Bayes vs Skinner: Classifying Motivated Learning Under Uncertainty

How to cite

Kosenko, Andrew; Neligh, Nathaniel. 2026. "Bayes vs Skinner: Classifying Motivated Learning Under Uncertainty." AEA RCT Registry. AEARCTR-0018443. https://doi.org/10.1257/rct.18443-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Andrew Kosenko* , Nathaniel Neligh

* Primary investigator

Registry ID
AEARCTR-0018443
DOI
https://doi.org/10.1257/rct.18443-1.0
First registered
Last updated
Status
In development
Trial dates
2026-04-27 – 2026-12-31
Countries
Private
JEL codes
C91
Registry record
http://www.socialscienceregistry.org/trials/18443

Abstract

We propose novel generalizations of two prominent theories of learning - Bayesian updating and operant conditioning - that provide differential, non-overlapping testable implications, based purely on which observables people learn from, and allow for separation and identification of learning types. We propose an experiment designed to generate maximal conflict between the theoretical predictions of these theories, to distinguish between them. This incentivized experiment will allow us to classify observed behavior in various contexts according to our generalizations of the two theories. It will also allow us to identify the distribution of learning modalities in the subject population under different conditions.

Intervention

We propose to conduct an incentivized laboratory experiment at the economics lab at the University of Tennessee, Knoxville. The experiment will test our proposed generalizations of Bayesian learning and reinforcement learning - the Generalized Signal-based learning (GS) and Generalized Reward-based (GR) learning paradigms.

Experimental design

The experiment will consist of five treatment arms. In each arm, subjects will complete a baseline task, and one of the following five variant tasks: 1) ambiguity, 2) higher stakes, 3) stronger signals, 4) abstract signals, and 5) cognitive load.

Each session is divided into a number of (e.g., thirty) sets of rounds, and each set is divided into three rounds. In each treatment, 15 rounds will be of the baseline task, and the other 15 will be one of the four variations. Whether the baseline or variant task is first is randomized for each individual.

Each round involves one decision is between two lotteries, A and B, followed by some information. Each of these lotteries pays off with either a high probability for the "good" option or low probability for the "bad" option. The state of the world, which determines which of two option is good, is drawn at the beginning of a set of rounds, with either state equally likely ex-ante. The participants are not informed of the state, so they have an incentive to learn through the provided
information; the incentives are such that knowledge of the state raises expected payoff. In addition to the primary payoff from their chosen lottery, there is a potential secondary payoff every period, which comes from a lottery C. This payoff is given with some probability independently of the state or any action from the player (or of anything else). While the choice of a player will ultimately determine their chance of getting the primary payoff from lottery A or B, participants are not informed of the payoffs from their chosen lottery during the experiment.

Design details

Randomization method
The randomization is within-subjects for each arm: whether the baseline task is first or second in that treatment. This randomization, along with all of the lottery payoffs and signals, will be handled by the computer (specifically by the Qualtrics platform using pseudo random number generation). The self-selection into treatments is also random - while students self-sign up for an experiment, they are not informed of the particular experiment they will be taking part in at the time of the signup.
Randomization unit
Individual
Clusters
up to 500 individuals
Observations
up to 500 individuals
Arms
up to 500 individuals
Primary outcomes
Proportions of GS and GR learning types. Then we will look at how these proportions compare across treatments and individuals. The individual level analysis will be used to categorize individuals as being more GR or more GS style learners. In addition to our primary analysis focusing on the distinction between GR and GS learning in conflict scenarios, we will also test specific comparable specifications of learning theories through structural estimation, using specific parametrized models (quantal response equilibrium for GS, and the generalized matching equation for GR).

Cited by

No citation data available for this registration yet.

Keywords