Professor Ido Erev

Research Summary

Dr. Erev and his co-workers study the effect of the economic environment on human behavior. They focus on the possibility that simple models of learning and response rules can clarify complex behavioral phenomena. One example involves the effect of rare (low probability) events. People behave as if they overweight these events in some settings (e.g., when buying insurance and lotteries), but underweight them in other settings (e.g., while driving). The research conducted by Dr. Erev and his associates shows that this pattern can be understood as the outcome of the joint effect of a mere presentation effect that drives planning decisions from descriptions, and the tendency to rely on recent personal experience in small decisions. This analysis and related observations are summarized below.

1. The value of descriptive models with general parameters over games

Experimental studies of the effect of experience on choice behavior appear to present a complex picture. Experience was found to quickly move behavior toward equilibrium in some games (situations), to have little effect in other games, and even to move behavior away from the equilibrium predictions in certain environments. These findings led many researchers to focus on situation-specific learning models (or models with situation-specific learning parameters).

The research conducted by Dr. Erev and his associates demonstrates the value of an alternative approach. In a series of papers they used computer simulations to show that surprisingly simple models of reinforcement learning that assume fixed parameters over games can reproduce the experimental results in a wide set of situations. Each of these papers focuses on a specific set of experimental games. For example, Roth and Erev (1995) focuses on games in extensive forms, Erev and Roth (1998) focuses on constant sum games, and Erev and Rapoport (1998) focuses on coordination games. While refining the basic arguments from paper to paper they try to insure that a new model, if necessary, will account to all the data captures by the models it replaces in addition to the new data that motivate them.

2. Measuring usefulness.

The main criticism of the attempt to use general models is the assertion that the assumption of general parameters over tasks and subjects can be statistically rejected. Dr. Erev and his associates answer this critic by noting that the models they propose are not meant to be accurate, these models suppose to be useful approximations. This answer was criticized on the ground that the concept “useful approximations” is vague. To address this criticism Erev, Roth Slonim and Barron (2002; 2006) proposed a method to measure of usefulness of a descriptive model. The proposed measure, referred to as the Equivalent Number of Observations (ENO), is an estimate of the expected size of the experiment that has to be run to obtain predictions that are more accurate than the models’ predictions.
To highlight the implications of this measure consider a situation in which an experiment is run sequentially, and after observing the behavior of the first n subjects you are asked to predict the behavior of subject n +1 based on your favorite model (that makes the point prediction M) and the mean behavior of the first n subjects (Xbar). Under reasonable assumptions the optimal prediction is Pred = w(M) + (1-w)(Xbar), where w =ENO/(n +ENO).
Erev et al.’s analysis shows that ENO is closely tied to well known concepts in statistics including Student t, Cohen d, regression analysis, and the minimum variance rule. The main contribution of this research is the demonstration that given a certain two-stage experimental design it is easy to see how these concepts can be used to measure useful of approximation. They also show that in the context of constant sum games the ENO of simple learning models is about 6. The ENO of equilibrium is close to 1.

Here is a link to two SAS programs that were used in Erev, Roth Slonim and Barron (2006). The first estimate the parameters of one of the model based on one data set. The second compute the model's ENO.

3. The difference between the intermediate- and the long-term

Roth and Erev (1995) analysis of simple learning models leads to positive and negative results. The positive results, discussed above, involve the predictability of behavior in the intermediate terms. Simple learning models can be used to approximate the experimental results, and this approximation is relatively robust to the details of the model. The negative result involves the difficulty in generalizing from an experiment to behavior in the long term. This analysis shows that small differences in the details of the model that have almost no implication on the predictions of the intermediate term, can have dramatic effect on the long term prediction. This observation was influential because much of the older learning literature focused on the long term. Roth and Erev’s analysis demonstrates the shortcoming of this focus, and the value of an alternative approach.

4. Adaptation, maximizations and reinforcement learning among cognitive strategies

Erev and Barron (2004) present a comprehensive review of experimental studies of learning in individual choice tasks under uncertainty. The review reveals robust deviations from maximization that can be described as indications of three effects: (1) A payoff variability effect: High payoff variability seems to move choice behavior toward random choice. (2) Underweighting of rare events: Alternatives with the highest payoffs most of the time are attractive even when they are associated with low expected return. (3) Loss aversion: Alternatives that minimize the probability of losses can be more attractive than alternatives that maximize the expected payoff.

Although the different tendencies can contradict and interact with each other, Erev and Barron’s analysis shows that their joint effect can be summarized with a simple model that assumes reinforcement learning among cognitive strategies. With a single set of four parameters, this model captures the 40 experimental conditions used to demonstrate the three types of deviations. Moreover, with the same parameters, the model provides good predictions of behavior in 39 other repeated choice conditions (including the games analyzed in Erev et al., 2002).

