Learning Under Uncertainty

Following the literature on belief updating, we will utilize a typical setup with two urns and two colors of balls in each urn. In particular, the urns will be called red and blue. Each urn will contain varying numbers of red and blue balls, such that the blue urn contains more blue balls than the red urn. One of the urns is selected by a computer to be the 'chosen urn.' Subjects are truthfully informed about the probability with which each urn is selected. Then the computer makes random draws (with replacement) from the selected urn. After some number of draws, we elicit subjects' beliefs regarding which of the urns has been selected.

In our planned experiment, the intervention consists of manipulating the prior probability of each urn being chosen, the composition of each urn, the realization of draws from the selected urn, and the number of draws after which beliefs are elicited. The urn compositions will be determined randomly by a computer ahead of the experiment. Based on the simulations, we rank scenarios presented to the participants as being more or less subject to the rounding bias and the direction of the bias. Specifically, for each set of scenarios (where a scenario includes the prior, the composition of each urn, the random draws, and the timing of decisions), we simulate decisions of agents that do not round and those that do. We then estimate the learning parameters for these 'non-rounders' and 'rounders' and determine the bias's direction and magnitude. The main hypothesis is that scenarios that are deemed to be susceptible to the positive rounding bias (based on our measure) will result in estimates that are different from scenarios that are susceptible to the negative rounding bias. In addition to the main focus on learning, we elicit measures of cognitive ability using progressive matrix reasoning tests from ICAR as well as a short version of the Big5 questionnaire. We hypothesize that higher cognitive ability will result in less frequent rounding and, as such, would be less subject to the influence of rounding bias on learning estimates. In addition, we intend to explore how the big 5 personality traits as well as personality types identified in Gerlach et al (2018) relate to the proclivity to round and (non) bayesian learning.

2024-09-25

2024-12-15

The primary variable of interest is the elicited belief about which of two urns is being chosen and the estimates of the (non) bayesian learning parameters.

Participants will be presented with a series of scenarios where they report their beliefs about probabilistic outcomes.

Each participant will see 12 scenarios. The first 9 will have a prior that is 'round' to 5 percentage points (e.g., 25, 30, 35 etc). The last 3 wil have priors that are not round (e.g., 31). Subjects will be matched in groups of 4 such that all four subjects in a group have the same sequence of priors across scenarios, but timing and draw realization are organized in a 2x2 factorial design. Specifically, 1 & 2 have the same draws but different timing of decisions (e.g., after 2nd and 5th signal vs after 1st and 8th signal). Same for 3&4 but with different draws. 1&3 have the same timing but different draws. 2&4 have the same timing but different draws. We do this because one of the IV estimators we plan to use will rely on signal strength.

Randomization done by a computer.

Individual

No

120

120

120

N/A