What is the most statistically efficient way to do off-policy optimization with batch data from bandit feedback? For log data generated by contextual bandit algorithms, we consider offline estimators for the expected reward from a counterfactual policy. Our estimators are shown to have lowest variance in a wide class of estimators, achieving variance reduction relative to standard estimators. We then apply our estimators to improve advertisement design by a major advertisement company. Consistent with the theoretical result, our estimators allow us to improve on the existing bandit algorithm with more statistical confidence compared to a state-of-theart benchmark.
Do new migration opportunities for rural households change the nature and extent of informal risk sharing? We experimentally document that randomly offering poor rural households subsidies to migrate leads to a 40% improvement in risk sharing in their villages. Our model of endogenous migration and risk sharing shows that risky and temporary migration opportunities can induce an improvement in risk sharing enabling profitable migration. Accounting for improved risk sharing, the migration experiment increased welfare by 12.9%. However, permanent declines in migration costs improve outside options for households and can lead to reductions in risk sharing. The short-run experimental results for migration subsidies can differ from the longer-run impacts of a policy that permanently subsidizes migration.
Do new migration opportunities for rural households change the nature and extent of informal risk sharing? We experimentally document that randomly offering poor rural households subsidies to migrate leads to a 40% improvement in risk sharing in their villages. We explain this finding using a model of endogenous migration and risk sharing. When migration is risky, the network can facilitate migration by insuring that risk, which in turn crowds-in risk sharing when new migration opportunities arise. We estimate the model and find that welfare gains from migration subsidies are 42% larger, compared with the welfare gains without spillovers, once we account for the changes in risk sharing. Our analysis illustrates that (a) ignoring the spillover effects on the network gives an incomplete picture of the welfare effects of migration, and (b) informal risk sharing may be an essential determinant of the takeup of new income-generating technologies.
Standard tests and condence sets in the moment inequality literature are not robust to model misspecifcation in the sense that they exhibit spurious precision when the identifed set is empty. This paper introduces tests and confidence sets that provide correct asymptotic inference for a pseudo-true parameter in such scenarios, and hence, do not suffer from spurious precision.
We document that an experimental intervention offering transport subsidies for poor rural households to migrate seasonally in Bangladesh improved risk sharing. A theoretical model of endogenous migration and risk sharing shows that the effect of subsidizing migration depends on the underlying economic environment. If migration is risky, a temporary subsidy can induce an improvement in risk sharing and enable profitable migration. We estimate the model and find that the migration experiment increased welfare by 12.9%. Counterfactual analysis suggests that a permanent, rather than temporary, decline in migration costs in the same environment would result in a reduction in risk sharing.
We document that an experimental intervention offering transport subsidies for poor rural households to migrate seasonally in Bangladesh improved risk sharing. A theoretical model of endogenous migration and risk sharing shows that the effect of subsidizing migration depends on the underlying economic environment. If migration is risky, a temporary subsidy can induce an improvement in risk sharing and enable profitable migration. We estimate the model and find that the migration experiment increased welfare by 12.9%. Counterfactual analysis suggests that a permanent, rather than temporary, decline in migration costs in the same environment would result in a reduction in risk sharing.
This paper is concerned with possible model misspecification in moment inequality models. Two issues are addressed. First, standard tests and confidence sets for the true parameter in the moment inequality literature are not robust to model misspecification in the sense that they exhibit spurious precision when the identified set is empty. This paper introduces tests and confidence sets that provide correct asymptotic inference for a pseudo-true parameter in such scenarios, and hence, do not suffer from spurious precision.
Second, specification tests have relatively low power against a range of misspecified models. Thus, failure to reject the null of correct specification does not necessarily provide evidence of correct specification. That is, model misspecification tests are subject to the problem that absence of evidence is not evidence of absence. This paper develops new diagnostics for model misspecification in moment inequality models that do not suffer from this problem.
Each year, more than two million U.S. households have an eviction case filed against them. Many cities have recently implemented policies aimed at reducing the number of evictions, motivated by research showing strong associations between being evicted and subsequent adverse economic outcomes. Yet it is difficult to determine to what extent those associations represent causal relationships, because eviction itself is likely to be a consequence of adverse life events. This paper addresses that challenge and offers new causal evidence on how eviction affects financial distress, residential mobility, and neighborhood quality. We collect the near-universe of Cook County court records over a period of seventeen years, and link these records to credit bureau and payday loans data. Using this data, we characterize the trajectory of financial strain in the run-up and aftermath of eviction court for both evicted and non-evicted households, finding high levels and striking increases in financial strain in the years before an eviction case is filed. Guided by this descriptive evidence, we employ two approaches to draw causal inference on the effect of eviction. The first takes advantage of the panel data through a difference-in-differences design. The second is an instrumental variables strategy, relying on the fact that court cases are randomly assigned to judges of varying leniency. We find that eviction negatively impacts credit access and durable consumption for several years. However, the effects are small relative to the financial strain experienced by both evicted and non-evicted tenants in the run-up to an eviction filing.
Standard tests and confidence sets in the moment inequality literature are not robust to model misspecification in the sense that they exhibit spurious precision when the identified set is empty. This paper introduces tests and confidence sets that provide correct asymptotic inference for a pseudo-true parameter in such scenarios, and hence, do not suffer from spurious precision.
Charities often send annual “thank you letters” to express gratitude to donors, but seek to defray these costs by inviting additional donations or engagement. However, the additional asks may backfire if potential donors see the thank you message as “insincere” or “manipulative.” We test this trade-off by conducting a field experiment in cooperation with a leading charity in India. We find that an explicit ask for additional donations or even a request to follow the organization on Facebook reduces giving. However, these effects are not only heterogeneous, but asymmetric by past giving behavior. Recent, frequent, and higher monetary value donors react negatively to additional asks by reducing giving, but lapsed, infrequent, and lower monetary value donors react positively by giving more. Our results highlight that findings based on purely cross-sectional experiments may offer incomplete insight. We estimate that differentially targeted ask messages based on past donation behavior, data readily available to charities, can increase donations overall by 6-11%.