bumrah in a moustache and glasses with fake nose now playing for england

Current draft of the paper can be found here: alexanderclyde.com/Documents/JM...

In the buyer-seller story, the seller is eventually unable to incentivize any type to buy any good at a worthwhile price. Splitting across dimensions would have no effect if the agent is fully rational. This result demonstrates that narrow inference can be quantitatively significant. (15/15)

Finally, I explore what happens when the number of action dimensions grows large. In cases like the human capital story, the firm can eventually get all types to invest in skills at vanishing cost. (14/15)

These results have relevance for whether the principal wants to present decisions jointly or separately. They want to salami-slice decisions in the human capital story and bundle decisions in the buyer-seller story. (13/15)

Narrow Inference leads the agent to over-estimate the marginal effect of buying either good on the total price. For any prices, there are then fewer types who want to buy. This leaves the principal worse-off and wanting to implement that fewer agents buy either good. (12/15)

Conversely, consider the case where the principal is a seller and the agent a buyer. The principal is selling goods, such as computer and software, that have a jointly determined price. Here the actions have immediate benefits to the buyer, but some cost to the principal from production. (11/15)

In cases like the human capital-worker story, where actions have some predictable (non-wage) cost to the workers for some ultimate benefit in terms of productivity to the firm, the principal benefits and implements that on every dimension more types choose the investments. (10/15)

It also allows us to make easy comparisons to the optimal mechanism when the agent is fully rational. Using the characterization, I provide a taxonomy of cases demonstrating when the principal benefits and loses out from facing an agent who does narrow inference. (9/15)

In the paper, I consider this problem in a more general setting. I show how we can characterize the principal’s optimal incentive mechanism when the agent makes narrow inference. This result is useful because it provides a recipe for us to solve for their optimal mechanism in examples. (8/15)

This can lead to a distortion in the worker's beliefs resembling the classic 'omitted variable bias' from introductory econometrics. For each action, narrow inference fails to control for the other action dimensions. The agent then overestimates the effect of any particular skill on wages. (7/15)

In line with this, under narrow inference workers form beliefs about the wage gains from learning each skill separately. They compare the average wages of workers in the firm who have the skill in question vs those who do not. This is illustrated by the following DAG. (6/15)

However, there is some work in experimental and behavioural economics suggesting that people (1) neglect correlation in data they use to form beliefs when there are many dimensions/variables (2) bracket choices into smaller sub-problems. (5/15)

Inferring how human capital affects wages can require workers to have a sophisticated causal understanding. It also requires that workers understand how their different choices affect wages jointly as one big decision problem. (4/15)

An example to illustrate the idea: consider a firm designing its wage structure. They face a worker who has to choose whether to make human capital investments in technical skills and/or managerial skills. Different types of workers sort into investing in different skill combinations. (3/15)

Our understanding of the incentives we face is fundamentally about what we think the consequences of our actions are for outcomes we care about. Therefore, accounting for limited causal understanding is a first order concern if we’re designing incentives. (2/15)