Driscoll-Kraay panel data with lagged dependent variable (LDV)

[Filipe Costa]

Hi everyone,
I have a panel with monthly financial markets data, including returns, volatility, sentiment and a few dummy variables to capture crisis, crashes and other particular events. The data is composed by 15 countries (N=15) and 300 monthly observations (T=300). There is cross-sectional dependence, heteroscedasticity and country fixed effects. I'm also concerned with two additional issues: serial correlation and endogeneity.

After reading extensively about panel models, I've chosen Driskoll-Kraay (xtscc) to deal with the data. Something like:

xtscc depvar ret vol sent, fe

From my understanding xtscc is able to deal well with serial correlation and the other problems reported above. However, I was thinking about the possibility of adding a lagged dependent variable (LDV) and run a model like:

xtscc depvar ret vol sent L.depvar, fe

I believe that the fact T is large relative to N, allows for the introduction of a LDV without problems, but not sure. Here are my questions:

1. Do you believe xtscc is the best option to fit the data and its characteristics, or would you run something else like xtreg?
2. Do you believe I can add the LDV to a xtscc fit without creating econometric issues?
3. Should I do Arellano-Bond dynamic model to fit the LDV, instead?

[Maxence Morlet]

Your data are well suited for xtscc as T>N by far in your case.

However, xtscc will not protect you from the Nickell bias (1981); there will be a substantial bias on your lagged DV.

If I were you, I would report results from xtscc, but also apply Sebastian Kripfganz's xtdpdgmm to address the Nickell bias issue and report these results as well.

[Joro Kolev]

Nickell's bias disappears as T grows large. See these lecture notes http://fmwww.bc.edu/EC-C/S2013/823/E...n05.slides.pdf and in particular slide number 4, for an accessible and short treatment with a formula for the bias.

So OP can do everything as described, and there is no obvious econometric problem in this plan.

Of course the computer is not going to break if an instrumental variable estimator like the Anderson-Hsiao or the Arellano-Bond estimator is employed.

[Filipe Costa]

Thank you for the answers.
I tested the model for fixed effects and I found that with an LDV there's no FE. In fact, pooled seems the best for that case. I'm not convinced of it.
The reason to use the LDV is because my depvar is an AR(1) variable. Does it really make sense to use the LDV? Of course the model fit is way larger with the LDV, but I guess that's not that important.