[Carlo Lazzaro]

Wojciech:
I would go with 2).
I'm not sure I understand your statement about creating fixed effect in random effect specification: what I get from you code 3) (I would skip code 4) altogether) is that -i.year- and -i.industry_numeric- are two categorical predictors.

[Wojciech Gulkowski]

Carlo, thank you for your suggestion.

From our previous discussion and also from the study that I attached I understood that adding factor variables such as i.year and i.industry create fixed effects. For example when we talked about two different codes that yield equivalent results (as far as coefficients go) in post #20. Also the researchers from the paper I attached used pooled OLS as you told me yet somehow they achieved the fixed effects so I guess it must have been exactly by adding dummy variables such as i.year and i.industry. So I understand that similar result can be achieved by adding fixed effects to random effect model by using i.year and i.industry dummies?

[Carlo Lazzaro]

Wojciech:
the example in my reply #20 shows that you can get the same (shared) coefficient with -xreg,fe- and pooled OLS: in that case, the fixed effect was -i.idcode-.
This in not to sat that whenever you add a categorical variable as a predictor you estimate a fixed effect.
Actually, Stata will give you back the -fe- for the -panelid- only via:

predict fixed_effect, u

[Wojciech Gulkowski]

Hi Carlo,

So I'm wondering why the papers that I cited said they are including industry fixed effects when adding industry dummies to OLS regressions (as you deducted they used OLS). Is there any other way to include industry fixed effect in OLS if not by using the industry dummy?

I have also a question on a different matter. I would like to address potential endogeneity problem in my data. In my case it means that on one hand litigation risk may cause firms to hold cash but on the other plaintiffs may target firms with higher cash holdings so that is a matter of causality. I'd like to prove that it is the litigation risk that causes cash holdings to increase and not the other way round. My idea is to approach the problem like this:

1) I get significant estimates of both litigation risk (lit_risk) and lagged litigation risk (lit_risk_L1) on cash holdings, that is positive relationship between litigation risk in year X and X-1 and cash holdings in year X. The estimates are more significant for lagged litigation risk (at 1% for lagged vs 5% and 10% for non-lagged litigation risk) which suggests that firms may take some time (i.e. 1 year) to respond to litigation risk with increasing their cash holdings.

2) In order to do a robustness check and see if my hypothesis of causality works I'd like to run the regression again with lagged cash holding dependent variable. I expect to find no significance in litigation risk and other control cash determinants I use. Thus if there is no significant relationship between cash holdings in year X-1 and volume of litigation in year X (which proxies my litigation risk variable) I would be able to conclude that it is the litigation risk that increases the cash holdings (as identified above) and not plaintiffs targeting firms with high cash holdings?

Please let me know if my interpretation makes sense and if I could run such robustness check for this causality issue. Thank you!

[Wojciech Gulkowski]

I also just came up with another idea. I switched between my dependent and independent variables. So now DepV is litigation risk in year X and IndepV is cash holding in year X-1. The code and results look like this:

. xtreg lit_risk ln_cash_L1 size lev mtb nwc rd growth cf cf_vol_5y industry_sigma acq capex ndi nei div i.year,fe vce(cluster id)

Fixed-effects (within) regression               Number of obs     =      3,082
Group variable: id                              Number of groups  =        351

R-sq:                                           Obs per group:
     within  = 0.0036                                         min =          1
     between = 0.1028                                         avg =        8.8
     overall = 0.0654                                         max =          9

                                                F(23,350)         =       1.14
corr(u_i, Xb)  = 0.2083                         Prob > F          =     0.3044

