I am an Associate Professor of Economics and Education at Teachers College, Columbia University and Co-Chair of the Education Technology Initiative at Abdul Latif Jameel Poverty Action Lab (J-PAL) at MIT.
With generous support from Schmidt Futures, Citadel, and the Gates Foundation, I have started a research lab, The Learning Collider. Our lab partners with large-scale technology companies and organizations collecting high-dimensional data to open up their data to researchers and to innovate, test, and scale interventions that promote social mobility. We currently focus on education, employment, and housing, with a particular attention to applications of machine learning, algorithmic fairness, and related tools. We are fortunate to work with some of the largest edtech, hiring, and housing platforms in the world.
Please feel free to ask me about our work and partnerships.
Email: bergman [at] tc [dot] columbia [dot] edu Phone: (212) 678-3932 Office: 417 Thorndike Curriculum Vitae (pdf)
Over the last 40 years, efforts to desegregate schools have largely been undone and intra-district programs have limited scope to stem the resulting rise in segregation. This is the first paper to study the short-run and long-run impacts of an inter-district desegregation program on the minority students given an opportunity to transfer to majority-white school districts. Students who are given the opportunity to transfer districts attend schools that are 73 percentage points more white than schools they would have attended. Transferring students have higher test scores, and, over the longer run, an increase in college enrollment by 8 percentage points. At the same time, there is an increase in special education classification and arrests, which are largely for non-violent offenses. Both the benefits and the risks of the desegregation program accrue to male students.
Low-income families in the United States tend to live in neighborhoods that offer limited opportunities for upward income mobility.
One potential explanation for this pattern is that lowincome families prefer such neighborhoods for other reasons,
such as affordability or proximity
to family and jobs. An alternative explanation is that families do not move to high-opportunity
areas because of barriers that prevent them from making such moves. We test between these
two explanations using a randomized controlled trial with housing voucher recipients in Seattle
and King County. We provided services to reduce barriers to moving to high-upward-mobility
neighborhoods: customized search assistance, landlord engagement, and short-term financial assistance.
The intervention increased the fraction of families who moved to high-upward-mobility
areas from 14% in the control group to 54% in the treatment group. Families induced to move to
higher opportunity areas by the treatment do not make sacrifices on
other dimensions of neighborhood quality and report much higher levels of neighborhood satisfaction. These findings
imply that most low-income families do not have a strong preference to stay in low-opportunity
areas; instead, barriers in the housing search process are a central driver of residential
segregation by income. Interviews with families reveal that the capacity to address each family's needs
in a specific manner, from emotional support to brokering with landlords to financial assistance, was critical to the program's success. Using quasi-experimental analyses and comparisons to
other studies, we show that more standardized policies--increasing voucher payment standards
in high-opportunity areas or informational interventions--have much smaller impacts.
We conclude that redesigning affordable housing policies to provide customized assistance in housing
search could reduce residential segregation and increase upward mobility substantially.
Abstract: School choice may allow schools to "cream skim"
students perceived as easier to educate. To test this,
we sent emails from fictitious parents to 6,452 schools
in 29 states and Washington, D.C. The fictitious parent
asked whether any student is eligible to apply to the
school and how to apply. Each email signaled a randomly
assigned attribute of the child. We find that schools are
less likely to respond to inquiries from students with poor
behavior, low achievement, or a special need. Lower response
rates to students with a potentially significant special
need are driven by charter schools. Otherwise, these
results hold for traditional public schools in areas
of school choice and high-value added schools.
In looking for the best workers over time, firms must balance exploitation
(selecting from groups with proven track records) with exploration
(selecting from under-represented groups to learn about quality).
Yet modern hiring algorithms, based on "supervised learning"
approaches, are designed solely for exploitation. In this paper,
we view hiring as a contextual bandit problem and build a resume
screening algorithm that values exploration by evaluating candidates
according to their upside potential. Using data from professional
services recruiting within a Fortune 500 firm, we show that this
approach improves both the quality (as measured by eventual offer
and acceptance rates) and the diversity of candidates selected for
an interview relative to the firm's existing practices. The same
is not true for traditional supervised learning based algorithms,
which improve quality but select far fewer minority applicants.
In an extension, we show that exploration-based algorithms are
also able to learn more effectively about simulated changes in
applicant quality over time. Together, our results highlight the
importance of incorporating exploration in developing hiring
algorithms that are potentially both more efficient and equitable.
