Using the Fraction of Missing Information to Identify Auxiliary Variables for Imputation Procedures via Proxy Pattern‐mixture Models

摘要: Summary In many surveys, imputation procedures are used to account for non-response bias induced by either unit or item non-response. Such optimised (in terms of reducing bias) when the models include covariates that highly predictive both response and outcome variables. To achieve this, we propose a method selecting sets in regression determine cells one more variables, using fraction missing information (FMI) as obtained via proxy pattern-mixture (PMM) model key metric. In our variable selection approach, use PPM obtain maximum likelihood estimate FMI separate candidate look point at which changes level off further auxiliary variables do not improve model. We illustrate proposed approach empirical data from Ohio Medicaid Assessment Survey Service Annual Survey.