Abstract
A logit model of the incidence of long-term unemployment is estimated using a random sample of completed unemployment spells in New Zealand over the period 1988–97. The regressions results are sensible, but the predictive power of the model seems to depend on the share of long-term unemployed in total unemployment. In particular, if one-third or less of annual observations are made up of long-term unemployment spells, the model is incapable of predicting any such spells, unless it is constrained to do so.