iiinova — where your digital self keeps evolving

Noah Greifer @noahgreifer.bsky.social · 13h

options(useFancyQuotes = FALSE)

Noah Greifer @noahgreifer.bsky.social · 09/10/2026

New vignette on the #rstats WeightIt website describing the various weighting methods available and how to choose among them: ngreifer.github.io/WeightIt/art...

ngreifer.github.io

Weighting Methods

0275

Noah Greifer @noahgreifer.bsky.social · 08/10/2026

I assume they normalized the density to have the same area in each category. Makes sense because the size of each group is impertinent to the observation.

110

Noah Greifer @noahgreifer.bsky.social · 06/10/2026

I guess what I was thinking was a bit out of scope, maybe closer to what rlang offers, which isn't package development per se, but contains general-purpose tools R package authors would use in their packages. This View is more about infrastructure and architecture.

020

Noah Greifer @noahgreifer.bsky.social · 06/10/2026

Already discovered a new tool to use from this ({patrick}), great addition! Wondering if packages that support argument checking like {checkmate} belong here?

110

Noah Greifer @noahgreifer.bsky.social · 04/10/2026

This is one of my favorite statistical facts, which I never see referenced:

6616

Noah Greifer @noahgreifer.bsky.social · 03/10/2026

I recently learned about Tweedie regression... could be useful here instead.

120

Noah Greifer @noahgreifer.bsky.social · 02/10/2026

And if you're thinking "Okay but I don't want to assume linearity of the predictor in my beta regression" I have something exciting for you in the works...

260

Noah Greifer @noahgreifer.bsky.social · 29/09/2026

CRAN: Comprehensive 🦭🦭chive Network

140

Noah Greifer @noahgreifer.bsky.social · 18/09/2026

Oh yeah I guess ordinal_weightit() throws away empty outcome levels... maybe I should change that. Thanks for running it!

111

Noah Greifer @noahgreifer.bsky.social · 18/09/2026

Why am I suddenly so passionate about the BART internals? Wouldn't you like to know...

static.klipy.com

Emoji Rubbing Hands

ALT: Emoji Rubbing Hands

040

Noah Greifer @noahgreifer.bsky.social · 18/09/2026

BART is really nice for a frequentist who doesn't like priors because the BART priors are so far removed from the estimand posterior that it doesn't feel like you're prior-ing yourself into a finding. Also the default priors are sensible. So you get a flexible model with a posterior almost for free.

3110

Noah Greifer @noahgreifer.bsky.social · 18/09/2026

I believe WeightIt::ordinal_weightit(., br = TRUE) is the only implementation of ordinal PO regression in R (brglm2 has bracl() which is ordinal AC). If you have your code and its easy to run, would you try it with that?

210

Noah Greifer @noahgreifer.bsky.social · 18/09/2026

I am inclined to agree with @jaredhuling.bsky.social that some kind of penalization could prevent this. Two identical thresholds mean an -Inf MLE for the log difference between two thresholds, which is how they are parameterized in MASS::polr().

110

Noah Greifer @noahgreifer.bsky.social · 18/09/2026

Nice demonstration! So it looks like the thresholds around the empty cell are identical and take the same value when the empty level is omitted. This does feel like a bad feature of the model since emptiness in the sample doesn't imply emptiness in the population.

100

Noah Greifer @noahgreifer.bsky.social · 18/09/2026

This is an amazing resource generally! Thanks so much for sharing!

020

Noah Greifer @noahgreifer.bsky.social · 17/09/2026

This is just a thought experiment, but I haven't really seen it discussed. I imagine the two approaches could yield different results. And what would happen with two adjacent categories with no responses? The threshold between them (if kept) would be unidentified.

630

Noah Greifer @noahgreifer.bsky.social · 17/09/2026

Let's say I have a discrete 1-5 scale, and I want to model it with ordinal regression. Normally, this would involve estimating 4 threshold parameters. Let's say no one selects option 2. Should I retain the empty category and estimate its upper threshold? Or collapse to 4 levels w/ 3 thresholds?

552

Noah Greifer @noahgreifer.bsky.social · 17/09/2026

BART was my gateway drug. Beauty will be borne of this addiction…

0141

Noah Greifer @noahgreifer.bsky.social · 17/09/2026

Guys... I might be becoming a Bayesian...

12481

Noah Greifer @noahgreifer.bsky.social · 16/09/2026

As a package writer, comments like this are what make it all worth it!

110

Noah Greifer @noahgreifer.bsky.social · 27/08/2026

I'm using it in a new R package I'm working on :)

000

Noah Greifer @noahgreifer.bsky.social · 26/08/2026

Can you explain this one to me as a non-Simpsons-watcher?

