Hi Everyone,
I have a question regarding the parallel lines assumption of ordered logit regressions when using weighted multi-country survey data and hierarchical regression models.
In more detail, my DV is a four point likert scale. I have survey data from 17 countries, encompassing some 17,000 respondents. Despite the small number of level-2 units, the log-ratio test indicates that rather than ignoring the nested quality of the data, I should use a hierarchical model in place of a simple ologit. I am, therefore, using a meologit estimator. I am weighting the data (iweight) and include a country-level variable.
When I run a Brant test on the unweighted, non-hierarchically estimated data, I find that the parallel lines assumption has been violated. 'Brant, detail' does not work on the hierarchical or the weighted data. From what I can tell reading the manual and looking online, there is not a good way to test if the PL assumption is also violated when using the hierarchical model with sample weights.
I am curious about a few things:
1) Is there an equivalent to the Brant test that works in my particular case?
2) In the absence of a Brant test that works on a multilevel estimator and with weighted data, should I just assume that because the assumption is violated and try another estimator and/or restructure the DV?
3) If yes to 2, I would welcome any suggestions on where to start. My gut reaction is to restructure the DV to a binary indicator and use a logit estimator.
Thanks for any assistance.
Kind regards,
Eric
I have a question regarding the parallel lines assumption of ordered logit regressions when using weighted multi-country survey data and hierarchical regression models.
In more detail, my DV is a four point likert scale. I have survey data from 17 countries, encompassing some 17,000 respondents. Despite the small number of level-2 units, the log-ratio test indicates that rather than ignoring the nested quality of the data, I should use a hierarchical model in place of a simple ologit. I am, therefore, using a meologit estimator. I am weighting the data (iweight) and include a country-level variable.
When I run a Brant test on the unweighted, non-hierarchically estimated data, I find that the parallel lines assumption has been violated. 'Brant, detail' does not work on the hierarchical or the weighted data. From what I can tell reading the manual and looking online, there is not a good way to test if the PL assumption is also violated when using the hierarchical model with sample weights.
I am curious about a few things:
1) Is there an equivalent to the Brant test that works in my particular case?
2) In the absence of a Brant test that works on a multilevel estimator and with weighted data, should I just assume that because the assumption is violated and try another estimator and/or restructure the DV?
3) If yes to 2, I would welcome any suggestions on where to start. My gut reaction is to restructure the DV to a binary indicator and use a logit estimator.
Thanks for any assistance.
Kind regards,
Eric
