A longstanding distinction in psychology is between implicit and explicit preferences. Implicit preferences are ordinarily measured by observing non-choice data, such as response time. In this paper we introduce a method for inferring implicit preferences directly from choices.We ask subjects to state preferred sentences for two criminals whose cases are presented side by side. We find that there is a negligible difference in average sentence given to black and to white defendants, when the defendant is presented alongside someone of the other race. However black defendants receive 1-year longer sentences when the defendants are presented alongside someone of the same race. We interpret this data as demonstrating two facts: the absence of any explicit racial preference, but the existence of an implicit racial preference (against black defendants). Data on order effects and response time are consistent with this interpretation. The necessary assumption is that implicit preferences toward an attribute (e.g. gender, race, sugar) have a stronger effect when the attribute is mixed with others, and so the decision becomes less “revealing” about one's preferences. We discuss reasons why preferences would have this property, advantages and disadvantages of this method relative to other measures of implicit preferences, and application to measuring implicit preferences in racial discrimination, self-control, and framing effects.

Signaling models are often used to explain the behavior of firms, politicians, and employees as attempts to persuade an observer. However in many models the effort spent on persuasion has no effect on the observer’s average action, because in equilibrium the observer rationally adjusts for the effort exerted. We show that when the receiver makes a binary decision, and signals are received with noise, signaling will in fact cause a systematic change in the receiver’s choices, and the change will be in favor of the sender’s preferred action. This occurs because the receiver will use a threshold rule, the threshold leads to bunching of signals just above the cutoff, and the bunching causes a skew distribution of posteriors, increasing the probability of the receiver taking the sender’s preferred action. The paper additionally contributes an analytical solution to a class of noisy signaling models with many applications.

It is commonly argued that the brain aggregates information in a hierarchical fashion. In this paper I point out that hierarchical aggregation of information will give rise to certain predictable imperfections in inference,
consistent with well-known features of perceptual illusions and decision biases. I consider a setup with two modules: both infer some unobserved value, each has private information, but the second module additionally
observes the first module’s posterior. As a whole this system will fail to aggregate information efficiently, in particular it predicts two particular features of decision-making: (1) the influence of irrelevant associations (framing
effects), and (2) the avoidance of dominated options. This combination of properties distinguishes the model from either random preferences or inattention.

Decisions made in laboratories have often been found to be sensitive to details of how the choice sets are structured and ordered. I provide a framework for describing these effects, in which preferences can vary depending on the history of choice sets, but which restricts preferences to be consistent among decisions that are made simultaneously. Within this framework I introduce a simple comparison effect, in which being exposed to a larger value along some dimension makes a person less sensitive to differences along that dimension (expressed in terms of the marginal rate of substitution). This effect predicts the existence of many otherwise disparate patterns of anomalous choice (contrast and anchoring effects, response range effects, joint vs separate reversals, decoy, scope neglect and common difference effects). I also discuss conditions under which this behaviour could reflect Bayesian inference from the composition of the choice set.


(2013) Handicapping Politicians: the Optimal Majority Rule in Incumbency Elections
(Joint with Ines Moreno de Barreda, Francesco Caselli, and Massimo Morelli) 
Economica (forthcoming) / NBER Working Paper version
We present a model of electoral competition between an incumbent and a challenger in which everything is symmetric except that voters receive more information about the quality of the incumbent than that of the challenger. The information is received with noise, and is subject to manipulation through costly effort by the incumbent. In equilibrium we show that this model predicts an incumbency advantage, such that incumbents are more likely to be elected than challengers even when their qualities are drawn from identical symmetric distributions, and elected by voters with rational expectations. We also show that a supermajority re-election rule, which sets a threshold for re-election somewhere greater than 50%, improves welfare, mainly through discouraging low-quality politicians from sending high signals

(2009) Leader Behaviour & the Natural Resource Curse (with Francesco Caselli) 
Oxford Economic Papers 61 (4) 628--650 
We discuss political economy mechanisms which can explain the resource curse, in which an increase in the size of resource rents causes a decrease in the economy’s total value added. We identify a number of channels through which resource rents will alter the incentives of a political leader. Some of these induce greater investment by the leader in assets that favour growth (infrastructure, rule of law, etc.), others lead to a potentially catastrophic drop in such activities. As a result, the effect of resource abundance can be highly non-monotonic. We argue that it is critical to understand how resources affect the leader’s "survival function," i.e. the reduced-form probability of retaining power. We also briefly survey decentralised mechanisms, in which rents induce a reallocation of labour by private agents, crowding out productive activity more than proportionately. We argue that these mechanisms cannot be fully understood without simultaneously studying leader behaviour.


In experiments subjects regularly trade off time and money at inconsistent rates, apparently becoming less sensitive to money when considering the purchase of a more expensive item. In this paper I introduce a demand function that matches this behaviour, and derive the predictions for equilibrium market structure, and discuss market evidence. The demand function is derived from a utility function which has context-dependent sensitivity to different goods (e.g., time, money). In market equilibrium with unit total demand the model predicts three phenomena: higher cost goods tend to have higher markups; higher cost goods tend to have greater price dispersion; and higher cost goods will have a larger number of sellers. The first two facts are commonly observed in the empirical IO literature. Finally I introduce a novel dataset of 3,500 costs and prices from a branch of a chain drugstore, with cross-section evidence in support of the theory.

Accounting for the Resource Curse
A substantial literature has found a robust negative correlation between economic growth and the share of income coming from resource exports, often interpreted as a 'curse of natural resources'. I decompose aggregate growth into growth rates in the resource and non-resource sectors, and investigate how far the original correlation can be explained by other economic phenomena: (i) lower growth in the resource sector, (ii) reversion to the mean in resource extraction, and (iii) resource-funded capital accumulation. I conclude that about half of the resource curse is accounted for through these channels.