eDNAjoint: An R package for interpreting paired or semi-paired environmental DNA and traditional survey data in a Bayesian framework

https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.70000

#statstab #295 The Fallacy of the Null-Hypothesis Significance Test

Thoughts: "the [..] aim of a scientific experiment is not to precipitate decisions, but to make an appropriate adjustment in the degree to which one accepts, or believes, the hypothesis"

#NHST #Bayes #ConfidenceIntervals #pvalues #significance #testing #hypotheses #likelihood #critique #fallacy

http://stats.org.uk/statistical-inference/Rozeboom1960.pdf

I got an email from the author promoting this benchmark comparison of #Julialang + StanBlocks + #Enzyme vs #Stan runtimes.

StanBlocks is a macro package for Julia that mimics the structure of a Stan program. This is the first I've heard about it.

A considerable number of these models are faster in Julia than Stan, maybe even most of them.

https://nsiccha.github.io/StanBlocks.jl/performance.html

„calling something logic doesn’t make it so. Calling someone rational doesn’t make it so“

I’ve been thinking for a while that, as someone who works on human rationality and rational argument, I should write a blog post on what that actually means (and, maybe more importantly, doesn‘t mean).

in the meantime, though, I found much to agree with in this piece:

Title: The magical thinking of guys who love logic
https://theoutline.com/post/7083/the-magical-thinking-of-guys-who-love-logic

What do you answer when a statistician asks you whether Bayesian methods are "sanctioned" or "legitimate" or "established" or similar?

When reporting a credibility interval, maybe also you, like I, are sometimes undecided between a 95%, a 90%, and an 89% interval (the last is common in the Bayesian literature). Well it turns out that the 89% interval has the following special property – for what it's worth:

Knowing whether the true value is within or without the 89% interval, corresponds to almost exactly *0.5 shannons* of uncertainty (more precisely 0.4999 Sh). That is, the uncertainty is half that of a 50% credibility interval, measured on the log-scale of Shannon information.

The 90% interval corresponds to 0.469 Sh. The 95% one, to 0.286 Sh.

So if one reports 50% and 89% credibility intervals, one is reporting 1 Sh and 0.5 Sh of uncertainty.

The remarks above don't pretend to be more than a curiosity :)

We have updated our paper "Advances in projection predictive inference" with Yann McLatchie, Sölvi Rögnvaldsson, and Frank Weber https://arxiv.org/abs/2306.15581 (to appear in Statistical Science). The paper reviews recent advances leading to more complete and faster workflow for projection predictive variable selection. We think we managed to clarify some parts in this revision.

I'm giving a half day Bayesian model selection tutorial in StanCon 2024 Oxford Monday 9th September https://mc-stan.org/events/stancon2024/. The tutorial will be aimed for those who are already familiar with Stan (running parallel with @rlmcelreath's Introduction to Stan tutorial). In addition to a basic overview, what would you like to hear?

`posterior` package (Tools for Working with Posterior Distributions) has new CRAN release which includes additional Pareto khat diagnostics support and corresponding vignette illustrating the use for diagnosing reliability of estimating any expectation with empirical mean https://mc-stan.org/posterior/articles/pareto_diagnostics.html

See other enchantments in the NEWS https://cran.r-project.org/web/packages/posterior/news/news.html

`priorsense` package for prior and likelihood sensitivity analysis is now on CRAN. It works with Stan (brms, rstan or cmdstanr) fits, or just posterior draws, as long as the log prior density and log likelihood evaluations (at the locations of the posterior draws) are available.

See more in Noa Kallioinen's post at
https://discourse.mc-stan.org/t/priorsense-1-0-is-now-on-cran/35635

It was great to have @MineDogucu visiting Aalto! She gave a great talk, answered plenty of audience questions, and we talked for hours about teaching Bayes

CmdStanR v0.8.0 Released

Major
- fix to support CmdStan 2.35 coming next week
- use any CmdStanFit object as inits

Other
- show_messages and show_exceptions arguments
- format argument to the unconstrain_draws()
- CmdStan arguments save_metric and save_cmdstan_config
- documentation for CmdStanR global options
- psis_resample and calculate_lp arguments added to Pathfinder

See more at https://discourse.mc-stan.org/t/cmdstanr-v0-8-0-released/35200

Is climate science biased by gender?

The Guardian writes in a recent article (https://www.theguardian.com/environment/article/2024/may/08/world-scientists-climate-failure-survey-global-temperature):
"Female scientists were also more downbeat than male scientists, with 49% thinking global temperature would rise at least 3C, compared with 38%."

Think about the implications of that sentence: assuming that all surveyed climate scientists are experts on the topic with essentially the same scientific information, they come to different conclusions depending on their gender. One way of interpreting this is that climate science is not objective, which would be troubling.

When we put uncertainty bars (from a Bayesian log-linear model) on the percentages reported by The Guardian the message becomes less troubling in one sense (see figure below).

It looks as if the data is consistent with a fraction of roughly about 40-45% of climate scientists -- irrespective of gender -- thinking that temperature will rise by more than 3°C, which is troubling enough.

More on this on the blog: https://manywordsandnumbers.org/2024/05/12/is-climate-science-biased-by-gender-%f0%9f%94%a5/
#climatecrisis #bayes #uncertainty

My Nabiximols case study https://users.aalto.fi/~ave/casestudies/Nabiximols/nabiximols.html has a nice set of PIT-ECDF plots that illustrate typical cases. PIT stands for probability integral transform, that is, the cumulative density or probability up to the observation. Given a well calibrated predictive distributions, PIT values should have uniform distribution.

The first one shows PIT-ECDF plot for a binomial model with too many PIT values close to 0 and 1, which indicates too narrow predictive distributions.

Bob Carpenter told he had successfully used @mcmc_stan 's Pathfinder with "a big poli sci model with 1M latent parameters and 12M binary observations in 12 groups".

(For those who don't know Pathfinder yet, see https://jmlr.org/papers/v23/21-0889.html)