Conndence Intervals on Likelihood Estimates for Estimating Association Strengths

Sparse data causes errors in the maximum-likelihood estimates of event probabilities that are often large enough to render measures of association such as pointwise mutual information useless for small sample sizes. This squib describes a procedure for estimating event probabilities that produces a conndence interval estimate rather than a point estimate. Using these conndence intervals in calculations of measures of association results in reasonable association strength rankings even if the data is drawn from very small sample sizes.