Sign of coefficients change after first-differencing

[Diego Cervantes]

Good evening,

I'm working with a dynamic panel data model. I first ran a fixed-effects regression (by using unit dummies) with xtpcse with the c(a) option in order to account for autocorrelation;

xtpcse gini100 l.gini100 findex flab gdpgrowth unemployment uniondensity trade socx d1 d2 d3 d4 d5, c(a)

However, in order to correct for Nickell bias, I tried running a first-differenced model with the reg command:

reg D.(gini100 l.gini100 findex flab gdpgrowth unemployment trade uniondensity socx), noconstant cluster(countrycode)

In the first regression, the coefficient for the lagged dependent variable as well as my main variable of interest (findex) were both significant and positive, as expected from the literature. However, in the second one they are both negative and only the lagged dependent variable is significant. The R-squared also drops dramatically from 0.9797 to 0.1198.

Does the interpretation of the coefficients change after first-differencing? Am I doing something wrong with how I'm coding? Or do the coefficients in the second regression represent the actual relationship without the bias?

Thank you.

[Phil Bromiley]

You'll increase your chances of a useful answer follow the FAQ on asking questions – provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

The first model includes variables that don't appear in the second model so it is not surprising you get different results.