Ols, re, fe

[Xin Wen]

Hi guys,

I met a problem with OLS, RE and FE. I present the results obtained from OLS, RE and FE for two dataset. For OLS, the dataset from 2013-2019, while RE and FE, from 2013-2015 short panel dataset. The reason for not including 2017 and 2019 into the panel because of the some data is not collected in these two years.

The reviewer comments that "

The authors correctly indicate on page 2 that cross-sectional data can produce estimates that are “potentially affected by omitted variable bias." Given that is true, all of the OLS analyses that rely on cross-sectional data should be removed from the paper. The estimates are biased. The authors should solely focus on the analyses using the panel data. Correspondingly, the descriptive statistics should be for the panel data set not the cross-sectional sample (Tables 1 and 2).

On page 12, the authors indicate that “the RE model provides more stable parameter estimates.” However, they provide no empirical support for this model specification over a FE model. I would suggest a Hausman specification test." "


In my analysis, i focused on RE due to the small within unit change as Hausman specification test is not reliable in this setting. only 300 sample have changes over the survey...

For this thing, i dont know how to revise....

[Carlo Lazzaro]

Xin:
I think that reviewer is correct.
Some comments follow:
1) you do not report the way you specified you OLS. In addition, you do not detail the postestimation tests you performed on your OLS. As an aside, if your OLS suffers from omitted variable bias, I do not see the reason how this nuisance can be eliminated when you switch to panel data regression;
2) if you actually have panel data, why bothering yourself with OLS?
3) saying that your coefficients are more stable under -xtreg,re- is technically meanigless. You should justify if the -re- estimator was consistent and efficient in your case and why it outperformed the -fe- estimator.

[Xin Wen]

Hi Carlo,

Thanks. I present OLS, RE, FE. One of my interest of variable is gender (or share of daughter)... to some extent or in some cases, it is not suitable to use FE. Despite i report it. For OLS, i meant it might be biased. Thus, perfrom a serious test to support the OLS is not biased, such as Oster, diegert etc... I am not sure this is correct? Or if i change my arguement that do not enphasis it? To be honest, my concern is that Hauseman is not reliable and stable when the within unit change is small. In this situation RE is preferred.

Best,
Xin

[Xin Wen]

The reason for presenting OLS is that the time is longer... and for comparison with RE and FE... the results are the same...

[Carlo Lazzaro]

Xin;
you can reply to the reviewer that, due to the too small variation of time-vaying variables you decided to switch to -re-.
We all know that -hausman- is more fragile that it is expected to be. You may want to consider the community-contributed module -xtoverid- to test whether (or not) -re- is actually the way to go in your case. That said, you should be informed that, being a bit old-fashioned, -xtoverid- does not support -fvvarlist- notation (see -xi- prefix as a possible fix).
Eventually, why not getting rid of OLS altogether in the revised draft of your paper and focusing on -xtreg- only?

[Xin Wen]

Thanks Carlo, i say 'too small variation of time-vaying variables' and provide literature to support it... the review say 'no'...
For ols, I dont want to drop because of the time period used in OLS is from 2013, 2015, 2017 and 2019, not only 2013 and 2015. The panel is only about 2013 and 2015. One important control is only collected in 2013 and 2015 for the same person... In this situation, i would like to emcompass more sample and having longer time.. (not sure this make sense?)

[Carlo Lazzaro]

Xin:
1) go -xtoverid- and, if -re- is actually the way to go, include -xtoverid- results in the revised submission;
2) what above may serve as additional support to justify your econometrics strategy;
3) I see the reason you invoke to include OLS, but I fear that reviewer would find mixing OLS and panel data regression kinda confusing. Acting frequently as a reviewer, when I see an unjustified eccess of methods, I start becoming concerned about authors' level of familiarity with all that stuff. All in all, if you have panel data, you go panel data regression (even with T=2).

[Xin Wen]

Hi Carlo, I tried xtoverid. The p is less than 0.05. it does not work. Yes, I present Panel data regression, and rely on it for the main analysis. Just perent OLS for comparison and clarify that showing the it because it has long time...

STILL DO NOT KNOW HOW TO FIGURE THIS PROBLEM OUT...

[Carlo Lazzaro]

Xin:
please do not shout (this forum as its own etiquette). Thanks.
That said:
1) -xtoverid- result rejects the null that -re- is the way to go. Therefore, you should switch to -fe-;
2) include your OLS, if you prefer.

[Xin Wen]

The interest of var is gender.. if i use FE, it means that i should drop this variable in some regession?

[Carlo Lazzaro]

Xin:
if you're interested in a time-invariant categorical variable but -fe- is the way to go, you may want to take a look at The Stata Blog » Fixed effects or random effects: The Mundlak approach

[Xin Wen]

Thanks, very appreciated!