reproducibility check [Nature]

While reading the Nature article Swarm Learning, by Warnat-Herresthal et [many] al., which goes beyond federated learning by removing the need for a central coordinator, [if resorting to naïve averaging of the neural network parameters] I came across this reporting summary on the statistics checks made by the authors. With a specific box on Bayesian analysis and MCMC implementation!

more air for MCMC

Aki Vehtari, Andrew Gelman, Dan Simpson, Bob Carpenter, and Paul-Christian Bürkner have just published a Bayesian Analysis paper about using an improved R factor for MCMC convergence assessment. From the early days of MCMC, convergence assessment has been a recurring (and recurrent!) question in the community. First leading to a flurry of proposals, [which Kerrie, Chantal, and myself reviewwwed in the Valencia 1998 proceedings], and then slowly disintegrating under the onslaughts of reality—i.e. that none could not be 100% foolproof in full generality—…. This included the (possibly now forgotten) single-versus-multiple-chains debate between Charlie Geyer [for single] and Andrew Gelman and Don Rubin [for multiple]. The later introduced an analysis-of-variance R factor, which remains quite popular up to this day, in part for being part of most MCMC software, like BUGS. That this R may fail to identify convergence issues, even in the more recent split version, does not come as a major surprise, since any situation with a long-term influence of the starting distribution may well fail to identify missing (significant) parts of the posterior support. (It is thus somewhat disconcerting to me to see that the main recommendation is to move the bound on R from 1.1 to 1.01, reminding me to some extent of a recent proposal to move the null rejection boundary from 0.05 to 0.005…) Similarly, the ESS may prove a poor signal for convergence or lack thereof, especially because the approximation of the asymptotic variance relies on stationarity assumptions. While multiplying the monitoring tools (as in CODA) helps with identifying convergence issues, looking at a single convergence indicator is somewhat like looking only at a frequentist estimator! (And with greater automation comes greater responsibility—in keeping a critical perspective.)

Looking for a broader perspective, I thus wonder at what we would instead need to assess the lack of convergence of an MCMC chain without much massaging of the said chain. An evaluation of the (Kullback, Wasserstein, or else) distance between the distribution of the chain at iteration n or across iterations, and the true target? A percentage of the mass of the posterior visited so far, which relates to estimating the normalising constant, with a relatively vast array of solutions made available in the recent years? I remain perplexed and frustrated by the fact that, 30 years later, the computed values of the visited likelihoods are not better exploited. Through for instance machine-learning approximations of the target. that could themselves be utilised for approximating the normalising constant and potential divergences from other approximations.

cost(s) of living

Yesterday, Andrew posted an announcement for a postdoc position in Paris, at the national medical research institute (INSERM) on Bayesian approaches to high throughput genetic analyses using nonlinear mixed effect models and the comments went ballistic about the low salary attached to this postdoctoral position, namely 2600€ – 3000€. As I have already commented on the rather stale clichés on French academics, let me briefly reflect on the limitations of comparing 3000€ a month in Paris with say $5000 a month in New York City. (Which seems to be at the high end of US postdoc salaries.) First, the posted salaries are “gross” but the French one already excludes the 25% taxes paid by the employer. I do not know if this is the case in the US. Second, comparing absolute values makes little sense imho. Even if the purchasing power parity is about one between France and the US, I think the long term cost of living [as opposed to visiting for a week] is lower here than there. If only because the amount is similar to, if higher than, the starting academic salaries and around the median salary. Interestingly, the same appears to be true for the US, if less favourably for the postdocs there.

simplified Bayesian analysis

A colleague from Dauphine sent me a paper by Carlo Graziani on a Bayesian analysis of vaccine efficiency, asking for my opinion. The Bayesian side is quite simple: given two Poisson observations, N~P(μ) and M~P(ν), there exists a reparameterisation of (μ,ν) into

e=1-μ/rν  and  λ=ν(1+(1-e)r)=μ+ν

vaccine efficiency and expectation of N+M, respectively, when r is the vaccine-to-placebo ratio of person-times at risk, ie the ratio of the numbers of participants in each group. Reparameterisation such that the likelihood factorises into a function of e and a function of λ. Using a product prior for this parameterisation leads to a posterior on e times a posterior on λ. This is a nice remark, which may have been made earlier (as for instance another approach to infer about e while treating λ as a nuisance parameter is to condition on N+M). The paper then proposes as an application of this remark an analysis of the results of three SARS-Cov-2 vaccines, meaning using the pairs (N,M) for each vaccine and deriving credible intervals, which sounds more like an exercise in basic Bayesian inference than a fundamental step in assessing the efficiency of the vaccines…

stratified MCMC

When working last week with a student, we came across [the slides of a talk at ICERM by Brian van Koten about] a stratified MCMC method whose core idea is to solve a eigenvector equation z’=z’F associated with the masses of “partition” functions Ψ evaluated at the target. (The arXived paper is also available since 2017 but I did not check it in more details.)Although the “partition” functions need to overlap for the matrix not to be diagonal (actually the only case that does not work is when these functions are truly indicator functions). As in other forms of stratified sampling, the practical difficulty is in picking the functions Ψ so that the evaluation of the terms of the matrix F is not overly impacted by the Monte Carlo error. If spending too much time in estimating these terms, there is not a clear gain in switching to stratified sampling, which may be why it is not particularly developed in the MCMC literature….

As an interesting aside, the illustration in this talk comes from the Mexican stamp thickness data I also used in my earlier mixture papers, concerning the 1872 Hidalgo issue that was printed on different qualities of paper. This makes the number k of components somewhat uncertain, although k=3 is sometimes used as a default. Hence a parameter and simulation space of dimension 8, even though the method is used toward approximating the marginal posteriors on the weights λ¹ and λ².