Academic journal article The Psychological Record

Human Performance on a Signal Detection Task: Discriminability and Sensitivity to Reinforcement

Academic journal article The Psychological Record

Human Performance on a Signal Detection Task: Discriminability and Sensitivity to Reinforcement

Article excerpt

In a conditional discrimination procedure, a subject emits one of two responses in the presence of a conditional stimulus. One response is correct and the other is incorrect, and the correct response is signaled by the conditional stimulus. An example of a common conditional discrimination procedure is the matching-to-sample (MTS) task, which is a discrete-trials procedure. On each trial, the subject is presented with one of two sample stimuli (the conditional stimuli) and two comparison stimuli, and reinforcement is typically contingent upon selecting the comparison stimulus that matches the sample. Another common type of conditional discrimination is a signal detection task, on which the detectability of differences between conditional stimuli is varied, and the subject must indicate whether or not a difference was detected. For example, the conditional stimulus might consist of two lines of slightly different lengths, and the subject must indicate whether the two lines are the same or different.

There are several quantitative accounts that attempt to predict performances on signal detection tasks, including signal detection theory (Green & Swets, 1966; Macmillan & Creelman, 2005), choice theory (Luce, 1963), and Davison and Tustin's (1978) behavioral signal detection model. Detection theories typically include two key measures: discriminability and bias. Discriminability is an accuracy measure that captures the extent to which differences between the conditional stimuli are detectable. Bias is a measure of preference for one of the response options. An important assumption of most detection theories is that discriminability and bias are independent such that variations in discriminability are not accompanied by changes in bias. For example, consider a quality-control inspector who must choose to accept or reject products based on the presence or absence of defects. Here, the presence and absence of defects serve as conditional stimuli. When defects are easy to detect, discriminability is high, and accurate choices are likely to occur. When defects are difficult to detect, discriminability is low, and errors are more common. If the manufacturer imposes strict penalties when a large number of defective products are sold, then the inspector might develop a greater tendency to reject products (i.e., the inspector shows a bias toward rejecting products). If discriminability and bias are independent, as most detection models assume, then the inspector's bias toward rejecting products might be influenced by the penalties but should not be affected by changes in the detectability of defects.

An explicit account of how biases are produced was provided by Davison and Tustin's (1978) signal detection model. Davison and Tustin based their model on the generalized matching law (Baum, 1974; Davison & McCarthy, 1988), which is a quantitative account of behavior on concurrent schedules. A concurrent schedule is a choice procedure on which a subject can respond at either of two alternatives that are simultaneously available. Responses on the two alternatives are occasionally reinforced, and the ratio of reinforcers programmed for the two alternatives varies. Subjects typically distribute responses among the two alternatives in a manner that closely approximates the distribution of reinforcers. The relation between the allocation of responses and reinforcers is described mathematically by the logarithmic form of the generalized matching equation:

log([B.sub.1]/[B.sub.2]) = alog([R.sub.1]/[R.sub.2]) + logc (1)

In Eq. 1, [B.sub.1] and [B.sub.2] are responses at alternatives 1 and 2, respectively, and [R.sub.1] and [R.sub.2] are reinforcers obtained at each of those alternatives. The parameter a is an estimate of the sensitivity of the response ratios to changes in the reinforcer ratios. The parameter c is an estimate of inherent bias. Unlike bias in detection models, which is a systematic preference produced by experimental manipulations, inherent bias in Eq. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.