The internight variability in REM latency in 92 drug-free inpatients with major depressive illness was recorded for 4 consecutive nights and subsequently assessed. Individual coefficients of variation in REM latency [CV = (standard deviation of mean REM latency for 4 recording nights/4-night mean REM latency) X 100] ranged from 5.1 to 121.7, with a mean of 37.0 (SD = 27.3) and a median of 27.4. CV was positively correlated with both age (p less than 0.05) and age at onset of depressive illness (p less than 0.01). Male patients showed more variability in REM latency than female patients (p less than 0.05); likewise, the subgroups of patients who either were incapacitated or had bipolar II illness showed greater variability in REM latency in comparison with the remainder of the sample (p less than 0.05). When the entire patient sample was stratified by CV into three equal subgroups, the subgroup of patients defined by the highest CV presented the longest sleep latency (p less than 0.05) and the shortest REM latency (p less than 0.0001). No other clinical or polysomnographic correlates of REM latency variability were noted nor was REM latency variability related to severity of illness, other subtypes of illness, or clinical response to antidepressant therapy. In selecting REM latency data for assessment of diagnostic sensitivity, the use of the shortest REM latency from at least 3 consecutive nights yielded a higher sensitivity (74%-81%) than did the use of any one individually specified night (50%-56%) or different internight means (49%-52%). The same conclusion applied when patient age was taken into account. These results have implications for standardizing the use of REM latency as a biological correlate in major depression.