0
$\begingroup$

I have a problem that has been plaguing me over the past three days. I have an unbalanced panel. T=52, N=39, obs. = 1565. I am using a fixed effects model with time and country level fixed effects. This works fine. The issue lies with serial correlation. The Durbin-Watson Test returns a value of 0.5-0.7, depending on which specification of regression I use. I also ran a Wooldridge test as it is more suited to panel data. It yields the same result. There is not only AR1, but also AR2 and AR3 serial correlation as suggested by the PACF plot. An Arrelano-Bond test suggests the same thing.

I have considered several options to deal with this. I have read a lot in a short period of time, as such, my knowledge might be very limited. I am sorry for this.

  1. Use Clustered standard errors, clustered at country level: Not feasible for AR>1.
  2. System GMM: Not feasible due to large T
  3. FGLS assuming an AR3 process to model the errors with FEs (by including time/country dummies): I cannot assume that the model is correctly specified, as such, FGLS will be biased. Thus, not feasible.
  4. FE Model with two lagged dependent variables: Kills any significance of other covariates, likely due to collinearity.
  5. FE Model using Driscoll-Kraay SEs.

At the moment, my favourite is option 5. Are there any methods I am missing? Am I exluding methods wrongly? I would also be thankful for any guiding words to think about such a problem as it is causing me a massive headache at the moment.

KR Christoph

Improve this question
$\endgroup$
2
  • $\begingroup$ Couldnt you work on the first differences of the dependent variable? But option 5 seems to be the best option. $\endgroup$ Commented Feb 26 at 17:05
  • $\begingroup$ @GuilhermeMarthe, first differencing removes autocorrelation that is due to presence of a unit root. Any other type of autocorrelation will not be removed by first differencing (it could be reduced, increased or stay the same). $\endgroup$ Commented Feb 26 at 17:37

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.