Past Experience Influences Judgment of Pain: Prediction of Sequential Dependencies

Past Experience Influences Judgment of Pain: Prediction of Sequential Dependencies Benjamin V Link (link@colorado.edu) Dept. of Computer Science, University of Colorado at Boulder Boulder, CO 80309 USA Brittany Kos (brittany.kos@colorado.edu) Dept. of Computer Science, University of Colorado at Boulder Boulder, CO 80309 USA Tor D. Wager (tor.wager@colorado.edu) Dept. of Psychology and Institute of Cognitive Science, University of Colorado at Boulder Boulder, CO 80309 USA Michael Mozer (mozer@colorado.edu) Dept. of Computer Science and Institute of Cognitive Science, University of Colorado at Boulder Boulder, CO 80309 USA Abstract Recent experience can influence judgments in a wide range of tasks, from reporting physical properties of stimuli to grading papers to evaluating movies. In this work, we analyze data from a task involving a series of judgments of pain (discomfort) made by participants who were asked to place their hands in a bowl of water of varying temperature. Although trials in this task were separated by a minute in order to avoid sequential dependencies, we nonetheless find that responses are reliably influenced by the recent trial history. We explore a space of statistical models to predict sequential dependencies, and show that a nonlinear autoregression using neural networks is able to predict over 6% of the response variability unrelated to the stimulus itself. We discuss the possibility of using decontamination procedures to remove this variability and thereby obtain more meaningful ratings from individuals. Keywords: els Sequential Dependencies; Judgment Mod- Introduction When asked to make absolute judgments in an experimen- tal setting individuals use anchoring or primacy: informa- tion presented earlier in time serves as a basis for making judgments later in time (Tversky & Kahneman, 1974). The need for anchors is due to the fact that individuals are poor at or possibly incapable of making absolute judgments and instead must rely on reference points to make relative judg- ments (Laming, 1984 ; Parducci, 1965 ; Stewart, Brown, & Chater, 2005). The literature in experimental and theoretical psychology exploring sequential dependencies suggests that reference points change from one judgment or rating to the next in a systematic manner. Teachers are cognizant of potential drift when grading pa- pers and the necessity of comparing early papers to those graded later. Sequential dependencies arise in a myriad of common tasks, such as responding to surveys, questionnaires, and evaluations. A relatively unexplored field of sequen- tial effects involves online recommendation engines. Net- flix, Amazon, and Google consistently recommend products through advertisements that they think you would be inter- ested in buying. Could these recommendation engines be improved by observing how you are rating products sequen- tially? By mitigating the influence of recent judgments, rec- ommendation engines could make more meaningful and ac- curate predictions for what products you are interested in. Even small improvements in these engines can mean large in- come increases. By having the best recommendation engine you not only sell more products, but you draw more users. Carefully controlled laboratory studies of sequential de- pendencies, dating from the 1950’s (Miller, 1956), consist of rating unidimensional stimuli, such as the decibel level of a tone, or the length of a line. These studies suggest that across many such domains, responses convey not much more than two bits of mutual information with the stimulus (Stewart et al., 2005). Various types of judgment tasks have been studied including absolute identification, where the individual’s task is to specify the value of the stimulus level (e.g., 10 levels of loudness), magnitude estimation, where the task is to esti- mate the magnitude of the stimulus which could vary contin- uously along a dimension, and categorization, where the task requires the individual to label stimuli by range. Due to the large size of responses in absolute identification and catego- rization tasks, and because individuals aren’t usually aware of the discrete stimuli in absolute identification tasks, there isn’t a qualitative difference among tasks. Typically, feedback is provided in absolute identification and categorization tasks. Without this feedback, explicit anchors against which stimuli can be assessed wouldn’t exist. The consequences of sequential effects can be complex. Normally, on trial t of an experiment, trial t − 1 has the largest influence on ratings and earlier trials—t − 2, t − 3, and so forth—have successively diminishing influences. Both the stimulus and response on a previous trial can have an effect, which makes sense if individuals formulate a response to the