Ben Bolker
LightNews LightNews
bbolker.bsky.social
Ben Bolker
@bbolker.bsky.social
2.2K followers 570 following 690 posts
Ecology, evolution, epidemiology, statistics (mixed models). McMaster University, Hamilton, Ontario https://math.mcmaster.ca/bolker
Posts Replies Media Videos
bbolker.bsky.social
Great! People cite this as "C. Walck, “Handbook on statistical distributions for experimentalists,” Particle Phys. Group, Fysikum, Univ. Stockholm, Stockholm, Sweden, Tech. Rep. SUF-PFY/96-01, 2007" (690 GS hits @ scholar.google.ca/scholar?cite... (my colleague ended up citing Johson&Kotz, I think)
scholar.google.ca
November 26, 2025 at 7:12 PM
bbolker.bsky.social
zoölogical! naïve !
November 26, 2025 at 6:54 PM
bbolker.bsky.social
For what it's worth lme4 does have it already, but only in limited cases (only for single, scalar random effects terms IIRC)
November 26, 2025 at 1:47 PM
bbolker.bsky.social
On the other hand, here's a Harvard u/g saying Harvard should shift focus from theory/fundamentals to skills development (!!) (suggesting that CS majors shouldn't bother with a basic course on algorithms ??) www.thecrimson.com/article/2025... (I get 'help us prepare for the real world', but ...)
Harvard Is Training Us for a World That No Longer Exists | Opinion | The Harvard Crimson
At the end of the day, Harvard doesn’t need to end its liberal arts focus — it just needs to modernize the process. Harvard exists to train future leaders. Let’s make sure we equip them with the skill...
www.thecrimson.com
November 23, 2025 at 4:22 PM
bbolker.bsky.social
any chance Dr. Cobey would be willing to post slides?
November 20, 2025 at 4:38 PM
bbolker.bsky.social
Does everyone know about this one already? Meng 2018. “Statistical Paradises and Paradoxes in Big Data ...” Annals of Applied Statistics doi.org/10.1214/18-A...
Statisticians are increasingly posed with thought-provoking and even paradoxical questions, challenging our qualifications for entering the statistical paradises created by Big Data. By developing measures for data quality, this article suggests a framework to address such a question: “Which one should I trust more: a 1% survey with 60% response rate or a self-reported administrative dataset covering 80% of the population?” A 5-element Eulerformula-like identity shows that for any dataset of size n, probabilistic or not, the difference between the sample average Xn and the population average XN is the product of three terms: (1) a data quality measure, ρR,X, the correlation between Xj and the response/recording indicator Rj ; (2) a data quantity measure, √(N − n)/n, where N is the population size; and (3) a problem difficulty measure, σX, the standard deviation of X. This decomposition provides multiple insights: (I) Probabilistic sampling ensures high data quality by controlling ρR,X at the level of N −1/2; (II) When we lose this control, the impact of N is no longer canceled by ρR,X, leading to a Law of Large Populations (LLP), that is, our estimation error, relative to the benchmarking rate 1/√n, increases with √N; and (III) the “bigness” of such Big Data (for population inferences) should be measured by the relative size f = n/N , not the absolute size n; (IV) When combining data sources for population inferences, those relatively tiny but higher quality ones should be given far more weights than suggested by their sizes.
November 19, 2025 at 4:57 PM
bbolker.bsky.social
can someone remind me what MCID is? (I know I could probably look it up ...)
November 18, 2025 at 9:22 PM
bbolker.bsky.social
This is super-cool (ha). (1) Link to a preprint? (2) [stats pedantry] I'd prefer a statement about the strength of interaction (change in slope wrt °C / year of age) to a statement about dichotomous "stat sig for group X, non-sig for group Y" [sorry] Gelman & Stern. 2006. doi.org/10.1198/0003....
doi.org
November 18, 2025 at 2:37 PM
bbolker.bsky.social
To be a little more specific, the proposal would be "the prior on the intercept should be regularizing rather than flat when the model has a link function that restricts the domain" (e.g. Normal(0,4) for logit/probit/cloglog link, Normal(mean(y), 4) for log link?)
