In many cases the distribution of saccadic reaction times (SRT) deviates considerably from a unimodal distribution and may often exhibit several peaks. We present a statistical approach to determining the number and form of the individual peaks. The overall density of the reaction times fi
t, i 1 F F F w obtained in w dierent experiments with the same subject is described as the sum of u basis functions xk
tY k 1 F F F u with dierent weights and an error term. A change in the experimental conditions is assumed to cause a change in the weights, not in the basis functions. We minimize the square of the dierence (measured data minus approximation), divided by the error of the data. Incrementing u step by step we determine the necessary number of basis functions. This method is applied to data of six subjects tested in dierent saccade tasks. We detect ®ve dierent modes: two in the range 80±140 ms (express modes), two in the range 145±190 ms (fast-regular mode) and ...