Amy D Willis

We really appreciate your support, Kendra!!

Estimating Fold Changes from Partially Observed Outcomes with Applications in Microbial Metagenomics We consider the problem of estimating fold-changes in the expected value of a multivariate outcome observed with unknown sample-specific and category-specific perturbations. This challenge arises in h...

Thx for checking! Co-author Teichman joined the team after v1. Here's what I'll be using:

David S Clausen, Sarah V Teichman, & Amy D Willis. 2025+. "Estimating Ratios of Means of Multicategory Data Observed with Sample and Category Perturbations." Biometrika, In Press. arxiv.org/abs/2402.05231.

Using Relative Abundance Data to Estimate of Multiplicative Differences in Mean Absolute Abundance A differential abundance method for the analysis of microbiome data. radEmu estimates fold-differences in the abundance of taxa across samples relative to "typical" fold-differences. Notably, it does ...

Documentation:

statdivlab.github.io/radEmu/

Blog post:

statdivlab.github.io/blog/article...

4/4

Feel guilty about replacing your zeroes with pseudocounts?

Fed up with checking robustness to said pseudocount?

Not sure what it means biologically to have a difference of 0.34 in averages of CLR-transformed data?*

radEmu 🕶️🦤

3/

*Spoiler: It means nothing. It's inherently not interpretable.

Oh yeah, ICYMI...

radEmu is a method for differential abundance (or differential expression). It's especially well-suited to genomics and microbiome studies.

Also, we made some common cases superdooper fast last week, so pull changes ⚡🏎️

Happy holidays, y'all! ❤️🦤❤️🦤❤️🦤❤️🦤❤️🦤❤️🦤❤️

2/

I am delighted to share that the radEmu manuscript was accepted today to Biometrika 🥳 Sending gratitude to our users and reviewers for their enthusiasm and support!!

Huge congrats to @davidandacat.bsky.social and Sarah Teichman on all their hard work and brilliance 😻❗

1/

With many thanks to @svteichman for bringing this to my attention and making these figures

Thanks for corroborating!!

We've seen again and again that ALDEx2's p-values can be either very conservative (= underpowered) or very anticonservative (= wrong), but this points to something more problematic, such as a mistake in how they're being calculated.

Thanks for corroborating!!

We've seen again and again that ALDEx2's p-values can be either very conservative (= underpowered) or very anticonservative (= wrong), but this points to something more problematic, such as a mistake in how they're being calculated.

ALDEx2's p-values are unusually strongly correlated with its effect sizes. Has anyone else noticed this? That... shouldn't happen.

Data is actual shotgun coverages, ~8000 "taxa", 57 samples.

Let's do better science together by using agreed-upon definitions! Woohoo!

4/4

Same article but without the paywall

(sorry, I didn't think they'd paywall it! duh Amy...)

github.com/statdivlab/p...

3/

There's some more info on confounders in microbiome studies (and how they differ from precision variables) in this perspective we wrote last year:

www.nature.com/articles/s41...

2/

A screenshot from a retweet dated March 2021. Original tweet from David + Cat: "A confounder is a variable that has a causal effect on an outcome of interest and is associated with a (possibly causal) predictor of interest. It is not any process/phenomenon/feature of reality that makes a study more difficult to conduct or interpret." Comment on retweet from Amy: 'Happy Definition Monday! “Confounder” is an overused term in microbial ecology & next-gen sequencing, often used to mean “something that makes science difficult”. Here’s what it actually means — thanks for the great reminder, @davidandacat! 🙌🏻💯'

I had this conversation again with someone today, so posting a screenshot of an old t**** for those who missed it the first time around -- or want a refresher!

1/n

With much gratitude to everyone who supported the wonderful #STAMPS summer course at the MBL, I share that STAMPS will not be offered in 2026.

💔😿

pixi is awesome. That is all.

@titus.idyll.org you're all over this, right?

Are you looking for a #tenuretrack or #openrank job in #biostatistics or #statistics? @fredhutch.org is hiring! Seattle is a wonderful place to grow your career; the bio/stat & science scene here is world-class; and you will work w/ amazing UW Biostat students. Apply!
apply.interfolio.com/176402

dude you can just text me with this stuff ❤️

% of microbiome (& changes) aren't identifiable from HTS, but stacked barplots are helpful for generating clues esp if big shifts

data snooping has implications for error rate control. there isn't a hypothesis you're interested in?

good luck & enjoy xx

A screenshot of the header for the tutorial linked in the post

In other news, I just had the great pleasure of teaching at #EBAME10 in Brittany 🇫🇷🥐.

For this workshop, I made a tutorial showing how to use radEmu with the taxonomic profiling output of #anvio 😍😻

Workflow available for everyone on our blog: statdivlab.github.io/blog/article...

❤️❤️

GitHub - statdivlab/radEmu Contribute to statdivlab/radEmu development by creating an account on GitHub.

Happy Friday from the StatDivLab!

We just made updates to radEmu to make it faster & more stable than ever.

Thanks to all our users -- we're excited to keep supporting it.

github.com/statdivlab/r...

Interpretable and robust microbial differential abundance, woohoo!

Thanks for your interest, @genomesevolve.bsky.social ! As a field, *on average*, I feel we have moved...

1. away from obsessing over alpha and beta diversity comparisons
2. towards comparisons that are less sensitive to rare/undetected species (than diversity)

So 📈

🥳

Wisdom from @titus.idyll.org on the final day of #STAMPS2025: "There's free as in beer, and there's free as in kittens. #Bioinformatics software is free as in kittens. You have to love and care for them or they... well... yeah."

The question was "...*on the same sequencing run*?"

Given that a Poisson regression with log link targets the same parameter as your NB regression, I'd be curious to see the coverage of robust Wald CI's. `rigr` wraps this, so does `raoBust`, so should be easy to add.

(Sorry -- I'm at a workshop today or I'd do it myself)

GitHub - statdivlab/raoBust: Generalized Linear Models with robust and non-robust Wald and Rao (score) tests Generalized Linear Models with robust and non-robust Wald and Rao (score) tests - statdivlab/raoBust

raoBust doesn't invert score tests @nlaroy.bsky.social, but it does implement (model-misspecification) robust score tests, which are amazing for inference.

Feel free to open a feature request. We'll see what we can do.

github.com/statdivlab/r...

The StatDivLab is getting excited for #STAMPS2025 and working on our lectures for 🤩 Stats Day 😻.

We ran this think-pair-share activity last year and, wow, the ensuing discussion was *very* engaged 😹😅

Reach out if you'd like to attend! Woods Hole, MA, July 14-24 2025. #microbiome #dataanalysis

Agreed that (eg) MAG assembly vs taxonomic estimation makes a huge difference to your answer. Also, please definitely read the Conflict of Interest statement for the shallow shotgun paper and note how much of its claims rely on bioinfomatic subsampling *and not actual shallow sequencing data*

Best workshop you can ever follow

That's wonderful to hear, Isabelle!! Thanks for sharing, too. Hope to see you there in 2026.