Contrast of marginal predicted mean

[Enrique Hernandez]

Dear all,

I am estimating the following model:

melogit tradeoff c.vote_share##c.lfirstp  || cntrynum: vote_share

I have estimated the marginal predicted mean across values of lfirstp for two values of the variable vote_share (1 and 51) using the margins command.

margins, at(lfirstp=(-2.4(2)2.4) vote_share=(1 51))

What I want to do next is to estimate the difference in the predicted mean (and the corresponding statistical significance of the difference) for values vote share 1 and 51 at different values of lfirstp. I thought something like this would work.

margins r1.vote_share#r51.vote_share, at(lfirstp=(-2.4(.4)2.4)) level(95)

However, trying to estimate this leads to the following error:

vote_share:  factor variables may not contain noninteger values

Do you have any idea about how I could do this?

Thanks!

[Marcos Almeida]

It seems the command is trying to consider vote_share a categorical variable in order to apply contrasts, but it is a continuous variable, hence the error message. In short, with "r." you're supposed to have a categorical variable.

[Enrique Hernandez]

Dear Marcos,

Thank you for your answer.

Yes, the "r." only works for categorical variables.

I wonder if there is any way to perform such a test with a continuos variable. That is, comparing the predicted mean of two specific values of a given continuos variable.

[Dimitriy V. Masterov]

I am not entirely sure I get exactly what you have in mind, but you can apply contrasts to margins:

sysuse auto, clear
reg price mpg weight
margins, at(weight = (1000(1000)5000) mpg = (10 20))
margins, at(weight = (1000(1000)5000) mpg = (10 20)) contrast(atcontrast(ar._at) wald)

You can also post the results of margins and then manipulate the coefficients with test, lincom and nlcom as if they were estimation results:

margins, at(weight = (1000(1000)5000) mpg = (10 20)) post coefl
test _b[2._at] = _b[6._at]
test (_b[2._at] = _b[6._at]) (_b[3._at] = _b[7._at])