Amit Chowdhry, MD, PhD

Constant sens/spec assumption (see Dawid 1976, Moons and Harrell 2003, Guggenmoos 2000) + pre-test probability scores (eg Wells’) = Information Mismatch

Preprint: arxiv.org/pdf/2503.15382

Here’s our new preprint led by graduating URMC MSTP candidate @samweisenthal. There has been previously published work which shows that sensitivity and specificity are not independent of covariates.

doi.org/10.48550/arX...

@samweisenthal.bsky.social

Principles of Biostatistics Principles of Biostatistics is a classic biostatistics textbook that I have enjoyed using to help teach students biostatistics previously, and I was excite

Pagano's Principles of Biostatistics does an excellent job with explaining conditional probabilities in an understandable way for clinicians

academic.oup.com/jrsssa/artic...

The easiest way to understand what assumptions are being made comes from understanding conditional probabilities.

Especially in this data-heavy era, we would argue that we should switch the way medical education teaches conditional probability – from calculating sensitivity and specificity using 2x2 tables to teaching conditional probabilities.

In this paper, we describe an example of understanding this dependence on covariates matters.

Epidemiology courses traditionally taught for medical students typically assume that sensitivity and specificity are independent of patient factors and are true constants. In fact, this is not the case. The clinical consequences of this have not been discussed in detail.

Here’s our new preprint led by graduating URMC MSTP candidate @samweisenthal. There has been previously published work which shows that sensitivity and specificity are not independent of covariates.

doi.org/10.48550/arX...

Highly recommend this episode. One of their best

Having finally arrived at a time in my career with sci papers under review at big journals (many just collaborations but watching them submit & resubmit in ANNOYING portals made worse by AI QA) I think we should just publish in bioRX and journals should request the papers they want. #medsky #cansky

I don't like "machine learning 'builds on' statistics." A machine learning model *is* a statistic.

It's like building a range rover, saying it's a form of rovers, and that rovers "build on" cars.

"Builds on" is a political phrase to sell more range rovers.

Was recently reminded of David Hand's alternative missing data taxonomy renaming the (in)famous taxonomy MCAR/MAR/MNAR by Donald Rubin to NDD/SDD/UDD. I am not generally a fan of renaming things, but this might be the exception

Source: rss.org.uk/training-eve...