5. Decisions from experience and their limited correspondence to decisions from experience

Barron and Erev (2003) explore situations in which the information available to decision makers is limited to feedback concerning the outcomes of their previous decisions. The results reveal that experience in these situations can lead to deviations from maximization in the opposite direction of the deviations observed when the decisions are made based on a description of the choice problem (as in Kahneman & Tversky, 1979). Experience was found to lead to a reversed common ratio/certainty effect, more risk seeking in the gain than in the loss domain, and to an underweighting of small probabilities. Only one of the examined properties of description-based decisions, loss aversion, seems to emerge robustly in these “feedback-based” decisions. These results are summarized with a simple model that illustrates that all the unique properties of feedback-based decisions can be a product of a tendency to rely on recent outcomes.

Hertwig et al. (2004) examine if the difference between description and feedback documented by Barron and Erev is a result of the source of the information or the number of decisions made in each case. They study decisions from sampling. The subjects were allowed to sample they possible payoff distributions and than make a single choice. The results replicate the trend found in Barron and Erev. Thus, it seems that this trend is a function of the fact that the source of information was the decision maker personal experience.

6. Perceptual games and safety dilemmas.

Erev and Gopher (1999) summarizes five years of research that generalizes the basic reinforcement learning model to address perceptual and attention control decisions. This research includes a descriptive signal detection (SDT) model that assumes reinforcement learning among perceptual cutoff strategies. Erev (1998) shows that this model can account for all robust violations of SDT. This model was found to outperform alternative models (Barkan, Zohar & Erev, 1999). In addition this model was extended to address 2-person games and evaluate its implications in safety dilemmas (Erev et al., 1995; Gopher et al., 2001) and consensus games (Gilat et al., 1997; Meyer et al., 2003)

7. The economics of small decisions

It is important to note that in the experimental paradigms considered above the decision makers took only few seconds to consider each problem, and the difference between the expected payoffs of the different alternatives was only few cents. These properties of this basic research suggest that the regularities summarized above might not occur when people make high stakes “big decisions” based on careful evaluation of the different alternatives. Erev and his associates believe that this possible shortcoming of the generalizability of their research does not mean that it has limited practical implications. Many natural behaviors involve “small decisions” that resemble the decisions studied in the research summarized above, and in many cases these small decisions can be highly consequential.

In order to evaluate the practical implications of their basic research Erev and his associates have recently started a line of research that focuses on small decisions outside the laboratory. The problems they considered include: Using safety devices (Yechiam, Erev & Barron, 2003), running red lights (Perry, Haruvy & Erev, 2002), selecting routes (Erev, Barron & Remington, 2004), entering a web site (Erev & Haruvy, 2004), trying to cheat in an exam (Erev et al., 2004), looking at the keyboard while typing (Yechiam, Erev, Erev-Yehene & Gopher, 2004), selecting among menu based and code based data analysis procedures (Yechiam, Erev & Parush, 2004), and the decision to try to solve a math problem (Erev, Erev & Markovich, 2004).

Each of these studies highlights a particular undesired behavior that seems robust but, under the current analysis, can be easily modified. In most papers Erev and his associates test this optimistic prediction in a laboratory and a field experiment. For example, Erev, Ingram, Raz and Shany (2004) focus on the problem of cheating in exams. They first show that (1) this problem, like many other rule enforcement problems has two Nash equilibria; and (2) the tendency to underweight rare events (discovered by Barron and Erev, 2003) moves behavior toward an equilibrium in which students are motivated to cheat. Under the current analysis this problem can be addressed with a more efficient allocation of the enforcement resources. Specifically small but high probability punishments in the beginning of the exam are expected be more effective than low probability high punishments.
At the initial stage of the study this prediction was evaluated (and supported) in a laboratory experiment. At the second stage they tested the implications of the suggested solution on midterm exams in the Technion. Currently proctors are asked to prepare a map of the class at the beginning of the exam. This activity is important to collect evidence to justify heavy punishment if cheaters are detected. Its main cost involves the fact that the proctors are busy at the beginning of the exams. Thus, the preparation of the map reduces the probability that early attempts to look at the neighbors’ notebooks will be gently punished (e.g., by a move to the first row). The current analysis suggests that this cost is higher than it seems. To evaluate this hypothesis Erev et al. compared two conditions. The control condition was similar to the method used today. In the experimental condition the proctors were asked to delay the preparation of the map by 50 minutes. The results show that this minimal manipulation significantly and clearly reduced the perceived cheating.

8. The coexistence of overconfidence and conservatism

In an older research project Erev, Wallsten, Budescu, (1994) found that two of the most important characteristics of human judgment: overconfident and conservatism, can be predicted by a simple model assuming that responses in both cases reflect some well-calibrated knowledge and random noise. This finding is important because alternative explanations that consider one phenomenon at a time imply that the two phenomena are inconsistent. As in the case of learning, the general model leads to interesting implications that cannot be derived from narrower models. For example, Wallsten, Erev, Budescu and Ditriech (1998) use this model to derive that optimal weighting of n assessments.

Erev and Hertwig (in preparation) try to relate the noisy response concept presented in Erev, Wallsten and Budescu (1994) with the recency concept highlighted in Erev and Barron (2003). This analysis suggests that the joint effect of these two basic principles can be used to map the condition under which people overweight and underweight rare events.

Edit page