subject i=1,2,..., n=5,000 the location of the point is at their potential outcome pair value { Y_i(0), Y_i(1) }.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

Hazard contrasts fail condition B in general - for this reason they are often said to be non-causal. Contrasts of marginal survival rates, on the other hand, do not fail condition B and so are said to be causal.

100

Arman Oganisian @stablemarkets.bsky.social · 01/08/2026

The definition of a causal effect within a potential outcome (PO) framework is a (A) comparison of POs under different interventions (B) within *a common set of units*.

100

Arman Oganisian @stablemarkets.bsky.social · 22/07/2026

• And when you're computing an "expected" policy effect for, say, New York City - what distribution are you actually averaging over? After all, there is no infinite super-population of New York Cities to sample from.

020

Arman Oganisian @stablemarkets.bsky.social · 22/07/2026

• The dangers of conflating a) posterior draws b) posterior predictive draws and c) auxiliary draws used for monte carlo integration.

130

Arman Oganisian @stablemarkets.bsky.social · 22/07/2026

The paper touches on: • Differences between individual, conditional, sample-average, and population-average effects and how to obtain posteriors for each properly. • When cross-world modeling is and is not avoidable.

120

Arman Oganisian @stablemarkets.bsky.social · 22/07/2026

However, blind implementation with off-the-shelf software can lead to 1) unintended conflation of sample and population-level estimands and 2) unwitting imposition of unfalsifiable cross-world assumptions.

110

Arman Oganisian @stablemarkets.bsky.social · 22/07/2026

Nonparametric Bayes is handy: you get automatic inference for just about any unknown - counterfactuals, parameters, etc. - with very flexible models.

110

Arman Oganisian @stablemarkets.bsky.social · 22/07/2026

Happy to see this paper out. It deals with a practical question in Bayesian causal inference: “𝘈𝘳𝘦 𝘺𝘰𝘶 𝘳𝘦𝘢𝘭𝘭𝘺 𝘤𝘰𝘮𝘱𝘶𝘵𝘪𝘯𝘨 𝘦𝘴𝘵𝘪𝘮𝘢𝘵𝘦𝘴 𝘰𝘧 𝘵𝘩𝘦 𝘤𝘢𝘶𝘴𝘢𝘭 𝘦𝘴𝘵𝘪𝘮𝘢𝘯𝘥 𝘺𝘰𝘶 𝘵𝘩𝘪𝘯𝘬 𝘺𝘰𝘶 𝘢𝘳𝘦 𝘢𝘯𝘥 𝘶𝘯𝘥𝘦𝘳 𝘵𝘩𝘦 𝘢𝘴𝘴𝘶𝘮𝘱𝘵𝘪𝘰𝘯𝘴 𝘺𝘰𝘶 𝘵𝘩𝘪𝘯𝘬 𝘺𝘰𝘶'𝘳𝘦 𝘮𝘢𝘬𝘪𝘯𝘨?” Short answer: you may not be. www.degruyterbrill.com/document/doi...

1325

Arman Oganisian @stablemarkets.bsky.social · 21/05/2026

I think a few years ago they tried running a winter session but canceled due to low enrollments - it’s a tricky time with holidays etc. But they might try again so keep an eye out!

010

Arman Oganisian @stablemarkets.bsky.social · 21/05/2026

Somewhat biased (i will be teaching a session) but this has been going on annually for several years and is one of the best causal trainings: web.cvent.com/event/551f13...

web.cvent.com

Home - 2026 Penn Causal Inference Summer Institute. Online registration by Cvent

2026 Penn Causal Inference Summer Institute

120

Arman Oganisian @stablemarkets.bsky.social · 14/03/2026

If relevant, this paper lays out some alternative inverse-weighted and discrete-time approaches which can be implanted in standard software (nothing bayesian). tinyurl.com/4tb6emm7

031

Arman Oganisian @stablemarkets.bsky.social · 14/03/2026

In this paper we framed the waiting time to next event (initiation of next trt or death) as a competing risk model nested within a causal g-computation approach. doi.org/10.1093/bios... Similar data structure. Turns out you just need to model the sequence of cause-specific hazards.

231

Arman Oganisian @stablemarkets.bsky.social · 07/03/2026

New paper in press at Biometrics by PhD Candidate Esteban Fernández-Morales 1) Develops Bayesian spike & slab and horseshoe models for causal inference under spatial spillover 2) Analyzes Philly's 2017 beverage tax accounting for cross-border shopping arxiv.org/pdf/2501.08231

1203

Arman Oganisian @stablemarkets.bsky.social · 04/03/2026

I agree it's very easy to come up with stories about some confounder. Again it's trivially true that observational studies will have confounding. But even something like "diet" and "health-seeking behavior may be too vague to be productive in my view.

000

Arman Oganisian @stablemarkets.bsky.social · 04/03/2026

...you have a big monster hanging out in front of you that can make you cheat on everything ...contrast, aperture, lighting. It's your duty to decide how much you're going to let yourself fall down that well.

000

Arman Oganisian @stablemarkets.bsky.social · 04/03/2026

Jack: I'm a believer that new technology needs new responsibility. If you're a photographer you have photoshop now. And if you're going to call yourself a photographer and dedicate your life to it...

100

Arman Oganisian @stablemarkets.bsky.social · 04/03/2026

Jack: it's going to get worse because the tech that's available now it's all about making things less labor intensive. As tech goes more down the line of "you don't need to sing and tune, we can sing and tune for you. press this button." It's going to get worse.

100

Arman Oganisian @stablemarkets.bsky.social · 04/03/2026

Conan: whether it's in music, comedy...it's all preparation. You sound like an old man. When young people ask "how do I do what you do"...I tell them you have to work really hard & always prepare and have your shit together and really have it locked down. It's not a fun message.

100

Arman Oganisian @stablemarkets.bsky.social · 02/03/2026

These are great - Greenland has done seminal work in this area! I think some differences here is formal causal inference within the PO framework, emphasis on nonparametric models, and practical implementation.

030

Arman Oganisian @stablemarkets.bsky.social · 02/03/2026

Instead saying “In my experience patients with renal toxicity tend to get the second line therapy more often and also tend to have worse cardiac events… and yet the authors did not adjust for factors related to renal toxicity” is very productive.

100

Arman Oganisian @stablemarkets.bsky.social · 02/03/2026

If you’re not willing to “think through the author’s research design” don’t agree to the review request. It’s not about saving the work. It’s about making specific / assessable critiques of that work. A vague “there could be unmeasured confounding” review comment is not scientifically productive.

200

Arman Oganisian @stablemarkets.bsky.social · 02/03/2026

Okay going to plug my upcoming short course at ACIC: bsky.app/profile/stab...

010

Arman Oganisian @stablemarkets.bsky.social · 02/03/2026

( Randomization doesn’t help because while it ensures that Y(1), Y(0) ind. A it does not ensure that Y(1) ind. Y(0) )

010

Arman Oganisian @stablemarkets.bsky.social · 02/03/2026

There’s def friction in that identification of ITEs require stronger assumptions. Specifically, models for the dependence btwn Y(1) & Y(0) within a subject Not only are we missing the counterfactual, but even randomization does not help. So we can only rely on is subjective clinical judgement.

110