Dynamic panel vs. fixed effects estimation?

[Sarah Germer]

I am currently analyzing the following panel data set with STATA on a daily base:

• N = 324 companies
• T = 252 trading days
• 6 social media variables from Facebook and Twitter data (6 for Facebook, 6 for Twitter; e.g., daily answer times, number of posts and replies)
• 2 financial performance variables from CRSP data (daily abnormal return, idiosyncratic risk)

I tried both fixed effects estimation (xtreg fe) and panel vector autoregression (pvar) but neither of the approaches yields satisfying results, i.e., very few significant effects. I also tried the Arellano-Bond approach (xtabond) but was not quite sure about the endogenous and predetermined regressors. Varying these yielded very few significant results.I also varied the operationalization of the social media and financial variables and looked at sub-samples (e.g., single industries, particular time frames, Facebook vs. Twitter sample) etc.

Apparently, for large T panels, the bias apparent for fixed effects estimation - the rationale for dynamic panel analysis - declines with time and eventually becomes insignificant, thus rendering a consistent fixed effects estimator. In comparison with other studies, I would assume that my T is rather large, thus a fixed effects estimation might be more sensible than a panel vector autoregression.

Any thoughts on this topic?

Thanks,
Sarah

[Carlo Lazzaro]

Sarah:
with such a long T dimension, I would consider -xtregar- instead of -xtreg-.
As an aside, evary regerssion model shopuld be aimed at giving a fair and true view of the data generating process: statistical significance of predictors is less relevant.

[Sarah Germer]

Hi Carlo, thanks for your reply! I looked into xtregar and it looks suitable.

However, what do you think about dynamic panel estimation, which, principally, seems to make sense considering the data structure I have?

Thank you!

[Carlo Lazzaro]

Sarah:
dynamic panel data models (-xtabond- command in Stata) are devised for datasets with many panels and few measurement instances when the u_i is correlated with the lags of the regressand.
As per your description, I do not think that -xtabond- is suitable for your data.