Undergraduates in five classes were asked to rate their instructors on four general items, and on twenty-four behavior-specific items describing instructor involvement. Three reliability coefficients differing in error score assumptions were calculated for different numbers of students and items, to permit comparison of the reliabilities of both types of items. Each of the three coefficients was appropriate for a different situation. Internal consistency reliability was appropriate when no generalization was desired beyond a given set of students. Rater reliability was best when generalization beyond a given item or set of items was not needed. General reliability was appropriate when generalization was desired over both items and students. The general reliability of the behavior-specific items was higher than that of the general items, if there were relatively few students compared to items. Rater and internal consistency reliabilities exhibited the opposite results. The halo effect for the behavior-specific items was not significantly less than the general items, but the relationship was in the predicted direction. Finally, there was less leniency bias on the behavior-specific items, as predicted. (The twenty-eight items are appended). (CP)