Choosing Your Pond: Location Choices and Relative Income

We conducted a survey experiment of medical students who participated in the 2017 National Resident Matching Program (NRMP) before they submitted their preferences to the algorithm. Eligible medical students from 27 accredited medical schools in the US were invited to participate in an online survey.

The survey comprised of the following groups of questions:
1) Choice set: elicits the names of the two favorite programs that the individual is considering for her/his order ranking submission

2) Prior beliefs: elicits perceptions about the cost of living and the earnings rank in the cities where these two programs are located

3) Feedback: provides subjects with feedback related to their perceptions on both of these attributes for both locations. We randomize the value of this feedback in a non-deceptive way by randomizing the data source used to compute these statistics. For example, for feedback on earnings ranking we use either the American Community Survey or Current Population Survey; for cost of living the Regional Price Parity index or the Cost of Living index.

4) Posterior beliefs: Re-elicit perceptions about the cost of living and the earnings rank.

5) Elicit the individual’s expected rank submission (between the two programs).

6) Basic demographic questions: gender, age and relationship status (definition comparable to Luttmer (2005)), and number of children.

The baseline survey was closed before the deadline for rank order submissions to the NRMP. Respondents were paid a $10 Amazon gift card for completing the survey and submitted their university email for receipt and verification. The day after the submission deadline, we followed up directly with the respondents. In the follow-up survey we first asked the final rank order submitted to the NRMP, then elicited perceptions about cost of living and earnings rank again (to measure persistence in learning), and additional program characteristics not included in the baseline due to space and time constraints. The follow-up survey was closed the day before Match day. Respondents were offered a $5 Amazon gift card for completing the follow-up survey.

In addition to the main experiment, we conducted an auxiliary survey experiment on Amazon Mechanical Turk to test the robustness of our results and to rule out potential mechanisms. The survey was identical to the baseline survey described above with two differences: the survey was framed as a hypothetical choice of moving to one of two locations; at the end of the survey, we elicited perceptions on additional location characteristics.

2017-01-06

2017-10-02

First stage: Revision in beliefs (posterior - prior) for earnings ranking and cost of living measured in baseline survey.
Main outcome: Expected rank order submission from baseline survey. That is, a dummy variable that equals 1 if respondent expected to rank first program higher than the second one.

Dummy variable constructed from a 6-point likert scale from question asking "As of this moment: of the two programs discussed so far, which one do you expect to rank higher for the NRMP?" Where options range from 1-"Very likely [Program A]" to 6-"Very likely [Program B]". Expected rank equals one if choice between 1 and 3 (meaning preferring Program A). This definition is comparable with their choice measured in the follow-up survey listed below.

- Rank order submission from the follow-up survey: dummy variable indicating respondent chose Location A over Location B
- Intensity of expected rank from baseline survey (6 point scale).
- Expects to live happier life at Location A (robustness)
- First stage: Long-term revision in beliefs (posterior[follow-up survey]-prior) for earnings ranking and cost of living.

In this information-provision experiment, all individuals receive a signal about a particular variable (cost of living or relative income), but we randomize between one of two signals that they may receive. Then, we can compare the behavior of individuals who, by chance, received a higher signal about a specific variable (e.g., showing a signal that the cost of living is 10% higher than the U.S. average versus showing that same individual a signal that the cost of living is 5% higher than the U.S. average). To do this in a non-deceptive fashion, we computed the statistics shown to the subjects using two alternative data sources, both of them valid, and we cross-randomized which of the two sources were shown to each individual.

Immediately after respondents provided their prior beliefs on both earnings rank and cost of living, they were shown two messages: one page with statistics about the cost of living in the two cities being considered and a second page with statistics about the earnings rank in each of the two cities. The following message is a sample of the feedback page for cost of living: “Los Angeles- Long Beach-Anaheim, CA metro area is 17.0% more expensive than the U.S. average. The Champaign-Urbana, IL metro area is 6.6% cheaper than the U.S. average.” The following message is a sample of the feedback page for earnings rank: “With your individual annual earnings of $54,000, you would be richer than 57.9% of Los Angeles-Long Beach-Anaheim, CA’s population. With your individual annual earnings of $54,000, you would be richer than 60.3% of Champaign-Urbana, IL’s population.” In both of these feedback pages, individuals were asked to take a moment to review the information carefully and were alerted that the information was only going to be shown once. We did not allow respondents to continue to the next page until at least 10 seconds had elapsed.

The sources were randomized between individuals; that is, we used the same cost of living source for the two cities being considered by each individual, and the same earnings data source for the two cities. As a result, individuals were randomly assigned to one of four treatment groups. For cost of living estimates, the two sources used were the Regional Price Parity (RPP) data by the Bureau of Economic Analysis and the Cost of Living Index (COLI) data compiled by the Council for Community and Economic Research. For the earnings rank feedback, the two sources used were the American Community Survey (ACS) and the Current Population Survey (CPS), both conducted by the U.S. Census Bureau. Therefore, the four information treatment groups were: (RPP, ACS), (RPP, CPS), (COLI, ACS), (COLI, CPS).

Individuals were debriefed in the feedback messages on the name of the source of the information that they received.

Randomization done by survey platform (Qualtrics)

Individual

No

Individual level randomization. Same as number of observations below.

Main survey: 1,000 Auxiliary survey: 1,300

Equally distributed among 4 treatment groups, around 250 in each for main experiment (325 in auxiliary).

We propose that, when individuals are deciding where to live, they care about the position in the income distribution in their prospective location. To test this hypothesis, we develop a new methodology to estimate preferences over location characteristics that combines choice data, survey data and information-provision experiments. We implement this methodology with a sample of 1,080 senior medical students who participated in the National Resident Matching Program. These subjects were choosing the cities to which they would move and live in for the next five years. We provide suggestive evidence that individuals care about what their relative income would be in a prospective city, and that those preferences have substantial heterogeneity by relationship status. Using a separate subject pool, we replicate our findings and provide tests for the underlying mechanisms.

https://www.nber.org/papers/w23615