Logistic GLMM coefficients blow up when adding nested random effects? I have some data (here) that I was hoping to analyse using lme4.  It consists of a binary outcome variable that has been calculated at a number of locations in each of a number of images, which themselves come from two different image databases of differing qualities.  Each outcome is calculated based on two different algorithms, and the aim is to compare the two algorithms.  
I had hoped to run them through glmer as follows:
library(lme4)

data = read.csv("funnycoeffsdata.csv")

model.glmm = glmer(outcome~algorithm + (1|db/image/locationid),data=data,family=binomial)
print(summary(model.glmm))

However, the coefficients I am getting out of this seem to me to be crazily huge, but without outlandishly huge standard errors:
Fixed effects:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept)  26.7815     0.7394   36.22   <2e-16 ***
algorithm    -8.7567     0.2939  -29.80   <2e-16 ***

This is totally different from the result I get from a generalized linear model, 
model.glm = glm(outcome~algorithm,data=data,family=binomial)
print(summary(model.glm))

...which seem like sensible coefficients, and which agree with the log odds ratios that I calculate manually:
Coefficients:
            Estimate Std. Error z value Pr(>|z|)    
(Intercept)  1.79533    0.09717   18.48   <2e-16 ***
algorithm   -0.71806    0.05933  -12.10   <2e-16 ***

I am stumped as to what is going wrong?!?!  I wondered if it might relate to the fact that some of the random effects have zero standard deviation:
Random effects:
 Groups                Name        Variance Std.Dev.
 locationid:(image:db) (Intercept) 765.4    27.67   
 image:db              (Intercept)   0.0     0.00   
 db                    (Intercept)   0.0     0.00   
Number of obs: 5346, groups:  locationid:(image:db), 2673; image:db, 48; db, 2

...but if that is the case, I'm not sure how to get round that, as I need the locationids to be nested within the images?
Any ideas would be hugely appreciated.  
 A: Finally I've cracked it.  The answer is in the huge standard deviation of the random effects.  If I plot the distribution of possible link values for the two groups given these parameters (and for algorithm=1 vs algorithm=2), I get something that looks like this:

If I relabel this with the corresponding probabilities in the response -- let's just label .01, and .99 now, it is clear that almost all of the mass of the distribution is sat above probabilities close to zero or one:

Shifting the centre of the normal distribution to the right or left can no longer be interpreted as change in probability that corresponds to the shift in log odds.  Instead, it has the effect of changing the amount of mass sat above zero, and sat above one, and modifies the probability expressed in the response in this way.  If I make the approximation then that the expected probability response will be roughly equal to mass of the normal distribution that is sat above zero:
> pnorm(c(18,9),mean=0,sd=28)
[1] 0.7398416 0.6260572

...I get values that are pretty close to the actual proportion of success I see for the two algorithms:
[1] 0.7459783 0.5888515

