Arkajyoti Saha

Assistant Professor at Department of Statistics at UCI ICS @UCIbrenICS, Alumni of UW Statistics @UWStat, JHU Biostat @jhubiostat

[Arkajyoti Saha @ArkajyotiSaha2]

Consider a common practice in genomics where researchers threshold the elements of the gene correlation matrix to extract groups of genes (co-expression networks). How do we test if the groups are uncorrelated? I, @daniela_witten & Jacob Bien answer this 1/https://t.co/uj4Ya2MSw7

[Arkajyoti Saha @ArkajyotiSaha2]

The Department of Statistics at the University of California, Irvine, is hiring a tenure-track assistant professor! Come join us in Southern California! Apply by December 1, 2024, for full consideration. See details below:
https://t.co/NBRA19bQUi

[Arkajyoti Saha @ArkajyotiSaha2]

@lmwebr Thanks so much!

[Arkajyoti Saha @ArkajyotiSaha2]

I'm thrilled to share that I've started as an assistant professor in the Department of Statistics at UC Irvine's Donald Bren School of ICS (@UCIbrenICS). This would not have been possible without my mentors @daniela_witten, Jacob Bien, @datta_science, and @nilanjan10c!

[Arkajyoti Saha @ArkajyotiSaha2]

@skdeshpande91 Thanks so much!

[Arkajyoti Saha @ArkajyotiSaha2]

You can use our R package to implement the method.

We would appreciate any feedback :) 8/end

https://t.co/suvAMwVNRJ

[Arkajyoti Saha @ArkajyotiSaha2]

2) Ignoring selection leads to an inflated type 1 error in standard settings, but in our case, it causes a loss of power. This occurs as the selection event in standard settings favors the alternative (e.g., file drawer problem), whereas we select hypotheses that favor null. 7/

[Arkajyoti Saha @ArkajyotiSaha2]

1) In standard settings, the selection event involves the mean; for us, it involves the covariance matrix. Thus, we cannot utilize existing machinery. To compute the p-value efficiently, we develop an analog of the polyhedral lemma for selective inference on the covariance. 6/

[Arkajyoti Saha @ArkajyotiSaha2]

Selective inference is a popular tool for testing data-driven hypotheses. Considering the amount of literature on selective inference, it's natural to ask, "What is new in this work?" Two things set our work apart from the existing literature: 5/

[Arkajyoti Saha @ArkajyotiSaha2]

This is a case of double-dipping. We selected the hypothesis (the groups to test for correlation) based on the data & then used the same data to test the hypothesis. In our paper, we solve this within the selective inference framework by conditioning on the selection event. 4/

[Arkajyoti Saha @ArkajyotiSaha2]

What's wrong here? We selected these specific groups for testing, because of the low sample correlation between them. A low sample correlation results in weak evidence against the null hypothesis. Since it is difficult to reject our selected null, the p-value is conservative. 3/

[Arkajyoti Saha @ArkajyotiSaha2]

Will traditional hypothesis tests work here? No! Take data where all variables are correlated, apply a threshold to the sample correlation matrix to identify groups, & then test to see whether the groups are correlated. Despite the groups being correlated, the p-value is high! 2/

[Arkajyoti Saha @ArkajyotiSaha2]

Consider a common practice in genomics where researchers threshold the elements of the gene correlation matrix to extract groups of genes (co-expression networks). How do we test if the groups are uncorrelated? I, @daniela_witten & Jacob Bien answer this 1/https://t.co/uj4Ya2MSw7

[Lukas Weber @lmwebr]

Published version of our paper on nnSVG - method to identify spatially variable genes - is now available from Nature Communications @NatureComms ! 🎉 https://t.co/LSajh4WXW7

[Abhi Datta @datta_science]

New preprint led by @ArkajyotiSaha2 on embedding Random Forests within traditional generalized mixed effects models for binary spatial data. We relax the linear fixed effect assumption and propose RF-GP combining Random Forests with Gaussian Processes. 1/
https://t.co/uktFb1tCvr

[Arkajyoti Saha @ArkajyotiSaha2]

Huge congratulations @daniela_witten !!!!! I feel privileged to have such an amazing postdoc mentor!!

[COPSS @COPSSNews]

Congratulations to the COPSS Presidents’ Award winner, @daniela_witten! #JSM2022

[Daniela Witten @daniela_witten]

#JSM2022 is right around the corner. 📅🤩🥳

Jacob Bien made a website to help identify talks of interest, based on your citation network: https://t.co/LVnBosTjfN

And an R package to create a customized JSM schedule: https://t.co/I9QcDofhgy

@AmstatNews @InstMathStat

[Abhi Datta @datta_science]

Paper on GLS-style Random Forests for spatial data led by @ArkajyotiSaha2 is now published. Combines the strengths of random forests for non-linear regression with Gaussian processes for modeling structured spatial dependence via use of GLS loss. 1/3
https://t.co/JloUCFYd9n

[Arkajyoti Saha @ArkajyotiSaha2]

I know of 943 memes on successfully defending a thesis. All of them are FACTS. Personal experience.

[Abhi Datta @datta_science]

Congratulations to my advisee @ArkajyotiSaha2 (co-advised with @nilanjan10c) on successfully defending his PhD thesis @jhubiostat today! 🎉

[Abhi Datta @datta_science]

New paper on calibrating PM2.5 data from the SEARCH low-cost monitoring network in Baltimore. We use a linear-regression based gain-offset model that reduces bias and variability of the raw or lab-corrected low-cost data .