# Individual Differences in Explaining Noisy Data

- Daniel R. Little,
*Indiana University*
- Richard M. Shiffrin,
*Indiana University*

## Abstract

In science, we design our inference approaches to trade off fit to
observed data (models are good that fit well) and complexity (models or
explanations that fit or explain everything are bad). Here, we examine how
observers balance fit and complexity by asking observers to estimate causal
models for noisy data. Specifically, participants are shown a number of
scatterplots that vary in the number of data points shown, the noise added to the
true function and the complexity of the true function. For each set of noisy data
points, participants estimate a function which best captures their guess at the
causal explanation between the input and the output. A generative psychological
model combining Bayesian model selection and Gaussian process regression is used
to examine individual differences in biases toward simple explanations. Our
results indicate that some participants prefer simple polynomial, rule-based
explanations and others prefer distance-based, similarity explanations.

Back to Table of Contents