Nathaniel B
LightNews LightNews
banner
bechhof.bsky.social
Nathaniel B
@bechhof.bsky.social
1.8K followers 710 following 46 posts
statistics obsessor, economics enjoyer, occasional engineer, founder at TrovBase
Posts Replies Media Videos
bechhof.bsky.social
I indulged in a bit of fun doing a first pass of a regularized model for homicide declines with ChatGPT, using @jeffasher.bsky.social's data from jasher.substack.com/p/why-i-thin... to estimate "underlying" rates of homicide decline by city.
Scatter plot of 31 U.S. cities (excluding “Total”) showing raw vs. Bayesian MAP–regularized percent change in homicides from 2024 to 2025. Raw changes are on the x-axis, regularized changes on the y-axis. Each point is labeled by city, with small arrows connecting any labels placed more than 10 pixels from their marker. A caption below summarizes the log‐ratio transformation, hierarchical Normal(μ,τ²) model with an Inverse‐Gamma penalty on τ², and the posterior‐mean back-transform to percent change. Screenshot of Jeff Asher’s Substack titled “Murder is down a lot in big cities.” Below the title is a subtitle: “2025 vs 2024 YTD murders in top 30 most murderous cities.” A table lists four columns—City, YTD 2024, YTD 2025, and % Change—and shows, in descending order of decline, that Denver fell from 29 to 13 homicides (–55.2 %), Birmingham from 61 to 33 (–45.9 %), Cleveland 54→30 (–44.4 %), San Antonio 50→30 (–40.0 %), Las Vegas 54→35 (–35.2 %), Atlanta 50→33 (–34.0 %), Dallas 90→60 (–33.3 %), Indianapolis 70→49 (–30.0 %), and so on.
June 13, 2025 at 1:33 AM
bechhof.bsky.social
There's just something special about running a proper Bayesian MCMC script that I can't replace with anything else after doing it for almost 10 years
A screenshot of a Bayesian statistical model training interface showing NUTS sampler progress. The display includes information about a model with 45 subgroup columns, training on 52,380 of 68,850 rows. The sampling is being performed using PyMC NUTS (JAX-compiled) with 4 chains running in parallel. The table shows the current progress of each chain with metrics including number of draws (ranging from 601-644), step size (0.01), gradient evaluations, sampling speed (approximately 1.14-1.22 draws/s), elapsed time (8:49), and remaining time estimates.
May 13, 2025 at 4:21 AM
bechhof.bsky.social
Let it be known far and wide that @haroldpollack.bsky.social follows up (including custom cup!) when you answer his call for tracking down clean data. He really is the top notch conversationalist you know from his many podcast/media appearances
Nathaniel Bechhofer and Harold Pollack posing for a selfie with the custom cup that Pollack gave as a reward for linking a spreadsheet with 2024 presidential election results by Congressional district.
March 14, 2025 at 10:09 PM
bechhof.bsky.social
Okay I did not realize I was that early here; past me made a great call
Screenshot indicating I'm user #43,396 on blue sky
September 18, 2024 at 7:03 PM
bechhof.bsky.social
Just one more book
Larry Wasserman’s book All of Nonparametric Statistics
May 20, 2023 at 6:20 PM