Logit Model Error Message

[gabrielle20]

Dear all,

I am trying to model a logit model to predict the determinants of banking crises. However one particular variable, DURATION, seems to be causing problem and the error message "Quasi-complete separation: DURATION>0 perfectly predicts binary response failure" comes up whenever I try and compute the model.

DURATION in this context measures the time in years that have past since the banking crisis.
CRISIS is defined as 0 when there is no crisis and 1 during the years of crisis.

I read from the EVIEWS User's Guide that "This error occurs when one of the regressors contains a separating value for which all of the observations with values below the threshold are associated with a single binary response, and all of the values above the threshold are associated with the alternative response. In this circumstance, the method of maximum likelihood breaks down."

and that

"The only solution to this problem is to remove the offending variable from your specification. Usually, the variable has been incorrectly entered in the model, as when a researcher includes a dummy variable that is identical to the dependent variable."


However, I am quite sure I have not measured DURATION wrongly and have entered the right number of years (since a crisis occured), but I do see how the nature of the variable is associated with the CRISIS measurement. Although the EVIEWS User Guide advises to "remove the offending variable from your specification", the DURATION variable however is quite important to the model and I cannot do so.

Therefore, could I please seek your advise on possible steps I could do to counter this?


Thank you very much

[EViews Glenn]

The issue is that the parameters of your model are not identified with DURATION in the specification. By choosing the parameter on DURATION to be +/-infinity, I can perfectly predict the dependent variable. Another way of saying this is that if you tell me that DURATION is > 0 then I can tell you precisely whether your dependent variable is a 0 and I don't need parameters or a model to do so.

The only things you can do are (as the manual suggests) not use DURATION, or alternately get more data such that the sample separation no longer occurs.

Absent that, you can simply say that DURATION > 0 perfectly predicts that your dependent variable is 0.