Private tutoring centers have increasingly become a central feature of American K--12 education, yet there is little evidence on whether access to them improves student outcomes. Using business records on tutoring establishments linked to administrative data covering over five million public school students in Texas, this paper estimates the short- and long-run impacts of private tutoring access. Exploiting variation from the entry and exit of tutoring centers near students' high school campus, I show that tutoring access has no effect on state standardized test scores or high school graduation, but raises college admission exam scores, shifts students toward higher-quality and out-of-state colleges, and increases their early-career earnings. These gains are entirely concentrated among students who are not economically disadvantaged. An original survey of Texas parents confirms that tutoring participation responds to nearby centers and documents wide gaps between advantaged and disadvantaged families in both tutoring use and spending. The rapid expansion of the private tutoring industry has thus widened socioeconomic gaps in both educational investment and students' long-run outcomes.

We study how large language model (LLM) based recommendations affect evaluations of uncertainty. We elicit certainty equivalents for (i) lotteries that share the same reduced-form winning probability but vary in cognitive demands because they are compound, and (ii) an Ellsberg-style ambiguous lottery. Across all lottery types, valuations from participants who receive LLM-based recommendations before each valuation exhibit second-order stochastic dominance relative to valuations from participants who do not. Hence, LLM-based recommendations leave mean valuations unchanged but compress the distribution, reducing the incidence of extreme valuations. Consistent with this pattern, LLM-based recommendations reduce the intensity of risk, complexity, and ambiguity aversion, while leaving the fraction of participants classified as averse unchanged.

This paper studies how households with school age children respond to China's nationwide ban on private academic tutoring. Using a nationally representative household panel on education expenditures and a difference-in-differences design, we show that the ban led to a modest decline in participation in and spending on private commercial tutoring, with an accompanying large increase in out-of-pocket parental spending on in-school education services. As a result, the share of household consumption spent on education rose, indicating reallocation across educational inputs rather than a reduction in total education investment. The behavioral responses vary across household types. Private academic tutoring spending decreased among households below the top income quartile but increased substantially among high-income households. Correspondingly, point estimates suggest that students from high-income households were more likely to rank at the top of their classes after the policy. Meanwhile, increases in out-of-pocket in-school education spending were similar across income groups. Overall, our findings highlight substitution toward alternative forms of academic support under the ban and suggest that the policy may have widened inequality in access to effective educational inputs.

We present the first large-scale empirical evidence on the effects of adopting a dynamic matching mechanism, in replacement of the Immediate Acceptance (IA) mechanism, on matching stability in college admissions in China. In 2007, the Inner Mongolia Autonomous Region introduced the "Real-time Dynamic Mechanism", which allowed college applicants to change their college choices as many times as they want during a restricted time interval while seeing their tentative admission outcome when they made each choice. Using administrative data on test scores and admission outcomes of the universe of National College Entrance Exam (NCEE) takers from 2005 to 2011, we construct measures of justified envy, an indicator of matching stability. We use a generalized difference-in-differences framework and, in contradiction to the theoretical and experimental predictions from previous studies, find no evidence that the real-time dynamic mechanism improved matching stability in the first four years after its implementation. Our findings suggest that the real-time dynamic mechanism is much less effective in eliminating justified envy than the parallel mechanism, a hybrid of IA and the Deferred Acceptance (DA) mechanism, which is now widely adopted in other provinces in China.

The past few years have seen a shift in many universities' admission policies from test-required to either test-optional or test-blind. This paper uses laboratory experiments to examine students' reporting behavior given their application package and the school's interpretation of non-reported standardized test scores. We find that voluntary disclosure is incomplete and selective, supporting both the incentive of partial unraveling (students with higher scores are more likely to report) and the incentive of reverse unraveling (students facing a better school's interpretation of non-reporting are less likely to report). Subjects exhibit some ability to learn about the hidden school interpretation, though their learning is imperfect. Using a structural model of student reporting behavior, we simulate the potential tradeoff between academic preparedness and diversity in a school's admission cohort. We find that, if students have perfect information about the school's interpretation of non-reporting,  test-blind is the worst and test-required is the best in both dimensions, while test-optional lies between the two extremes. When students do not have perfect information, some test-optional policies can generate more diversity than test-required, because some students with better observable attributes may underestimate the penalty on their non-reporting. This allows the school to admit more students that have worse observable attributes but report. The results are robust to a variety of extensions, including when schools have access to alternative signals of academic ability and standardized test score is a noisy but sufficiently informative measure of student ability.

We study how voluntary disclosure responds to a correlated public signal when receivers are partially naive. We develop a framework in which the public signal shifts naive receivers' interpretation of silence, and derive a closed-form disclosure threshold nesting classical unraveling. A favorable public signal can crowd out disclosure, leaving receivers with less information than absent the signal. We test this in a laboratory experiment varying the correlation between private state and public signal. Disclosure falls as the public signal becomes more favorable, and the decline is amplified by stronger correlation. Receiver guesses upon non-disclosure rise correspondingly, departing from the rational benchmark toward the public signal. Strategic silence can thus crowd out the public signal itself, and transparency can be lower with it than without. The effect is non-monotonic in the correlation: strongest when the signal is informative enough to anchor beliefs but too noisy to substitute for disclosure.

We experimentally study how uncertainty about the true state shapes reporting behavior when payoffs increase in the reported outcomes. In a chip-draw experiment, subjects report the number on a randomly drawn chip. In one treatment they observe the number directly, while in the other they observe only a signal that restricts the set of possible values without identifying the true number. We find that imperfect information substantially reshapes the distribution of reports: when the true state is uncertain, the majority of subjects report the highest value consistent with their signal. At the same time, imperfect information does not increase the frequency of income-maximizing reports. These results indicate that imperfect information can systematically change the distribution of reported outcomes through a mechanism that differs from direct misreporting of a known state.

We use laboratory experiments to examine strategic disclosure and deception in a sender–receiver game modeling the data economy. Firms (senders) are privately informed of their quality and decide how to disclose it to consumers (receivers). They may signal higher quality through exaggeration, but face penalties if misrepresentation is detected. First, we vary penalty severity and returns to perceived quality, and find that exaggeration declines with stronger penalties and lower returns. We then compare two verification regimes: exogenous detection, which mimics probabilistic data breaches, and endogenous detection, in which consumers (receivers) incur a cost to verify the firm’s claim mimicking privacy audits. Contrary to equilibrium predictions, exaggeration is more prevalent under exogenous detection than under endogenous detection. This is driven by the finding that some receivers function as watchdogs and have strong desire to detect regardless of detection cost.