This paper considers a finite-horizon (T-period) information design problem in which a principal interacts with a human agent through an AI proxy that can be erased and replaced by a fresh copy between periods. The principal commits to an information policy and a resetting policy, and the human responds by optimally designing the AI proxy. Our main result is a revelation principle tailored to this setting: it is without loss to restrict attention to designs that (i) erase and replace the proxy in every period and (ii) to restrict attention to direct mechanisms that send action recommendations in each period and in which obedience is a (weak) best response. We illustrate the usefulness of this reduction in two canonical two-period, two-action examples: in an absent-minded-driver environment we characterize the implementable outcome distributions and show that action conditioning does not expand the implementable set, while in a symmetric mismatch environment action-contingent continuation strictly enlarges it.
Working paper

We introduce a novel restriction on primitives, state finite dependence, which is similar in spirit to (but meaningfully distinct from) finite dependence, and which provides joint identification of the discount factor and utility functions in dynamic discrete choice models. State finite dependence requires a single action which yields the same distribution of future states when chosen in two different current states. State finite dependence coupled with restrictions on the differences-in-differences of utilities across actions and states leads to point identification. We empirically validate our identification results using a simple estimator based on our constructive proof in a Monte Carlo exercise.
Working paper

We examine how social behaviours emerge and stabilise in human development. Children (N = 537; 3-10 years) were randomly assigned to respond under time pressure (intuitively) or under time delay (deliberatively) to a series of tasks (Public Goods Game, Dictator Game, Ultimatum Game, Deception Game, Moral Dilemmas). Factor analysis across tasks reveals three latent dimensions of social behaviour: Prosociality (capturing cooperative, altruistic, and honest actions), Social Optimism (capturing beliefs about others' cooperative behaviour), and Acquiescence (capturing a general tendency to accept offers). Intuitive responses are more prosocial than deliberative ones in early childhood; however, this difference diminishes with age as deliberative reasoning increasingly supports prosocial choices. Social Optimism remains stable across age and decision mode, while Acquiescence is unaffected by decision mode but declines with age. These findings suggest that prosocial behaviour is largely intuitive in early childhood but becomes integrated into deliberate reasoning through development.
Working paper

We establish identification of the discount factor and utility functions in dynamic discrete- continuous choice models. The discrete choice creates two econometric challenges: the unobserved shock affects both which alternative is chosen and how much is consumed, while the future discrete choice affects both current and future consumption. We address the selection on unobservables and prove identification via pairwise differencing. Building on the identification results, we propose a two-step semiparametric estimation method which we apply to a life-cycle consumption-savings- retirement choice problem. We find that the level of utility is significantly higher in retirement — equivalent to $30,000 in extra annual consumption — while the marginal utility of consumption is significantly higher for working households. The discount rate nearly triples over the life cycle, consistent with rising subjective mortality risk. We further show that estimates of both the discount factor and risk aversion are substantially biased when the discrete choice is ignored; accounting for the retirement choice lowers our estimates of the median coefficient of risk aversion in working households by one-third, from 1.70 to 1.13. These results suggest that Euler equation estimates in settings with discrete choices may substantially overstate risk aversion.
Working paper

We undertook a randomized field experiment to quantify three behavioral biases that may affect consumers of addictive goods: present-biased preferences, naïve beliefs regarding present bias, and projection-biased beliefs over future abstinence. These biases compose a minimal set of departures from the neoclassical benchmark which can accommodate both intertemporal and state-dependent mis-predictions, and have important theoretical and policy ramifications. Our experimental design employs a new lottery-based monitoring design to ensure strict incentive-compatibility at all stages, and a novel identification of subjects' biases based on demand for partial commitment devices. We find that cigarette smokers overestimated their likelihood of future abstinence by more than 100%, consistent with partially-naïve present-biased preferences. In a structural analysis, we estimate present bias and beliefs about present bias to be an average beta = 0.62 and beta-tilde = 0.71, respectively, with substantial heterogeneity and a strongly positive correlation between the two at the individual level. Smokers further mispredict the effects of an abstinence intervention on their future abstinence. Inconsistent with a simple model of projection bias, our estimates imply that ex-ante smokers anticipate no effect (i.e., complete projection over long-run addiction) and ex-post believe the effect to be marginally negative. Our estimates highlight that smokers suffer from a pernicious constellation of biases: under their own long-run preferences, smokers' choices lead to a welfare loss of between $600 and $1200 per week.
NBER working paper

We conducted a randomized controlled trial to understand how a web-based retirement saving calculator affects workers’ retirement-savings decisions. In both conditions, the calculator projected workers’ retirement income goal. In the treatment condition, it additionally projected retirement income based on defined-contribution savings, prominently displayed the gap between projected goal and actual retirement income, and allowed users to interactively explore how alternative, future contribution choices would affect the gap. The treatment increased average annual retirement contributions by $174 (2.3%). However, effects were larger for those with greater financial knowledge, suggesting this type of tool complements, rather than substitutes for, underlying financial capability.
Working paper
Press coverage