November 18, 2025 at 2:31 PM
bbolker.bsky.social
(2/?) I do get it that using flat priors for fixed effects forestalls a lot of grief from users about "why are my answers different from when I run this in lme4" ... (Also, I understand the reason for data-dependent priors but feels icky ... cran.r-project.org/web/packages... )
Note on data-based priors Because the scaling is based on the scales of the predictors (and possibly the outcome) these are technically data-dependent priors. However, since these priors are quite wide (and in most cases rather conservative), the amount of information used is weak and mainly takes into account the order of magnitude of the variables. This enables rstanarm to offer defaults that are reasonable for many models.
November 18, 2025 at 2:21 PM
bbolker.bsky.social
(1/?) One of those nice "obvious in hindsight" things. Think the rstanarm folks would be open for discussion? (brms is amazing but I prefer the rstanarm default priors which are typically more sensible/regularizing; e.g. prior pred sims are impossible with default brms priors.)
November 18, 2025 at 2:17 PM
bbolker.bsky.social
Huh, interesting. Would love pointers to literature/examples if you have any handy ...
November 17, 2025 at 2:09 PM
bbolker.bsky.social
Why is flat beta (i.e. {1,1}) not tractable? (Domain is bounded and thus this prior is proper, unlike flat priors for distributions with infinite domains ...) Maybe I'm missing something? (Of course perfectly reasonable to disagree whether B(0.5,0.5) or B(1,1) or something else is best ...)
November 17, 2025 at 2:07 PM
bbolker.bsky.social
reformulate()
November 17, 2025 at 3:04 AM
bbolker.bsky.social
according to the docs, by default it's Beta(0.5, 0.5), highest posterior density interval. sdorairaj.r-universe.dev/binom/doc/ma...
Package 'binom' reference manual
sdorairaj.r-universe.dev
November 16, 2025 at 1:28 PM
bbolker.bsky.social
PS I realize I could probably have vibe/LLM-coded this faster, but I had fun (and procrastinated) writing artisanal code ...
November 15, 2025 at 8:37 PM
bbolker.bsky.social
Huh. Surprised you'd imagine it was 2x higher (sorry if your point was hyperbolic for entertainment value ...) CAD hasn't been > $1.10USD since 1950 ... (data from fx.sauder.ubc.ca/etc/CADpages..., © 2024 Prof Antweiler, University of British Columbia) gist.github.com/bbolker/0bf3...
Time series of CAD from 1950 to 2025. Value fluctuated between $0.90 and $1.05 USD from 1950-1975, $0.60 to $0.90 1975-2005, briefly exceeded par around 2010, now back down to about $0.75 USD
November 15, 2025 at 8:29 PM
bbolker.bsky.social
reminded me of this guy en.wikipedia.org/wiki/Keron_T... (there were a lot of jokes about "he took the A train" ...)
Keron Thomas (c. 1975 – July 22, 2013)[2] was primarily known for posing as New York City Subway motorman "Regoberto Sabio" on May 8, 1993, and operating an A train in revenue service for over three hours, when he was sixteen years old.
November 13, 2025 at 2:05 AM
bbolker.bsky.social
Thinking about vibe-coding something to take BibTeX & (1) remove interior braces (2) Brace-protect [& case-correct] Proper/CAPITALIZED keywords (from a def file with regexps) (3) brace-protect/italicize species names (ditto). Anyone done this already? Like this: web.archive.org/web/20010707...
(Thanks to Rolf Sander for the shell script and for corrections.)
web.archive.org
November 11, 2025 at 1:01 AM
bbolker.bsky.social
follow-up: I guess you're referring to conclusions arising from Friston et al (2022) doi.org/10.1038/s415... ? That paper looked cool from a modeling point of view but I guess I'm not surprised it's non-identifiable. Can you point me to a published/public critique?
Dynamic causal modelling of COVID-19 and its mitigations - Scientific Reports
Scientific Reports - Dynamic causal modelling of COVID-19 and its mitigations
doi.org
November 10, 2025 at 1:20 PM
bbolker.bsky.social
Scraped data from the plot ... gist.github.com/bbolker/d6a6... Scraping is imperfect because the plot is a fairly crappy JPG (hard to distinguish overlapping points). Dashed line is 156 particles per liter, nominal normal/superspreader cutoff (points fall on either side due to scraping error)
plot of scraped data from Edwards et al 2021. On a logarithmic y scale, regression slopes are very similar. Red points are superspreaders (nominally > 156), black are 'normal' (nominally < 156)
November 9, 2025 at 5:46 PM