"Incorrect" p-value during probit regression when probability of a response is zero I am running a probit regression with glm.  I got a non-significant value (p~0.98) for an interaction that other statistical methods suggested should be significant.  Looking into it, I think it's because 1 of the cells had zero positive ("YES") outcomes.  When I artificially change a data point so that there's one positive outcome, I get a significant result of p < 0.02, even though this actually decreases the difference in the effect of the cue ("Cue") between the two conditions ("Condition").
The code in R is glm(Response~Cue*Condition, family=binomial(link=probit), data=data) 
The data look like this:
Condition = 0
        Response:    0   1  
  Cue: 0            49   3  
       1            25  25  

Condition = 1  
        Response:    0   1  
  Cue: 0            48   3       
       1            47   0 

The counts aren't quite exact but it shows the pattern, which is that the subject responds "appropriately" 50% of the time to Cue 1 in Condition 0 but does not do so in Condition 1 (and in fact, responds slightly more often in this condition to the incorrect cue, although this isn't significant).  Changing the last row of counts to 46 1 gives the significant interaction I was expecting.
My intuitions are hazy, but I imagine this has something to do with the probit function being undefined for probabilities of 0 or 1.  Does this seem correct?  If so, why does it return a non-significant value rather than a warning?  More to the point, is there a way to deal with this?  I can imagine building in a correction that gives a non-zero estimate for the probability based on the number of trials, but is there a standard, implemented solution?
 A: Let me address a few preliminary issues first.  As a terminological matter, there is no such thing as "logistic regression with a probit link".  Rather, there is the generalized linear model that can use a variety of different links (of which, logit is one, and probit is another), and different distributions of $Y$.  Thus, you are doing probit regression.  The maximum likelihood methods used to fit these models are not troubled by 0's, unless there's something very weird about your data that I don't recognize.  When and whether your statistical software returns a warning vs. results will depend on many factors, importantly, which software you are using, but as I say, it's not clear that anything is amiss.  
An interaction means that the nature of the effect of one covariate depends on the level of another.  For example, $Y$ may increase as $X_1$ increases when $X_2$ is low, but decrease as $X_1$ increases when $X_2$ is high.  However, that is a different issue than the main effect of a covariate (cf., "decreases the difference between the conditions").  It is quite possible that that "cell" (n.b., I know nothing of the structure of your data or the design of your study) has a great deal of leverage.  Without knowing more about your data, my first guess would be that you should accept the results that the software are returning.  I do believe that arbitrarily changing your data to get the results you expect is generally considered a no-no.
A: We don't have enough information to really know what is going on, how are you computing the p-value? what do the graphs of your data look like? etc.  One possibility is the Hauk-Donner effect (you can google for the term to see many discussions of this).  Another possibility is complete separation in the interaction term (which may indicate that you are overfitting the model for the data given). 
