Problem with konfound command to get the ITCV

[Aurore Burietz]

I want to examine the impact of unobservable confounding variables on my regression results.
As such, I use the command konfound and I manage to get the RIR output but not the ITCV.

Ma dataset consists in observations at the loan level and looks like:

unique_key AllInSpreadDrawnbps IG rel_withoneleadmin_5y IG_relwithoneleadmin5y etc.
17931 275.00 0 1 0
57707 100.00 1 0 0
28179 75.00 1 1 1
24660 425.00 0 1 0
1887 250.00 0 1 0
128593 500.00 0 1 0
160148 600.00 0 1 0
15129 375.00 0 0 0

My regression model is the following:
reg AllInSpreadDrawnbps IG_relwithoneleadmin5y rel_withoneleadmin_5y IG lnamt log_mat_month d_collateral rev term spec_obj club_deal synd_size i.ctry i.sic1 i.year, robust

Then I enter the following command:
konfound IG_relwithoneleadmin5y, sig(0.05)

And I get the following output (as well as the graph with the threshold and the estimate not reported below):
__________________________________________________ __________________________________________________ _________
Linear regression Number of obs = 24,518
F(103, 24412) = .
Prob > F = .
R-squared = 0.5274
Root MSE = 125.31

----------------------------------------------------------------------------------------
| Robust
AllInSpreadDrawnbps | Coefficient std. err. t P>|t| [95% conf. interval]
-----------------------+----------------------------------------------------------------
IG_relwithoneleadmin5y | 45.5331 3.16541 14.38 0.000 39.3287 51.73749
rel_withoneleadmin_5y | -41.5446 2.48028 -16.75 0.000 -46.4061 -36.6831
IG | -134.8248 3.59735 -37.48 0.000 -141.8758 -127.7738
lnamt | -18.55069 .8089811 -22.93 0.000 -20.13634 -16.96503
log_mat_month | 14.07543 2.545693 5.53 0.000 9.085714 19.06514
d_collateral | 34.99248 2.673828 13.09 0.000 29.75161 40.23334
rev | -55.28328 4.106865 -13.46 0.000 -63.33298 -47.23357
term | 24.9439 4.58205 5.44 0.000 15.96281 33.925
spec_obj | 36.14087 1.863601 19.39 0.000 32.4881 39.79364
club_deal | -25.94557 3.981516 -6.52 0.000 -33.74959 -18.14156
synd_size | -39.56289 1.27087 -31.13 0.000 -42.05387 -37.07191
...
ctry FE
...
sic1 FE
...
year FE
...
_cons | 408.7034 32.90041 12.42 0.000 344.2166 473.1903
----------------------------------------------------------------------------------------

The default framework for the konfound command is now RIR (Robustness of
Inference to Replacement). To generate ITCV, enter: konfound ind_var, indx(IT)
where ind_var is the focal predictor.

For variable IG_relwithoneleadmin5y

Robustness of Inference to Replacement (RIR)

RIR = 21176

To nullify the inference of an effect using the threshold of 6.204 for
statistical significance (with null hypothesis = 0 and alpha = .05), 86.37%
of the (45.533) estimate would have to be due to bias. This implies that to
nullify the inference one would expect to have to replace 21176 (86.37%)
observations with data points for which the effect is 0 (RIR = 21176).
_IG_relwithoneleadmin5yuncond_rycv invalid name
r(198);

end of do-file

r(198);
__________________________________________________ __________________________________________________ ________

To solve the problem, I have tried alternative commands as below:
konfound IG_relwithoneleadmin5y, indx("IT")
konfound IG_relwithoneleadmin5y, indx(IT)
konfound IG_relwithoneleadmin5y, indx("IT") uncond(0)

I have also tried to change the name of my variable to remove "_" hence running the model with IGrelwithoneleadmin5y.

I always get the same error message and no information on the ITCV parameter.

For information, I am using StataSE 17.0.

Thank you in advance for your help!

[Andrew Musau]

I don't use the command konfound (from SSC), but I think I can spot the problem. Stata variable names cannot exceed 32 characters. See

help limits

It appears that the command creates an internal variable by adding a suffix "uncond_rycv" to the variable name. This suffix is 11 characters long, so that means that your variable name cannot exceed 32-11 = 21 characters. The variable "_IG_relwithoneleadmin5y" is 22 characters long.

di length("IG_relwithoneleadmin5y")

Res.:

. di length("IG_relwithoneleadmin5y")
22

So rename variables that are longer than 21 characters in length, e.g.,

rename IG_relwithoneleadmin5y IG_relw1leadmin5y

[Aurore Burietz]

Thank you very much Andrew, shortening variable names has solved the problem and I manage to get both the RIR and the ITCV with the konfound command!