iiinova — where your digital self keeps evolving

my assumption is that multilevel models need more structure than just a random intercept to beat glmnet since there a lambda such that ridge and random intercepts are equal

@avehtari.bsky.social curious if any of these have come up in your work

010

Returning to this! Do you know of any empirical applications of this estimator or sample code?

100

alex hayes @alexpghayes.com · 15/05/2026

i'm looking for interesting teaching examples where (1) random effects/multilevel models predict better than glmnet (2) glmnet predicts better than random effects #rstats #statistics

3115

I appreciated reading the original comment. I see www.sciencedirect.com/science/arti... suggests bootstrapping for estimating risk of predictive models; do you have more extensive advice about model selection for clinical models somewhere?

sciencedirect.com

Internal validation of predictive models: Efficiency of some procedures for logistic regression analysis

The performance of a predictive model is overestimated when simply determined on the sample of subjects that was used to construct the model. Several …

000

Reposted by alex hayes

Thomas Dietterich @tdietterich.bsky.social · 13/05/2026

We are implementing a similar policy at @arxiv.bsky.social. If there is incontrovertible evidence of LLM slop in a paper, this means the authors did not take the time to read the LLM output and we can't trust anything else in the paper. Penalty is 1 year ban from arXiv followed by...

15412139

alex hayes @alexpghayes.com · 13/05/2026

I'm curious what the other signatures are!

100

alex hayes @alexpghayes.com · 10/05/2026

i wrote up some notes! what's a good email?

i'm quite excited about this work and how it allows us to handle noise in networks with spectral tools we still have some time before we submit to a journal, so questions, comments and feedback are very welcome!

010

our theory covers weighted networks observed with additive noise in simulations we show our approach also works for: - networks with missing edges - ego-centric network data - aggregated relational data this is because we know how to estimate the spectrum well under these noise processes

this leads to a general approach for handling noise in low-rank networks 1. choose a spectral estimator appropriate to your noise process 2. plug in your low-rank estimate of E[A] for A in a typical two-stage least squares estimator

so far this is intellectually interesting. but it turns out it is also immensely useful the latent contagion estimator only requires an estimate of the spectrum of E[A] to construct crucially, we can often get a good estimate of E[A] from a noisy observation of A

why are peer and latent contagion two-stage least squares estimators equivalent in low-rank networks? the network A concentrates quickly around E[A]. estimates of E[A] also concentrate quickly around E[A] in sufficiently dense networks the noise in A or the estimate around E[A] gets averaged away

this is not a generic result; it requires low-rank structure in your network! think: random dot product graphs the converse is also true: a peer contagion estimator can also be used to estimate coefficients under a latent contagion data generating process

if you use latent contagion as working model to derive an estimator, and then use that estimator on data from a peer contagion, everything just works the incorrectly specified latent contagion estimator is asymptotically indistinguishable from the correct peer contagion estimator!

2. the second, and potentially much larger problem, is that contagion in a latent space might not be causally meaningful contagion in a latent space might not correspond to a "real" contagion! this is where things get really cool

there's two problems with modeling contagion in a latent space 1. these models can have vanishing signal-to-noise ratios. this derailed the project for 18 months while we figured out what was going on, and lead to a whole other manuscript explaining the problem bsky.app/profile/alex...

what can you do if the network is noisy? one idea is to use a latent space network model and model peer effects as happening in a latent space then, the presence or absence of individual edges matters less, and instead you need to be able to estimate the latent structure of the network

Screencap of paper introduction, in particular the formula for latent and peer contagion network autoregressive models

120

typically, peer effects estimators assume that you have a perfectly accurate observation of a social network but most people who have worked with network data know that measuring networks is challenging! you often only see a noisy variant of the network

120

alex hayes @alexpghayes.com · 17/04/2026

new print! keith levin and i came with a new approach to estimating peer effects in noisy, low-rank networks i'm very excited about our idea because the approach works with a variety of different noise processes arxiv.org/abs/2605.03204

Screenshot of the first page of the pre-print, showing the title, abstract and first paragraph of the text.

1102

Was having some bluesky issues, sorry for the multiple responses there!

010

alex hayes @alexpghayes.com · 16/04/2026

Done! Thanks for all your work on Quarto github.com/quarto-dev/q...

020

alex hayes @alexpghayes.com · 16/04/2026

This is far outside my wheelhouse but I'll see if I can get an agent to make some reproducible examples

alex hayes @alexpghayes.com · 16/04/2026

Not sure how helpful this is, but here's the commit of Claude/Gemini fixes for axe-core violations on my blog github.com/alexpghayes/...

i'd love to hear about it!

100

currently i'm just copy-pasting the axe-core warnings i get into claude and hoping it knows how to resolve things better than i do it seems like the things claude is doing are fixes patching html / css directly rather than changes to the content of my qmd files

000

gotcha. one thing that is hard to tell is how much accessibility failures are coming from choices i'm making vs quarto side things is it helpful to you if i report accessibility warnings i get about my website?

210