Read on Twitter
I really appreciate this episode’s bird’s eye view of SEM vs regression and overall agree that it often makes sense to just work in SEM.
But there’s a BUT. There are some important situations where regression comes to the rescue, so here we go:
1/11 https://twitter.com/quantitudepod/status/1348985575852961794
But there’s a BUT. There are some important situations where regression comes to the rescue, so here we go:
1/11 https://twitter.com/quantitudepod/status/1348985575852961794
Most importantly, there are loads of undergraduate and junior graduate students out there doing research for the first time. They have only learned the GLM and no, I don’t think it’s generally a good use of time to try to bring them up to speed on SEM at this stage.
2/11
2/11
You have 8-10 months to do an honours thesis. You should be learning about the whole research process, and some minor upgrades in stats so you can do your analysis.
Regression-appropriate research questions are essential here.
3/11
Regression-appropriate research questions are essential here.
3/11
Same goes for Masters theses for the most part IMO. You have 2 years for this, and your program covers the GLM in your first year, maybe some multivariate but probably not SEM. I see so many students with masters projects that need stats they don’t know. This serves no one.
4/11
4/11
Counterpoint... learning SEM can be part of the process; sure I get that. Tell me how many masters defences have you been to where you felt the student could articulate the nature of the model they tested beyond superficial details?
5/11
5/11
Rushing to do analyses the *more correct* complicated way means a lost opportunity for a junior student to grapple with fundamentals of the GLM in their first major project.
Next you know they’ll fall in with the mixture modellers and get into all kinds of trouble.
6/11
Next you know they’ll fall in with the mixture modellers and get into all kinds of trouble.
6/11
Second, I’m not sure how much of research needs to be elegant theory-driven hypothesis testing, which is what the focus of SEM was in this episode: set up *your model*, make it falsifiable, and test like the wind.
7/11
7/11
I love me some big theories, but they’ve also gotten us into trouble and we have a replication crisis to show for it.
Regression is a great tool for narrower questions about single parameters that may scale up to a bigger model later.
8/11
Regression is a great tool for narrower questions about single parameters that may scale up to a bigger model later.
8/11
And there are tools like equivalence testing to resolve the “is this Beta bigger than that Beta” question and related problems.
9/11
9/11
The last thing that comes to mind is that regression is really practical for questions focusing on parameter sensitivity that are data-intensive.
Specification curve analysis. SSVS. Different types of cross-validation.
10/11
Specification curve analysis. SSVS. Different types of cross-validation.
10/11
So to sum up, I heart regression and so should you. Unless you’re a full professor still using the GLM to answer SEM-appropriate questions because you can’t be bothered to learn it. I’ll deal with you in peer review.
/end
/end
