Large errors for logistic regression, what does it mean?

Question

I run the following scripts in r for mixed effect logistic regression.

textbook.usage.glm <- glmer(textbook.usageSession ~ session.week * condition.player  + (1|group.name),family="binomial",data=dfDSP)
summary(textbook.usage.glm)

And I got the following results.

 Generalized linear mixed model fit by the Laplace approximation 
 Formula: textbook.usageSession ~ session.week * condition.player + (1 |      group.name) 
 Data: dfDSP 
 AIC   BIC logLik deviance
 37.27 45.19 -13.64    27.27
 Random effects:
 Groups     Name        Variance Std.Dev.
 group.name (Intercept)  0        0      
 Number of obs: 36, groups: group.name, 2

 Fixed effects:
                                    Estimate Std. Error z value Pr(>|z|)
 (Intercept)                            -68.00   14338.57  -0.005    0.996
 session.week                            17.31    3584.64   0.005    0.996
 condition.playerDEFAULT                 75.63   14338.57   0.005    0.996
 session.week:condition.playerDEFAULT   -20.09    3584.64  -0.006    0.996

 Correlation of Fixed Effects:
        (Intr) sssn.w c.DEFA
 session.wek -1.000              
 cnd.DEFAULT -1.000  1.000       
 s.:.DEFAULT  1.000 -1.000 -1.000

The results look pretty suspicious. The errors are very large, and all the p values are of the same. If the data itself does not have problems, what might be the reason?

Answer 1

It may be that your model is completely determined/separated. That is, there is some variable for which certain values perfectly predict failure (i.e. all observations with those values have a value of 1 in the DV). The fact that your sample is only 36 obs also signals to that possibility. It also happens when failures are extremely rare.

I know that glm() in r warns you about cases being completely determined.

Things you could try are:

1) Try looking at crosstabulations of your IVs against your DVs to look for problematic variables.

2) Try looking at your DV to see if failure is extremely rare. The best scenario would be with 50% of the observations being failures.

UCLA also has a good online resource regarding this issue, with even some r examples