Previous research on driver drowsiness detection has focused primarily on lane deviation metrics and high levels of fatigue. The present research sought to develop a method for detecting driver drowsiness at more moderate levels of fatigue, well before accident risk is imminent. Eighty-seven different driver drowsiness detection metrics proposed in the literature were evaluated in two simulated shift work studies with high-fidelity simulator driving in a controlled laboratory environment. Twenty-nine participants were subjected to a night shift condition, which resulted in moderate levels of fatigue; 12 participants were in a day shift condition, which served as control. Ten simulated work days in the study design each included four 30-min driving sessions, during which participants drove a standardized scenario of rural highways. Ten straight and uneventful road segments in each driving session were designated to extract the 87 different driving metrics being evaluated. The dimensionality of the overall data set across all participants, all driving sessions and all road segments was reduced with principal component analysis, which revealed that there were two dominant dimensions: measures of steering wheel variability and measures of lateral lane position variability. The latter correlated most with an independent measure of fatigue, namely performance on a psychomotor vigilance test administered prior to each drive. We replicated our findings across eight curved road segments used for validation in each driving session. Furthermore, we showed that lateral lane position variability could be derived from measured changes in steering wheel angle through a transfer function, reflecting how steering wheel movements change vehicle heading in accordance with the forces acting on the vehicle and the road. This is important given that traditional video-based lane tracking technology is prone to data loss when lane markers are missing, when weather conditions are bad, or in darkness. Our research findings indicated that steering wheel variability provides a basis for developing a cost-effective and easy-to-install alternative technology for in-vehicle driver drowsiness detection at moderate levels of fatigue.