Dynamic panel data can solve the serial correlation problem or not?

[Shovkat Khodjaev]

Hi everybody,

I ran the Wooldridge test to check for autocorrelation in panel data, and the result showed that I had a problem with serial correlation. I used fixed effect and robust function but it didn't help. I want to ask, is it possible to solve the serial correlation problem using dynamic panel data? If not, I would be very grateful if someone could advise me how I can solve the serial correlation problem.

Thank you!

[Carlo Lazzaro]

Shovkat:
if you're dealing with a static, N>T panel dataset, just add -robust- or -vce(cluster clusterid)- options to your code.
Otherwise, please (as per FAQ), provide more details. Thanks.

[Shovkat Khodjaev]

Dear Carlo Lazzaro,

Thank for your response. Sorry for the lack of details. I am doing a fixed effects regression with an option robust model where I have 1 vegetation index for 3 years. My dependent variable is crop yield. My final goal is to evaluate how accurately the vegetation index estimates crop yields. My command is below:

xtreg yield_ndvi avg8_ndre, fe robust

I would like control for autocorrelation in my regression model.


My main question is how I can apply dynamic panel data when I have a limited number of variables. I know about fixed -robust- or -vce(cluster clusterid) but many published articles indicate that this option does not control autocorrelation and recommend using dynamic panel data to control autocorrelation. Can you please advise me on how I can apply dynamic panel data if I have a limited number of variables?

Thank you!

[Carlo Lazzaro]

Shovkat:
admittedly, I've never heard of this facet of dynamic panel data model.
Just taking one step aside, I think that the main issue here is that you seem to have one predictor only.
How can the result of whatever regression be informative with such a scant specification?

[Shovkat Khodjaev]

Dear Carlo Lazzaro,

Thank for your response. Sorry for the lack of details. So, the main purpose of my paper is to demonstrate the accuracy of vegetation indices based on publicly available satellite data. I have submitted it to an agricultural journal, but one of the econometric reviewers wants to see how I control autocorrelation. The main problem is that I am a remote sensing specialist and have no econometric background. On this forum, I am looking for a way to solve this problem because my knowledge is not enough to solve it myself.

Thank you!

[Carlo Lazzaro]

Shovkat:
as per your code, I assume that you're dealing with a N>T (N=cross-sectional dimension; T=time-series dimension), static, panel dataset.
Therefore, if you have at least 30 panels, you can safely add -robust- or -vce(cluster clusterid)- options to your code to take both heteroskedasticity and/or autocorrelation.
That said, I do not know whether or not the right-hand side of your regression equation hosts other predictors than the -avg8_ndre- independent variable you show in your code.
As you know, predictors should give a fair and true view of the data generating process.

[Shovkat Khodjaev]

Dear Carlo Lazzaro,

Thank for your response. I tested the accuracy of several vegetation indices (independent variables) for measuring actual yields at different times of the period. This procedure was to determine which of our proposed independent (vegetation index) provides the highest accuracy of crop yields. I have several independent variables which are correlated with each other therefore I run the model for each independent variable separately. Right now there are 232 panels and one independent variable in my model.

[Carlo Lazzaro]

Shovkat:
so you have a N>T panel dataset.
Therefore, you can go -xtreg- and add -robust- or -vce(cluster clusterid)- options to your code to take both heteroskedasticity and/or autocorrelation into account.
I do not comment on the other issues reported in your question.

[Jeff Wooldridge]

I'm with Carlo. I think for what you want to do, a lagged dependent variable isn't appropriate -- although others might disagree. But serial correlation is not a serious issue unless it is so strong that the estimated equation isn't capturing a meaningful relationship. Include time period dummies, use fixed effects, and cluster your standard errors. How many time periods?