Sudip Paul writes:
I’ve written a blog post arguing that the philosophical reference class problem — which comparison group is “right” for assigning a probability to an individual case — is, once the grouping structure is fixed, the same estimation problem that partial pooling was built to handle. The bias-variance tradeoff between narrow and broad reference classes is resolved by multilevel shrinkage rather than by picking a level.
The post includes a literature review. One finding that may interest you: I couldn’t find any published work that explicitly connects the reference class problem to partial pooling in multilevel models, despite the machinery being standard. The closest prior work connecting the two comes from David Poole’s group at UBC, published in AI venues. The philosophical literature (Hájek, Wallmann, Roth & Tolbert) discusses the problem extensively without arriving at partial pooling.
I replied with some links:
– In chapter 1 of Bayesian Data Analysis we discuss the foundations of probability and the idea of a reference set. In the third edition, see page 12 where we explicitly use the term “reference set” and sections 1.6 and 1.7 for empirical examples.
– I have a post from 2018 called “Bayesians are frequentists,” where I write, “the Bayesian prior distribution corresponds to the frequentist sample space: it’s the set of problems for which a particular statistical model or procedure will be applied.”
– And another post from around the same time, “The idea of replication is central not just to scientific practice but also to formal statistics . . . Frequentist statistics relies on the reference set of repeated experiments, and Bayesian statistics relies on the prior distribution which represents the population of effects.” This is cool because we’re connecting three concepts: the Bayesian prior, the frequentist reference set, and real-world replications. This relates to multilevel modeling too, in that studies of real-world replications will be done using meta-analysis, which uses multilevel modeling.
– And here’s a 2011 article, “Bayesian statistical pragmatism,” where I write, “Bayesian probability calibration is closely connected to frequentist probability statements, in that both are conditional on ‘reference sets’ of comparable events” and “Bayesian probability, like frequentist probability, is except in the simplest of examples a model-based activity that is mathematically anchored by physical randomization at one end and calibration to a reference set at the other.”
Bayesian inference is all about partially pooling toward the prior, and “the prior” represents a population or reference set of relevant problems.
In short, I agree with Sudip Paul, and it’s super-frustrating that many people don’t seem to realize this point.
Paul responds:
Having read through the references you sent — I think the gap my post is trying to fill is the last step: connecting what you’ve been saying about priors and reference sets to the philosophical reference class problem by name — Venn, Reichenbach, Hájek — and framing partial pooling as the resolution to what philosophers have treated as a foundational problem for 150 years. You’ve been making the statistical point for years, but the philosophers discussing the reference class problem haven’t found it (Hájek’s 2007 paper doesn’t cite any multilevel modeling literature, and Wallmann and Williamson’s 2017 survey of four approaches to the problem doesn’t include partial pooling). The two conversations have been happening in parallel without connecting.
In the months since the above exchange, Sudip Paul has expanded his post into a paper. He adds:
Beyond the post, the paper adds a decision-theoretic argument (via Stein inadmissibility, every selection rule in the philosophical literature is dominated by shrinkage in total risk), shows that the classical proposals — Reichenbach, Carnap, Kyburg, Pollock — are limits or special cases of the same pooling apparatus, and works through the exchangeability bridge from de Finetti to the multilevel model.
I like to call the Dawid/Rubin/Gelman style of Bayesian statistics “pragmatic Bayesian.” It’s laid out right on page 1 of BDA1 and hasn’t changed. Step 3 in their 3-step process for Bayesian workflow is:
It’s the go back and fit a better model if the first one fails part that makes this unusual from a traditional Bayesian (or frequentist) perspective. It was a comfortable place to be for me coming from machine learning and natural language processing and industry, where of course we’re going to fix a model if its predictions are bad.
Ironically, the traditional Bayesians fell into two opposing camps. The first is the “objective Bayesians” (or “cringing Bayesians” as Andrew might say), who wanted to have uninformative or “reference” priors because they didn’t really believe in using prior information. The second camp contains the “subjective Bayesians”, who want to put down whatever they “believe” in priors. The key issue with both the objective and subjective Bayesians is that neither of them believe in checking the calibration of their posterior predictive inferences.
It’s precisely the calibration part of posterior inference that Phil Dawid wrote about. I like Gneiting and Raftery’s elboration and fine grained analysis of calibration. The whole concept doesn’t get much of a look-in by that name in BDA, being only mentioned in two places. Chapter 1 mentions it around record linkage, but that was not in the original edition. On page 128, the authors discuss calibration of means as a kind of Bayesian measure of bias, noting that if the model is correct, Bayesian posteriors are calibrated (in expectation) by construction. I wrote a blog post on this topic, Bayesian models are calibrated by definition, ages ago, because it’s the basis for simulation-based calibration checks.
