When to use the "small" option in xtabond2

[Manuel Almodovar]

Dear Stata users,

Regarding the xtabond2 command, Roodman (2009) suggests using the "small" option to correct the covariance matrix estimate. In this regard, my question is: when is a sample considered small? Does it refer to a specific number of observations? Does it depend on the proportion between groups (N) and time (T)? If possible, I would like to have a reference for using the "small" option.

Thank you for your responses

[George Ford]

https://www.statalist.org/forums/forum/general-stata-discussion/general/1697840-dynamic-panel-model-for-large-t-and-small-n

Nickell, Econometrica1981 is probably a decent starting point. There are some ways to quantify bias in T.

[Manuel Almodovar]

Thank you very much for the information. I will read the article carefully.

[George Ford]

I've seen rules of thumb of 10 or even 20, but nothing cited to support it. I think these come from calculations from Nickell, where the bias can be quantified from two parameters I think.

[Sebastian Kripfganz]

I think there is some confusion here. The "small" option has nothing to do with the small-T bias of the fixed-effects estimator in dynamic panel models. It is simply a degrees-of-freedom correction for the standard errors, as is done by many other commands (either automatically or by a similar "small" option), not just panel data commands.

In a simple OLS regression, this ensures that the estimator of the error variance is unbiased in finite samples (instead of just being consistent asymptotically); see any econometrics textbook. With instrumental variables, this unbiasedness argument actually no longer works. However, a small-sample correction of the standard errors is typically still applied because it tends to improve the accuracy of statistical tests in finite samples (although this is not guaranteed).

You can always use the "small" option, no matter how small or large your sample is. If your sample is very large, it will not make any difference. If your sample is very small, the above arguments apply.

[Manuel Almodovar]

Perhaps it should only be used when it causes a significant change in standard errors. When I say "use" I mean report it in a paper.
Would this be the right strategy?