Re: Predictive model for rare event

Posted by Bruce Weaver on Jul 30, 2011; 1:07pm
URL: http://spssx-discussion.165.s1.nabble.com/Predictive-model-for-rare-event-tp4649637p4649727.html

To avoid over-fitting a logistic regression model, one should have 15-20 events per model parameter (although sometimes that is relaxed to 10 events per parameter).  "Event" = the outcome category with the lower frequency.  For more info, see:

   http://www.class.uidaho.edu/psy586/Course%20Readings/Babyak_04.pdf
   http://biostat.mc.vanderbilt.edu/wiki/Main/ManuscriptChecklist -- see "Overfitting and lack of model validation"

HTH.

Paul wrote

Hi everyone,

I would like to build a predictive model(Logistic regression or decision
tree or NNet . But the frequency of the event I am predicting is extremely
small less than 1%.
In fact here is my frequency distribution of my dependent variable
Churn_ind
Churn_ind=1  150(0.75%)
Churn_ind=0  19850(99.25%).
Questions:
Q1: What is the minimum sample size to run a reliable model?
Q2: What model could best fit this type of distribution where my event is
less than 1% and in this case 150 out of 20000 ?
Q3:Is there a minimum N sample size when running a decision Tree I tried
to run a decision but got no results.
I turning to the group here to seek for ideas . Your assistance is more
more than welcome.

Thanks,

Paul

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD