Need help to solve heteroskedasticity in panel data.

[Jari Hoog]

Hello,

I have panel data and I'm using a fixed effects model. I want to see whether heteroskedasticity is a problem in my regressions. Since I read that I cannot use the command rvfplot for panel data, I decided to manually scatter plot the residuals against fitted values.

After running my regression: xtreg Prop_crime_rate Dropout_rate Dropout_rate_1 Avg_dispo_inc Avg_resolved, fe (estimating a relationship between property crime rate and dropout rate)

I entered the following commands:

1. predict dropoutP, xb
2. predict s1, residual
3. scatter s1 dropoutP

A scatterplot of the residuals was created, showing me that heteroskedasticity is present. The scatterplot is shown in the image below.

After seeing this, I ran the regression: xtreg Prop_crime_rate Dropout_rate Dropout_rate_1 Avg_dispo_inc Avg_resolved, fe vce (robust) thinking that robust standard errors will solve the problem.

To check whether heteroskedasticity is gone, I ran the same commands to scatter the residuals, hoping to see homoskedastic residuals.:

1. predict dropoutPR, xb
2. predict s12, residual
3. scatter s12 dropout PR

The exact same scatterplot as before appears. So my question is, did nothing change after adding robust standard errors to my regression? Or is the problem now solved, but it does not show in the scatterplot, because the residuals do not actually change?

If the problem is still present, what can I do to solve it?

Thanks in advance!

[Clyde Schechter]

Using robust standard errors does not eliminate the heteroscedasticity. What it does is assure that the standard errors of the regression output are valid in spite of the heteroscedasticity. This is generally all that people care about.

If you really want to eliminate the heteroscedasticity, you have to change the dependent variable. Sometimes a log transformation will accomplish this when the pattern looks like the one you show (error variance increases in a funnel-like shape as predicted value increases.) But, of course, this approach should only be used if a linear relationship between the log of your outcome variable and the predictors is a reasonable model of reality.

[Jari Hoog]

Thanks Clyde, I guess just using robust will do the trick.