Degree

Doctor of Philosophy (PhD)

Department

Engineering Science

Document Type

Dissertation

Abstract

Recently, data-based fragility models have been implemented to explain multi-hazard hurricane building fragility. However, deficiencies in current fragility models exist. Typically, categorical dependent variables are modeled as continuous numerical variables, which may result in inefficient standard errors for estimated coefficients and the models may result in probabilities greater than one or less than zero. Additionally, published models are limited to inference and interpretation of main variable coefficients without consideration of interaction terms between the variables and do not consider evaluation of model performance.

This dissertation addresses these deficiencies in the development of predictive data-based fragility models for multinomial categorical damage states (DS) and binary collapse/non-collapse using proportional odds cumulative logit and logistic regression models, respectively. The models are fitted as a function of hazard parameters and their interactions (Chapter 2) and as a function of hazard parameters, building attributes, and their interactions (Chapter 4). Chapter 3 develops numerical imputation diagnostic and comparison approaches for missing binary data, providing a methodology to evaluate imputation techniques that maximize field reconnaissance data for integration in Chapter 4. Fragility model prediction accuracy is evaluated using “leave-one-out” cross-validation (LOOCV), expressed in terms of the cross-classification rate (CCR). Model inputs are physical damage and building attribute data collected in coastal Mississippi following 2005 Hurricane Katrina and high-resolution, numerical hindcast hazard intensities from the Simulating Waves Nearshore and ADvanced CIRCulation (SWAN+ADCIRC) models.

For models excluding building attributes, maximum significant wave height is a significant DS predictor, while maximum 3-second gust wind speed, maximum surge depth, and maximum water speed are significant collapse predictors. Model prediction accuracy ranges from 81% to 87%. For models including building attributes, maximum 3-second gust wind speed, maximum significant wave height, maximum water speed, foundation type, number of stories, and the interaction of both maximum water speed and maximum significant wave height with number of stories are significant DS predictors. Building attributes, maximum surge depth above local ground, maximum water speed, maximum significant wave height, foundation type, number of stories, and the interaction of maximum significant wave height with number of stories are significant collapse predictors. Model prediction accuracy ranges from 84% to 90%.

Date

5-24-2018

Committee Chair

Friedland, Carol

Available for download on Friday, May 24, 2019

Share

COinS