A few questions on dynamic panel regression model estimation

[Nick Cain]

Hello List,

I've got a time-series cross sectional data set of US states that has a spatial component so I'm interested in estimating a dynamic spatial panel model.

My dependent variable is wind capacity additions per state year and my regressors are various factors that may influence wind power development.

• Although most of my regressors vary by state and by year, at least one is time-invariant (but varies by state).
• Another one of my regressors, which represents technological improvements to wind machines, is the same across all the states in my panel (although it varies by time).
• The Hausman test is nonsignificant and thus a random effects model seems appropriate.

I've run a dynamic panel using the XSMLE package and Hausman-Taylor panel using the xthtaylor command.

Given these somewhat unique aspects of my data, are these estimation approaches acceptable? Is there an alternative estimation model that anyone can recommend?

Thanks!

-nick

[Jeff Wooldridge]

Nick: A former student of mine, Valentin Verdier, has a paper on estimating dynamic models with spatial dependence. His focus is on improving efficiency by using spatial lags as additional instruments. He is currently in the economics department at North Carolina. Sebastian Kripfganz and Claudia Schwarz (http://www.kripfganz.de/) have a nice paper on allowing time-invariant variables, and Sebastian also has estimators that explicitly model the spatial lags. You might have to combine the two approaches.

[Nick Cain]

Dear Jeff,

Thanks for the resources -- I am looking at Dr. Verdier's work right now and it's very helpful (and thanks also for your test for autocorrelation, I'm using that too!

From reviewing previous posts on the list, it appears that the cluster-robust approach can compensate for both heteroskedasticity and for autocorrelation. Is that correct or should additional tools (e.g., lags) be used?

Thanks again!

-nick