iiinova — where your digital self keeps evolving

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

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

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

alex hayes @alexpghayes.com · 02/02/2026

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

At long last @zotero.org has an Android app!

181

alex hayes @alexpghayes.com · 08/12/2025

I just finished @gelliottmorris.com's delightful Strength in Numbers and I'm excited to read more about opinion research and polling Let me know if you have recommendations for followup reading!

050

alex hayes @alexpghayes.com · 25/11/2025

Keith and I strengthened our lower bounds for randomized experiments in the linear-in-means model to a general minimax result Updated results available at arxiv.org/abs/2410.10772

0192

alex hayes @alexpghayes.com · 15/11/2025

Even if you don't find the main figure concerning, do you find the contrast with pre-registered RCT z-scores concerning? It's hard to imagine a world in which this contrast is benign

140

alex hayes @alexpghayes.com · 18/10/2025

I missed that an expanded version of Philip Stark's "Pay No Attention to the Model Behind the Curtain" was published a few years back Needless to say it's very good and worth a read link.springer.com/10.1007/s000...

Screenshot of the first page of "Pay No Attention to the Model Behind the Curtain" by Philip Stark

070

alex hayes @alexpghayes.com · 15/07/2025

~~ making sense of academic statistics ~~ i wrote about the confusing relationship between statistics and data analysis, and also about how statistics relates to science #statistics #rstats #datascience www.alexpghayes.com/post/making-...

Screenshot of the text of the linked blogpost 1/4

Screenshot of the text of the linked blogpost 2/4

Screenshot of the text of the linked blogpost 3/4

Screenshot of the text of the linked blogpost 4/4

1511219

alex hayes @alexpghayes.com · 21/03/2025

This capability isn't directly built into the software, but you can just check if the adjustment criterion holds for C with respect to: - A on M, and - A and M jointly on Y. So you can check identification without needing to apply graphical criteria yourself! www.degruyter.com/document/doi...

alex hayes @alexpghayes.com · 10/03/2025

Sunset across Lake Mendota in Madison

050

alex hayes @alexpghayes.com · 12/02/2025

Enjoying Kino Zhao's recent HDSR paper on a way to think about statistical assumptions I'm also pleased that a stats venue is publishing this kind of work hdsr.mitpress.mit.edu/pub/qasl4fza...

Screenshot of the following paragraph of text from the paper: "Under the perspectivist framework I have been sketching in this section, however, the goal of making these contentious modeling assumptions is not to faithfully describe the world but to prescribe a particular point of view. There is no tension that needs resolving. The description of the world underlying statistical models has to be understood from the perspective that is sketched by the statistical assumptions, which means that the same description may no longer be considered adequate when understood from a different perspective. This is not a limitation of the statistical perspective, but a property of all perspectives."

171

alex hayes @alexpghayes.com · 10/02/2025

My work on network regression and mediation in latent space models is now published at JMLR! jmlr.org/papers/v26/2...

Screenshot of http://jmlr.org/papers/v26/23-1317.html

2256

alex hayes @alexpghayes.com · 21/01/2025

latexdiff-vc is a magic way to see what changes coauthors made to a LaTeX manuscript latexdiff-vc --git --pdf -r {my_last_commit} -r HEAD paper.tex www.mankier.com/1/latexdiff-vc

Screenshot of a PDF of an academic manuscript, with markup denoting revisions between two git commits

071

alex hayes @alexpghayes.com · 15/12/2024

Last night I finished "Code" by @charlespetzold.bsky.social, which starts by explaining how to physically implement logic gates with electromagnets and ends with a full-fledged explanation of a computer Undoubted one of the best books I've ever read and I only wish I'd read it sooner

A softcover copy of the book "Code" by Charles Petzold

4563

alex hayes @alexpghayes.com · 15/11/2024

~~ new blog post ~~ i wrote about the meme that academic code is bad, what i think is achievable, and why i don't think we should be trying to get academics to write software for production would love to hear what folks think! #rstats #pydata www.alexpghayes.com/post/what-i-...

84912

alex hayes @alexpghayes.com · 05/11/2024

Minor update to {vsp} just hit CRAN #rstats - Bug fixes: get_hubs() functions working again - Now automatically permutes B matrix to make it as diagonal as possible, such that Y and Z factors match up (to the degree possible) rohelab.github.io/vsp/news/ind...

alex hayes @alexpghayes.com · 30/10/2024

experiment idea for someone in poli sci: assess whether the 2nd, 3rd, etc postcard promising to publicize turnout still has a positive impact on voter turnout www.cambridge.org/core/journal...

240

here's another one on the importance of visualizing data vs index/time: these are permutations of the same data! Petruccelli, Joseph. “Using a Quartet and Variations to Teach Data Analysis.” MSOR Connections 7, no. 2 (May 2007): 20–23. doi.org/10.11120/mso....

a phylogenetic example where parameter estimates from a brownian motion model are all the same despite the trees being different doi.org/10.1111/2041...

the datasaurus dozen, which is the original quartet on steroids Matejka, Justin, and George Fitzmaurice. “Same Stats, Different Graphs: Generating Datasets with Varied Appearance and Identical Statistics through Simulated Annealing.” CHI 2017 doi.org/10.1145/3025...

