Journal of Survey Statistics and Methodology最新文献

Propensity score (PS) methods, including inverse probability of treatment weighting (IPTW) analysis, are increasingly applied to complex survey data in geriatric studies to infer causal effects. However, the comparative effectiveness of various PS estimation methods, particularly novel machine learning algorithms, has not been thoroughly explored when complex survey data are involved. We conducted a comprehensive simulation study to compare the following six PS estimation methods in IPTW analysis: Logistic Regression, Covariate Balancing Propensity Score, Generalized Boosted Model, Classification and Regression Tree, Random Forest (RF), and Super Learner. We considered 12 scenarios with varying treatment effects, degrees of non-linearity and non-additivity in the associations between covariates and the exposure, and levels of PS overlap. The performance of these six methods was assessed in terms of mean relative bias, root mean square error, and coverage probability. The results showed a similar performance across all methods when PS overlap was strong. However, RF consistently outperformed the other methods when PS overlap was not strong and under non-additive and non-linear scenarios. The results suggest RF to be a more effective approach for PS estimation than the other proposed methods when applying IPTW analysis to complex survey data for population average treatment effects. The methods were applied to data from the Medicare Beneficiary Current Survey for years 2002-2019 to estimate the impact of hospice use on end-of-life healthcare costs. Findings from the real-world example show that hospice use was significantly associated with reduced end-of-life healthcare costs of Medicare Beneficiaries.

We present methodology for creating synthetic data and an application to create a publicly releasable synthetic version of the Longitudinal Aging Study in India (LASI). The LASI, a health and retirement survey, is used for research and educational purposes, but it can only be shared under restricted access due to privacy considerations. We present novel methods to synthesize the survey, maintaining three nested levels of observation-individuals, couples, and households-with both continuous and categorical variables and survey weights. We show that the synthetic data maintains the distributional patterns of the confidential data and largely mitigates identification and attribute disclosure risk. We also present a novel method for controlling the risk and utility tradeoff for the synthetic data that take into account the survey sampling rates. Specifically, we down-weight records that have a high likelihood of being uniquely identifiable in the population due to unique demographic information and oversampling. We show this approach reduces both identification and attribute risk for records while preserving better utility over another common approach of coarsening records. Our methods and evaluations provide a foundation for creating a synthetic version of surveys with multiple units of observation, such as the LASI.

Government statistical offices worldwide are under pressure to produce statistics rapidly and for more detailed geographies, to compete with unofficial estimates available from web-based big data sources or from private companies. Commonly suggested sources of improved health information are electronic health records (EHRs) and medical claims data. These data sources are collectively known as real world data (RWD) because they are generated from routine health care processes, and they are available for millions of patients. It is clear that RWD can provide estimates that are more timely and less expensive to produce- but a key question is whether or not they are very accurate. To test this, we took advantage of a unique health data source that includes a full range of sociodemographic variables and compare estimates using all of those potential weighting variables, versus estimates derived when only age and sex are available for weighting (as is common with most RWD sources). We show that not accounting for other variables can produce misleading, and quite inaccurate, health estimates.

Typical design-based methods for weighting probability samples rely on several assumptions, including the random selection of sampled units according to known probabilities of selection and ignorable unit nonresponse. If any of these assumptions are not met, weighting methods that account for the probabilities of selection, nonresponse, and calibration may not fully account for the potential selection bias in a given sample, which could produce misleading population estimates. This analysis investigates possible selection bias in the 2019 Health Survey Mailer (HSM), a sub-study of the longitudinal Health and Retirement Study (HRS). The primary HRS data collection has occurred in "even" years since 1992, but additional survey data collections take place in the "off-wave" odd years via mailed invitations sent to selected participants. While the HSM achieved a high response rate (83 percent), the assumption of ignorable probability-based selection of HRS panel members may not hold due to the eligibility criteria that were imposed. To investigate this possible non-ignorable selection bias, our analysis utilizes a novel analysis method for estimating measures of unadjusted bias for proportions (MUBP), introduced by Andridge et al. in 2019. This method incorporates aggregate information from the larger HRS target population, including means, variances, and covariances for key covariates related to the HSM variables, to inform estimates of proportions. We explore potential non-ignorable selection bias by comparing proportions calculated from the HSM under three conditions: ignoring HRS weights, weighting based on the usual design-based approach for HRS "off-wave" mail surveys, and using the MUBP adjustment. We find examples of differences between the weighted and MUBP-adjusted estimates in four out of ten outcomes we analyzed. However, these differences are modest, and while this result gives some evidence of non-ignorable selection bias, typical design-based weighting methods are sufficient for correcting for it and their use is appropriate in this case.