Gologit2 with Fixed Effects

[Fabian Borges]

Hello. I have a question regarding the use of fixed effects with fgologit2.

I'm analyzing the determinants of attitudes on who should be covered in a cash transfer program from children. I planned on using ordered logit because the dependent variable (cashtranscov) ranges from 1 (transfers should only cover children in extreme poverty) to 4 (transfers should cover all children). The data is from a survey of seven countries, so I'm using country fixed effects (i.PAIS).

gologit2 cashtranscov mujer Quintiles educacion formality laborincome P7_Ninos_2 b2.edadcat bonorecip foodaid urbano evangelico etnia_blanco i.PAIS, or

The ordered logit failed the Brandt test, so I tried gologit2. It corrected for the variables that failed the Brant test.

gologit2 cashtranscov mujer Quintiles educacion formality laborincome P7_Ninos_2 b2.edadcat b4.vote_choice2 ineqviews taxrich P15_VAL_MANODURAr bonorecip foodaid urbano evangelico etnia_blanco i.PAIS, or autofit lrforce store(mod1)

Step 18: Constraints for parallel lines are not imposed for
Quintiles (P Value = 0.01343)
educacion (P Value = 0.00356)
1.edadcat (P Value = 0.00006)
evangelico (P Value = 0.00201)
4.PAIS (P Value = 0.00775)
5.PAIS (P Value = 0.00000)
6.PAIS (P Value = 0.00005)
7.PAIS (P Value = 0.00000)

Is it a problem that running gologit2 leads to differing coefficients for some of the country dummy variables? Should I impose constraints on all of them? Should I not impose constraints on any?

Is there something else I should be doing?

I looked into feologit, but as far as I could tell it was for panel data. My survey is only cross-sectional.

Any help and all suggestions would be welcome,
Fabian

[Luca Calianno]

Hi Fabian,
All of your steps seem to make sense.
I don't think it is a problem if some of the coefficients differ across models (even more so if the ologit does not satisfy the proportional odds assumption!). You can also use a Mundlack type of adjustment (add regional averages of the control variables).
feologit works and can be used! Just set xtset PAIS or feologit cashtranscov ..., group(PAIS)

[Richard Williams]

First, I recommend using autofit(.01) or something even stricter. You are running a lot of tests and just by chance alone some violations could show up as being statistically significant even if the population parameters are not. It looks like your model would be slightly more parsimonious then.

Second, if I ever get ambitious enough, I might add an option to gologit2 to test all the parameters of a factor variable simultaneously rather than one by one. In your case that would result in all the PAÍS variables being constrained to meeting proportional odds assumption or all not being so constrained. You could do such tests by specifying the code yourself.

Or, just go ahead and unconstrain 2.PAIS and 3.PAIS. That is what I would probably be tempted to do.

Finally there are a lot of ordinal models around so consider if something else, like feologit, might be better.