Job ad: Postdoc to work with me & @rogierk.bsky.social at the Donders on lifespan development questions from Sept onward!
Profile: independent, good quant skills, interested in theory-driven work
Please share widely
www.careersatradboudumc.com/vacancies/po...
#Postdoc #AcademicJobs
Looking forward to that! I was very positively surprised by NumPyro, as it uses full NUTS and still is many times faster than Stan even for exactly the same problem. In this case, however, another issue with Stan is that it only allows pure NUTS, so hybrid sampler were impossible.
we could think of on several realistic models, and the speedup is somewhere between nada and 10x. Oftentimes closer to the latter. In real life that might be the difference between possible or not. Another lesson learned is that NumPyro is crazy fast! Paper link: arxiv.org/abs/2603.29647
Still, nobody has shown how to do DSEM with binomially distributed data. In this paper we fix this by introducing a hybrid NUTS-Gibbs sampler, which alternates between Pólya-Gamma sampling and NUTS with exact Kalman marginalization. We compare the algorithm to all alternatives
Ecological momentary assessment data are everywhere in the social and medical sciences. DSEM offers a great way of analyzing them, by coupling vector-ARIMA models for each individual through Bayesian priors, under the assumptions that two people in general at least share some similarity.
Still, in one of my simulations Metropolis-within-Gibbs was faster than NUTS-Kalman, as shown in the figure below. I'm curious how much of this is due to Stan. People have told me that NumPyro can potentially do this a lot faster, but I don't really believe it until I've tried.
Yes, doing the matrix operation in the Kalman filter the "textbook way", this algorithm became very inefficient for models with many indicators. However, I found some numerical tricks which reduced the dimension of the matrix inverse inside the Kalman filter, and then it got a lot better.
Kalman filter. It still targets the correct posterior, but can be orders of magnitude more efficient for some widely used models. Other times it is just a little bit more efficient. Open-source Stan implementation (@mc-stan.org) will be on OSF in the afternoon. Preprint: doi.org/10.48550/arX...
Dynamic structural equation modeling is very popular for ecological momentary assessment data, but the original algorithm contains an unnecessary bottleneck making latent variable modeling challenging. Here I propose the NUTS-Kalman algorithm, which replaces within-level sampling with a
well done Lior.
There seems a subset of scientists who don’t understand
probability and are convinced that statisticians are just pedantic killjoys determined to take their significant results away from them.
observation deviates from what we would expect at the given timepoint. The coefficients beta[i] are equal across timepoints by design, so this does not equate them per se. Asparouhov et al. in the original DSEM paper are a bit vague about this, but I think what we do what they call latent centering.
Thanks! The term b[i]x[i,t-1] is there to catch up systematic variations with some time-varying predictor x[i,t-1]. We subtract it for the same reason as we subtract alpha_i, namely to get rid of systematic variation. Hence, what's left inside the parentheses is the residual, i.e., how much the
The No-U-Turn Sampler, utilizing Hamiltonian Monte Carlo, was very efficient for estimating these models, and the paper is accompanied by annotated Stan code which practitioners can modify to their needs.
Out in Multivariate Behavioral Research today, open access:
doi.org/10.1080/0027...
Ecological momentary assessment data are everywhere these days. In psychology, dynamic structural equation models (DSEMs) are particularly attractive for analyzing such data. In this paper Ethan McCormick and I show how you can easily incorporate nonlinear trends and cycles using splines.
Hi Daniel, Creating visuals here, graphing over there? You can now create graphs and run analyses, all within BioRender. Our AI reads your raw spreadsheet, detects variables, and formats your data so you can generate publication-ready visuals in seconds. From scatterplots to heatmaps, regressions to ANOVA—no manual entry, no reformatting, no extra software required.
Would could possibly go wrong with using AI to read raw data to generate graphs that will put straight into articles for publication in scientific journals? 🤷♂️
and then do some tiny changes that removed the issue and kept the package on CRAN. Morale of the story: CRAN rules are there for a reason, and it's possible to figure things out if you just try obsessively enough.
Now time for resubmission!
Oh joy! BDR found a subtle warning on Fedora with gcc15 with very strict compiler flags that threatened to kick my galamm #rstats pkg out of CRAN by Saturday. After a week of headscratching I managed to set up my old Ubuntu laptop to reproduce the error :-)
Very interesting from @VidalDidac - MUCH higher reliability for structural neuroimaging measures with longer follow-up time rather than more follow-ups or higher n. 2.-year follow-up requires 4 times higher n than 6-year follow up. @LCBC_UiO direct.mit.edu/imag/article...
Tomorrow last day for early-bird registration to PrefStat 2025, 2nd International Summer School on Preference Learning for Ranking and Ordinal Data (www.prefstat.org)! In Oslo, 30.06 – 04.07. Register here: www.integreat.no/events/publi...
@valeriavitelli.bsky.social @ocbe.bsky.social @uio.no
I think I can check all the boxes. Feel free to DM me if you still need this.
Wow!
og dermed får mindre frihet til å disponere egen inntekt og det blir mindre penger å investere i teknologi som vi trenger for å kunne konkurrere med USA og Kina. Jeg er redd vi ikke er noe bedre enn resten av Europa, og det er synd. Jeg synes denne beskriver det bra: www.politico.eu/article/euro...
Jeg tror Norge lider av en forestilling om at velferdsstaten ikke bare skal være et sikkerhetsnett, men en forsikring som skal ordne alt, og at alt kan løses med større offentlig sektor. Strømstøtten og sykelønnsordning er gode eksempler. Prisen er høy, fordi alle må betale forsikringspremie (skatt)
it's almost a black box algorithm, self-tuning with minimal user input. Rankings remain tricky combinatorial objects so it's still quite computationally heavy, but we hope this is a step forward in making Mallows models more practically applicable.
arxiv.org/abs/2412.13644
so here we've instead developed nested sequential Monte Carlo algoriths, SMC^2 among friends. We derive the algorithms for a very general case of the Mallows model, and test them on complete rankings, top-3 rankings, pairwise preferences, and clustering. An additional advantage of SMC^2 is that
Preprint! The Bayesian Mallows model is a very flexible model for analyzing rank and preference data, and has been applied across a large number of domains. In many cases, however, the data naturally arrive sequentially in time. Existing Metropolis-Hastings algorithms scale poorly in this case
Thanks! I'll check it out.
Does anyone know of a good quantitative methods textbook for the social sciences? Ideally not too tied to a given analysis program, and not too much ANOVA stuff. I'm teaching a course which covers experimental design, multiple regression, mixed models.
Thanks!