This study examines whether conditional lottery incentives, where the probability of winning increases with continued participation, can motivate panelists to reengage in a randomly sampled, quarterly online survey. Respondents were randomly assigned to one of three groups: a control group (CG) receiving the standard & euro;10 incentive, treatment group 1 (T1) with an additional lottery offering ten Amazon vouchers worth & euro;500, and treatment group 2 (T2) introducing conditionality by linking the probability of winning to participation in successive waves. This design tests whether reengagement depends not only on incentive size but also on respondents' perceived ability to influence their odds of winning. Results show no statistically significant treatment effects. Neither the high-value lottery nor its conditional variant increased participation. Subgroup analyses reveal no systematic differences across prior participation, education, income, or country of origin. These findings confirm earlier evidence on the limited effectiveness of lotteries in survey research but extend it by testing conditionality as a motivational mechanism. Although conditional incentives strengthen the conceptual link between effort and reward, restoring agency alone appears insufficient to overcome deeper motivational barriers in quarterly online panels.
Computer-Assisted Recorded Interviewing captures audio recordings of survey interviews in real time, producing a distinctive form of unstructured paradata-question-level audio files that capture the question-answer sequences between the interviewer and respondent. These recordings offer insights into the survey response process, enabling a deeper understanding of how responses are formed and delivered. This paper explores the use of machine learning to automate the processing of recordings and to compare interviewer-respondent interactions in video interviews with those of in-person and telephone interviews in the production environment of a large national survey of US households. We first applied machine learning algorithms to transcribe the audio recordings and generate indicators describing interviewer-respondent interactions. We then assessed differences across the three modes in two interviewer behaviors indicative of adherence to standardized interviewing protocol (reading questions verbatim and maintaining question meaning) and in three interaction outcome measures indicative of respondent difficulty answering survey questions (number of the respondent's turns, duration of the respondent's first turn, and the total duration of turns with speech). We found significant differences in interviewer question reading behaviors and in the duration of the respondent's first turn across the three modes, but the differences were small. Additionally, the question itself emerged as the primary source of variability across all outcome measures. This paper demonstrates a viable method for using recordings to efficiently assess interviewer-respondent interactions for research purposes.
We address the problem of selecting a fixed-size sample from a finite population when the goal is to reduce the estimation error as much as possible. Our approach builds on the idea of choosing samples that are nearly optimal with respect to the values of the study variable, even though these values are not observed in advance. Instead, we rely on predictions derived from auxiliary information, for example, through statistical or machine-learning models, and construct sampling designs that are optimal for these predicted values. We show that this strategy is remarkably effective: even when the predictions are imperfect, the resulting designs tend to be very close to the unknown optimal design based on the true values. To understand why this happens, we study the geometric structure underlying the space of all feasible sampling designs and use this perspective to develop practical algorithms. Simulation experiments indicate that the proposed methodology achieves noticeably lower estimation error than classical unequal-probability sampling methods commonly used in the literature, including balanced procedures such as cube and classical fixed-size designs like conditional Poisson sampling.
Missing data poses fundamental challenges for causal discovery in observational studies, particularly when inferring directed acyclic graphs that represent underlying cause-and-effect relationships. Traditional approaches to handling missing data, such as complete-case analysis, can introduce biases and reduce the reliability of causal inferences. While multiple imputation via chained equations represents the current standard for handling missing data, its effectiveness for causal discovery is limited by both the absence of complete observations and the lack of principled methods for aggregating causal graphs across imputed datasets. This work introduces a novel unified framework that integrates causal structure learning directly into the imputation process. The key innovation lies in using preliminary adjacency information to estimate missingness propensity scores, which are then incorporated into the imputation model together with source indicator variables to generate more accurate estimates of missing values. The conditional independence test is implemented with source indicators included in the conditioning set, and by pooling test statistics rather than graph structures across multiple imputations, the proposed approach addresses the traditional challenge of graph aggregation while maintaining statistical rigor. Simulation studies demonstrate that the proposed framework achieves better performance compared to the traditional approach across all relevant metrics. The proposed framework is adapted to handle complex survey data by incorporating survey weights and survey design in both data imputation and conditional independence tests for causal discovery. When applied to childhood anthropometric survey data from Mali, the proposed approach produces results that better align with expected outcomes. By combining imputation and causal discovery in a unified process while accounting for survey design, it provides survey researchers with a practical tool for integrating and analyzing incomplete, complex survey data from multiple sources or time points.
We conducted a pre-registered between-subject randomized field experiment in the 2022 Survey of Consumer Finances, an interviewer-administered survey, to examine how different pre-paid incentives affect interview appointments, survey response rates, interviewer contact attempts needed to obtain an appointment or a survey response, time in the field, and item response rates. We assess the effects of: (i) increasing the total dollar value of the pre-paid incentive (from $5 to either $10 or $15), (ii) using two evenly-sized pre-paid incentives instead of one, controlling for the total dollar value (two payments of $5 instead of one $10), and (iii) using two unevenly sized pre-paid incentives, controlling for the total dollar value ($5 followed by $10 and vice versa, instead of one $15). All treatment groups had at least 1,200 households. Our evidence indicates that a single $15 pre-paid incentive increases survey response rates, without increasing time in the field or contact attempts, and without decreasing item response rates. Splitting the $15 into two payments increases time in the field without improving survey or item response rates, regardless of whether the first payment is larger or smaller than the second. Our results imply that the total incentive amount, delivered early in the field period, may be more important to individuals than the number of times they receive an incentive.
