Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • -predict, mu- or -predict, mu fixedonly- after -melogit-

    Hello,

    I am using multilevel mixed-effects logistic regression -melogit- and I want to present the results using predicted probabilities. However, I am getting very confused about whether or not I want to use the -fixedonly- option when predicting the probabilities or not. I have found very useful information about how to estimate the different probabilities - for example from here:
    http://www.ats.ucla.edu/stat/stata/f..._xtmelogit.htm

    However, I haven't really seen an explanation of what effect this has on interpretation. How does the interpretation of the predicted probability differ if you use -predict, mu fixedonly- instead of -predict, mu- ?

    For example, say I have multiple 0/1 (failure/success) results on a number of subjects and I think that the test result depends on the location of the testing (each person was tested 3 times at each location), I would use the command:

    melogit result i.location || subject:

    There's a statistically significant difference between location, and I want to present this as predicted probabilities. What I really want is the probably of someone, anyone (i.e. not necessarily in the study), having a success. So, do I use -predict, mu- or -predict, mu fixed only- ? I am thinking that in order to generalise outside of the study population I need to include the random effects, so use -predict, mu-. Is that correct?

    Another example... Say that I am now only interested in people who have had a failure at some stage - i.e. I want to know the probabilty of sucess, given that they have had a failure at some stage. So a similar model, but restricted to those who have ever had a failure:

    melogit result i.location if ever_fail==1 || subject:

    This time, I would like the probably of a success for a person given that they have had a failure at some stage, so I think that the between subject variability is not of relevance. So, in this instance, I am thinking that I would want -predict, mu fixedonly-. Would you agree?

    I would really appreciate any help you can offer on this. I'm managing to get myself very muddled.

    Many thanks,

    Gillian



  • #2
    Hi,
    this older post concerns also my question. The link for UCLA does not work anymore, however it is archived for reference:
    https://web.archive.org/web/20160305..._xtmelogit.htm

    I think this no longer holds since Stata 15 does compute the effect even without the option "mu fixedonly". My question is then: what exactly does happen to the random effects in the model without the mu fixed option? What is margins doing here? And is there still any reason to use the "mu fixedonly" option or is it just better for generating the most accurate and realistic predictions to use the regular command without this option?
    Best wishes

    (Stata 16.1 MP)

    Comment

    Working...
    X