Bob:
Interesting point that we don’t talk much about calibration in BDA. We don’t talk much about external validation more generally! I guess that we were following the statistical tradition of analyzing the data right in front of us. Rubin did discuss calibration in his famous 1984 paper but by the time we were writing BDA, he wasn’t so interested in the topic either. I got back into calibration through Aki: we were trying to improve DIC, which led us to WAIC, which led us to cross validation as a more general principle. And we do have some of that in chapter 7 of BDA3. Also there’s an example in chapter 2 of modeling rare cancer rates—it derived from a classroom demo I did many years ago, and then I worked it out as a Bayesian analysis, and we do calibration: we fit the model to the first half of the data and use it to predict the second half. Phil and I did a similar leave-some-out calibration in our first paper on radon modeling.
From a teaching perspective, one flaw in both those examples is that they both have the purpose of showing that the Bayesian inference gives calibrated predictions. This is a fine lesson, and it counters naive beliefs about “unbiasedness,” but one thing I didn’t discuss is the use of external calibration for model criticism.
> Bayesian models are calibrated by definition
To be fair, that seems trivial enough to not deserve much discussion. Given a model and conditional on the data, if we assume that the prior distribution is correct the posterior gives the correct probability density for the parameters. If the parameters do indeed follow that prior distribution the assumption above is always warranted.
Carlos:
The reason why the statement, “Bayesian models are calibrated by definition,” is worth discussing is that the full version is, “Bayesian models are calibrated when averaging over the prior predictive distribution”—which reminds us that they’re not in general calibrated if the prior or data model is wrong, or if you don’t average over the prior. All that is mathematically trivial too, but mathematically trivial results can deserve discussion if they affect practice.
Averaging over the prior predictive distribution? Something like ABC method?
In this case can we can that abc is approximately calibrated under the correctly specified model?
Aminreza: This doesn’t require ABC. It’s just saying that if our data really comes from simulating from the model (generate parameters from the prior, then data conditional on the parameters), then our posteriors are calibrated in expectation. We don’t get a worst-case guarantee.
> they’re not in general calibrated if the prior or data model is wrong, or if you don’t average over the prior.
My comment was in the context of whether it’s true that Bayesian posterior intervals have the right coverage when the prior and data model are correct (and why that’s, apparently, not discussed much in the book). If it had been about whether it’s true that Bayesian posterior intervals do not have the right coverage when the prior and data model are not correct that seems even more trivial but I agree with the broader point that using a wrong model not as good as using a correct model (and expecting otherwise is a recipe for disappointment or worse).
> Bayesian inference is all about partially pooling toward the prior, and “the prior” represents a population or reference set of relevant problems.
One may also say that Bayesian inference is all about how we represent uncertainty and how it changes as we incorporate additional information.
I see how I could make it about a reference set of relevant problems, just like I could make the concept of 37% probability to be about the reference set of things that have 37% probability. (I guess that with some effort I could even make it about partially pooling toward the prior.) That interpretation may be useful in some cases but doesn’t seem the essence of Bayesian inference in the usual sense of the term.
Andrew:
Working my way through the various connected readings, it did strike me that this has relevance for the question of how to model intersectional identities.
On this quote from Paul:
“You’ve been making the statistical point for years, but the philosophers discussing the reference class problem haven’t found it (Hájek’s 2007 paper doesn’t cite any multilevel modeling literature, and Wallmann and Williamson’s 2017 survey of four approaches to the problem doesn’t include partial pooling).”
Hajek doesn’t cite multilevel modeling literature because it’s not relevant to the argument he’s making. The metaphysical-epistemological split doesn’t have anything to do with estimation per se: it’s about the whole process of choosing a single number to report. Hajek opens with this description of the split:
“Here it is important to distinguish a metaphysical problem and an epistemological problem. The former problem arises because it seems that there should be a fact of the matter about the probability of X; what, then, is it? The latter problem arises as an immediate consequence: a rational agent apparently can assign only one (unconditional) probability to X: what, then, should that probability be?”
After going through his various arguments, he then makes two moves: probability is fundamentally relative, so there is no “one (unconditional) probability” to assign to X, dissolving the metaphysical problem; we are still left with the choice of which conditional probability to assign to X, which leaves the epistemological problem open. The reason I bring this up, and the reason I phrased it this way, is that by choosing partial pooling with some model, you’ve already made the choice of which probability to report–P(X | everything in the model). The Bayesian machinery gets you the optimal estimate for it, given the model, but it doesn’t make the choice itself go away. As Paul points out in the paper, but maybe doesn’t quite realize the consequences of, this simply bumps the reference class problem up a level, and it’s just as unsolved there as it was before, because:
1. If partial pooling fixes the conditional to P(X | everything in the model); and
2. You choose what goes into the model; then
3. You’re choosing which conditional to assign;
which leads you back to the epistemological version of the problem, as I understand Hajek to have conceived it, leaving it unsolved. If most people don’t realize how partial pooling addresses what remains of the reference class problem, it might be, well, because it doesn’t. :)
P.S. I understand the paper itself has been flagged as AI-generated (by me!), but I thought I’d respond, since it manages to raise some interesting philosophical points that seem worth the discussion, the questionable authorship notwithstanding.
P.P.S. I wonder whether there’s a connection to Wolpert’s NFLT somewhere lurking in here.