                                     (Std. Err. adjusted for 351 clusters in id)
--------------------------------------------------------------------------------
               |               Robust
      lit_risk |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
---------------+----------------------------------------------------------------
    ln_cash_L1 |   .2871454   .2416618     1.19   0.236    -.1881465    .7624374
          size |   .9466028   1.284871     0.74   0.462    -1.580437    3.473643
           lev |  -1.245571   2.094912    -0.59   0.553    -5.365772     2.87463
           mtb |  -.0587954   .1572341    -0.37   0.709     -.368038    .2504471
           nwc |  -6.998818   6.905282    -1.01   0.311    -20.57989    6.582249
            rd |  -.4961855   1.254047    -0.40   0.693    -2.962602    1.970231
        growth |   .3997085   3.108338     0.13   0.898    -5.713663     6.51308
            cf |   1.664195   4.430782     0.38   0.707    -7.050113     10.3785
     cf_vol_5y |    3.25787   6.381671     0.51   0.610    -9.293377    15.80912
industry_sigma |   13.45013   15.11897     0.89   0.374    -16.28533    43.18558
           acq |  -1.521907   2.190983    -0.69   0.488    -5.831057    2.787242
         capex |   15.76633   8.196191     1.92   0.055    -.3536515    31.88631
           ndi |   1.189985   2.272429     0.52   0.601    -3.279349    5.659319
           nei |  -2.074384   4.103369    -0.51   0.614    -10.14475    5.995979
           div |   1.048681   1.680184     0.62   0.533    -2.255846    4.353209
               |
          year |
         2011  |   .1661361   .7749815     0.21   0.830     -1.35807    1.690343
         2012  |  -.7488583   .6333224    -1.18   0.238    -1.994455     .496738
         2013  |   .7570477   .7931244     0.95   0.340    -.8028415    2.316937
         2014  |   .2143333    .854488     0.25   0.802    -1.466244     1.89491
         2015  |   .1920137   1.215992     0.16   0.875    -2.199558    2.583585
         2016  |   1.169417    2.74959     0.43   0.671     -4.23838    6.577213
         2017  |   .0080952   1.386499     0.01   0.995    -2.718823    2.735013
         2018  |  -.3366292   1.253065    -0.27   0.788    -2.801114    2.127855
               |
         _cons |  -3.459696   12.65699    -0.27   0.785    -28.35303    21.43363
---------------+----------------------------------------------------------------
       sigma_u |  19.769966
       sigma_e |  15.104291
           rho |  .63143318   (fraction of variance due to u_i)
--------------------------------------------------------------------------------

.

As you can see the model is insignificant as well as the individual variables. Particularly cash holding is insignificant in explaining the litigation risk. Please let me know which approach would be better to show that it is the litigation risk that increases cash holdings and not the other way round. The one presented in 2) of post #34 or the one presented in this post.

Thank you.

[Carlo Lazzaro]

Wojciech: provided that yours is niot may research field, the option 1) in #34 seems more palatable.
That said, if you think that endogeneity is actually an issue with your data, you should skim through the literature in your research field and see how others set up an instrumental regression when facing your very same problem.

[Wojciech Gulkowski]

Carlo,

thanks. You mean this option, right?

2) In order to do a robustness check and see if my hypothesis of causality works I'd like to run the regression again with lagged cash holding dependent variable. I expect to find no significance in litigation risk and other control cash determinants I use. Thus if there is no significant relationship between cash holdings in year X-1 and volume of litigation in year X (which proxies my litigation risk variable) I would be able to conclude that it is the litigation risk that increases the cash holdings (as identified above) and not plaintiffs targeting firms with high cash holdings?

You wrote "option 1)" so it got me confused as I mentioned both 1) and 2) in post #34.

The thing is that I don't have much time and cannot afford to collect any new data. I would like to do the best I can with what I have.

[Carlo Lazzaro]

Wojciech:
I meant:

1) I get significant estimates of both litigation risk (lit_risk) and lagged litigation risk (lit_risk_L1) on cash holdings, that is positive relationship between litigation risk in year X and X-1 and cash holdings in year X. The estimates are more significant for lagged litigation risk (at 1% for lagged vs 5% and 10% for non-lagged litigation risk) which suggests that firms may take some time (i.e. 1 year) to respond to litigation risk with increasing their cash holdings.

[Wojciech Gulkowski]

Ok I see. However, option 1) (what you quoted) is the main experiment I run. So in order to see if there is no opposite effect (i.e reverse causality) would option 2) be correct (what I quoted)? I understand that you are not expert in this field but I mean from econometrics' interpretation standpoint.

[Carlo Lazzaro]

Wojciech:
again, see if what you have in mind as 2) is in line with the literature of your research field,

[Bao Tran]

Hi there,

How can I add an IF statement to the Firm and Year FE? I have tried to run the following but it did not work. Thanks!

------------
xtset SIC_n Year
xtreg W_EARN5 RSIZE RRDS RAD i.Year, fe if XRD != 0

[Carlo Lazzaro]

Bao:
see https://www.statalist.org/forums/help#stata 12.1 What to say about your commands and your problem

Thanks.