Previously reported acoustic reflex threshold data from normal and hearing-impaired subjects indicate that the effect of stimulus bandwidth on reflex thresholds is altered by sensorineural hearing loss. It is this change in the "bandwidth effect" that forms the basis for predicting hearing loss from reflex threshold data. Three predictive procedures were compared for 17 normal and 60 hearing-impaired ears. All methods correctly identified most hearing losses but none of the methods accurately estimated magnitude of hearing loss. Two methods were characterized by a high rate of false positives. The third was tailored to minimize false positives (6%) and maintain a high rate (93%) of predicting hearing losses greater than 32 dB while making no attempt to make finer discriminations. This more conservative approach minimizes serious predictive errors while identifying a high proportion of clinically significant hearing losses.