On the Relationship Between Missing Information and Missing Data in the 2012 NAMCS Physician Workflow Mail Survey

• Activity Number: 645
• Type: Contributed
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Government Statistics Section
• Author(s): Qiyuan Pan* and Rong Wei
• Companies: CDC/NCHS/DHCS and CDC/NCHS/ORM
• Keywords: Multiple imputation ; fraction of missing information (?) ; sufficient number of imputations ; missing data ; NAMCS

Abstract

In Multiple Imputation (MI), the fraction of missing information, ?, defined as a certain function of the constituents of the variance of the MI estimator, has from the beginning played a key role in deciding how large the number m of imputations should be (Rubin 1987). Rubin defines the relative efficiency of m as RE = (1+?/m)-1/2, and concludes that only a small m, e.g. m = 2 or 3, is needed in MI for ??0.3. In recent years, however, researchers have presented evidence that much greater m values are needed in MI. A better understanding of ? helps explain why Rubin's RE may give an m that is apparently sufficient but may actually be too small for many important statistical inferences. To date little research on ? has been done using real survey data. In the present study, ? was determined for 4, 10, 20, and 29% of missing data using the 2012 National Ambulatory Medical Care Survey (NAMCS) Physician Workflow Mail Survey. The results and their implications add to our understanding of ? as well as the relevance of the ?-based RE in determining a sufficient m.