Schools closed extensively during the COVID-19 pandemic and occur in other settings,
such as teacher strikes and natural disasters. This paper provides some of the
first experimental evidence on strategies to minimize learning loss when schools close.
We run a randomized trial of low-technology interventions - SMS messages and phone
calls – with parents to support their child. The combined treatment cost-effectively
improves learning by 0.12 standard deviations. We develop remote assessment innovations,
which show robust learning outcomes. Our findings have immediate policy
relevance and long-run implications for the role of technology and parents as partial
educational substitutes when schooling is disrupted.
Abstract: To investigate the role of information frictions in low-income families' housing choices, we experimentally varied the availability of school quality information on a nationwide website of housing listings for families with housing vouchers. We use this variation, together with detailed data on families' search behaviors and location choices, to estimate a model of housing search and neighborhood choice that incorporates imperfect information and potentially biased beliefs. Having data from both the treatment and control groups allows us to estimate families' prior beliefs about school quality and each group's apparent valuation of school quality. We find that imperfect information and biased beliefs about school quality cause low-income families to live in neighborhoods with lower-performing, more segregated schools, and that ignoring this information problem would lead to biased estimates of families' valuation of school quality. Providing school quality information causes families to choose neighborhoods with schools that have 1.5 percentage point higher proficiency rate on state exams. Families not only observe school quality with noise, but systematically underestimate school quality conditional on neighborhood characteristics. If we had assumed full information, we would have estimated that the control group valued school quality relative to their commute downtown by less than half that of the treatment group.
Using Predictive Analytics to Track Students: Evidence from a Seven-College Experiment (Just updated)
with Elizabeth Kopko and Julio Rodriguez
Tracking is widespread in U.S. education. In post-secondary education alone, at least 71% of colleges use a test to track students. However, there are concerns that the most frequently used college placement exams lack validity and reliability, and unnecessarily place students from under-represented groups into remedial courses. While recent research has shown that tracking can have positive effects on student learning, inaccurate placement has consequences: students face misaligned curricula and must pay tuition for remedial courses that do not bear credits toward graduation. We develop an alternative system to place students that uses predictive analytics to combine multiple measures into a placement instrument. Compared to colleges’ existing placement tests, the algorithm is more predictive of future performance. We then conduct an experiment across seven colleges to evaluate the algorithm’s effects on students. Placement rates into college-level courses increased substantially without reducing pass rates. Adjusting for multiple testing, algorithmic placement generally, though not always, narrowed gaps in college placement rates and remedial course taking across demographic groups. A detailed cost analysis shows that the algorithmic placement system is socially efficient: it saves costs for students while increasing college credits earned, which more than offsets increased costs for colleges. Costs could be reduced with improved data digitization, as opposed to entering data by hand.
(Accepted, Journal of Political Economy)
Published and Forthcoming Papers
This paper studies information frictions between parents and their children, and how these affect
human capital investments. I provide detailed, biweekly information to a random sample of
parents about their child's missed assignments and grades and find parents have upwardly-biased
beliefs about their child's effort. Providing additional information attenuates this bias and
improves student achievement. Using data from the experiment, I then estimate a persuasion
game between parents and their children that shows the treatment effect is due to a combination
of more accurate beliefs and reduced monitoring costs. The experimental results and policy
simulations from the model demonstrate that improving the quality of
school reporting or providing frequent information to parents about their child's effort
in school can produce gains in achievement at a low cost.
(Forthcoming, Journal of Labor Economics)
Truancy predicts many risky behaviors and adverse outcomes. We use detailed
administrative data to construct social networks based on students who miss
class together. We simulate these networks to show that certain students
systematically coordinate their absences in the observed data. Leveraging
a parent-information intervention on student absences, we find spillover
effects from treated students onto peers in their network; excluding
these effects understates the intervention's cost effectiveness by 19%.
We show there is potential to use these networks to implement costly
interventions more efficiently. We develop an algorithm that incorporates
spillovers and treatment-effect heterogeneity identified by
machine-learning techniques to target interventions more efficiently given a budget constraint.
