Resources for learning how to interpret xtmixed output

[butler bj]

Hello everyone,

I have a mixed factorial design — 2 (between subjects) x 2 (within subjects). I can easily just run a repeated measures anova, but I am trying to learn how to fit regression models, so I'm trying to analyze this data using xtmixed (now called mixed). I can fit the model fine, but my questions lie in interpreting the results — should I be interpreting the regression coefficients, margins, and contrasts as I would a simple OLS regression? If possible, I would love to be pointed to a resource that explains this (in further detail than the STATA manuals). I am considering purchasing Multilevel and Longitudinal Modeling Using Stata, Third Edition, but I fear this may be overkill for my simple design.

Thank you

[Joseph Coveney]

If you're concerned about interpretation, then I believe that Rabe-Hesketh and Skrondal would not be overkill for you.

Here's another aspect where the two volumes are helpful: deciding what to model (or what to leave out of the model). It's not necessarily trivial, and with mixed, you've got a lot of choices to make in that regard, even for a supposedly simple experimental design such as you describe.

[butler bj]

If you're concerned about interpretation, then I believe that Rabe-Hesketh and Skrondal would not be overkill for you.

Here's another aspect where the two volumes are helpful: deciding what to model (or what to leave out of the model). It's not necessarily trivial, and with mixed, you've got a lot of choices to make in that regard, even for a supposedly simple experimental design such as you describe.

Awesome, thanks for your response.

[Nathan E. Fosse]

I agree, Rabe-hesketh is a great resource for multilevel modeling in Stata. On issue not addressed is moving from the output to publishable papers.

Caveat regarding saved estimates (never mentioned in any book I've read)
The caveat here is that that Stata's saved eclass "random effects parameters" (varying-intercepts/slopes) are saved as logarithms of standard deviations. I have no idea why this is the case. If you are using these coefficients in tables and matrices be sure to backtransform them into variances or into standard deviations. Ben Jann has an example using the estout ado he wrote.

Try out the applied following overview (with emphasis on issues faced while producing knowledge):
I found these to be far easier. We most often want to know how to apply our data analyses and when to do them correctly. I am a big fan of the following short book (no Stata commands, however).
Nezlek, J. B. (2011). Multilevel Modeling for Social and Personality Psychology. Los Angeles, CA: Sage Publications.

Another book, to remind oneself that there remains much confusion regarding multilevel models:
All that said, I still enjoy the Rabe-Hesketh books; however, it goes quickly into questions, so brush up on your Greek symbols, remind yourself that the left-hand side (the outcome) of a mixed model is often presented with multiple subscripts (although this entirely conceptual).

Finally, remind yourself that there's a lot of confusion regarding the meanings of random effects, especially in research. Are we varying slopes? Intercepts? Both? And, no, we are never looking at effects qua effects in such models.

For the latter two sources of confusion, I recommend Gelman and Hill's book, which clarifies and graphs output (and includes a bit of Stata code).
Gelman, A., & Hill, J. (2007). Data analysis using regression and multilevel/hierarchical models. Cambridge; New York: Cambridge University Press.

[Mehreen Ansari]

Hello every one,

I am just a beginner researcher with so much (almost everything) to learn. I am working on my mixed/multilevel model for data of two levels i.e. firm level nested into Country level.
By using the xtmixed command on Stata 16, I now have my tables outreged but the problem here is the interpretation of those results. As i have looking at papers but unable to exactly understand how the interpretations for multilevel are different from OLS or panel regression results? Help regarding that is needed.

thanking in anticipation.