On the Origins of Data: Sensitivity to Sampling in Inductive Inferences

Dan Navarro


The study of inductive reasoning typically presents people with problems of the following form: "Objects A, B and C are known to possess property P. How likely is it that object D also possesses property P?" In much of the theoretical literature, such problems are characterised as "inference from givens" (A, B and C), with little if any consideration given to the process by which such facts came to light. Yet in many real world situations the manner in which facts are put together is just as informative as the facts themselves: what we believe about D depends not just on the truth of facts A, B and C, but on how such facts were selected, and perhaps even upon what agenda we think underpins this selection. That is, the sampling method for facts matters.

In this talk I discuss experiments examining the effect that sampling assumptions have upon inferences. The core of this talk discusses category based induction problems, and presents experiments showing how people reason differently when facts are presented as a helpful hint versus when they are perceived to be randomly generated truths, and show that people's inferences closely match the predictions of standard Bayesian models. Time permitting, I will talk about how specific agendas shape people's inferences about non-randomly selected facts, how sampling assumptions affect basic categorization tasks, and how cognitive models can capture these effects.