Well the question is pretty straight forward. I have a ordinal dependent variable with an independent (ordinal) variable and two control variables on the first level, country of residence at the second level and the region of each country at the third level.
If it is not possible to decompose the variance, what should I report, and how? Is the variance at each level the variance that is explained in that specific level? Or how do I report and interpret the outcomes?
The stata output is attached!
Since you have a random intercept for county and region (counties nested in regions) then the variance indicates the amount of variability between counties, and between regions, in the outcome (truestep). Both levels are significant (positive CIs), suggesting that there is a significant amount of between county and between region variability in the outcome.
You could also allow these higher levels to vary in education and age to see if there is some random variation in truestep.
I'm not sure how to quote you? Anyway, so if I understand you correctly: since both are significant and show the variability between countries and regions, should I interpret the number (0.28 and 0.12) as more and less explained variance? And is that at each level i.e. at the country level, 28% variance is explained and 12% is explained at the region level? Or is this a wrong assumption?
These are estimated variances of the random intercepts for the respective levels. They variances are in the metric of the dependent variable. I would personally keep it simple and say that we conclude that these intercepts vary from country to country and region to region. If you want to get a bit more fancy, you could say something about the magnitude difference of the variance over the standard error. That said, I suggest that you rerun the model reporting the coefficients instead of the ORs.
If you really want to explore the random effects components more between regions and between countries, you can type: predict randint1 randint2, reffects and then graph the results by country/region by typing:
graph hbar (mean) randint1, over(country) ytitle ("random intercepts by country")
graph hbar (mean) randint2, over(region) ytitle ("random intercepts by region")
Well since my professor explicitely told us he prefers odds ratios for all ordinal analyses, I rather go with odds ratios for this assignment :)
Anyway, I'll keep it simple as you suggested! Thanks a bunch since you are (literally!!) the only one on the internet that has now provided an answer on this question! Google, YouTube etc. all came up blank in terms of interpretation of Ordered Multilevel logistics!
I've tried most things I could to get predicted probabilities and the graphs. However, for nearly all postestimation options, it gives the same (single) error.
. predict randint1 randint2, reffects
(calculating posterior means of random effects)
(using 7 quadrature points)
could not compute empirical Bayes' means;
missing values were returned by the evaluator
r(322);
Oops, I gave you the command for a continuous mixed model. After running meologit you type:
predict re*, remeans
graph hbar (mean) re1, over(country) ytitle ("random intercepts by country")
graph hbar (mean) re2, over(region) ytitle ("random intercepts by region")
It's not just the predict that doesnt work (at all), neither does margins or any other post-estimation test for that matter.
. predict re*, remeans
(calculating posterior means of random effects)
(using 7 quadrature points)
could not compute empirical Bayes' means;
missing values were returned by the evaluator
r(322);
Apparently it's either a bug, or just works very occasionally:
This one saying that the me fucntion just doenst work well with large datasets, but the solution does not work for me either: http://hsphsun3.harvard.edu/cgi-bin/lwgate/STATALIST/archives/statalist.1307/Author/article-73.html
This one tries to figure out the same problem as me concluding that it can't be done untill stata comes with a solution: http://www.statalist.org/forums/forum/general-stata-discussion/general/664037-meologit-problems-with-post-estimation-prediction
This appears to be an issue with your data. I know this code works. Try this example:
webuse tvsfpors
meologit thk prethk cc##tv || school: || class:
predict re*, remeans
graph hbar (mean) re1 , over( school ) ytitle ("random intercepts by school")
graph hbar (mean) re2 , over( class ) ytitle ("random intercepts by class")
It is an issue with my data, true. But it most likely is the 'issue' of being a large dataset (50.000+ respondents). The me function doesn't like that and although the evaltype(gf0) option should fix that, it doesn't.
Thanks a bunch though with all the help you've given! As soon as I find a solution to the current big data problem I'll post it here again!
To whom It still may concern
The below page shows two examples including two-level and three-level model of a multilevel ordered logistic regression analysis and explaining the output.
https://www.stata.com/manuals13/memeologit.pdf
How to get ICC, in the multilevel multinomial logit model in R or STATA? ?(2 level) Thanks!
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Probability
- Probability Theory
- Reliability