Dear all,
In my thesis, I run a logit model on imputed data followed by a prediction of robust standard errors in order to detect outliers in all of my M=10 datasets. Outliers are defined as rsta > 3. I wanted to check whether my results changed when these outliers were removed (just as a sensitivity test, as it is not recommended to remove outliers). However, these models encounter convergence troubles with multiple variables predicting failure/success perfectly. This results in a drastic reduction of observations (2700 to ca. 800).
My code is as following:
The results from one of the imputed datasets produce the following:
It would be great to get some advice on how I can evaluate the effects of outliers in my model in any other way than this! I already tried the firthlogit to handle complete separation better, but this does not support clustered standard errors, which I, unfortunately, need to include.
In my thesis, I run a logit model on imputed data followed by a prediction of robust standard errors in order to detect outliers in all of my M=10 datasets. Outliers are defined as rsta > 3. I wanted to check whether my results changed when these outliers were removed (just as a sensitivity test, as it is not recommended to remove outliers). However, these models encounter convergence troubles with multiple variables predicting failure/success perfectly. This results in a drastic reduction of observations (2700 to ca. 800).
My code is as following:
Code:
mi xeq: logit violent_success lagfpe lagpolyarchy lnlaggdp lnlagpop lnlagmilper oceania asia europe americas coldwar, vce(cluster country_name); predict rsta, rsta; drop if rsta > 3; logit violent_success lagfpe lagpolyarchy lnlaggdp lnlagpop lnlagmilper oceania asia europe americas, vce(cluster country_name);
Code:
Logistic regression Number of obs = 842
Wald chi2(0) = .
Prob > chi2 = .
Log pseudolikelihood = 0 Pseudo R2 = 1.0000
(Std. err. adjusted for 51 clusters in country_name)
---------------------------------------------------------------------------------
| Robust
violent_success | Coefficient std. err. z P>|z| [95% conf. interval]
----------------+----------------------------------------------------------------
fpe | -4363.704 . . . .
polyarchy | 7930.141 . . . . .
gdp | -322.5747 . . . . .
pop | -1940.156 . . . . .
milper | 47.987 . . . . .
oceania | 5060.001 . . . . .
asia | 0 (omitted)
europe | 0 (omitted)
americas | 9373.118 . . . . .
coldwar | -1260.82 . . . . .
_cons | 24333.16 . . . . .
---------------------------------------------------------------------------------
Note: 834 failures and 8 successes completely determined.
