Solved – Autocorrelation in panel data

autocorrelationdataset

I have a random effects panel data model with 4 variables, 10 years of data and 946 observations describing different financial values for different companies. The panel is balanced. The model looks something like (written without subscripts t for time and j for company): $ln(A)=b0+b1(ln B) + b2 (ln C)…+u$

However, when I run my model in Eviews, the given Durbin-Watson statistic shows that I have positive autocorrelation in my model (D-W statistic value is lower than d-lower (dL). I have found out that there are numerous ways to correct your model for pure autocorrelation, but I have run into trouble when implementing any of them.

What I have tried so far:

  1. First, I tried Generalized Differencing (also known as Generalized Least Squares). I didn't find any information as to whether this was a suitable method for panel data and had some considerable trouble retaining the natural-logarithmic functions of my model (since some values changed into negative values and taking logs from negatives is impossible), but even after implementing this to my best of knowledge the autocorrelation was not eliminated. In fact, instead of positive autocorrelation I now had negative autocorrelation.
  2. Secondly, I found that Newey-West standard errors can eliminate autocorrelation. However, Eviews doesn't support this function for panel data, which made me believe that, perhaps, this method does not work for panel data. Would it work?
  3. The third solution I found from some more complicated papers. Basically, the authors of the theory (Arellano and Bond 1991) include a lagged dependant variable (model turns into a dynamic model) and then estimate a final model with Generalized Method of Moments. However, I couldn't find much information on how to create such a lagged variable (and which one of my variables should be lagged), nor how big the lag should be be and overall how to do this in practice.

My questions:

  • How do I eliminate autocorrelation in panel data?
  • Which methods really work for panel data? All of the ones mentioned above or none? Maybe there is some additional way to do this?
  • How do I bring one of those methods into practice? Not everyone uses Eviews, but it is easy to see a formula, but much harder to explain to the computer how you want this formula solved.

Reference:

Arellano, M. and Bond, S. (1991) Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations. Review of Economic Studies, 58(2): 277-297 [DOI]

Best Answer

You might want to check out this paper: Bertrand, Marianne, Ester Duflo, and Sendhil Mullainathan. 2004. "How Much Should We Trust Differences-in-Differences Estimates?" Quarterly Journal of Economics. 119(1): 249--275. [prepub version]

The paper recommends using a block bootstrap, probably not canned in EViews, or clustered standard errors, which I imagine would be. Lastly, you could calculate a simple before-and-after difference, which also works well according to the authors.