How does limited attention to attributes shape stochastic choice?

Two rules share attention but aggregate it differently

Both rules rest on the same primitives: an attention probability for each attribute, and an attribute-specific preference ordering over alternatives. Attention is assumed independent across attributes, and the decision maker is generally attentive, meaning each attribute is more likely to be considered than not. What separates the rules is what happens after attention lands on an attribute.

The Multiplicative Attention Rule (MAR) is tops-only within each attribute: an alternative earns credit only on attributes where it ranks first, and those credits are multiplied across attributes. An alternative that ranks second everywhere receives no positive reward from any attribute. The Additive Attention Rule (AAR) instead averages attention across attributes and credits an alternative by its full ordinal rank within the menu, so being second is worth something.

That difference in aggregation is the paper's central explanatory device. The authors characterize each rule with axioms on observable choice data and show that, in both cases, the attention parameters and attribute-based preferences are uniquely identified from that data. The identification is economical: preference orderings can be recovered from binary comparisons under singleton attribute sets, and attention parameters from singleton alternative sets under singleton attribute sets.

Multiplying tops-only credit produces default-heavy choice

MAR's formula makes deferral a structural outcome rather than an add-on. In the paper's three-alternative, two-attribute example, with attention probabilities 0.6 and 0.9 and different top alternatives on each attribute, the three listed alternatives receive probabilities 0.03, 0.18, and 0.01. The default alternative absorbs the residual, 0.78.

The mechanism is that an alternative must be top-ranked on the attributes that draw attention, and the credits for those attributes multiply. When attributes point to different winners, the decision maker is pulled in conflicting directions and often ends up outside the listed menu. The paper describes this as choice deferral driven by the cognitive load of attending to too many attributes.

The authors' examples sharpen the point. With a Pareto dominant alternative present, high attention probabilities and few relevant attributes can still let that alternative beat the default: with two attributes at 0.9 and 0.85, the dominant alternative gets 0.765 against a default probability of 0.2275. Lower both attention probabilities to 0.6 and the dominant alternative falls to 0.36 while the default rises to 0.56. Add more attributes at high attention, and the dominant alternative drops to 0.4698 against a default of 0.530196. These are model computations under stated parameter values, not empirical estimates.

Averaging ranks reproduces attraction without changing attributes

AAR's behavioral signature is the Attraction Effect. Adding a decoy that is dominated by one alternative but not by another can reverse the choice ranking of the two original alternatives while the relevant attribute set stays fixed across menus. The paper states that MAR can deliver this only when the attribute set itself changes, because MAR is a stochastic version of a tops-only rule.

The paper also shows why a simpler normalization fails. A rule that divides by menu size rather than by the rank sum N(S) = |S|(|S|+1)/2 retains averaging and full-rank credit and even reproduces the same attraction reversal condition, but it is not a genuine stochastic choice rule. With attention weights 0.4 and 0.5, the probabilities sum to 0.9 on a three-alternative menu and to 1.125 once a fourth alternative is added, forcing the default probability to −0.125. The rank-sum normalization is what keeps the probabilities valid as the menu grows.

Additive linearity rules out the compromise effect

The same linearity that lets AAR capture attraction prevents it from capturing the Compromise Effect. In the paper's configuration where a middle alternative z is ranked between x and y on both attributes, making z the most likely choice would require attention weights to satisfy mutually exclusive conditions, γ(b) > γ(a) and γ(a) > γ(b). No parameter values satisfy both.

MAR fails on the Compromise Effect for a different reason: it is tops-only, and a compromise alternative is never at the top of any attribute. The authors note that accommodating the effect would require a non-linear, for example concave, transformation of ranks, which lies outside the additive class they characterize. This is a stated scope limit of the two rules, not a claim that no attribute-based model could produce the effect.

Observable axioms pin down both rules exactly

For MAR, Theorem 1 states that a stochastic choice rule satisfies the listed axioms, including independence across attribute sets, finiteness, uniformity, dominance, binary distinctness, and a transitivity condition, if and only if it is a MAR. The proof builds attribute-specific linear orders from binary comparisons under a single attribute, then uses dominance to show that only top rank matters, and independence to split and reassemble probabilities into the multiplicative form.

Uniqueness follows directly. If two parameter families both represent the same choice data under MAR, singleton menus force the attention probabilities to coincide, and any disagreement in the attribute orderings would force an attention probability equal to one half, contradicting the requirement that it exceed one half. AAR receives a parallel axiomatic characterization, with its own axioms and its own uniqueness result.

The paper also establishes regularity in attributes for MAR: adding relevant attributes weakly lowers the probability that a given alternative is chosen, which is the formal counterpart of choices becoming harder as more attributes compete for attention.

Attribute count and attention levels decide who wins

The paper's applied discussion follows from its own variables. A firm selling a Pareto dominant product benefits when the platform limits the number of relevant attributes and when attention to those attributes is high, because MAR multiplies attention across the attributes where the product ranks first. The authors describe negotiating with a platform over which filters consumers may use as a way to lower the relevant attribute count.

The examples also show robustness in relative terms. With 99 attributes at attention 0.9, an alternative that tops every attribute has probability 0.9^99, about 0.0000295, while the rival's probability is approximately zero; adding one attribute favoring the rival lowers the first alternative's absolute probability but leaves its relative advantage intact. Absolute choice probabilities and relative rankings therefore move differently, and the paper keeps them separate.

The practical priority the paper supports is conditional rather than universal: whether a dominant alternative or the default wins depends on the number of relevant attributes and on how far attention probabilities sit above one half. The same model that explains default-heavy behavior under many attributes also predicts a dominant alternative winning when attributes are few and attention is high.

Independent attention is the model's key restriction

The model assumes attention is independent across attributes, and the authors flag this as a limitation in their own examples: nutritional attributes and fruit type in a cereal choice are correlated, which the framework does not allow. They name dependent attention, where high attention to one attribute raises attention to a complementary one, as future work.

The results are characterizations of choice rules under stated axioms, illustrated with parameterized examples rather than estimated on data. The probabilities in the examples are model outputs for given attention weights and orderings, so they show what the rules imply, not what consumers were observed to do. Within that scope, the paper's distinctive contribution is the comparison it makes possible: multiplicative tops-only aggregation buys deferral and default choice, additive rank aggregation buys attraction with fixed attributes, and neither buys the compromise effect.

Sources used in this answer