Large regression coefficients-Am I doing something wrong?

[Ellen Veelema]

Dear all,

I am doing a regression using first the command

xtset FormFilingID QDate, quarterly

(Stata gives this:
panel variable: FormFilingID (unbalanced)
time variable: QDate, 2015q1 to 2017q4
delta: 1 quarter)

xtreg abs_DACC Disclosure Ln_Assets Ln_marketvalue Loss IPO restate BIGFOUR1 Male MB CashFlow, re

This is the output:

Random-effects GLS regression Number of obs = 7,221
Group variable: FormFilingID Number of groups = 842

R-sq: Obs per group:
within = 0.0705 min = 1
between = 0.0720 avg = 8.6
overall = 0.0494 max = 9

Wald chi2(10) = 530.64
corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000

--------------------------------------------------------------------------------
abs_DACC | Coef. Std. Err. z P>|z| [95% Conf. Interval]
---------------+----------------------------------------------------------------
Disclosure | -6.113946 4.300984 -1.42 0.155 -14.54372 2.315828
Ln_Assets | 28.32825 2.935873 9.65 0.000 22.57404 34.08245
Ln_marketvalue | -45.19214 3.350653 -13.49 0.000 -51.7593 -38.62498
Loss | -28.15692 5.830906 -4.83 0.000 -39.58529 -16.72856
IPO | 7.605641 6.595344 1.15 0.249 -5.320996 20.53228
restate | -3.810372 6.753607 -0.56 0.573 -17.0472 9.426455
BIGFOUR1 | 11.81093 10.89348 1.08 0.278 -9.539912 33.16176
Male | 5.253551 13.4826 0.39 0.697 -21.17186 31.67896
MB | -.0025431 .0124751 -0.20 0.838 -.0269939 .0219077
CashFlow | -7.009103 .3883925 -18.05 0.000 -7.770338 -6.247867
_cons | 121.357 18.82433 6.45 0.000 84.462 158.252
---------------+----------------------------------------------------------------
sigma_u | 142.79481
sigma_e | 162.28735
rho | .43636702 (fraction of variance due to u_i)
--------------------------------------------------------------------------------

As you can see the coefficients are large, that is why I think that I am doing something wrong. How can interpret these results? I have done a random effects regression, but fixed effects does not make much of a difference.
Previous to the regression analysis, I changed the data set from 'wide' to long, for making a variable. Maybe it is an idea to change it back? How can I do that easily, without loss of data.
Or do I have to eliminate outliers for example. I do not know what to do now, so I am hoping that someone could provide me with some help. Thanks a lot!

[Nick Cox]

Each coefficient has units of measurement, units of response / units of predictor, even if the units are the same and so cancel. So, whether the coefficients look or even are large depends on the relationship between the variables and both units.

Your output is hard to read because you didn't use CODE delimiters (FAQ Advice #12) but I don't see anything implausibly large.

[Ellen Veelema]

Thanks for your answer, I would look at the units of measurement.
I have still two more things:

- The independent variable keeps being insignificant. With panel data, with a number of observations of 7221 (after telling stata to ignore missing values), I think this is weird.
I want to check whether this is also the case when I am changing it back to its normal format 'wide'. How can I do this? Or is there something I can do?

- I have three dummy variables (Public/Male/BIGFOUR1) that I pasted directly from excel into the data editor. When I ran the following command

xtreg DACC Disclosure Ln_Assets Ln_marketvalue CashFlow Loss IPO restate MB Public Male BIGFOUR1, fe

these variables are omitted from the regression output table, which is bad because these are important variables for my hypothesis testing. What is going on here and how can I fix it?

Many thanks.

[Nick Cox]

Panel data deserves to be in long layout. It is hard to see why you think a change of layout will help you here. At a minimum what you would be doing is fitting a quite different model with fewer observations.

I don't know which variable you're referring to as the independent variable. I see 11 predictors here. I don't know why you're jumping back and forth between fixed and random effects.

I think you need most of all to talk to someone in your field familiar with this kind of data, and sorry, but that's not me. I can't even guess what DACC is.

[Ellen Veelema]

Thanks for your message!