Receiving 'no coefficients in common' message when attempting to run Hausman test

[Zach Goldberg]

Greetings,

I'm analyzing survey data with Stata 14. My dependent variable is a 7-point ordinal item. Our statistics teacher recommended that we estimate both ordered logistic and OLS models and use the Hausman test to see whether the more efficient one (i.e. OLS) can be used. So I did the following:

regress Increase_NumImmigrants ideology  party_ID

estimates store REG

ologit Increase_NumImmigrants ideology  party_ID

estimates store ORD

hausman REG ORD

It returns the following error message:

no coefficients in common; specify equations(matchlist)
for problems with different equation names.

My teacher's powerpoint specified these exact instructions so I'm not sure where I'm going wrong. Any help would be much appreciated. Thanks in advance!

* Example generated by -dataex-. To install: ssc install dataex
clear
input float Increase_NumImmigrants double ideology float party_ID
6 5 1
4 5 7
4 1 4
7 4 6
2 5 4
5 4 4
7 2 1
5 6 6
4 4 1
1 4 4
6 1 1
2 7 5
7 4 3
4 1 1
3 4 4
6 1 2
4 2 7
5 3 1
7 1 2
4 5 6
6 4 1
4 7 3
3 5 5
6 3 3
6 1 1
4 4 7
4 1 1
3 4 2
5 2 1
6 3 4
2 2 1
4 2 1
7 1 1
5 2 1
6 3 3
4 7 1
1 7 4
2 6 2
5 1 1
4 1 1
4 1 1
4 1 2
6 1 1
4 4 1
6 1 1
7 1 1
7 1 1
2 4 4
5 2 1
4 4 4
7 1 1
4 4 4
5 5 2
4 3 3
7 1 1
4 2 1
1 7 1
5 1 1
4 1 1
5 1 7
6 4 1
7 4 2
1 5 6
4 1 7
7 4 1
7 1 1
6 4 1
1 4 7
4 4 1
4 4 1
4 6 3
5 1 1
5 4 2
7 2 1
6 4 2
6 1 1
4 4 2
6 1 1
6 2 2
7 1 1
6 1 2
4 7 2
3 2 2
6 2 1
6 2 1
5 4 3
3 1 1
1 1 1
4 6 2
5 7 7
2 2 2
2 5 4
7 1 1
4 2 1
7 2 1
4 2 1
7 1 1
4 4 4
4 1 1
5 2 3
end
label values ideology lcself
label def lcself 1 "Very liberal", modify
label def lcself 2 "Somewhat liberal", modify
label def lcself 3 "Closer to liberals", modify
label def lcself 4 "Neither liberal nor conservative", modify
label def lcself 5 "Closer to conservatives", modify
label def lcself 6 "Somewhat conservative", modify
label def lcself 7 "Very conservative", modify

[Clyde Schechter]

Well, unfortunately your example data is not compatible with the code you show. Your example data has no variables Y, X1 or X2, and I can't figure out which of the variables in your example correspond to these roles. I assume that you have actually tried to run this code with the correct variable names (or that you have run it on a data set that contains variables named Y, X1, and X2.

The problem you are encountering arises because -ologit- and -regress- name their regression equations differently, so that when -hausman- invokes -suest- to compare their results, it can't figure out what to compare with what.

Here I use your example data and run an OLS regression and an ordered logistic regression and compare their results with -suest-. This is all that -hausman- does: it is a wrapper for -suest-. But -hausman- does not properly figure out the names of the equations to pass to -suest-, so we have to bypass it to get an actual Hausman test result. The particular regressions I have run are for demonstration purposes and probably have little or nothing to do with your intended regressions.

regress Increase ideology
estimates store REG
ologit Increase ideology
estimates store ORD

suest REG ORD, coefl
test [REG_mean = ORD_Increase_NumImmigrants], common

Added: All of that said, I am perplexed at the use of a Hausman test to choose between a least squares regression and an order logit model. I have never seen that before. And it doesn't make sense to me. The -ologit- model's residual variance is, by definition, constraint to (pi^2)/3, whereas that of -regress- is unconstrained. So the coefficients are on different scales (unless, by chance the residual variance from -regress- turns out to be (pi^2)/3. So I don't see how the two models could possibly give equivalent results, and the only interpretation I could think of for a non-significant result would be a Type II error.

If somebody out there understands this better than I do and can explain this way of using Hausman, I would appreciate it.

[Zach Goldberg]

Hey Clyde,

Thanks for your helpful response. I should clarify that my professor only suggested using the Hausman test when the ordinal dependent variable exceeds (roughly) 4 values. Whether that rule of thumb is 'correct' or not I can't possibly say--I'm only a novice with this.

[Zach Goldberg]

Can anyone else weigh in here?