New blog post, in which I discuss the error of "the difference between stat significant and not is not itself stat significant". It often causes people to post-hoc try to explain things that are easily just due to the standard error of estimates, andrewpwheeler.com/2025/07/28/t...

I recently received a question on Poisson vs OLS models for a dose response relationship, and posted the exchange to my blog. Long story short even with count data OLS models can make sense, it depends on the functional form. andrewpwheeler.com/2025/05/28/a...

Paper with Scott Jacques on stripping sensitive information from narratives is out, www.qualitativecriminology.com/pub/zhiuy6jg.... Ultimate goal to make it easier for qual people to share there data for replication. Has links to python code and uses open source models.

After a long, long wait (and a transition from academia to private sector), our paper is finally out! www.sciencedirect.com/science/arti... We used a unique source of data tracking residents of Milwaukee from the 1980s through the 2020s to examine the role of alcohol abuse on gun violence.

Got to love the speed of academic publishing. A paper that I helped collaborate on in 2021 is, just now, getting published. A full 4 years later, and also I am no longer in academia. But tack on another +1 to the GoogleScholar page I guess!

Ah yes, this drowning risk cluster is *checks notes* the entire nation of Iran.

Remember the 2024 election? Are you NOT sick of hearing predictions about who will win? Want to revisit 11/4?! Probably not. But I wrote up some thoughts on my novice approach to building a poll aggregation model during this last cycle. gmcirco.github.io/blog/posts/p...

Might regret, but genuinely curious: It's my understanding that most/all "big" LLMs in genAI today are transformer-based. This gives speed/scaling advantages, but has some drawbacks. (Hallucinations, AI "sheen" on images, etc) Other ML methods (GAN for images, for ex) balance the other way, right?

I've been looking at some data about the flow of "crime guns" (seized guns that were used in a crime) across state borders using ATF data. What I find is pretty interesting - although may not come to a surprise for many: gmcirco.github.io/blog/posts/c...

This article from @nytimes.com is making the rounds. The main premise: Is it still worth it to learn how to code in an increasingly AI-centric world? My person opinion is "Yes, but..." www.nytimes.com/2024/11/24/b...

Where do crime guns in your state come from? In many places, NOT the same state the gun was registered in! For example, more than 80% of seized crime guns in New York were bought or registered somewhere else:

Bit of a preview of something I've been working on - revisiting Project Green Light Detroit. One thing I always wanted to delve a bit deeper into was the relationship between PGLD and the proactive police presence. Below is a plot showing the pretty stark change post-Green Light.

Here's a blog post about the one important skill I *didn't* learn in grad school: SQL! Here I talk a bit about using SQL inside of R, and how to use DuckDB to speed up analysis of large datasets. Looking back, something I wish I had learned a lot earlier. gmcirco.github.io/blog/posts/d...

It's been a while - but a little blog post here about synthetic data simulation with spatial data. I detail a bit of the how and why of generating synthetic crime data using data from Hartford, CT on robberies and gas stations. gmcirco.github.io/blog/posts/s...

After that little exchange about missing the Livejournal days, I decided to bring 'em back (but better, 'cause Dreamwidth). So here's a quick train post about steeking my career. b-e-x.dreamwidth.org/641679.html

A slightly more dry blog post here In short, if you use SMOTE to address class imbalance (which, arguably, you should not ever) be sure not to evaluate your model metrics on a hold-out set derived from the SMOTE dataset. Your predictions will be WILDLY optimistic. gmcirco.github.io/blog/posts/t...

A guide to BlueSky for Scientists * Common questions * Links to resources * An explanation of feeds * A directory of science feeds Please share with scientists on BlueSky! Written by me and @markrubin.bsky.social 🧪 #stats #PsychSciSky #neuroscience

Someone recently asked me how to use #rstats to map journeys of individuals across the United States. It's actually pretty easy, but requires a bit of work. Here's a quick blog post that walks through how to do this using the `sf` package in R.

Here's me dipping my toes into language models and NLP. In this quickie blog post I write about using pre-built @huggingface.bsky.social sentence transformers to aid searching through medical narratives: gmcirco.github.io/blog/posts/n...

How should we operationalize deployment time in police calls for service? Read all about it in the latest publication from our Evidence-Based Policing #ebp research program at the NSCR. crimesciencejournal.biomedcentral.com/articles/10....

Very excited to announce that Andy Wheeler and I won first place in NIJ's "Innovations in Measuring Community Perceptions" challenge nij.ojp.gov/funding/inno...

My new favorite toy: gam(y ~ unit + s(time) + s(time, by = unit)) Effectively "borrows strength" across units to inform predictions, while still allowing unit-specific deviations from the "usual" to come through. Black lines = s(time) Dashed = unit + s(time, by = unit) Colored = combined

I'm a data guy and also a coffee nerd. So when James Hoffman released data from his Great American Coffee Taste Test I had to look at it: gmcirco.github.io/blog/posts/c... tl;dr: There are some small gender preferences for a funky coffee, but most of the effect is age or expertise-based!

A brief blog post on synthetic controls. Specifically, what we mean what we talk about "micro" approaches. A walkthrough using microsynth and some data I generated on NYPD's 2014 Operation Impact. gmcirco.github.io/blog/posts/s...

Me and (social-less Andy Wheeler) released our approach for NIJ's Innovations in Measuring Community Perceptions Challenge. We propose a mail + web-based push survey to minimize costs, and use regression-based techniques to improve small area estimates. Check it out www.crimrxiv.com/pub/p2pxki1g...