Following Andrew, here is my (Aki’s) list of published papers and preprints in 2023 (20% together with Andrew)
Published
-
Manushi Welandawe, Michael Riis Andersen, Aki Vehtari, and Jonathan H. Huggins (2023). Robust, Automated, and Accurate Black-box Variational Inference. Journal of Machine Learning Research, accepted for publication.
arXiv preprint arXiv:2203.15945. -
Alex Cooper, Dan Simpson, Lauren Kennedy, Catherine Forbes, and Aki Vehtari (2023). Cross-validatory model selection for Bayesian autoregressions with exogenous regressors. Bayesian Analysis, accepted for publication.
arXiv preprint arXiv:2301.08276. -
Noa Kallioinen, Topi Paananen, Paul-Christian Bürkner, and Aki Vehtari (2023). Detecting and diagnosing prior and likelihood sensitivity with power-scaling. Statistics and Computing, 34(57).
Online
arXiv preprint arXiv:2107.14054.
Supplementary code.
priorsense: R package -
Martin Modrák, Angie H. Moon, Shinyoung Kim, Paul Bürkner, Niko Huurre, Kateřina Faltejsková, Andrew Gelman, and Aki Vehtari (2023). Simulation-based calibration checking for Bayesian computation: The choice of test quantities shapes sensitivity. Bayesian Analysis, doi:10.1214/23-BA1404.
arXiv preprint arXiv:2211.02383.
Code
SBC R package -
Erik Štrumbelj, Alexandre Bouchard-Côté, Jukka Corander, Andrew Gelman, Håvard Rue, Lawrence Murray, Henri Pesonen, Martyn Plummer, and Aki Vehtari (2023). Past, Present, and Future of Software for Bayesian Inference. Statistical Science, accepted for publication. preprint
-
Marta Kołczyńska, Paul-Christian Bürkner, Lauren Kennedy, and Aki Vehtari (2023). Trust in state institutions in Europe, 1989–2019. Survey Research Methods, accetped for publication.
SocArXiv preprint doi:10.31235/osf.io/3v5g7. -
Juho Timonen, Nikolas Siccha, Ben Bales, Harri Lähdesmäki, and Aki Vehtari (2023). An importance sampling approach for reliable and efficient inference in Bayesian ordinary differential equation models. Stat, doi:10.1002/sta4.614.
arXiv preprint arXiv:2205.09059. -
Petrus Mikkola, Osvaldo A. Martin, Suyog Chandramouli, Marcelo Hartmann, Oriol Abril Pla, Owen Thomas, Henri Pesonen, Jukka Corander, Aki Vehtari, Samuel Kaski, Paul-Christian Bürkner, Arto Klami (2023). Prior knowledge elicitation: The past, present, and future. Bayesian Analysis, doi:10.1214/23-BA1381.
arXiv preprint arXiv:2112.01380. -
Peter Mikula, Oldřich Tomášek, Dušan Romportl, Timothy K. Aikins, Jorge E. Avendaño, Bukola D. A. Braimoh-Azaki, Adams Chaskda, Will Cresswell, Susan J. Cunningham, Svein Dale, Gabriela R. Favoretto, Kelvin S. Floyd, Hayley Glover, Tomáš Grim, Dominic A. W. Henry, Tomas Holmern, Martin Hromada, Soladoye B. Iwajomo, Amanda Lilleyman, Flora J. Magige, Rowan O. Martin, Marina F. de A. Maximiano, Eric D. Nana, Emmanuel Ncube, Henry Ndaimani, Emma Nelson, Johann H. van Niekerk, Carina Pienaar, Augusto J. Piratelli, Penny Pistorius, Anna Radkovic, Chevonne Reynolds, Eivin Røskaft, Griffin K. Shanungu, Paulo R. Siqueira, Tawanda Tarakini, Nattaly Tejeiro-Mahecha, Michelle L. Thompson, Wanyoike Wamiti, Mark Wilson, Donovan R. C. Tye, Nicholas D. Tye, Aki Vehtari, Piotr Tryjanowski, Michael A. Weston, Daniel T. Blumstein, and Tomáš Albrecht (2023). Bird tolerance to humans in open tropical ecosystems. Nature Communications, 14:2146. doi:10.1038/s41467-023-37936-5.
