Brant tests: What should I conclude from Brant tests about the proportional odds assumption of ordered logit as the results conflict?

[Robert Murphy]

Results: The control compared to treatment group 1 (who saw life satisfaction information) I find that the proportional odds assumption is not violated when I look at preference (1^st, 2^nd, 3^rd preference) for anxiety/depression (which is my main interest) but are violated when look at preferences for self-care and for pain (although violation seems “small”).
The control compared to treatment group 2 (who saw day affect information) I find that the proportional odds assumption is not violated when I look at preference (1^st, 2^nd, 3^rd preference) for anxiety/depression (which is my main interest) and not violated for self-care and for pain but are violated when look at preferences for pain (although violation doesn’t seem to be on the information provision variable.

Before I tested the assumption the results of the ordered logit indicate that co-efficient on INFO is significant and negative for anxiety/depression, is significant and positive for self-care and non-significant for pain.

I am unsure what to conclude about the proportional odds assumption? Details are below.

Prediction: Provision of information on the association of health states with SWB to participants before they appraise health states will result in a lower preference for anxiety/depression than other health states.

Experiment A randomized control trial in which the control group judges the health states as normal, one treatment group judges the health states after provision of information on the association of the health states with life satisfaction and the second treatment group judges the health states after provision of information on the association of the health states with day affect.

Choice: Participants are provided with a description of some problems with self-care, moderate pain/discomfort and moderate anxiety/depression. Participants made a choice between health states, from most to least preferred to live in by entering 1 for most preferred, 2 for second preference and 3 for third preference.

Model: The dependent variable is an ordered outcome variable so choices are examined using an ordered logit (ordered probit) model of the form in Equation 1.
P (C_j) = β₀ + β₁INFO _i + β₂DEMO_i + β₃EMPLOY_i + β₄HEALTH_i + β₄LS_i + ε_i

Output of brant test is shown below; first for LS and then for DA.

DA Info	ANXIETY/DEPRESSION				LS info	ANXIETY/DEPRESSION
	chi2	p>chi2	df			chi2	p>chi2	df
All	14.61	0.201	11		All	7.54	0.754	11
1.groupdum	0	0.979	1		1.groupdum	0.03	0.853	1
2.sex	0.34	0.562	1		2.sex	1.5	0.22	1
age	0.02	0.901	1		age	1.58	0.209	1
agesqu	0.14	0.708	1		agesqu	1.5	0.221	1
1.reldum	0.02	0.888	1		1.reldum	0.9	0.342	1
1.mainactdum	0	0.97	1		1.mainactdum	0	0.97	1
1.highedudum	2.9	0.089	1		1.highedudum	0.1	0.754	1
2.highedudum	6.78	0.009	1		2.highedudum	2.36	0.124	1
3.highedudum	2.2	0.138	1		3.highedudum	0.1	0.752	1
yourhealthvas	0.4	0.527	1		uvas	0.5	0.482	1
ls	2.75	0.097	1		ls	0.85	0.356	1
DA Info	SELF CARE				LS Info	SELF CARE
	chi2	p>chi2	df			chi2	p>chi2	df
All	18.48	0.071	11		All	20.8	0.035	11
1.groupdum	0.07	0.799	1		1.groupdum	2.14	0.143	1
2.sex	0.71	0.401	1		2.sex	6.47	0.011	1
age	0.9	0.344	1		age	1.41	0.234	1
agesqu	1.55	0.214	1		agesqu	1.01	0.315	1
1.reldum	0.27	0.603	1		1.reldum	0.08	0.773	1
1.mainactdum	1.34	0.246	1		1.mainactdum	1.97	0.161	1
1.highedudum	0.22	0.64	1		1.highedudum	3.47	0.062	1
2.highedudum	1.06	0.304	1		2.highedudum	0.67	0.415	1
3.highedudum	0.01	0.919	1		3.highedudum	2.61	0.106	1
yourhealthvas	0.01	0.93	1		uvas	0.11	0.741	1
ls	2.13	0.144	1		ls	4.74	0.029	1
DA Info	PAIN				LS Infor	PAIN
	chi2	p>chi2	df			chi2	p>chi2	df
All	22.22	0.023	11		All	20.8	0.035	11
1.groupdum	1.78	0.183	1		1.groupdum	2.14	0.143	1
2.sex	0.42	0.519	1		2.sex	6.47	0.011	1
age	0.01	0.913	1		age	1.41	0.234	1
agesqu	0.13	0.718	1		agesqu	1.01	0.315	1
1.reldum	0.09	0.762	1		1.reldum	0.08	0.773	1
1.mainactdum	3.02	0.082	1		1.mainactdum	1.97	0.161	1
1.highedudum	0.02	0.895	1		1.highedudum	3.47	0.062	1
2.highedudum	0.21	0.65	1		2.highedudum	0.67	0.415	1
3.highedudum	0.14	0.711	1		3.highedudum	2.61	0.106	1
yourhealthvas	0.53	0.468	1		uvas	0.11	0.741	1
ls	5.3	0.021	1		ls	4.74	0.029	1

[Robert Murphy]

Brant tests and proportional odds assumption: can anyone help answer this post?

[Joseph Coveney]

I am unsure what to conclude about the proportional odds assumption?

Brant tests and proportional odds assumption: can anyone help answer this post?

1. Even with artificial data, you ought to be able to reject the null hypothesis of no violation of the assumption at least a couple of times, at least somewhere, especially with the Brant test. So, why not just forgo the testing ceremony, and start out with the proposition that, with natural data, the proportional odds (parallel lines) assumption is always violated, at lease in the small?

2. Once you're past that, you can use the methods and advice shown in the hyperlinks given in this post to judge for yourself whether the violations are such that you need to be concerned, and to describe the practical implication of the magnitude of the violation for your study's results and their interpretation. If the violations are gross enough to require separate regression coefficients for the cutoffs, then you can use the user-written gologit2 (SSC) to fit an ordered-categorical model that relaxes the assumption for the affected predictors.

[Robert Murphy]

I will try that Joseph. Thanks Robert