Working Papers

Will Artificial Intelligence Get in the Way of Achieving Gender Equality?

with Daniel Carvajal and Siri Isaksson

The promise of generative AI to increase human productivity relies on developing skills to become proficient at it. There is reason to suspect that women and men use AI tools differently, which could result in productivity and payoff gaps in a labor market increasingly demanding knowledge in AI. Thus, it is important to understand if there are gender differences in AI-usage among current students. We conduct a survey at the Norwegian School of Economics collecting use and attitudes towards ChatGPT, a measure of AI proficiency, and responses to policies allowing or forbidding ChatGPT use. Three key findings emerge: first, female students report a significantly lower use of ChatGPT compared to their male counterparts. Second, male students are more skilled at writing successful prompts, even after accounting for higher ChatGPT usage. Third, imposing university bans on ChatGPT use widens the gender gap in intended use substantially. We provide insights into potential factors influencing the AI adoption gender gap and highlight the role of appropriate encouragement and policies in allowing female students to benefit from AI usage, thereby mitigating potential impacts on later labor market outcomes.

Bridging the Gender Gap in Access to STEM through In-Exam Stress Management

with Marcela Gomez-Ruiz

Why women underperform relative to men in high stakes exams while excelling under lower stakes remains a puzzle. Previous research suggests differential responses by gender to exam pressure as a plausible explanation. We examine the effects of a unique stress management intervention assigning applicants to a popular STEM program in Uruguay to in-exam stress reappraisal and meditation exercises. The exercises aim to reframe stress as something that can enhance performance rather than hinder it. We find that treated women leave fewer unanswered questions, resulting in a 0.25 SD boost in performance. As a result, 10% more women are admitted into the program.

Innocuous Exam Features? The Impact of Answer Placement on High-Stakes Test Performance and College Admissions

with Erika Povea

Multiple-choice tests are widely used across the world to allocate education slots and select job candidates. Albeit a seemingly innocuous feature, we provide evidence that the placement of the correct answer affects test outcomes. We leverage the random allocation of entrance exam booklets containing different answer distributions from nearly 100,000 applicants to Colombia’s top public university. We find that applicants are 5% less likely to answer correctly when the correct answer is in D, the last option of the choice set. Furthermore, encountering a one SD higher share of correct answers in D (6.7 pp above the mean) in mathematics, reduces overall performance by 0.01 standard deviations and lowers the chance of admission to applicants’ preferred major by 3%. Heterogeneity analyses using linked data from all higher education attendees in Colombia indicate that male applicants with a higher share of Ds in math are less likely to enroll in college altogether. Our results are driven by test-takers’ tendency to choose D less frequently than A, B and C, a pattern we also find using the worldwide PISA test and the entrance exam to Brazilian universities. Considering the lifelong implications of college access and initial field of study choice, our findings emphasize that subtle factors can exert a more significant influence than expected on test results, disproportionately affecting unlucky test takers.

Strategic Decisions Have “Major” Consequences: Gender Differences in College Major Choices

with Molly Hawkins

Does the way students navigate college admissions processes contribute to gender gaps in major choices? We study the major choices of applicants to Colombia’s largest public university who marginally miss the cutoff for their intended major. We find that women scoring just below the cutoff submit additional less-preferred majors and enroll sooner compared to similar men who intended the same major. Men instead reapply later for admission. As a result, a gender gap in potential salaries of 4.1-5.6% emerges at the cutoff. We find that differential responses to just failing generate gender gaps in enrollment, major choice, and earnings potential.

Failing to follow the rules: Can imprisonment lead to more imprisonment without more actual crime?

with David Harding, Jeffrey Morenoff, and Shawn D. Bushway
Revision requested at The Journal of Human Resources
Listed on SSRN’s Top Ten download list for: Corrections & Sentencing Law (February-April, 2023)

We find that people involved in low-level crime receiving a prison sentence are more likely than those with non-prison sentences to be re-imprisoned due to technical violations of parole, rather than due to new crimes. We identify the extent and cost of this incapacitation effect among individuals with similar criminal histories using exogenous variation in sentence type from discontinuities in Michigan Sentencing Guidelines. Technical violations disproportionately affect drug users and those first arrested as juveniles. Higher re-imprisonment adds one-quarter to the original sentence’s incapacitation days while only preventing low-severity crime, suggesting that prison is cost-ineffective for individuals on the margin.

How does relative performance feedback affect beliefs and academic decisions? Evidence from a field experiment

I conduct field and lab-in-the-field experiments, with students preparing for a college entrance exam, to identify how receiving relative performance feedback affects students’ beliefs, performance and academic decisions. Previous work in education shows that students’ behavior responds to correcting absolute performance beliefs but is unable to identify what type of adjustments students make that may be behind these decisions. To shed light on this question, I elicit beliefs from all students about relative performance in weekly practice tests and provide feedback to treated students about their quartile in the score distribution at a test preparation center in Colombia. Combining the panel dataset collected from the experiment with administrative data, I show that low-performing students become discouraged by feedback. On average, they reduce investments and effort in practice tests, which leads them to the decision of not taking the exam they are preparing for. However, their beliefs are not different from low-performers in the control group, which suggests that there is a mismatch between incentivized beliefs and beliefs revealed by observed behavior. Female and male students become discouraged in different ways, with males reducing investments and females keeping effort up but deciding to not take the college entrance exam. Overall, my results shed light on the potential discouragement effects of informational interventions on students with low academic performance.

AEA RCT Registry: AEARCTR-0006970
Experimental Instructions

The Value of a Signal: Information Processing among Students Outside the Lab

Many decisions in education require students to correctly assess their relative performance and process informational signals. While we know from laboratory studies that individuals process information in a biased way, we do not know enough about how information processing takes place in real-life settings. This paper provides the first estimates of information processing in the setting of students preparing for a high-stakes college entrance exam, following the protocols used in lab studies. I document that updating does not differ from the behavior observed in the lab as students update conservatively relative to the Bayesian benchmark and their updates do not differ statistically in response to a positive or a negative relative performance signal. The value of the relative performance signal is highest for students whose absolute score is in the middle of the distribution, where it is harder to assess whether one’s score is above or below the median. Female students’ beliefs are more responsive to the signal than male students’ beliefs, which seems to be explained by more underconfident performance priors among women. However, receiving the signal does not affect an external measure of confidence in gaining admission to the university they are applying for. Overall, my results suggest that providing a relative performance signal can help students with high uncertainty better assess their relative performance.

AEA RCT Registry: AEARCTR-0006970
Belief elicitation protocol

Willingness to Pay for Formal Job Attributes: A Discrete Choice Experiment In Colombian Mom-and-Dad Stores

with Amalia Rodríguez

Informal workers in developing countries are vulnerable to economic shocks since their jobs lack health and retirement benefits. Using a discrete choice experiment on a random sample of 2,000 ”mom-and-dad” store owners and employees in Colombia, we estimate their willingness to pay (WTP) for formal jobs benefits, namely health and retirement plans. We find that, on average, workers are willing to forego up to 28% and 37% of their earnings to access formal health and retirement benefits, respectively. In contrast to previous research, which suggests that workers’ low WTP for benefits induces them to become informal under expansions of free or subsidized health insurance programs, our WTP estimations suggest that workers actually highly value these benefits.