sigma_u=0 in random effects model

[Thana Keda]

Hello,

I am new in this forum and i hope that someone will be able to help!

I am currently trying to evaluate the timely effect of the covid crisis on web sales, using social media increased use as an explanatory variable. My dataset is actually not extensive and N=11.

However, when running my panel data, Hausman test provides a p-value > 0,05, which indicate -re- as preferred. When running the Breusch and Pagan test, once again, the p-value is > 0,05 and is =1.

Y1 stands for the firms used in my datased while social_media is the reported use of the latter.

Here's the code i used:

Does this mean that i have to use pooled OLS? or is there another trick that could be used instead?

Thank you in advance for your answers!

[Joro Kolev]

You are pushing the notion of "large sample" quite a bit with your 11 observations...

But I wonder how is it possible that -xtreg, fe- calculates sigma_u at 5.3, while -xtreg, re- calculates sigma_u at 0... Maybe somebody else can explain this mistery.

[Carlo Lazzaro]

Thana:
in addition your F-test on u_i=0 in -fe- does niot reach statitsical significance too.
Therefore, you should switch to OLS; that said, with such a scant sample size I do not think that you can safely go any farther that descriptive statistics.

[Joseph Coveney]

I am currently trying to evaluate the timely effect of the covid crisis on web sales, using social media increased use as an explanatory variable.

Isn't your regression model is missing a predictor, namely, time?

It's also not clear whether your social-media-use predictor is a continuous (quantitative) measure of social media use or a yes-no indicator ("dummy") variable of increased use between the first and second measurement intervals.

Regardless, if you're interested in "timely effect", then maybe stick with the fixed-effects model, and ignore the preliminary testing stuff given the structure of your data (repeated measurements on whatever id represents) and that it's an observational study with all of the implications of what corr(u_i, Xb) = -0.2536 is hinting at.

But it you want to forge ahead with a mixed-effects regression model, then you'll want to use something like mixed with its small-sample-size adjustments (dfmethod() reml) in light of your small sample size.

[Thana Keda]

Thank you for your answer Carlo, i'm gonna try to look more into this

Joseph, social-media-use variable is continuous. Should i add a time dummy? going yes if covid year and no otherwise?

I will try to look into the mixed effects model, thank you very much!

[Joseph Coveney]

social-media-use variable is continuous. Should i add a time dummy?

Yes, you should add the indicator variable for time, that is, first and second measurements (time == 0 and time == 1, respectively). You might find that it alone explains some, or even most, of the variation observed in the Web sales outcome variable.

Consider assessing time × social-media-use interaction, too, if it makes sense in the context of your research question.

going yes if covid year and no otherwise?

I'm not certain what you mean here, but I assume that a time indicator variable and this COVID-year variable are both for the same condition, that is, the first measurement was made prior to the outbreak and the second measurement was made during the ongoing pandemic.