Hello Everyone! I have been experimenting with using #Quarto to call #julialang to run #multilevel models with #MixedModels. Unfortunately, my document is taking about 10-15 minutes to render with small data sets. I've found it difficult to understand the #julialang documentation on this issue, so would appreciate any "Explain to me like I'm 5" explanations of how to speed up #julialang. Code is here: https://agrogan1.github.io/multilevel-multilingual/. I am grateful for #julialang, just wish I could figure out the speed.

New on the blog: showcasing the immense hackability of #brms by extending a random intercept model with linear predictors on the standard deviation of the random intercept. Should you do it? Most likely not, but if you really really want, there is a way. Also the techniques shown are general and let you do a lot of other crazy stuff with brms. Happy for any feedback!
https://www.martinmodrak.cz/2024/02/17/brms-hacking-linear-predictors-for-random-effect-standard-deviations/

Exciting News! Our work on Reliability and Feasibility of Linear Mixed Models in Fully Crossed Experimental Designs published in AMPPS! #R #lme4 #MixedModels @Scandle & @letstido @universityofleeds

https://journals.sagepub.com/doi/10.1177/25152459231214454

We present #recommendations and a clear #pipeline for handling #random effects in the presence of non-convergent and singular models. No more reduced models causing first-type errors due to data pseudoreplication!

Recent #PeerReview work on #MixedModels and #GrowthModels had me going back to these classical texts on the genre:

1) https://journals.sagepub.com/doi/10.3102/01623737018004265

2) https://www.cambridge.org/core/journals/development-and-psychopathology/article/design-and-analysis-of-longitudinal-studies-of-development-and-psychopathology-in-context-statistical-models-and-methodological-recommendations/3F89776EE1F27BA4E4F49435AFBB7662

They are excellent tutorials on the most basic aspects of incorporating trajectories / functional change into such models. And the assumptions these require.

Another #PeerReview finished.

Paper ~ 7000 words
Review ~ 2000 words
Duration ~ 2 hours

I notice a long manuscript less, if it is well-written. Main point here once again:

#MixedModels are difficult to report. This paper offers a lot of detail on how to develop such a project and report it (especially Table 7): https://www.sciencedirect.com/science/article/pii/S0749596X20300061

The chapter in Hancock & Mueller's "The reviewer's guide to quantitative methods in the social sciences" is also very helpful.

I often look at papers where authors used a lot of effort to shoehorn a #LongitudinalAnalysis into a trajectory or #MixedModels that do not quite the job the team wants.

Analysing longitudinal data (esp. w time-varying covariates) via G-Estimation is an alternative for consideration:
https://journals.sagepub.com/doi/full/10.1177/25152459231174029 #Tutorial

The underlying thinking is not entirely different, but often one needs only a little step / laterality to get a new view on an analysis problem.

Another #PeerReview finished.

Paper ~ 7000 words
Review ~ 2000 words
Duration ~ 2 hours

You notice a long manuscript less, if it is well-written.

Main point here once again: #MixedModels are difficult to report.

This paper offers a lot of detail on how to develop such a project and report it (especially Table 7): https://www.sciencedirect.com/science/article/pii/S0749596X20300061

The chapter in Hancock & Mueller's "The reviewer's guide to quantitative methods in the social sciences" is also very helpful.

Another #PeerReview finished.

Paper ~ 4700 words
Review ~ 1500 words
Duration ~ 2 hours

The application of #MixedModels requires discussion of the decisions made in modeling as well as detailed reporting of a range of results.

This paper offers a lot of detail on how to develop such a project and report it (especially Table 7): https://www.sciencedirect.com/science/article/pii/S0749596X20300061

Unfortunately it is not #OpenAccess and no alternative version seems to be available

New Blogpost on Type-1 error in LMM/MixedModels

What happens if one does not include random-slopes?

https://benediktehinger.de/blog/science/lmm-type-1-error-for-1condition1subject/

Including an Interactive Demo: https://benediktehinger.de/interactive-pluto-notebooks/output/lmmType1_simple.html

this is mostly restating Barr et al. 2013 and making it freshly accessible - but I hope it helps!

@cogsci @cognition @eeg #statistics #MixedModels

If you want to control for the "repeat" in a repeated measures design using LMMs - you have to model that random slope!

---

y ~ 1 + cond + (1|subject)

does *not* control for within condition effects (except if you have only 1 trial per level per subject)

---

If this sounds relevant to you, I could prepare a blog-post + interactive demo

Take a weird dive into the Intraclass Correlation Coefficient (ICC) with my newest statistics meditation!

https://youtu.be/PqFJ2cggFfY

How can the ICC be a correlation and a proportion of variance at the same time? Zone out to this question, the chickens, and the roosters.

This is probably most interesting to you if you are already mildly motivated to think about the #ICC. #Statistics #Meditation #IntraclassCorrelation #MixedModels #MultilevelModels #Correlation #VarianceComponents #STEAM

Finally working to wrap my head around time-varying predictors. Read Wang & Maxwell 2015 today, which discusses detrending X and Y in #mixedmodels to remove the effects of time. Is this actually appropriate for the context of #aging research?

https://pubmed.ncbi.nlm.nih.gov/25822206/

Time for an #introduction:

I’m Ewan a statistician and researcher in Biostatistics & Health Informatics, King’s College London. I do things with numbers and mental health.

Right now, that includes: (1) mental-physical links in routine data; (2) wearable sensors to predict depression; (3) clinical trials of digital interventions.

I’m excited about #OpenScience, #Bayesian (#brms), #MixedModels, #rstats, #tidyverse, #Quarto, #DAGs.

Elsewhere: gravel cycling, ️, ️, breaking things in #Linux.