Failed Autocorrelation Command on Abnormal Data

[Shella Dini]

Hello, all. Please excuse my English, since I am not native. I am currently doing my thesis for bachelor degree.
I've found myself lacking in econometrics. For my thesis, I'm using time series and cross sections data over 41 firms in the span of 5 years. It is about the influence of employee spending, goods spending, and capital spending on performance quality. I've regressed and found the residual is not normally distributed and multicollinear presents. I've also run autocorrelation test using xtreg and xttest3 command, but it doesn't show anything. How to fix this? It is said, this can be fixed with clustered standard error. But I am just worried of the potential of violating of classical assumptions. Thank you

[Carlo Lazzaro]

Shella:
welcome to this forum.
You're seemingy dealing with a panel dataset.
With 41 panels (that is, firms) and heteroskedastic residuals, you should apply -robust- or -vce(cluster panelid)- standard errors.
About multicollinearity, I cannot say, as you do not post what you typed and what Stata gave you back (as recommended by FAQ).

[Felix Bittmann]

As an addition to Carlo's helpful response, you can also try wild cluster bootstraps if you are using a FE model, see: https://www.stata.com/manuals/rwildbootstrap.pdf

[Shella Dini]

Shella:
welcome to this forum.
You're seemingy dealing with a panel dataset.
With 41 panels (that is, firms) and heteroskedastic residuals, you should apply -robust- or -vce(cluster panelid)- standard errors.
About multicollinearity, I cannot say, as you do not post what you typed and what Stata gave you back (as recommended by FAQ).

Thank you so much for the response, Carlo. Yes, I am dealing with a panel dataset.
I've tried checking back the Stata output but apparently I could not open it back? Turned out I missed the first important thing, that is Starting the Log.
As the result I need to regress it again, which I haven't done because my professor wants the R square to be at least 60% (Mine is 13%). I'm currently scratching my head on how to make it reach 60%. I'll keep in mind about the -robust- or -vce(cluster panelid)- standard errors. Once again, thanks

[Shella Dini]

As an addition to Carlo's helpful response, you can also try wild cluster bootstraps if you are using a FE model, see: https://www.stata.com/manuals/rwildbootstrap.pdf

Thank you Felix. Yes, using the hausman test, it seems FE suits my panel dataset.

[Carlo Lazzaro]

Shella:
how it comes that your supervisor sets a target for the Rsq of your regression without discussing with you the data generating process and the predictors of your regression equation?

[Shella Dini]

Shella:
how it comes that your supervisor sets a target for the Rsq of your regression without discussing with you the data generating process and the predictors of your regression equation?

Well, she actually discussed the predictors of the regression equation beforehand, and she agreed with my predictors. So I ran the regression by myself, since she wanted us process it without her help (this is a disaster for students like me who lacks understanding in econometrics). She has set the requirements of the regression equation (free of autocorrelation, heterocedasticity, multicollinear, Rsq at least 60%, etc) but upon discovering the Rsq doesn't meet the requirement, she asked me to change the equation with another predictors. What predictors? That's for me to do my own research, as long as I meet the requirements, she said.

[Carlo Lazzaro]

Shella:
provided that the Rsq=0.60 still sounds a weird requirement to me, I would suggest you:
1) have a comprehensive knowledge of the data generating process of your regression;
2) as per FAQ, share wiith the list what you typed and what Stata gave you back;
3) try adding square predictors.

[Shella Dini]

Shella:
provided that the Rsq=0.60 still sounds a weird requirement to me, I would suggest you:
1) have a comprehensive knowledge of the data generating process of your regression;
2) as per FAQ, share wiith the list what you typed and what Stata gave you back;
3) try adding square predictors.

Carlo, thank you for your suggestions. I'm trying my best to educate myself. However, may I ask why the minimum requirement of Rsq 0.60 is weird? Wouldn't it just mean, my professor wants my model to at least predict more accurately? About the square predictors, should I put the square on all of my predictors?

My work process is attached below. If you don't mind, please share your thoughts on the matter.
Thank you

[Carlo Lazzaro]

Shella:
with three predictors only, you cannot go that far.
That said, there are panel data analyses with a low within Rsq (and there's nothing to do to fix this issue).
Squaring a given predictor can give you an idea about the non-linear relationship between that predictor and the dependent variable.
I would also try to log the y and see what happens.
As an aside, with >30 panels, I would consider -vce(cluster panelid)- standard errors.

[Shella Dini]

Shella:
with three predictors only, you cannot go that far.
That said, there are panel data analyses with a low within Rsq (and there's nothing to do to fix this issue).
Squaring a given predictor can give you an idea about the non-linear relationship between that predictor and the dependent variable.
I would also try to log the y and see what happens.
As an aside, with >30 panels, I would consider -vce(cluster panelid)- standard errors.

Carlo, thank you. I've fixed the predictors now. From three to seven. And I'm doing hausman test again but this showed up instead:

Note: the rank of the differenced variance matrix (4) does not equal the number of coefficients being tested (7); be sure this is what you expect, or
there may be problems computing the test. Examine the output of your estimators for anything unexpected and possibly consider scaling your
variables so that the coefficients are on a similar scale.

I've seen some post about it (https://www.statalist.org/forums/for...being-tested-3). But I still don't get the answer.
Here is what I've typed in Stata about that hausman test. Hope to hear your opinion on this. Thank you.

[Carlo Lazzaro]

Shella:
1) use cluster robust standard errors;
2) test -re- only via Stata community-contributed module -xtoverid- (type -search xtoverid-) to spot and install it, If the null id rejected, go -fe-.

[Shella Dini]

Shella:
1) use cluster robust standard errors;
2) test -re- only via Stata community-contributed module -xtoverid- (type -search xtoverid-) to spot and install it, If the null id rejected, go -fe-.

Thank you Carlo