(Organizational Behavior and Human Decision Processes)
We conduct a field experiment to understand how enrollment defaults affect the take up and
impact of an education technology designed to help parents improve student achievement. The
standard strategy schools use to introduce this system to parents--online signup--induces negligible
adoption. Simplifying the enrollment process modestly increases adoption, primarily among parents of
higher-performing students. Automatically enrolling parents dramatically increases adoption.
Automatic enrollment significantly and meaningfully improves student achievement. Survey results
suggest that automatic enrollment is uncommon, and that it may be uncommon because its impact is
unanticipated by policymakers. Surveyed superintendents, principals, and family engagement
coordinators overestimate the take-up rate of the standard condition by 38 percentage points and
underestimate the take-up rate of automatic enrollment by 31 percentage points. After learning the
actual take-up rates under each enrollment condition, there is a corresponding 140% increase in the
willingness to pay for the technology when shifting implementation from opt-in enrollment to opt-out
(Journal of Policy Analysis and Management)
There is increasing evidence that tax credits for college do not affect college enrollment. This may be because prospective students do not know about tax benefits for credits or because the design of tax credits is not conducive to affecting educational outcomes. We focus on changing the salience of tax benefits by providing information about tax benefits for college using a sample of over 1 million students or prospective students in Texas. We sent emails and letters to students that described tax benefits for college and tracked college outcomes. For all three of our samples---rising high school seniors, already enrolled students, and students who had previously applied to college but were not currently enrolled---information about tax benefits for college did not affect enrollment or reenrollment. We test whether effects vary according to information frames and found that no treatment arms changed student outcomes. We conclude that salience is not the primary reason that tax credits for college do not affect enrollment.
(Journal of Human Resources)
We partnered a low-cost communication technology with school information
systems to automate the gathering and provision of information on students'
academic progress to parents of middle and high school students. We sent
weekly, automated alerts to parents about their child's missed assignments,
grades, and class absences. The alerts reduced course failures by 28%,
increased class attendance by 12\%, and increased student retention,
though there was no impact on state test scores. There were larger
effects for below-median GPA students and high school students. We
sent over 32,000 messages at a variable cost of $63.
(Economics of Education Review)
Abstract: This paper uses school-district
data and a regression discontinuity design to study the
effects of making teachers' value-added ratings available
to the public and searchable by name. We find that classroom
compositions change as a result of this new information.
In particular, high-scoring students sort into the classrooms of published,
high-value added teachers. This sorting occurs when there is within school-grade
variation in teachers' value added.
Parents are one of the most powerful determinants of a child's education outcomes. However, behavioral and informational barriers impede parents' engagement with their children. Parenting is complex and limited cognitive bandwidth steers parents' attention away from education investments with long-run returns and toward tasks with immediate returns. Monitoring children is difficult because school-to-parent communication is poor, and parents have inflated perceptions of their child's performance. Poverty exacerbates these problems. This paper shows how the provision of timely, actionable information to parents can attenuate these barriers and improve parental engagement from kindergarten through high school. In particular, providing this information via text message can improve education outcomes at low cost.
As schools are making significant investments in education technologies
it is important to assess whether various products are adopted by their
end users and whether they are effective as used. This paper studies
the adoption and ability to promote usage of one type of technology
that is increasingly ubiquitous: school-to-parent communication technologies.
Analyzing usage data from a Learning Management System across several hundred
schools and then conducting a two-stage experiment across 59 schools to nudge
the use of this technology by families, I find that 57% of families ever use
it and adoption correlates strongly with measures of income and student achievement.
While a simple nudge increases usage and modestly improves student achievement,
without more significant intervention to encourage usage by disadvantaged families,
these technologies may exacerbate gaps in information access across income and performance levels.
This paper studies the ability to foster parent skills and resolve information problems as a means to improving student achievement.
We conducted a three-arm randomized control trial in which community-based organizations provided
regular information to families about their child's academic progress in one arm and supplemented
this with home visits on skills-based information in a separate arm. Math and English test scores
improved for the treatment arm with home visits. There are large effects on retention for both
groups during the year, though learning gains tend to accrue for students with average-and-above
baseline performance and students at the lower-end of the distribution appear marginally retained.
J-PAL Ed Tech
Columbia Committee on the Economics of Education
Email: bergman [at] tc [dot] columbia [dot] edu
Phone: (212) 678-3932
Office: 417 Thorndike
Curriculum Vitae (pdf)