Sequential Bayesian Regression for Multiple Imputation and Conditional Editing

摘要: Analysts faced with errors in data apply editing rules to fix erroneous data. These edits are deterministically assigned and may not be correct all cases. This dissertation presents a unified method multiply impute missing edit using sequence of Bayesian regression models. The techniques used an exact parallel for multiple imputation models presented allow different types subject several error mechanisms. is called Sequential Regression Multiple Imputation Conditional Editing (SyBRMICE) creates fully imputed edited sets. Desired analyses performed on each complete consistently set individually. Results from these combined the same combining imputation. resulting parameter estimates intervals will then correctly account incurred both processes. Development SyBRMICE was motivated by Project Connect (PC). 8 year longitudinal intervention study aiming reduce teen pregnancy STD rates select middle high schools Los Angeles area. Survey collected annually measure effectiveness interventions. A paper survey administered students as group classroom, student responses have five years. subset participated years repeated answers question student. Data found PC can categorized belonging one types. If variable such gender that should remain constant over time observed differ across surveys, this said inconsistent response. variable, age or ever having sexual intercourse, increase monotonically non-monotonic reporting pattern, monotonic Lastly if two more related variables give conflicting information, Models stochastically three presented. measures, monotone longitudinal, multivariate developed separately steps example larger unifying procedure. examples demonstrate flexibility customizability analysis consistent sets generated procedure compared results single deterministically-edited, complete-case set.