1) Economists calculate the value of a statistical life (VSL) by looking at fatal occupational injuries. The idea is to calculate how much compensation is needed to convince someone to take high fatal risk jobs.

• Here we are interested in finding how much someone values their own life. But we are using a proxy that is how much extra risk they are willing to take for extra compensation (or how much compensation is needed for extra risk). What kind of data problem does this resemble?
  • Note that it is ok to use proxies as long as we understand the ramifications and remember that they are not the ideal variable we would want.
• Someone with low initial wealth may be willing to take more risks for little extra compensation as the marginal value of money is higher for them (like taking care of their family) whereas someone with high initial wealth would need higher compensation to agree to take such risk. This could lead to lower-income individuals being selected for higher-risk jobs. What kind of data problem would this create in calculating VSL using fatal occupational injuries?

2) Give an example of a Prediction vs Causation problem. In other words, we are interested in delta Y (change in Y) but the algorithm (or intervention) is good at identifying (or selecting based on) Y itself.

• Remember the example in class: salad discount aimed to get people to buy more salad (increase delta Y) but instead targeted people who were already buying lots of salad (high Y) resulting in not increase in purchases but decrease in revenue.

3) Design an experiment for the example you gave in the previous question. In order to find the change in Y following a certain intervention, we need counterfactuals (like treatment and control groups). If an experiment is difficult to design for that example, think about how you can find a natural experiment in real life.

4) In class, we compared the performance of a judge and ML algorithm based on what a judge cared about in her utility function when releasing a defendant. Let’s say the government is interested in building a new community center and mall in a residential area. To do so, we would have to relocate some residents from the neighborhood.

• Come up with inputs for a utility function of the government with expected costs and expected benefits of building this community center.
• What kinds of data do we need to collect and what are some sampling and labeling errors we might encounter?

5) There is a debate about whether the Payment Protection Program loans went to businesses in need and how much it helped them survive the pandemic. Do you think the government could use Machine Learning to better target urgent aids in the future? In what ways, this is a problem of prediction, and in what ways it is a problem of causation?