Linear model accounting for fixed effects vs. ordered probit

ARW90

New Member
#1
I originally ran an ordered probit model on an ordinal dependent variable (values 1-12), however in order to account for fixed effects in my analysis I decided to run a simple OLS regression. Although I know that this goes against some of the assumptions of OLS, prior to accounting for fixed effects the analysis gives similar results to that of the ordered probit. However, once I account for fixed effects my results change massively.

The two variables i am interested in are dummy variables. In the ordered probit and the OLS regression without fixed effects variable A is insignificant with a near zero coefficient and variable B is significant with a negative coefficient. However, once I account for fixed effects variable A becomes very negative and significant whereas variable B becomes insignificant and near zero.

I am very confused about what is happening within my analysis. I understand what accounting for fixed effects has done but i did not expect it to change my coefficients so massively and as a result I am now quite confused. My first question is which model should I be more interested in interpreting? and secondly, is this a common occurrence when accounting for fixed effects?

Any help would be greatly appreciated. Thanks
 

jrai

New Member
#2
Is this a panel data? Please clarify what did you do when you say that you accounted for the fixed effects? How did you account for fixed effects using OLS?
 

ARW90

New Member
#3
Thanks for replying I apologise if I am not making much sense I am rather confused at the moment.

Yes it is panel data. I used the regress command on my ordinal variable originally and then I used xtreg........,fe in ordered to account for fixed effects.

Thanks.
 

jrai

New Member
#4
It is unreasonable to expect the output from two different models to be similar. You can account for the fixed effects within the ordered probit model as well. Calculate subject-specific means for each of the variables, merge those into the original data set, and then calculate deviations from those means. After that you can estimate the ordered probit using these newly calculated variables.