Read on Twitter
In late Dec I solicited suggestions for what a structural equation modeling (SEM) course should contain.
I'm grateful for everyone who engaged & replied.
I promised to summarize those in the new year, so here's a thread on what folks had to say. https://twitter.com/aidangcw/status/1341404133551693833?s=20
I'm grateful for everyone who engaged & replied.
I promised to summarize those in the new year, so here's a thread on what folks had to say. https://twitter.com/aidangcw/status/1341404133551693833?s=20
Starting with a general sentiment, several comments that basically said, "SEM isn't the answer to everything"
@bmwiernik offered a scathing "you don't need it" and he and others suggested regression is often all you need.
A few thoughts on this... https://twitter.com/bmwiernik/status/1341413658702843904?s=20
@bmwiernik offered a scathing "you don't need it" and he and others suggested regression is often all you need.
A few thoughts on this... https://twitter.com/bmwiernik/status/1341413658702843904?s=20
1st, what did I expect, of course twitter offers the nihilistic answer for any question.
2nd, I agree, you often don't need all of SEMs badass features.
3rd, and most importantly, knowing SEM isn't just about estimating complex models, it's so much more valuable than that...
2nd, I agree, you often don't need all of SEMs badass features.
3rd, and most importantly, knowing SEM isn't just about estimating complex models, it's so much more valuable than that...
Knowing SEM gives you much more control and flexibility in understanding and thinking about models you can test in data...more modeling flexibility can unlock new theoretical insights.
This point was made very nicely in a @quantitudepod episode here: https://twitter.com/quantitudepod/status/1348985575852961794?s=20
This point was made very nicely in a @quantitudepod episode here: https://twitter.com/quantitudepod/status/1348985575852961794?s=20
The @quantitudepod folks note that regression models are just special cases of SEM, but very limited, and why limit yourself?
I'd go further, though, and say that you need to add multilevel/mixed effects models into the mix, and MSEM is the most general framework.
I'd go further, though, and say that you need to add multilevel/mixed effects models into the mix, and MSEM is the most general framework.
This is the biggest argument for using Mplus as a modeling program, because it offers near seamless access to all the models that fall under the broader rubric of MSEM, of which basic SEM, MLM, and the generalized LM are all special cases.
The field needs to move here, IMO.
The field needs to move here, IMO.
Many other specific comments struck a cautionary tone.
The basic message is "with great power comes great responsibility", and this can't be emphasized enough.
SEM removes major guardrails that come with more limited methods, and it's easy to stray and fall off the cliff.
The basic message is "with great power comes great responsibility", and this can't be emphasized enough.
SEM removes major guardrails that come with more limited methods, and it's easy to stray and fall off the cliff.
Some other posts emphasized a model comparison approach, suggesting that you really should be comparing multiple models.
Also that there are often (MANY) equivalent models, and the only way you might be able to adjudicate these will be theoretical.
Also that there are often (MANY) equivalent models, and the only way you might be able to adjudicate these will be theoretical.
I found the discussion with @mijkenijk regarding the conceptual and empirical status of formative vs. latent variables to be particularly informative. https://twitter.com/mijkenijk/status/1341432442427645952?s=20
There are a lot of other specific comments that are really useful, but sort of narrow, and I'd recommend that you check out the thread's replies.
Thanks to everyone who shared.
Thanks to everyone who shared.
