Simple learning models can illuminate biased results from choice titration experiments

TL;DR

This paper demonstrates that common reinforcement learning models inherently produce biases in choice titration experiments, which can distort the estimation of utility equivalence points, urging reconsideration of such experimental conclusions.

Contribution

The study analytically and empirically shows that a broad class of learning models generate biases in titration procedures, impacting the interpretation of experimental results.

Findings

Reinforcement learning models produce titration bias.

Bias varies non-linearly with reward value.

Results suggest reconsidering inferences from titration experiments.

Abstract

The choice titration procedure presents a subject with a repeated choice between a standard option that always provides the same reward and an adjusting option for which the reward schedule is adjusted based on the subjects previous choices. The procedure is designed to determine the point of indifference between the two schedules which is then used to estimate a utility equivalence point between the two options. Analyzing the titration procedure as a Markov birth death process, we show that a large class of reinforcement learning models invariably generates a titration bias, and that the bias varies non-linearly with the reward value. We treat several titration procedures, presenting analytic results for some simple learning models and simulation results for more complex models. These results suggest that results from titration experiments are likely to be biased and that inferences based on the titration experiments may need to be reconsidered.