Hi researchers
I am struggling to understand complete separation for logistic regression in a SPSS output. Could someone please advise me in basic terminology as I'm new to research and statistics!
I have read field 2013 and Wilson and Lorenz 2015 and don't understand still!
Thank you
Tom
In short terms, if you have a predictor that completely separates the dependent variable then it occurs. What means complete separation is that by just looking at this variable, you can judge if dependent variable is 0 or 1 (say).
You can find many resources about this on internet including the following
http://www.ats.ucla.edu/stat/mult_pkg/faq/general/complete_separation_logit_models.htm
http://support.minitab.com/en-us/minitab/17/topic-library/modeling-statistics/regression-and-correlation/logistic-regression/what-are-complete-separation-and-quasi-complete-separation/
Key references are Allison (2008).[1] (Albert and Anderson 1984).[2]
Complete separation occurs whenever a linear function of X can generate perfect predictions of y. Put simply a frequency of zero in any single cell of the 2 by 2 table implies quasi-separation; two diagonally opposed zeros in the table means that the condition is complete.
The general principle is that whenever there is a zero in any cell of a 2 by 2 table, the maximum likelihood estimate of the logistic slope coefficient does not exist, This is most likely to be the case and more easily spotted when the sample is small but it can also occur and is less easily spotted with a large sample but the outcome is relatively infrequent and the categorical predictors (or maybe their interactions; even more difficult) have been divided into many subgroups. When this is the case an examination of a detailed cross-tabulation will reveal the source of the problem and then combining of categories usually provides a ready solution.
If you want to find it then you must cross tabulate for the full model including interactions and find the zeroes. Another way would be to put in the main effect for a predictor one a time, and then two at a time etc until this problem shows up.
Good hunting
[1] Allison P. D. (2008) Convergence failures in logistic regression., SAS Global Forum: Statistics and Data Analysis, Paper 306-2008, 1–11. Available at http://www2.sas.com/proceedings/forum2008/360-2008.pdf.
[2] Albert, A. and J. A. Anderson (1984) On the existence of maximum likelihood estimates in logistic regression models. Biometrika, 71, 1-10.
Hi
Thanks for the help and pointing me in the right direction. After viewing the first link from Mehmet it tells me how to identity complete separation in spss using Nagelkerke R Square, if it 1.000 it's a perfect fit.
Let's say my output Nagelkerke shows .400, what does that mean!
With Quasi-complete separation
Following the link provided says to look at
Nagelkerke
Classification Table of observed and predicted
Variables in the equation
I don't really know how to interpret my data to find complete separation or quasi-complete separation.
I appreciate your help on this!
Thanks
Tom
http://www.ats.ucla.edu/stat/mult_pkg/faq/general/complete_separation_logit_models.htm
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS