Modeling costly attention when prior beliefs are uncertain

Two kinds of uncertainty

Rational inattention studies how decision makers selectively acquire information when learning is costly. Rather than assuming that people know everything relevant to a choice, it asks which information is worth obtaining before acting.

Many such models nevertheless assume a precisely known prior: a probability distribution over possible states of the world. Emerson Melo’s preprint, Distributionally Robust Rational Inattention, relaxes that assumption. The decision maker must choose how to acquire information while allowing that the probabilities used to evaluate the decision could be misspecified.

The paper is a theoretical arXiv preprint, not an empirical study. Its contributions are mathematical characterizations and numerical illustrations, rather than measurements of how people actually allocate attention.

Making uncertain beliefs manageable

The model contains finitely many alternatives and possible states, with each action’s payoff depending on the state. In the baseline specification, information costs follow Shannon mutual information, which measures how informative choices are about states.

Instead of trusting one prior, the decision maker considers a set of possible priors surrounding a benchmark distribution. A phi-divergence measures how much another distribution differs from that benchmark. The chosen divergence determines the set’s shape; a radius limits the permitted discrepancy.

The decision maker maximizes the lowest expected payoff, after information costs, across those permitted priors. Here, “worst case” means worst within the specified uncertainty set—not every imaginable outcome. Both payoffs and information costs are evaluated using the candidate prior.

Melo first reduces this problem to an optimization over unconditional choice probabilities: the probabilities of selecting each alternative before conditioning on a particular state. He then uses Lagrangian duality, introducing variables associated with the ambiguity constraint and probability normalization, to obtain an equivalent finite-dimensional concave optimization problem.

Under the stated regularity conditions, this reformulation is exact rather than an approximation. It characterizes optimal choice probabilities and a worst-case prior. With Shannon information costs, the conditional choice rule retains the generalized multinomial logit structure of the standard model.

Which alternatives remain in contention?

A consideration set consists of alternatives assigned positive unconditional choice probability. The paper derives necessary and sufficient conditions for these sets under prior ambiguity.

One application is a stylized insurance-plan choice problem. Exactly one plan is a good match in each state, while all other plans provide the same lower payoff. Under Kullback–Leibler ambiguity, the robust consideration set contains every alternative included under the trusted-prior benchmark, potentially adding others.

The interpretation is specific: an alternative that looks unlikely to be best under the benchmark cannot be dismissed as confidently when the prior may change adversely.

In a five-plan numerical example, the trusted-prior solution includes two plans. Across the illustrated robustness settings, the set expands to include all five, and choice probabilities become less concentrated on the initially favored plan. This is a model result, not evidence that insurance customers necessarily examine more plans in practice.

The form of ambiguity matters

Different divergences produce different behavior. Kullback–Leibler ambiguity reweights states exponentially. Its constrained formulation is connected through duality to an entropic-risk formulation that penalizes departures from the benchmark. These are related formulations, not identical definitions; removing ambiguity recovers ordinary expected evaluation.

Another specification yields Average Value-at-Risk, a risk measure emphasizing high losses. In the insurance illustration, increasing robustness makes groups of plans acquire exactly equal unconditional choice probabilities in successive stages. The five-plan example reaches a uniform allocation at a finite parameter threshold. That uniform-policy conclusion depends on the example’s symmetric payoff structure, not on arbitrary decision problems.

The paper also studies Cressie–Read divergences. For a specified range of this family, evaluating prior ambiguity reduces to a one-dimensional threshold optimization for each candidate choice vector.

When Cressie–Read divergences instead govern attention costs, they can produce exactly zero conditional choice probabilities in particular states. This differs from excluding an alternative from the overall consideration set: an alternative can remain available overall but never be chosen in a particular state.

Scope and remaining questions

Two extensions preserve finite-dimensional formulations: one separates the divergences governing attention costs and prior ambiguity; another adds uncertainty about the baseline prior itself. With generalized attention costs, the optimized reference choice distribution need not equal the induced unconditional choice frequencies.

The results depend on the model’s finite setting, payoff assumptions and specified uncertainty sets. Importantly, a larger consideration set does not establish that more information is acquired. Melo identifies the effect on mutual information, empirical identification of prior ambiguity, and dynamic learning as questions for future research.