Effects of point-loss punishers on human signal-detection performance

Abstract

Three experiments using human participants varied the distribution of point-gain reinforcers or point-loss punishers in two-alternative signal-detection procedures. Experiment 1 varied the distribution of point-gain reinforcers for correct responses (Group A) and point-loss punishers for errors (Group B) across conditions. Response bias varied systematically as a function of the relative reinforcer or punisher frequencies. Experiment 2 arranged two conditions - one where an unequal ratio of reinforcement (5ratio1 or 1ratio5) was presented without punishment (R-only), and another where the same reinforcer ratio was presented with an equal distribution of point-loss punishers (R+P). Response bias was significantly greater in the R-only condition than the R+P condition, supporting a subtractive model of punishment. Experiment 3 varied the distribution of point-gain reinforcers for correct responses across four unequal reinforcer ratios (5ratio1, 2ratio1, 1ratio2, 1ratio5) both without (R-only) and with (R+P) an equal distribution of point-loss punishers for errors. Response bias varied systematically with changes in relative reinforcer frequency for both R-only and R+P conditions, with 5 out of 8 participants showing increases in sensitivity estimates from R-only to R+P conditions. Overall, the results indicated that punishers have similar but opposite effects to reinforcers in detection procedures and that combined reinforcer and punisher effects might be better modeled by a subtractive punishment model than an additive punishment model, consistent with research using concurrent-schedule choice procedures.

Keywords: humans; mouse-click; point-loss; punishment; signal detection.

Fig 1

A 2 × 2 matrix illustrating the four possible response outcomes in a two-alternative signal-detection task.

Fig 2

Predictions made by punishment versions of Davison and Tustin's (1978) GML-based model of signal detection. The effects of varied punisher ratio (top) and reinforcer ratio (bottom) on response bias (log b) are plotted for additive model predictions (dotted lines), and subtractive model predictions (dashed lines), when a  =  .9, log c  =  0, and q  =  1. Figure 2 (bottom) also plots the predicted changes in response bias when the relative reinforcer ratio is varied without punishment for errors (solid line).

Fig 3

An illustrative example of a “more blue” stimulus array with the response buttons presented during each trial in Experiment 1.

Fig 4

Discriminability (log d – top) and response bias (log b – bottom) are plotted over changes in relative reinforcer frequency (log R₁₁/R₂₂) for Group A (left) and relative punisher frequency (log P₂₁/P₁₂) for Group B (right) of Experiment 1. Individual participant data and the overall means are given.

Fig 5

Discriminability (log d – top) and response bias (log b – bottom) are plotted over the three sessions for Order 1 (R-only, R+P, R-only – left) and Order 2 (R+P, R-only, R+P – right) of Experiment 2. Individual participant data and the overall means are given.

Fig 6

Discriminability (log d – top) and response bias (log b – bottom) are plotted over changes in relative reinforcer frequency (log R₁₁/R₂₂) for the reinforcer-only conditions (left) and the reinforcer + punisher conditions (right) in Experiment 3. Individual participant data and the overall means are given.

Fig 7

Response bias (log b) is plotted over changes in relative reinforcer frequency (log R₁₁/R₂₂) separately for each participant in Experiment 3. Reinforcer-only conditions are presented as filled circles, while reinforcer + punisher conditions are presented as unfilled triangles. Results from least squares linear regression analyses are also presented for each participant for R-only and R+P conditions separately.