Categorical Data Analysis Using The SAS® System, 2nd Edition

8.8. Exact Methods in Logistic Regression

Until recently, there was no convenient alternative in situations where maximum likelihood estimation failed or small cell counts made the resulting maximum likelihood estimates inappropriate. However, it is now possible to compute parameter estimates, confidence intervals, and p-values for statistical tests using methodology based on exact permutation distributions. The key is conditioning on the appropriate sufficient statistic. The idea is not a new one, having been suggested by Cox (1970), but recent algorithmic advances in computing the exact distributions have made the methodology computationally feasible. Refer to Tritchler (1984) and Hirji, Mehta, and Patel (1987) for more details regarding these ...