Daniel Oberski

good example of a weak quasi-experimental design that is nonetheless convincing.

Not in regular lists, which would be a closer equivalent to the given R code IMO.

Interesting. How it thinks R should work *is* how Python works, and a lot of the coding benchmarks are very focused on python. Maybe that is the explanation?

PhD candidate On the next generation of structural equation modelling tools We are looking for looking for PhD candidate in Methodology and Statistics On the next generation of structural equation modelling tools

🚨 PhD Job Alert/PhD Vacature🚨

English:
Come work with Katrijn Van Deun, Katharina Loter, and me on "SEM2.0" in Tilburg!

Nederlands:
Kom met Katrijn Van Deun, Katharina Loter, and mij werken aan "SEM2.0" in Tilburg!

www.academictransfer.com/nl/jobs/3514...

Sending an email with this content to semnet (or wanting to) is all part of the circle of life Eiko. Generations of researchers have found out: semnet simply cannot change its nature, like in that scorpion story.

I feel an instrumental variable hype coming on!

This is one of the most exciting parts of our calendar. Apply to SICSS-ODISSEI!

wow, excellent news!

For those interested in latent class, latent profile, mixture modeling, model-based clustering, all of last year's materials, including R code, are available online: daob.github.io/latent-class...

Good to see you're still keeping up with socials during this time of reflection :p

We have slightly updated our white paper on the future of Social Science Infrastructure in Europe. zenodo.org/records/1456...

We would be very interested in hearing thoughts and opinions. The global data landscape is changing dramatically and we need to think critically about what we do next.

Good luck Michel.

Anderson collected a lot more data than Fisher actually used for his example application in the LDA paper. What we did was feed some of that other data to many models at OpenML as well as the original LDA. I don't remember looking at overconfidence, but I'm sure you're right.

I know you're joking but we actually did this. Linear discriminant analysis was still the best :D

European Social Survey (ESS) Netherlands Community Conference 2025 - ODISSEI – Open Data Infrastructure for Social Science and Economic Innovations With the first release of the 11th wave of the European Social Survey, we would like to announce a Community Conference of the European Social Survey Netherlands. The conference will take place on the...

🚨 Call for Papers: European Social Survey Netherlands Community Conference 🚨

We invite colleagues from all disciplines using the European Social Survey to leverage our understanding of the social world.

Submit your abstract by October 31💻

Learn more and apply👉 edu.nl/f9vmp

#Sociology #cssky

There is no 1-1 conversion between probit and logit coefficients. Probit coefs are easy to interpret if you standardize the LRV ("STDY"): the coefficient is the number of sd's increase for a one unit increase of the predictor. Easy to convert to probabilities by pulling the prediction through pnorm.

I.e. the sense is likely "more interpretable" rather than "obvious".

I think they don't mean "intuitive" in the sense of "obvious" but in the sense that it gives some extra information: From V and the pattern of zeroes you can tell which linear dependencies are taken care of. E.g. if there's both age and year of birth in X, those will show up in one column of V.

A "digital twin"?

Thanks, this is *exactly* what I have been looking for!

LISS has a basket option which automatically merges the time points/datasets you need. Is this what you mean?

(also the "time-invariant" background variables tend to be simple copies, except when they are actually repeated; so a single copy is often enough, depending on the time period)

I'm sure he would be delighted that he was able to bring down the number of his opponents to 1 in 5 in only 363 years!

(sorry for the wonky colors btw)

Quite. Hopefully it's all just measurement error from being confused by the question??

Line graph showing the percentage of people believing "the Sun goes around the Earth" in 12 European countries, for the years 1992, 2001, and 2005. The averages for these years were about 10%, 30%, and 30%, respectively.

Eurobarometer 38.1, 55.2, and 63.1. Provided without further comment.

i.e. I confused myself!

I may have been confused by "gold standard": if you're doing observational data-causal inference but say you'd prefer to have an experiment, that implies you think that experiments are better. That crossed wires in my brain with your criticism of papers claiming experiments are not better.

Ok thanks! I think I still don't fully understand the "temporarily embarrassed" bit, but those do sound like extremely vague arguments. (which issues, what complexity, what context, why is it important, how does something else account for context?)

Even though LLM seem to me to be able to do this job reasonably well & more flexibly, there could be another reason to use regular expressions, when statcheck is used to check articles submitted to a journal: it encourages people to write the statistics in APA style.

This is my experience too - seems to work fine as long as you give it precise instructions to use Python etc. There are some ways to force it into actually evaluating numbers even more.

Hi Julia, this sounds like an interesting discussion, but I do not really understand what you mean. Which articles are you referring to? What kind of arguments do they give? And what do you mean by "temporarily embarrassed experimentalists"?