🚨Job talk thread🚨

Title: What Can *Conformal Inference* Offer to Statistics?

Slides: lihualei71.github.io/Job_Talk_Lihua…

Main points:
(1) Conformal Inference can be made applicable in many #stats problems
(2) There are lots of misconceptions about Conformal Inference
(3) Try it!

1/n Conformal Inference was designed for generating prediction intervals with guaranteed coverage in standard #ML problems.

Nevertheless, it can be modified to be applicable in

✔️Causal inference
✔️Survival analysis
✔️Election night model
✔️Outlier detection
✔️Risk calibration

2/n

A thread on BUE & BLUE🔥

Gauss-Markov condition:
1) y=Xβ+ε
2) E[ε|X]=0
3) Cov(ε|X)=σ^2Σ

Standard GM:
4) Σ=I

The GM thm shows that OLS/GLS is BL(inear)UE.

Hansen (’20) shows it holds for all unbiased est (inc. nonlinear) w/ an elegant proof (tilted density + Cramer-Rao)

1/n

https://twitter.com/CavaliereGiu/status/1492454733948362753

Call F_2 & F_2^0 the classes of dists satisfying 1-3 & 1-4.
Hansen proves if \hat{β} is unbiased under F_2 for all Σ, then GLS (OLS) is BUE under F_2 (F_2^0).

An intriguing Q. raised by @jmwooldridge & @CavaliereGiu is

Does there exist nonlinear unbiased est under F_2?

2/n

New paper alert🚨#statstwitter

Conformal inference, often framed as a technique to generate prediction intervals, is also a tool for out-of-distribution detection. We studied marginal/conditional conformal p-values for multiple testing with marginal/conditional error control 1/n

https://twitter.com/stats_stephen/status/1384525910846558220

We consider the setting where a dataset of “inliers” is available. Existing outlier detection algorithms often output a “score” for each testing point indicating how regular it is.

But how to choose a cutoff to get guaranteed statistical error (e.g., type-I error) control? 2/n

Check out our new work on conformalized survival analysis w/@RenZhimei and Emmanuel Candès: arxiv.org/abs/2103.09763 Our method can wrap around any survival predictive algorithms and produce calibrated covariate-dependent lower predictive bounds (LPBs) on survival times. 1/n Survival predictive analysis is complicated by *censoring*, which partially masks the outcome. For example, the actual survival time is unknown for units whose event (e.g., death) has yet to happen. A common type is called the “end-of-study” censoring, illustrated below. 2/n