-
Gabriel Riutort-Mayol, Paul-Christian Bürkner, Michael R. Andersen, Arno Solin, and Aki Vehtari (2023). Practical Hilbert space approximate Bayesian Gaussian processes for probabilistic programming. Statistics and Computing, 33(17):1-28. doi:10.1007/s11222-022-10167-2.
arXiv preprint arXiv:2004.11408.
Pre-prints
-
Lauren Kennedy, Aki Vehtari, and Andrew Gelman (2023). Scoring multilevel regression and poststratification based population and subpopulation estimates. arXiv preprint arXiv:2312.06334.
-
Alex Cooper, Aki Vehtari, Catherine Forbes, Lauren Kennedy, and Dan Simpson (2023). Bayesian cross-validation by parallel Markov chain Monte Carlo. arXiv preprint arXiv:2310.07002.
-
Yann McLatchie and Aki Vehtari (2023). Efficient estimation and correction of selection-induced bias with order statistics. arXiv preprint arXiv:2309.03742.
-
Yann McLatchie, Sölvi Rögnvaldsson, Frank Weber, and Aki Vehtari (2023). Robust and efficient projection predictive inference. arXiv preprint arXiv:2306.15581.
-
Frank Weber, Änne Glass, and Aki Vehtari (2023). Projection predictive variable selection for discrete response families with finite support. arXiv preprint arXiv:2301.01660.
jd asked Andrew “which paper from 2023 do you like best?”, and I also find it difficult to choose one. I highlight two papers, but I’m proud of all of them!
“Detecting and diagnosing prior and likelihood sensitivity with power-scaling” is based on an idea that had been on my todo list for a very long time, and seeing that it works so well and can have practical software implementation was really nice.
In “Practical Hilbert space approximate Bayesian Gaussian processes for probabilistic programming” we didn’t come up with a new GP approximation, but we were able to develop simple diagnostics to tell whether we have enough basis functions. I just love when diagnostics can answer frequently asked questions like “How do I choose the number of basis functions?”
Aki:
This is great. We should get all our contributors to post on their favorite work from the previous year.
I was thinking about the same!
Nice! I just read the “Detecting and diagnosing prior and likelihood sensitivity with power-scaling” paper. I had heard of the ‘priorsense’ package but hadn’t looked into it. Initially it would seem this is best for using on models fit with packages like ‘rstanarm’ with the default priors or maybe GP’s where it is more difficult to think about the parameters. However, I always use prior predictive checks anyway, to make sure that the priors make sense on the outcome space. In the first case study on the body fat data, you write that prior predictive checks indicated that the priors were reasonable, yet you still found prior data conflict for one of the coefficient parameters. So is diagnosing prior and likelihood sensitivity something that you would recommend to add to the workflow in general? Even if prior predictive checks look reasonable, you may still want to adjust the priors after seeing these sensitivity checks?
Another question – recently I have been doing a lot more work on much more mechanistic models that more directly model laboratory experiments. In these cases I use pretty informative priors for some parameters, which is needed to help the models converge and also makes sense in controlled lab settings. I already know there is prior sensitivity. If I saw that in a check, I wouldn’t necessarily change the prior. In that scenario is there still a use of checking the prior and likelihood sensitivity using power scaling approach?
Also, I’ll just add that I really like the advice on not iteratively tuning priors to make diagnostic warnings disappear. I’ve seen quite a few people doing this on questions on the Stan forum, where priors are almost viewed as simply a knob that one turns to make warnings go away.
We are working on more examples and improving the priorsense and other packages to illustrate how priorsense can be used with so little additional work that it would be sensible to add it to the workflow in general. The new examples focus even more on the quantities of interest in different examples. First few examples are part of my BDA course material (still waiting some tool improvements). Yes, prior predictive check and prior and likelihood sensitivity analyses are complementary. The result of the sensitivity analysis is not necessary adjustment of prior, for example, if you trust your informative prior and see conflict or weak likelihood, maybe you have to think about your data collection more, or at least you know you need to justify your prior more carefully and report how the inference for the quantity of interest might change given a different prior.
Cool, thanks. That makes sense. I’ll try it out.