My presentation at #SAA2025 about modeling circadian rhythms in momentary affect is accessible on @OSFramework: https://t.co/VzYQfRtkZu. A summary thread of one of the papers discussed in the presentation is in the quote below:

A new semantic search engine for survey questions! This is based on our SurveyBot3000 model (to embed items) and 31k instruments from the APA PsycTests database.

To protect my sanity (almost literally) and to claim my night sleep back (literally) I decided to bite the 100-euro bullet today—🚬🚬🚬.

Me refusing to pay an additional €98/month on top of my current monthly €22 to upgrade my #ClaudeCode Pro plan to Max:

Are you doing #EMA research and wonder how to go about it? In recent work we've adressed some open questions and challenges, here is a brief summary of papers and materials. 🧵

For just $60, we classified half a million arXiv papers by institution with DeepSeek-V3🤯 You can now filter arXiv papers by organizations like DeepSeek, Meta, or OpenAI See which organizations are contributing the most to open science and research 🔎

We would like to thank Dr. Sacha Epskamp (@SachaEpskamp), Dr. Charles Driver (@CharlesDriverAU), and Dr. Michael D. Hunter (of Penn State, whom I could not find on Twitter) for their thoughtful reviews and comments which led to the major improvement to our paper. #PsychTwitter

If you have already read the paper, the major additions are in the discussion section and the footnotes, in which we give a critical account of how measurement error can affect SARMA models—and our findings—with additional considerations for automated model selection routines.

Our paper on day-of-week effects & weekly dynamics has been accepted for publication at @APA_Journals #PsychologicalMethods. Besides what is covered in the quoted thread, this version (on #PsyArXiv: https://t.co/IWg4JbNg3o) reflects further on model selection & measurement error.

For those curious about how o3-mini performs on multi-digit multiplication, here's the result. It does much better than o1 but still struggles past 13×13. (Same evaluation setup as before, but with 40 test examples per cell.)

1/3 Tutorial on exploring ecological momentary data is online at AMPPS, with: -Accessible ways to visualize data for better understanding -Models to get some first insights -Further reading boxes for more advanced topics -Reproducible pipeline you can run over your own data

Thanks for reading this #thread! In a subsequent working paper, we are exploring testing for the presence of cycles in multilevel (physiological) data, correctly estimating ψ and its CIs at levels 1 & 2, and understanding level-2 correlations with ψ. So stay tuned, #PsychTwitter!

And even if we let go of the supposed substantive interpretation of the cosinor parameters (& use it, e.g., to correct for daytime trends in AR models) it has very strong—and unrealistic—restrictions on the trend shape, making #CosinorAnalysis a bad choice for such scenarios too.

Having said the lack of substantive utility of peak timing in ESM data, one might consider salvaging the other parameters of the cosinor model, viz. the MESOR (quantifying the overall level) & the amplitude. However, the ESM night gaps undermine their substantive utility as well.

However, categorizing individuals into groups is not trivial at all: It involves arbitrary decisions re. when and how to set the thresholds, and what to do when the CI of ψ crosses those thresholds. (We do not report it in the paper, but ALMOST ALL CIs cross MULTIPLE thresholds!)

Depending on the item, 41-65% of our sample had peaks during typical waking hours, meaning that we can't interpret the peak offset parameter for up to 60% of individuals. At best only the individuals that fall within the same category of trend shapes can be meaningfully compared.

...and chances are the implied peak is projected during the night—& we have absolutely no data to corroborate whether the person is *actually* the happiest at, say, 3:00 AM. Thus, we are limited to studying the trends during the window of data collection, which can take 4 shapes:

Now comes the second, more fundamental issue: #ESM—literally "#ExperienceSampling"— data is only collected during (parts of) waking hours, & even if you believe we "experience" emotions while asleep, you cannot "sample" them. But the cosinor trend spans 24 hours of a full day...

With this mistake, the peaks of MANY individuals (up to 63% in a sample of 202) are mislocated by 12 hours, and changing the starting time of the cycle doesn't help. This could have been easily avoided by simply using the two-argument arctangent function (see also panel C above).

This simple fact has been overlooked in parts of the literature, which used the conventional arctangent function to obtain ψ. However, with this approach, individuals that peak 12 hours apart are not distinguishable, as the peak will always be estimated between 6:00PM and 6:00AM.