Binary Building Attribute Imputation, Evaluation, and Comparison Approaches for Hurricane Damage Data Sets

Abstract

Missing building attributes are problematic for development of data-based fragility models. Relative to other disciplines, the application of imputation techniques is limited in the field of engineering. Current imputation techniques to replace missing building attributes lack evaluations of imputation model performance, which ensure accuracy and validity of the imputed data. This paper presents two imputation approaches, along with imputation diagnostic and comparison approaches, for binary building attribute data with missing observations. Predictive mean matching (PMM) and multiple imputation (MI) are used to impute foundation type and number of stories attributes. The diagnostic approach, based on the logistic regression goodness-of-fit test, is used to evaluate the imputation model fit. The comparison approach, based on the percentage of correctly imputed observations, is used to evaluate the imputation model performance. A data set of single-family homes damaged by the 2005 Hurricane Katrina is used to demonstrate implementation of the methodology. Based on the comparison approach, PMM models showed 9% and 2% greater accuracy than MI models in imputing foundation type and number of stories, respectively.