170

a predictive quartet showing that models with the same prediction error can vary dramatically Biecek, Przemysław, Hubert Baniecki, Mateusz Krzyziński, and Dianne Cook. “Performance Is Not Enough: The Story Told by a Rashomon Quartet.” JCGS 2024 www.tandfonline.com/doi/full/10....

141

some quartets demonstrating how different conditional average treatment effects can all result in the same average treatment effect Gelman, Andrew, Jessica Hullman, and Lauren Kennedy. “Causal Quartets: Different Ways to Attain the Same Average Treatment Effect.” doi.org/10.1080/0003...

171

next up is a causal quartet showing that different causal mechanisms can lead to the same data D’Agostino McGowan, Lucy, Travis Gerke, and Malcolm Barrett. “Causal Inference Is Not Just a Statistics Problem.” Journal of Statistics and Data Science Education doi.org/10.1080/2693...

3135

the classic is anscombe's quartet, which shows that data with the same summary statistics can look very different when plotted Anscombe, F J. “Graphs in Statistical Analysis.” The American Statistician 27, no. 1 (February 1973): 17–21. www.jstor.org/stable/2682899

140

Fast and easy way to set your handle to your #quarto blog domain (I just did mine): - Settings > Change Handle > No DNS Panel > Copy DID - Create atproto-did file in blog - Add atproto-did file to Quarto resources list - Push Takes <5 minutes

0228

In terms of directly interpreting factors, I like to use heatmaps these days Figure from arxiv.org/abs/2212.12041

020

Less directly useful for interpretation, but pairs plots can be a helpful diagnostic, although exactly what to look for is still somewhat heuristic

So here's a way to visualize PCA juliasilge.com/blog/cocktai...

So the negative result does not assume RDPG structure (Thm 1). However, in RDPGs, controlling for latent positions can recover identifiability (Thm 2). Conceivably the Hoff model you studied with Guido could also be safe from asymptotic collinearity, but we didn't check!

So we have a limited result that says that treatments dependent on network structure can avoid the colinearity problem. We only consider RDPGs as a way to induce dependence between a nodal covariate and the network, but presumably other models can as well

Asymptotic collinearity is not quite an identification failure, but it can absolutely cause inconsistency! We prove it often makes OLS inconsistent. Simulations show that it also breaks QMLE and 2SLS/IV estimators

140

This breaks the linear-in-means model, because the contextual peer effect becomes colinear with the intercept. Surprisingly, this same intuition holds for the endogenous peer effect. We prove this occurs whenever iid nodal covariates are independent of the network

Bramoulle 2009 proved this, and it seemed like the problem was fixed Unfortunately, Keith and I show that intransitivity does not justify using the linear-in-means model You can have lots of intransitivity and still be unable to estimate most model coefficients

230

But then around ~2010 folks figured out that in most realistic social networks, linear-in-means models are in fact identified. The identifying condition is intransitivity, or open triangles. B and D below are mutual friends with C, but not connected. This identifies!

130

alex hayes @alexpghayes.com · 19/09/2024

New pre-print! Keith Levin and I show that the reflection problem is a lot worse than people think it is, and linear-in-means models are degenerate in a very concerning way arxiv.org/abs/2410.10772 #econsky #statssky

34914

zotero 7 features a nice ui update!

050

alex hayes @alexpghayes.com · 29/08/2024

Just published a new paper with @karlrohe.bsky.social! It's about how to co-cluster/co-factor citation networks! Pre-print version available on arxiv arxiv.org/abs/2408.14604 More details over on Twitter :) #statistics

0113

alex hayes @alexpghayes.com · 27/06/2024

there was some interesting discussion of this idea in www.bostonreview.net/articles/rac... that i have been sitting with for a while

alex hayes @alexpghayes.com · 25/06/2024

this was my first time playing with jax, which was fun, and along the way i also learned about randomized rounding, which is a useful way to sample without replacement while controlling marginal sampling probabilities

alex hayes @alexpghayes.com · 13/12/2023

woah i didn't understand how cool webr is until today the fact that i can embed #rstats in a quarto document, on a static website, for people to play with on their own 🤌🤌🤌 example: www.alexpghayes.com/post/2023-12...

screencap of the linked website, featured a sample of R code in the webR interface and the resulting plot generated from the code

48322

alex hayes @alexpghayes.com · 12/12/2023

there are two profs on my committee who don't respond to emails so i've started turning my subject lines into research clickbait and it's criminally effective

drake meme. top panel: subject: scheduling committee meeting. bottom panel: YOUR FAVORITE ESTIMAND MIGHT BE UNIDENTIFIED IN OUR MODEL, ACT NOW TO LEARN MORE

327470

alex hayes @alexpghayes.com · 28/11/2023

Woops axis-labeling error, should read "Error is self-assessed IQ" or "Overconfidence" on the y-axis. Correctly labelled figure

alex hayes @alexpghayes.com · 28/11/2023

they publish their data and it takes two seconds to note that overconfidence is on average decreasing in their data! they claim there is no dk because dk is about conditional variance as a function of skill, not conditional mean as a function of skill 🫠

250

alex hayes @alexpghayes.com · 28/11/2023

so musk shared a post claiming dunning-kruger is a statistical artefact and dk is once again all over internet the post argues against the original dk methodology by testing for a dk effect on simulated data which has "no dk effect present" and then is appalled to find a dk effect in this data 🙃

a scatterplot showing uniformly sampled data in [0, 1] by [0,1] unit grid