Humans as intuitive classifiers

Abstract

Mainstream decision research rests on two implicit working assumptions, inspired by subjective expected utility theory. The first assumes that the underlying processes can be separated into judgment and decision-making stages without affecting their outcomes. The second assumes that in properly run experiments, the presentation of a complete description of the incentive structure replaces the judgment stage (and eliminates the impact of past experiences that can only affect judgment). While these working assumptions seem reasonable and harmless, the current paper suggests that they impair the derivation of useful predictions. The negative effect of the separation assumption is clarified by the predicted impact of rare events. Studies that separate judgment from decision making document oversensitivity to rare events, but without the separation people exhibit the opposite bias. The negative effects of the assumed impact of description include masking the large and predictable effect of past experiences on the way people use descriptions. We propose that the cognitive processes that underlie decision making are more similar to machine learning classification algorithms than to a two-stage probability judgment and utility weighting process. Our analysis suggests that clear insights can be obtained even when the number of feasible classes is very large, and the effort to list the rules that best describe behavior in each class is of limited value.

Keywords: J/DM separation paradox; description-experience gap; the RUB assumption; underweighting of rare events; wavy recency effect.

Figure 1

Examples of studies of judgement and decision making with and without the J/DM separation.

Figure 2

The list of stimuli used by Erev et al. (2008b).

Figure 3

The screens in Trials 1 and 6 in one of the conditions studied by Erev et al. (2017), when the participant chose the left key.

Figure 4

The impact of experience on sensitivity to losses (from Erev et al., 2023). The experiment used a variant of the experimental paradigm described in the lower panel of Figure 1. It included 100 trials, and the participants were presented with the payoff from all options after each choice. The left-hand column presents the incentive structure, u is a random draw from the range 0 to 1 [that is, from u(0,1)]. The right-hand choice rate graphs present the prediction of the PAS model, described below.

Figure 5

Demonstration of the wavy recency effect (adapted from Plonsky and Erev, 2017). Participants selected repeatedly for 100 trials between two unmarked buttons and received feedback concerning the payoff from both the chosen and the forgone option following each trial. One option generated a payoff of 0 with certainty while the other was a risky gamble detailed in the legend. (A) Exhibits the choice rates of the gamble contingent on the gamble providing a gain at trial t; (B) exhibits the choice rates of the gamble contingent on the gamble providing a loss at trial t; and (C) presents the difference between the corresponding plots in (A,B). Thus, the wavy curves in (C) reflect the impact of an outcome generated by the gamble at trial t on its choice rate in subsequent trials. Positive values (on the Y-axis) imply “positive recency” and negative values imply “negative recency.” Data is averaged across 48 participants from Nevo and Erev (2012) and 80 participants from Teoderescu et al. (2013).

Figure 6

A thought experiment.

Figure 7

An example of a 4-state Markov chain that could determine the payoff from Top in Figure 6.

Figure 8

A decision tree analysis of the results in the first 15 trials of the thought experiment presented in Figure 6. The average payoff line presents the observed average payoff in each category. The question mark (?) implies that the training data do not include observation in the relevant branch.