Trouble with gologit2 with mi and svy data

[Tamara Richards]

Hello,

I am examining health care satisfaction with particular interest on those with neurological conditions. I am having trouble identifying the appropriate syntax for gologit2 with mi and svy/gsvy commands.

The commands for gologit2 work well when I do complete cases analysis using STATA 13. The proportional odds assumption is violated so I use gologit2 to create partial proportional odds model following steps highlighted in a paper by Richard Williams:

Mymodel runs: gsvy: gologit2 RecdCareSatisfaction AGECAT Gender MARSTAT EDUC NeuCon SelfPerceivedHealth SPerMHealth GLifeSat UnmetCare qualhealthcarerecd, autofit lrf
OR
gsvy: gologit2 RecdCareSatisfaction AGECAT Gender MARSTAT EDUC NeuCon SelfPerceivedHealth SPerMHealth GLifeSat UnmetCare qualhealthcarerecd, pl(NeuCon Gender MARSTAT GLifeSat EDUC) lrf

I do receive a warning with the output that I don't yet know if it is troubling because I'm concentrating on checking that the commands work. WARNING! 92 in-sample cases have an outcome with a predicted probability that is less than 0. See the gologit2 help section on Warning Messages for more information.

However, I can't seem to find the appropriate syntax when I apply multiple imputation to my survey data.

The code for my model works: mi estimate, dots or: svy: ologit RecdCareSatisfaction ib0.AGECAT 1.Gender ib0.MARSTAT ib0.EDUC ib0.INCOME 1.NeuCon ib0.SelfPerceivedHealth ib0.SPerMHealth ib0.GLifeSat 1.UnmetCare ib0.qualhealthcarerecd

My proportional odds assumption is violated and I am trying to use gologit2.

I have tried the following and received the output:
1. mi estimate, dots or: gsvy: gologit2 RecdCareSatisfaction AGECAT MARSTAT EDUC INCOME NeuCon SelfPerceivedHealth SPerMHealth GLifeSat UnmetCare qualhealthcarerecd

command prefix gsvy: not allowed
r(198);

2. mi estimate, dots or: cmdok: gsvy: gologit2 RecdCareSatisfaction AGECAT MARSTAT EDUC INCOME NeuCon SelfPerceivedHealth SPerMHealth GLifeSa
> t UnmetCare qualhealthcarerecd
command prefix cmdok: not allowed
r(198);

I have spent the past week reading every article by Richard Williams and the svy and mi commands seem to be supported for STATA. Is there something I am missing with the code for gologit2? Is there something lacking in my method? Any suggestions would be greatly appreciated.

[Tamara Richards]

My apologies, I’m using Stata 13 on Windows

[Richard Williams]

I've seen people use gologit2 with multiple imputation and I've seen people use it with svy:, but I don't think I've ever seen the two used together before.

You actually only need gsvy if you are using autofit. Since you aren't, you can just use svy. I think sntax 1 would then work.

I would not use MI with autofit. Instead, decide for yourself which variables to constrain or not constrain. You can get guidance from autofitting the non-imputed data. The problem with using autofit and MI is that different models might get estimated with different imputations, which is not what MI is designed to handle.

For those who don't know, gologit2 is available from SSC. The support page is

https://www3.nd.edu/~rwilliam/gologit2/index.html

[Tamara Richards]

Thank you Richard for your prompt reply.

Just to be sure I understand when you say I can get guidance from autofitting the non-imputed data. Am I correctly interpreting that I should first run my models without imputing the missing data to see which variables to constrain? Then constrain those variables for my mi data, avoiding the use of autofit and thus the gsvy command?

As suggested, using svy instead of gsvy did in fact produce an output for syntax 1. mi estimate, dots or: svy: gologit2 RecdCareSatisfaction AGECAT MARSTAT EDUC INCOME NeuCon SelfPerceivedHealth SPerMHealth GLifeSat UnmetCare qualhealthcarerecd.

However, I got this same message x 10 which I assume to be related to all the variables in my model: WARNING! have an outcome with a predicted probability that is
less than 0. See the gologit2 help section on Warning Messages for more information.

I am not sure what that means but will check the help section as suggested. I may have additional questions.

Thank you!

[Richard Williams]

See the gologit2 troubleshooting FAQ:on negative predicted probabilities:

https://www3.nd.edu/~rwilliam/gologit2/tsfaq.html

Yes, avoid using autofit with mi. You aren't morally bound to do exactly what autofit says, so you could, say, impose additional restrictions, or relax some, if you thought it appropriate. For example, I don't like it when there is a categorical variable and some categories have proportional odds imposed and some don't. I would do them all one way or the other.

[Tamara Richards]

Thank you again Richard. I imposed different restrictions and the final model: mi estimate, dots or: svy: gologit2 RecdCareSatisfaction AGECAT Gender MARSTAT EDUC INCOME UnmetCare qualhealthcarerecd, pl(AGECAT MARSTAT UnmetCare qualhealthcarerecd) lrf ran smoothly and without warnings.

Now on to interpretation and ensuring that I understand what I have actually done and that it is statistically sound.

[Richard Williams]

As noted in the help, suggested readings/ citations are

Williams, Richard. 2006. "Generalized Ordered Logit/ Partial Proportional Odds Models for Ordinal Dependent Variables." The Stata Journal 6(1):58-82. The published article is available for free at http://www.stata-journal.com/article...article=st0097.

Williams, Richard. 2016. 2016. "Understanding and interpreting generalized ordered logit models." The Journal of Mathematical Sociology, 40:1, 7-20, http://www.tandfonline.com/doi/full/...X.2015.1112384.

[Tamara Richards]

Got it, thanks!