I recently added some resources for students to my webpage (. It includes a guide to navigating the statistics job market, which I hope will be useful to those applying this year!.

RT @nevillesanjana: Delighted to share new work from our lab: .MultiPerturb-seq 🎛️❌📥📤. Over the last few years, we've been combining CRIS….

RT @davidliwei: Our latest computational model to unlock the power of single-cell perturbations (e.g., Perturb-seq), by i. quantifying part….

RT @macfound: Rina Foygel Barber (@UChicagoPSD) is a statistician and 2023 MacArthur Fellow investigating the theoretical foundations of fo….

RT @naturemethods: Introducing guided sparse factor analysis (GSFA), a statistical framework to detect changes in gene expression resulting….

Thank you @WhartonAnlytcs for supporting this work!.

RT @james_y_zou: We share our thoughts on how #datascience education can use + evolve w/ #LLMs We give examples/re….

I really like Mendeley Desktop's automatic BibTeX syncing feature. Now that Mendeley Reference Manager has superseded Mendeley Desktop, I am wondering if/when @mendeley_com will add this feature to Reference Manager. Is anyone else still using Mendeley Desktop for this reason?.

RT @johnomix: 🚨Our work on multi-modal, single-cell pooled CRISPR screens to study GWAS variant-to-function, STING-seq, is now out in @Scie….

Thank you very much to my co-authors Ziang Niu (@MaxwellAng1), Abhinav Chakraborty, and Oliver Dukes!.

In summary, a double robustness phenomenon provides hope for applying MX methods with the model for X learned in sample, and brings these methods in closer proximity to doubly robust and semiparametric inference.

We also looked at power, adjusting for Type-I error inflation using oracle calibration. We found GCM test to have somewhat higher power than dCRT, and lasso-based tests to have somewhat higher power than post-lasso-based tests.

We did not try very small samples or very discrete data, but we suspect in these scenarios that dCRT may have better Type-I error control than the GCM test.

Comparing dCRT to GCM test in numerical simulations, we found fairly similar Type-I error. Notably, switching the lasso out for post-lasso dramatically improves Type-I error control for both methods. The Maxway CRT ( inspired us to try this.

We also leveraged semiparametric theory to prove that the GCM test is asymptotically optimal against local partially linear (and generalized partially linear) alternatives.

In fact, we proved that the dCRT is asymptotically equivalent to another doubly robust conditional independence test: Shah and Peters' Generalized Covariance Measure (GCM) test (. The resampling distribution converges to N(0,1) in a conditional sense.

We formalized this intuition by proving a double robustness property for the dCRT. If the product of the errors in the model for X and the model for Y|X decays quickly enough, then the dCRT will control Type-I error.

We found that the Type-I error can be heavily inflated, but it depends on how well the model for Y given X is fit. A better approximation to E[Y|X] gives better Type-I error control!

We begin to address this gap by considering what happens to the dCRT ( when X is learned in sample.

MX methods now deployed in genetics ( and genomics (, usually fitting the model for X in sample. However, existing theory assumes either that the model for X is perfectly known or trained on a large unlabeled sample.

Do model-X methods (like knockoffs or CRT) still control Type-I error when the model for X is learned in sample?. See our new preprint (, read the thread below, and/or come to my ISSI talk ( next Wednesday to find out!