Causal induction in time

Abstract

Causes require time to propagate their effects. We can see stars at night because of the light they emitted hundreds of years ago. We can smell the fragrant aroma of baking bread because heat gradually changed the structure of the food, emitting particles that traveled on the breeze. In this thesis, I investigate how people use temporal information to make causal inferences. I propose a rational framework for causal induction based on continuous-time evidence, examine human performance in passive and active continuous-time causal learning tasks, and develop bounded rational accounts that can offer explanations for human causal judgments and intervention strategies.

Chapters 2 and 3 review previous theoretical frameworks on causal induction, and empirical work on the role of time in causal induction, respectively. Chapter 4 develops a rational framework for processing temporal evidence. It provides an explanation for how delays shape human causal induction and accounts for human causal judgments across seven different temporal causal learning tasks. Chapter 5 and Chapter 6 test how people passively or actively learn causal structures based on events unfolding in real time. I found people are capable temporal causal learners who successfully identify structures that involve generative and preventative relationships, as well as acyclic and cyclic connections. Nevertheless, the computational demands of normative learning could easily exceed human capacity. People’s causal judgments align better with an algorithm that approximates the normative solution via a simulation and local summary statistics scheme, suggesting the reliance on structurally local computation and temporally local evidence. People’s intervention decisions align better with a resource-rational model that emphasizes a balance between expected information and expected inferential complexity when choosing interventions. Chapter 7 shows that when given a limited period of observation, people not only focus on existing data, but also consider future possibilities, relying on extrapolated data to make inferences. This demonstrates the unique “continuing” feature of time, and how generalization plays a role in the utilization of temporal information. Chapter 8 synthesizes the findings of this thesis and proposes future research directions of causal learning in temporal contexts.