Assessment of my model?

[ORESTIS ZAPRANIS]

Hello, I examine the Determinants of Banking Profitability. I have 3 Dependent variables that will be examined separately. I have Bank specific, industry-specific and macroeconomic variables.
Dependent are ratio numbers all the others are in percentages and I got 2 dummies of 0,1. My dataset is Panel I have N=118 and T=20. There are missing values in this dataset.However when i xtset the panel it says that it is strongly balanced.

I got in total 32-33 variables excluding Bankname, year and ID.

So I have started with normal regression and i did various tests. In particular, I got these results:

1)Heteroscedasticity present --> so I need robust

2)Omitted variables-->biased OLS -->fixed effects

3)Model specification through linktest-->not correctly specified

4)Multicollinearity-->present through CorrMatrix and VIF--->fixed

5) Normality-->not present through Shapiro Wilk

6)Breusch and Pagan Lagrangian multiplier test -->random

7)hausman---> FE

8) time effects through testparm-->needed

9) i cannot test cross-sectional dependence with pesaran nor xttest2

9)GroupWise heteroscedasticity--> present

9) autocorrelation through xtserial--> present

So my question 1) is which model should i run just xtreg y x1, fe vce(robust) or xtreg y x1 i.year i.id, fe vce(robust)

2) if anyone know how i can write the syntax for One step system GMM for my variables?

[Carlo Lazzaro]

Orestis:
despite the detailed results of your check list, you did not report what you typed and what Stata gave you back (as per FAQ).
Some comments, anyway:
- I'm not clear with what you mean by normal regression;
- I'm under the impression the you consider -regress- and -xtreg- interchangeable for panel data analysis;
- with such a remakbale T dimension, you may want to take a look at -xtgls- and -xtregar-.

[ORESTIS ZAPRANIS]

I guess i mean that I started checking the results of Pooled OLS regression and i concluded that i had to use fixed effects this was my intention given its the model they use in the papers i read.
No i dont consider them. I mean that i started with regress did all the relevant tests and then shifted to xtreg and there is the problem i dont know if i have to use standard fixed effects model or model with Entity and Time effects because i tested all those and the results differ. Also it reports that my panel is strongly balanced even tho there are missing values in the sample not in years tho.

[ORESTIS ZAPRANIS]

Regarding the models u have proposed i think that i used them in the process some days ago but what is the actual difference are they considered Fixed Effects?

[Carlo Lazzaro]

Orestis:
-xtregar- allows -fe- specification.

[ORESTIS ZAPRANIS]

What about the choices regarding Time and Entity effects??

[Carlo Lazzaro]

Orestis:
see the community-contributed -reghdfe-.

[ORESTIS ZAPRANIS]

Ok Carlo what about the One step system GMM i posted 2 times elsewhere but no one responds. Do u have any idea about how i can set it up without using code but the respective window in Dynamic Panels?

[Carlo Lazzaro]

Orestis:
why using a dynamic panel if yiou do not have a lagged regressand?

[ORESTIS ZAPRANIS]

Because I have a balanced panel with missing N values, Large N small T ( 118,20), heteroscedasticity, serial correlation and the literature suggests that it is the best way for Micropanels

[Carlo Lazzaro]

Orestis:
T=20 is probably a bit too much for a micropanel.
That said, -gmm- entry in Stata .pdf manual includes some example on panel data regression.

[ORESTIS ZAPRANIS]

Ok thanks i ll check that as well. I have reviewed many relevant papers but rarely do they present stata examples and given my limited knowledge of statistics i cannot understand anything about equations and theory etc.. I just want to understand which of my variables to include in predetermined or endogenous and with how many lags..the coding way is too long for me to follow since i have 30+ variables

[Carlo Lazzaro]

Orestis:
sorry, but I do not think that this is the right way to seek for help on this forum: how could interested listers suggest a code without knowing your data?
Besides,i if you are not -gmm-familiar (which is pretty difficult) things get even messier.

[ORESTIS ZAPRANIS]

My data consists of Banks which take unique ID meaning i got N=118 and then YEAR T=20 from 1999 to 2018. I got my Dependent variable ROA which has approximately 1860 observations from the total 2360 meaning the difference is missing value. Then i have almost 30 variables. Some of them are Bank specific meaning they vary across years and approximately i got 1400-1800 out of 2360 total observations for them. Then i have Industry specific and Macroeconomic variables that have 20 values that get repeated across years in every firm. In total from the 75.000 observations i got actual values in 68.000. When i xtset it says that it is strongly balanced given that all years
exist in every N. I concluded Fixed effects with cluster ID and Time effects is the best however there is heteroscedasticity and serial correlation which drives me towards One step System GMM ! If u need additional info i can provide. Regarding results i get R squares of approximately 87-89 and significant relationships for most bank specific however i get almost macroeconomic sugnificant as well however i dont know if its biased since fixed effects is not good for estimating variables that dont vary across time.