Optimal Dynamic Allocation of Attention

Abstract

We consider a decision maker (DM) who, before taking an action, seeks information by allocating her limited attention dynamically over different news sources that are biased toward alternative actions. Endogenous choice of information generates rich dynamics: The chosen news source either reinforces or weakens the prior, shaping subsequent attention choices, belief updating, and the final action. The DM adopts a learning strategy biased toward the current belief when the belief is extreme and against that belief when it is moderate. Applied to consumption of news media, observed behavior exhibits an "echo-chamber" effect for partisan voters and a novel "anti echo-chamber" effect for moderates. DDMs, which lead to a continuous belief process, are more suitable for the problem of learning the properties of a data-generating process from a sequence of samples, as in clinical trials, or for the analysis of statistical information. Poisson models, which lead to discontinuous updating, are useful to model the discovery of individual pieces of information that are very informative, as is common in the political 10 Since the belief is a martingale, we have λαDividing by dt and letting dt → 0 yields (2.1).11 These examples are similar in spirit to Suen (2004) and Burke (2008), except that information is not costly in these models, which leads to a different characterization. In particular, dynamics has no significant effect in these models, in contrast to the point made here.

Consumer Rational (In)Attention to Favorable and Unfavorable Product Information, and Firm Information Design

“…Applications of rational inattention theory have been focused on macroeconomic contexts such as consumption-savings problems (e.g., Luo 2008; Maćkowiak and Wiederholt 2015; Sims 2006) and rigid pricing (e.g., Maćkowiak and Wiederholt 2009; Matějka 2016). Some recent works provide applications in microeconomic contexts, such as an investment game (Yang 2011), discrete choice for rationally inattentive consumers (Matějka and McKay 2015), and dynamic attention allocation for a consumer of news when sources of news are biased (Che and Mierendorff 2019). In the marketing literature, Joo (2020) develops an empirical framework based on rational inattention theory and applies it to new product introduction, while Natan (2020) applies a similar framework to choice in large assortments.…”

“…Following Sims’s pioneering work on rational inattention (Sims 2003, 2006), a number of studies in economics on costly information processing have adopted mutual information as the information cost function in a variety of settings, including consumption-savings problems (e.g., Luo 2008; Maćkowiak and Wiederholt 2015; Sims 2006), rigid pricing (e.g., Maćkowiak and Wiederholt 2009; Matějka 2016), monetary policy (Paciello and Wiederholt 2013), investment decisions (Mondria 2010; Yang 2011), and dynamic learning (Che and Mierendorff 2019; Steiner, Stewart, and Matějka 2017). Morris and Strack (2019) and Hébert and Woodford (2020) show that ex ante cost of acquiring information can arise from a sequential information sampling approach and, under certain conditions, the ex ante cost can be based on entropy.…”

The psychology and neuroscience of partisanship

“…Selective exposure is one likely cause of differing priors, fueling "rational" asymmetric updating. Attention is thought to play a large role in how humans and animals explore and decide (Gottlieb, 2018) and rational attention may explain partisan echo chambers and similar phenomena (Che & Mierendorff, 2019; see also Chapter 33). Researchers have also applied a Bayesian updating framework to multiple processes involving political information, such as perception of climate-related events (Ripberger et al, 2017), political persuasion (Coppock, 2016) and the formation of partisan attitudes (Gerber & Green, 1998).…”

Dynamically Aggregating Diverse Information

“…4 Gossner et al (2019) study the consequences of attention manipulations, where the agent is forced to these papers, we study a setting where the agent dynamically learns about many correlated attributes. 5 Che and Mierendorff (2019) and Mayskaya (2019) also consider choice from a prescribed set of information sources, but they focus on Poisson signals that confirm either one of two states.…”