120

Noah Greifer @noahgreifer.bsky.social · 26/08/2026

Or just boot::boot(., sim = "balanced") 😉

140

Noah Greifer @noahgreifer.bsky.social · 25/08/2026

Also these are making me realize I haven't watched The Simpsons since I don't get a single one! I guess I'm not the right person to make a Simpsons pun in my R package name 😔

120

Noah Greifer @noahgreifer.bsky.social · 25/08/2026

Since you're a Simpson nerd, help me pick a name for my new BART-based R package

641

Noah Greifer @noahgreifer.bsky.social · 10/08/2026

This really makes me wish I could do grad school over again as your student. Maybe I should just get a second PhD?

150

Noah Greifer @noahgreifer.bsky.social · 06/08/2026

Yes, I would say so! I'm just being persnickety since randomization inference wouldn't work for what I do.

110

Noah Greifer @noahgreifer.bsky.social · 06/08/2026

I don't know about "default"... it assumes the only randomness is in treatment assignment, but in many cases randomness comes from sampling or realization of an error term. The permutation test only allows you to make inference to the distribution of assignments for your one sample.

120

Noah Greifer @noahgreifer.bsky.social · 05/08/2026

Knowing that I said something profound enough to be in your blog post is an honor :) My own irreverent take on odds ratios: ngreifer.github.io/blog/an-odds...

ngreifer.github.io

An Odds Ratio Paradox – Noah Greifer

Noah Greifer’s website

3290

Noah Greifer @noahgreifer.bsky.social · 04/08/2026

Yup! It doesn't accept svydesign objects, but you can supply sampling to weights to almost all weighting methods via the `s.weights` argument to weightit(), and you can adjust for clustering in the outcome models using the `cluster` argument to, e.g., glm_weightit(). See here:

ngreifer.github.io

Estimating Effects After Weighting

110

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

Answers: 1. Censoring weights 2. Multilevel propensity scores

180

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

- M-estimation in subgroups Previously, using the `by` argument to estimate weights in subgroups meant you couldn't use M-estimation. Now, M-estimation is supported, which should facilitate subgroup and moderation analysis.

120

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

- Bias-reduced models Ordinal and multinomial models fit through ordinal_weightit() and multinom_weightit() now can be bias-reduced using corrections by Firth and @ikosmidis.com, which prevents infinite estimates and improves over MLE in bias and variance. M-estimation supported!

120

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

- Weights for continuous treatments Weights need estimates of the marginal density as well as the conditional density. Previously, the same density was used for both. Now, the marginal density is computed by marginalizing over the conditional density, yielding better weights.

110

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

- Multilevel PS Propensity scores can be estimated incorporating random effects to model, e.g., students nested within schools. The model formula is passed to lme4::glmer(). All treatment types are supported, as are BART-propensity scores with random effects through the stan4bart package.

ngreifer.github.io

Propensity Score Weighting Using Generalized Linear Models — method_glm

This page explains the details of estimating weights from generalized linear model-based propensity scores by setting method = "glm" in the call to weightit() or weightitMSM(). This method can be used...

240

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

- Censoring weights WeightIt now supports the estimation of censoring weights, as part of a point or longitudinal treatment. To identify that a variable is a censoring indicator, its model formula should have the .cens() marker, e.g., .cens(C) ~ X1 + X2. M-estimation SEs are supported!

ngreifer.github.io

Mark a censoring indicator in a model formula — .cens

.cens() marks a variable as a censoring indicator rather than a treatment, requesting inverse probability of censoring weights (IPCW). It is most often used on the left side of a formula supplied to w...

120

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

On top of this are many, many bug fixes and a dramatically expanded testing suite. See the full list here: ngreifer.github.io/WeightIt/new... I'll discuss the major new features below.

ngreifer.github.io

Changelog

110

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

I'm so happy to announce version 2.0.0 of my #Rstats package WeightIt is out on CRAN! New features: censoring weights, multilevel propensity scores, improved weights for continuous treatments, bias-reduced ordinal and multinomial models, M-estimation in subgroups Check out the website below!

ngreifer.github.io

Weighting for Covariate Balance in Observational Studies

Generates balancing weights for causal effect estimation in observational studies with binary, multi-category, or continuous point or longitudinal treatments by easing and extending the functionality ...

29136

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

Good guesses, but not quite. Maybe 🫥 would be a better clue for the first one?

110

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

I care about this a lot and put simultaneous (curvewise) confidence bands into my package adrftools by default. I think conf bands are useful for showing that the estimated curve itself shouldn't be taken seriously as the truth. I also include specific tests of a curve's flatness or linearity.

150

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

Exciting new updates to #Rstats WeightIt coming soon... 👀 Any guesses? Here are two clues in emoji form: 1. 💀 2. 🏫 (Read separately! no death in schools please)

240

Noah Greifer @noahgreifer.bsky.social · 03/08/2026

Fascinating and clear paper by @corymccartan.com and @melodyyhuang.bsky.social, greatly enhancing our understanding of how Bayesian Additive Regression Trees (BART) works and why it is so effective. A must-read for my fellow BART enthusiasts. #statssky #causalinference

13911