Voter polarization and extremism

We present a theory of endogenous policy preferences and political beliefs with boundedly rational agents who find it costly to process detailed information. Agents are otherwise fully rational, and they strategically choose how to process information. Their optimal solution is to update coarsely, ignoring less informative signals, and lumping together different histories into a single informational state. We consider an environment with a common prior, in which all agents prefer a moderate policy over extreme alternatives to the left or the right under this prior, and with common signals that would allow Bayesian-updating agents to asymptotically learn the state of Nature and to agree on the policy that is best for everyone. In this environment, we find sufficient conditions under which a majority of agents eventually become extreme and the population becomes polarized: some agents support the left policy, and some support the right policy.