Imputation in a Complex Survey of Cancer Care
*Yulei He, Department of Health Care Policy, Harvard Medical School
Alan M Zaslavsky, Department of Health Care Policy, Harvard Medical School
Keywords: Cancer, Missing data, Posterior predictive checking, Sequential regression multiple imputation, Survey
The Cancer Care Outcomes Research and Surveillance (CanCORS) Consortium is a multisite, multimode and multiwave study assessing the process and quality of care delivered to population-based cohorts of newly diagnosed and treated patients with lung and colorectal cancer. We use multiple imputation to handle block or item nonresponse in the CanCORS surveys. It would be difficult to impose a joint imputation model which accounts for the complex data features such as the various types of variables, boundaries of variables imposed by the survey questionnaire, and structured skip patterns caused by patients with distinct characteristics. Instead, we apply the sequential conditional regression imputation approach that characterizes the data by a series of conditional models for each incomplete variable. We use posterior predictive checking to assess the lack of fit of the imputation models.