We evaluate how nonresponse affects conclusions drawn from survey data and consider how researchers can reliably test and correct for nonresponse bias. To do so, we examine a survey on labor market conditions during the COVID-19 pandemic that used randomly assigned financial incentives to encourage participation. We link the survey data to administrative data sources, allowing us to observe a ground truth for participants and nonparticipants. We find evidence of large nonresponse bias, even after correcting for observable differences between participants and nonparticipants. We apply a range of existing methods that account for nonresponse bias due to unobserved differences, including worst-case bounds, bounds that incorporate monotonicity assumptions, and approaches based on parametric and nonparametric selection models. These methods produce bounds (or point estimates) that are either too wide to be useful or far from the ground truth. We show how these shortcomings can be addressed by modeling how nonparticipation can be both active (declining to participate) and passive (not seeing the survey invitation). The model makes use of variation from the randomly assigned financial incentives, as well as the timing of reminder emails. Applying the model to our data produces bounds (or point estimates) that are narrower and closer to the ground truth than the other methods.

No abstract

We use data from a serological study that experimentally varied financial incentives for participation to detect and characterize selection bias. Participants are from neighborhoods with substantially lower COVID-19 risks. Existing methods to account for the resulting selection bias produce wide bounds or estimates that are inconsistent with the population. One explanation for these inconsistent estimates is that the underlying methods presume a single dimension of unobserved heterogeneity. The data suggest that there are two types of nonparticipants with opposing selection patterns. Allowing for these different types may lead to better accounting for selection bias.

At least one co-author has disclosed additional relationships of potential relevance for this research. Further information is available online at http://www.nber.org/papers/w30880 NBER working papers are circulated for discussion and comment purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies official NBER publications.