08-13-2016 03:42 AM
I'm doing a Multilevel logistic regression with the proc glimmix statement.
For interpretation purposes, I build my model in the following steps (E.G these results are shown in a table)
(1) random intercept + level 1 variables
(2) random intercept + level 1 variables + interaction terms with gender (I want to look at the possible genderdifferences)
(3) random intercept + level 1 variables + level 2 variables
(4) random intercept + level 1 variables + level 2 variables + cross level interaction terms.
Now it appears that the third step is the best fitting model. However, I want to say something about the second step where I made the interaction terms (step 2 is better fitting than step 1). Can I do this if I mentioned: 'Although step 3 is the best fitting model, It could be interesting to look at model 2 where genderdifferences are exposed.
Thanks a lot!
08-15-2016 01:00 PM
If I could ask, how are you deciding the best fit?
Also, since you are fitting logistic models where marginal models suffer from regression to the mean, have you considered conditional models? (See Stroup's Generalized Linear Mixed Models textbook for more on this.)
08-18-2016 04:22 PM
Hi there! Thanks for the quick answer!
I Used the -2LL of each model and also the parameters!
Thanks for the tip, I didn't consider that but will defenitly look at it. However, my models are ready and finished.
Now, I have an other question that I would like to ask. Maybe you know the answer to this, but it seems to be a difficult subject.
If you estimate two models (both multilevel logistic regressions) to look at the association of the socioeconomic status and alcoholconsumtion. My model contains (1) moderate drinkers, (2) hazardous problematic drinkers and (3) harmful problematic drinkers. The last two groups are also used as one group 'problematic drinkers'.
In the first regression I want to look at the odds of being a problematic drinkers (in comparison to moderate drinkers).
In the second regression I want to have a closer look at the 'problematic drinkers' and therefore look at the odds of being a harmful problematic drinkers (in comparison to hazardous problematic drinkers).
Suppose that the log odds of women in the first regression is -0.45, and in the second regression -1.24. Could I say - with caution - that women seems to have a negative log odds of being a problematic drinker and that the logodds strengthen when the problematic drinking is becoming more dangerous (e.g. harmful drinking).
In other words, could I make ane comparisons between the two models, If I do so with some caution?
Thanks in advance!
08-24-2016 08:52 AM
With caution, yes. Something along the lines of ."..two models were fit to the existing data. Comparison of the log odds estimates showed <whatever you want to say, but basically it should be just 'larger' or 'smaller'>, within the current study."
You may want to include something about the variation around the estimates (confidence bounds), and adjust that for multiple comparisons. I would get really conservative here and use a Bonferroni adjustment, something I almost never use.