We introduce a model of two-sided statistical discrimination in which worker and firm beliefs are complementary. Firms try to infer whether workers have made investments required for them to be productive, and simultaneously, workers try to deduce whether firms have made investments necessary for them to thrive. When multiple equilibria exist, group differences are sustained by both sides of the interaction – workers and firms. Strategic complementarity complicates both empirical analysis designed to detect discrimination and policy meant to alleviate it. Affirmative action is much less effective than in traditional statistical discrimination models. More generally, we demonstrate the futility of policies that are designed to correct gender and racial disparities but do not address both sides of the coordination problem. We propose a two-sided version of “investment insurance” – a policy in which the government (after observing a noisy version of the employer’s signal) offers to hire any worker who it believes to be qualified and whom the employers does not offer a job – and show that it (weakly) dominates any alternative. The paper concludes by proposing a way to identify statistical discrimination by employers when beliefs are complements.
This study examines the impact on student achievement of high-dosage reading tutoring for middle school students in New York City Public Schools, using a school-level randomized field experiment. Across three years, schools offered at least 130 hours of 4-on-1 tutoring based on a guided reading model, which consisted of 1-on-1 read alouds, independent reading, vocabulary review, and group discussion. We show that, at the mean, tutoring has a positive and significant effect on school attendance, a positive, but insignificant, effect on English Language Arts (ELA) state test scores and no effect on math state test scores. There is important heterogeneity by race. For black students, our treatment increased attendance by 2.0 percentage points (control mean 92.4 percent) and ELA scores by 0.09 standard deviations per year – two times larger than the effect of the Promise Academy Middle School in the Harlem Children’s Zone and KIPP Charter Middle Schools on reading achievement. For Hispanic students, the treatment effect is 0.8 percentage points on attendance (control mean 92.0 percent) and 0.01 standard deviations per year on ELA scores. We argue that the difference between the effectiveness of tutoring for black and Hispanic students is best explained by the average tutor characteristics at the schools they attend.
This study examines the impact on student achievement of implementing management training for principals in traditional public schools in Houston, Texas, using a school-level randomized field experiment. Across two years, principals were provided 300 hours of training on lesson planning, data-driven instruction, and teacher observation and coaching. The findings show that offering management training to principals significantly increases student achievement in all subjects in year one and has an insignificant effect in year two. We argue that the results in year two are driven by principal turnover, coupled with the cumulative nature of the training. Schools with principals who are predicted to remain in their positions for both years of the experiment demonstrate large treatment effects in both years – particularly those with principals who are also predicted to implement the training with high fidelity – while those with principals that are predicted to leave have statistically insignificant effects in each year of treatment.
We estimate the impact of charter schools on early-life labor market outcomes using administrative data from Texas. We find that, at the mean, charter schools have no impact on test scores and a negative impact on earnings. No Excuses charter schools increase test scores and four-year college enrollment, but have a small and statistically insignificant impact on earnings, while other types of charter schools decrease test scores, four-year college enrollment, and earnings. Moving to school-level estimates, we find that charter schools that decrease test scores also tend to decrease earnings, while charter schools that increase test scores have no discernible impact on earnings. In contrast, high school graduation effects are predictive of earnings effects throughout the distribution of school quality. The paper concludes with a speculative discussion of what might explain our set of facts.
This article describes a randomized field experiment in which parents were provided financial incentives to engage in behaviors designed to increase early childhood cognitive and executive function skills through a parent academy. Parents were rewarded for attendance at early childhood sessions, completing homework assignments with their children, and for their child’s demonstration of mastery on interim assessments. This intervention had large and statistically significant positive impacts on both cognitive and non-cognitive test scores of Hispanics and Whites, but no impact on Blacks. These differential outcomes across races are not attributable to differences in observable characteristics (e.g. family size, mother’s age, mother’s education) or to the intensity of engagement with the program. Children with above median (pre-treatment) non cognitive scores accrue the most benefits from treatment.
We introduce a model in which agents observe signals about the state of the world, some of which are open to interpretation. Our decision makers use Bayes’ rule in an iterative way: ﬁrst to interpret each signal and then to form a posterior on the se-quence of interpreted signals. This ‘double updating’ leads to conﬁrmation bias and can lead agents who observe the same information to polarize: the distance between their beliefs can grow after observing a common sequence of signals. Such updating is approximately optimal if agents must interpret ambiguous signals and suﬃciently discount the future. If they are very patient but can only store interpretations of ambiguous signals, then a time-varying random interpretation rule (still double-updating) is approximately optimal. In a continuous (normally distributed) version of the model, we show that posterior beliefs never lose the inﬂuence of the prior and still always converge, but always converge to something that is inﬂuenced by the prior and early signals and so is wrong with probability one. Beliefs become arbitrarily accurate as the signal accuracy increases, but are always biased. We explore the model in an on-line experiment in which individuals interpret research summaries about climate change and the death penalty and report beliefs. Consistent with the model, not only is there a signiﬁcant relationship between an individual’s prior and their interpretation of the summaries; but more than half of the subjects exhibit polarizing behavior-shifting their beliefs further from the average belief after seeing the same summaries as all other subjects.
This paper describes randomized ﬁeld experiments in eighty-four urban public schools in two cities designed to understand the impact of aligned incentives on student achievement. In Washington DC, incentives were “horizontal” – provided to one agent (students) for various inputs in the education production function (i.e. attendance, behavior, interim assessments,homework, and uniforms). In Houston, TX, incentives were “vertical” – provided to multiple agents (parents, teachers, and students) for a single input (math objectives). On outcomes for which we provided direct incentives, there were large and statistically signiﬁcant effects from both treatments. Horizontal incentives led to increases in math and reading test scores. Vertical incentives increased math achievement, but resulted in decreased reading, science, and social studies test scores. We argue that the data is consistent with agents perceiving academic achievement in various subjects as substitutes, not complements, in education production.