Comparing coefficients across the models

[Ujjwal]

Dear Stata Forum,

I was trying to compare (or test the equality of) the coefficients

1. of the same specification (using 'xtreg') but for male and female samples using 'estimates store' and then 'suest' command. It seems that 'suest' command does not work with 'xtreg'. Is there any other way to compare between male and female parameters?

2. between the quantiles (using sqreg). Can I use 'suest' command to compare the parameters across the quantiles? If yes, any available codes would be very helpful.

Thanks,

Ujjwal,
Postgraduate researcher, Economics,
Leeds University Business School

[Clyde Schechter]

For your -xtreg- model, I would do this by running a single model on both sexes and then include sex#predictor interaction terms. The coefficient of sex#X in the model output is the difference between the male and female coefficients, and it even comes with a confidence interval, etc. If there are variables whose coefficients you are reasonably sure do not appreciably differ by sex, you can omit the interaction terms for those particular variables and lose little or nothing.

I do not know if sqreg is supported by -suest-, but even if it is, I would still recommend the interaction term approach.

[Roman Mostazir]

For number-1, you could analyze the whole dataset with both gender and add gender in the model. The coefficeint for gender produced by the model is the test for equality of the coefficients between male and female.

For number-2, have you considered Stata's quantile regression? See help qreg

Editted: qreg is not probably suitable for nested data.

[Attaullah Shah]

Clyde has suggested an easy and elegant approach. If you are still interested in comparing coefficients, you can try estimating the two regressions separately for each group. The differences in coefficients of the two groups of regressions are tested for statistical significance using the z-score calculated as




where B1and B2 refer to the coefficients of a given variable in the two groups of regressions; and e2 are the standard errors of B1 and B2,respectively.

[Ujjwal]

Thanks both of you for the suggestions. Actually I am running a two step model for fixed effect quantile regression for panel data- where in the first step I run fixed effect and in the second step, I am running quantile regression using Stata sqreg command. For panel data, sex dummies could be dropped as it does not changes over time - that is why I ran the model separately in male and female sample.

Thanks Attaullah, it seems an easier solution which I can do manually.

[Attaullah Shah]

If you you want to include dummies in fixed effects models, add the prefix i with the dummy variable, i.e. i.gender. There are other possibilities, such using the prefix xi: with reg command and then inlcude dummy variables to make the regression fixed effects, i.e. LSDV model