This study examined the relative predictive validities of several measures of pain intensity. Forty chronic pain patients completed 6-14 days worth of hourly pain ratings, which were averaged to obtain a measure of actual average pain intensity. These patients then made ratings, on 101-point numerical rating scales, of worst, least, and usual pain during the previous 2 wks, and of their current pain. A series of correlation coefficients were computed and regression analyses were performed to determine the individual or composite measures that best predicted actual average pain intensity. Consistent with previous research, the best single predictor of actual average pain intensity was patient rating of least pain in the previous 2 wks. Of all possible composites of usual, least, worst, and current pain ratings, the arithmetic mean of least and usual pain had the strongest relationship to actual average pain. The inclusion of ratings of most pain or current pain in any composite score actually weakened the relationship between the composite score and actual average pain intensity. These results suggest that, when clinicians or researchers wish to assess average pain among chronic pain patients, but cannot obtain multiple measures of pain over time, the most valid measure would be the arithmetic mean of patient-recalled least and usual pain.