Eric Pedersen

I am not very convinced by Dutch Book arguments in general, but I do think that they show that *if* the environment is approximately stationary with multiple opposed decision makers, then any learning rule that doesn't approximate Bayesian updating will be vulnerable to exploitation by other agents

There is a lot of work on misspecification, so we know that e.g. a posterior should minimize the KL divergence between the true model and the specified one, but I don't think I've seen much on the question of "how does a learner come up with a model to compare with data in the first place?"

I feel like pretty much everyone working on decision theory seems to have tacitly agreed that "we don't really have any handle on the question of how you should learn the data generating process in the first place, so let's not dwell on it"

There are some interesting counter-examples of when posterior updating shows terrible learning properties when the true data generating distribution is not in the support of the posterior (e.g. bactra.org/weblog/606.h...)

bsky.app/profile/eric...

I hadn't realized it until today, but there really isn't much of a line between "No see, you should have told it that it's an engineer in the prompt" and "That's not right; you need to twirl your wand windershins and pronounce it alo-HO-mora instead"

17. If I had the option to press a button and LLMs would not have been, and never would be, invented, I would push it without hesitation. In fact, I'd have a considerable, positive willingness to pay to push that button.

5. Most frequentist methods are just *fine* and there's no need to always go full luxury bayesian in every application.

Meme image of two Spidermans pointing at one another; one labeled "penalized smooths" and the other labelled "random effects"

Related

I feel like @economeager.bsky.social has already made this joke somewhere, but I wasn't able to track it down

It's pretty on the nose! Linear algebra is just so powerful that it makes sense to spend hours trying to find some way of cramming your problem into matrix format just so you can use it

Inspired by reading Jason Bramburger's truly excellent book "Data Driven Methods for Dynamical Systems" epubs.siam.org/doi/10.1137/...

Meme image from Parks and Rec of a character dressed as a dictator, with text on it. text on top: "Solving high dimensional system? SVD." Text on bottom of image: "Nonlinear dynamics? Believe it or not, linearize and then SVD."

Every applied math book about dynamics or regression in short:

That really has a "My first Squarespace site" vibe to it, doesn't it

`mgcv` does use a quasi-Bayesian approach for calculating it's standard errors, though, that's designed to at least partially account for uncertainty in estimation of variance parameters. I suspect using the SEs from the frequentist cov matrix would have worse coverage

Plot of CI coverages for glm.nb, mgcv with Wald Intervals, and mgcv with t intervals. T intervals have coverage very close to 95% for all levels of Phi

Out of curiosity, I tried this same setup with `mgcv`, and was pleasantly surprised with the results. Center column: mgcv using standard fit +/- 1.96*SE, right column is using the critical value from the t-dist with n-2 df. (I didn't have the compute time to compare to other methods besides glm.nb)

Fair. I can't argue with that

Gif of actor Buster Keaton with a very deadpan expression, tipping his hat to someone off-screen.

But what about Buster Keaton?

I’m also thrilled to announce that I will be starting a faculty position with @centreecomgmt.bsky.social @ U of Guelph this fall!

I will be recruiting students & postdocs for 2026 - please reach out or come find me at #CSEE2025 if you’re interested in theory, global change & fisheries ecology 🐟

A blue poster describing opportunities to join the Bernhardt Lab -- for more information please see https://www.bernhardtlab.org/join-us

✨ The Bernhardt Lab at the University of Guelph is recruiting graduate students for 2026! Join us! We have several fully funded grad positions available ✨

Please spread the word!

www.bernhardtlab.org/join-us

#CSEE2025

Slide showing info for the Canadian Community of Statistical Ecology. Google group email is ccse@googlegroups.com. Also contains a QR code link for the site: https://groups.google.com/g/ccse-

It's great to see a packed house at the #CSEE2025 "Forefront of Statistical Ecology in Canada" symposium! It's a great sign of strength of the statistical ecology community here

If you want to join the community and conversation, sign up to the Google Group!

"8 km-long plane terrifies, amazes residents of Montreal!"

The issue is that both Animate Dead and Sending are 3rd level spells. Anyone you'd try to hire to watch your dead farm hands is already making way more as a day trader using corn spot prices from Waterdeep.

The only necromancers for hire would be edgelord losers with a bone to pick

Between Sinners and Kpop Demonhunters, this is turning out to be a surprisingly good year for musicals.

The Dortmunder novels by Donald Westlake (starting with The Hot Rock) are fun light caper novels, featuring a crew of very competent crooks with no luck at all. I've read the first 3 so far and have enjoyed them all

Yep, the coverage statements for CIs for GAMs in mgcv are averaged "across the function" (which is, confusingly, not the same thing as a simultaneous interval for the whole function; @gsimpson.bsky.social's got a nice post on the difference between these fromthebottomoftheheap.net/2016/12/15/s...)

The coverage statements for CIs / p-values for penalized terms are averaged across the domain of the penalized regression function though; so e.g. individual CIs for specific random effect levels might be overly conservative or optimistic

ON CONFIDENCE INTERVALS FOR GENERALIZED ADDITIVE MODELS BASED ON PENALIZED REGRESSION SPLINES Generalized additive models represented using low rank penalized regression splines, estimated by penalized likelihood maximisation and with smoothness selected by generalized cross validation or sim....

It is possible to calculate valid p-values with penalization by adjusting degrees of freedom for penalization; see Wood 2006 (onlinelibrary.wiley.com/doi/10.1111/...), and mgcv. It relies on some standard asymptotic arguments for calculating the values, but generally good small sample properties

Graphs of time series of raising and falling violent crime rates in New Zealand, Britain, Canada, and Australia. A second line on each plot shows the rise and fall of concentrations of lead in the blood of preschool children, lagged by 19 years compared to the murder rate. For each graph, crime rates and lagged blood levels showed very similar patterns of rise and decline across decades

I found the cross-country comparisons of rise and fall in murder rates vs lead blood levels by Kevin Drum in Mother Jones to be some of the most convincing evidence for me

www.motherjones.com/kevin-drum/2...

Strong argument that it's leaded gasoline driving a lot of that increase and decline