Combining Information from Multiple Complex Surveys.

Combining Multiple Surveys;Missing Data;Complex Sampling Design Features;Synthetic Populations;Multiple Imputation;Health Insurance Coverage;Statistics and Numeric Data;Social Sciences;Survey Methodology

Dong, QiValliant, Richard L. ;

University of Michigan

关键词: Combining Multiple Surveys;    Missing Data;    Complex Sampling Design Features;    Synthetic Populations;    Multiple Imputation;    Health Insurance Coverage;    Statistics and Numeric Data;    Social Sciences;    Survey Methodology;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/93990/qidong_1.pdf?sequence=2&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Increasingly, many substantive research questions require a degree of information not adequately collected in a single survey. Fortunately, survey organizations often repeatedly draw samples from the same population for different surveys and collect data on a considerable number of overlapping variables. This dissertation presents a new method for combining multiple surveys from a missing data perspective. Two major improvements of the new method include: 1) adjusting for the incompatibility among different sample designs and 2) combining an unlimited number of surveys. The basic proposal is to simulate synthetic populations from which the respondents of each survey have been selected. In this process, different sampling designs of the multiple surveys will be taken into account.Once we have the synthetic populations, we could treat them as simple random samples with no complex sampling design features and borrow information across surveys to adjust for nonsampling errors or fill in the variables that are lacking in one or more surveys. Then, we can analyze each synthetic population with standard complete-data software for simple random samples and obtain valid inference by combining the point and variance estimates, first across synthetic populations within each survey, and then across multiple surveys. Existing methods borrowed from the disclosure risk field will be used to combine the synthetic populations from one survey; combining these results across multiple surveys will require the methods developed in this dissertation. The first study develops the combining rule when multiple surveys present and proposes a model-based method to impute the unobserved population. The 2006 National Health Interview Survey (NHIS) and Medical Expenditure Panel Study (MEPS) are combined to estimate health insurance coverage. The second study develops a nonparametric method to impute the unobserved population, which is used to generate synthetic populations for the NHIS and MEPS and produce combined estimates of health insurance coverage. The third study combines surveys with missing variables. A two-stage combining rule is developed to account for the uncertainty due to simultaneously imputing the missing variables and generating synthetic populations. The 2006 Behavioral Risk Factor Surveillance System is combined with the NHIS and MEPS to estimate health insurance coverage.