There are some track features you can get, like runtime or explicit lyrics, but the interesting ones like key, bpm or danceability are part of the audio features API which has been disabled since last year (developer.spotify.com/blog/2024-11...). DDP+Track analysis could still be very interesting.

I tried to do this in my computational methods class this spring and it turned out the most of the Spotify tracks/music API is now disabled, so you cannot easily get matching song-level data anymore, except from large dumps of popular songs, or by using alternative APIs using artist/title.

For a second I was really impressed that @richardfletcher.bsky.social brought out HUGE cardboard printouts for his slides ;-)

I'd argue that BA/MA theses don't need to be informative re an actual research question, but only inform us whether students actually learned something. Of course, running naive studies with tiny self-selected or convenience samples isn't helping the students in that regard either.

Interesting. Do the funders provide reasons for this requirement, ie are they science-related or legal? In many ethics/IRB (equivalent) discussions I've had in the past, it seemed that collecting less demographic information seemed to be preferable in general for privacy/data protection reasons.

What happens if you have NA in your demographics? Do papers get rejected if the reported mean age is based on 73% of the sample? I get the warmup argument, but why not warm up with motivating items? If however, your experiment's primary outcome is age or gender, put that item in ASAP ;-)

This is also what I learned and continue to teach. Sociodemographics rarely a predictor (or outcome!) of interest, at least in my courses...

I am a big fan of www.soscisurvey.de The basics are free anyway but university licenses of various types including for own servers exist as well, I think psychology at UZH even has one.
Plus, I've had very good experiences with support as well.

Good and timely stuff! I wonder how we can navigate the (very real) tradeoff between quality and reproducibility. Will reviewer #2 accept a drop of .2 in accuracy between gpt/gemini vs. local models if you make this decision transparent? What if results change? Just add another robustness check?

If you think of time as a sampling dimension (which we often do in many other contexts), it's not even a metaphorical population, but an occasion-specific sample. Which is also argued in the above literature regarding country-year data.

Not to be confused with Schöpfungshöhenselbstzweifel!

I sense a strong nudge to never reject anything 🤗

Why not keep the current funcs and switch between hidden and research API based on an argument or an API key being present in the background? Or use hidden API as a fallback when both can be used for the same task? I use traktok (thanks!) for teaching and hope for non-breaking changes ;-)

Predoc, actually.