fixed effects with lagged dependent variable?

[Frank Kumah]

Hi Everybody!
I am estimating the impact of ICT diffusion on economic growth using a panel data set of 25 countries covering 8 years.
I therefore use:
-gdp as dependent variable
-Mobile phone subscriptions, Internet users, fixed broadband subscriptions, ICT Imports as independent variable
and a few macroeconomic stability control variables.

Is it a good idea to use lags of my dependent variable in a fixed effects model?


Thanks in advance!

[Sebastian Kripfganz]

The conventional fixed-effects estimator will be biased in a data set with a small number of years when you include a lagged dependent variable. Potential solutions are to use a generalized method of moments (GMM) estimator, a maximum likelihood (ML) estimator, or a bias-corrected (BC) estimator; see:

GMM:

• https://www.statalist.org/forums/for...ent-conditions
• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference

ML:

• https://www.statalist.org/forums/for...c-panel-models
• Kripfganz, S. (2016). Quasi-maximum likelihood estimation of linear dynamic short-T panel-data models. Stata Journal 16 (4), 1013-1038

BC:

• https://www.statalist.org/forums/for...el-data-models

Having said that, 25 countries is also very small and the above methods may not work very well with such a small sample. If you are using economic growth instead of GDP as your dependent variable, you may not necessarily need a lagged dependent variable in your model. Sure, there might be some bias due to the neglected dynamics but you are avoiding a whole lot of other complications when you stick to the static model with your small sample.

[Frank Kumah]

Thank you Sebastian! Using GDP Growth instead of GDP makes perfect sense, I will do that.
Many previous papers in my research filed use GMM as an estimator for this relationship. But as you pointed out, I've read that GMM works best in small T and large N situations, so it's not optimal in my case.
In the context of my sample with small T and small N, would you also prefer a FE model or use GMM anyway?

[Joro Kolev]

With such a small sample you should try the simplest model possible. So if you do not have a very good reason to include lagged GDP, e.g., that all previous research have included it, stick to a simpler model.

Of course the computer will not break if you try a dynamic model, and you can compare and see whether your coefficients change much. Apart from what was suggested above, you can try the simplest dynamic model-- the Anderson Hsiao estimator.