Electronic Theses and Dissertations

Date

2021

Document Type

Thesis

Degree Name

Master of Public Health

Department

Health Studies

Committee Chair

Meredith Ray

Committee Member

Tit-Yee Wong

Committee Member

Yu (Joyce) Jiang

Abstract

Factorial design is one of the best ways to screen different factors in a system of study. It allows the study of the main effects and their interactions at the same time. We used the 2k-factorial design to screen different factors that affect the bacterial sensitivity to antibiotics. Through simulation studies, we examined the sensitivity and specificity of four different approaches to analyze the factorial data under different scenarios. The parametric ANOVA showed higher sensitivity and specificity for normally distributed data, but it failed largely for non-normal data. Rank-ANOVA had very high sensitivity and specificity for both normal and non-normal data when replicates were higher. Aligned-rank approach also had higher sensitivity and specificity for normal data, but not for non-normal data. However, it was very sensitive to outliers and thus failed to detect the significant interactions in our settings. Similarly, permutation approach failed to detect the significant interactions largely for non-normal distributions and needed higher number of observations for larger number of factors to run without error. Overall, rank-based ANOVA was the robust approach to analyze the factorial data without losing the statistical power under our simulation settings. We also used the above-mentioned approaches on our real data obtained from 2k factorial design. Model trimming options are not available in some of the functions that we used and therefore only full models were compared. Based on root mean squared error (RMSE) values all four approaches, except the permutation test, were similar in predicting the response variable. Permutation test was not applicable with all 7 factors and their interactions in the model due to a smaller number of observations, so reduced model without the factor temperature was used for the analysis. All four approaches yielded comparable p-value for the highest order interaction with 6 factors in the model but when 7 factors were in the model, rank ANOVA had different p-value than other approaches. Since, factorial data with larger number of factors are difficult to run with multiple replicates, assumptions of parametric approach may hold true. Non-parametric approaches like rank-based and aligned-rank-based approaches are more suitable for the purpose.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest

Share

COinS