STATISTICAL METHODS FOR REPEATED MEASURES IN EXPERIMENTAL GINGIVITIS WITH ADJUSTMENT FOR LEFT TRUNCATION DUE TO LOWER DETECTION LIMITS

Public Deposited
Analytics

Downloadable Content

Download PDF
Request Version for Screen Reader
Last Modified
  • March 19, 2019
Creator
  • Wekheye, Kelley
    • Affiliation: Gillings School of Global Public Health, Department of Biostatistics
Abstract
  • In the characterization of biomarkers measured repeatedly over time, there is a need to summarize the information contained in the multivariate data. In experimental gingivitis (EG), for example, biomarker levels change when the benefits of toothbrushing are withheld during an induction phase, then restored during a resolution phase. The pattern of change over time of biomarker levels associated with gingivitis could reflect change in various directions; therefore, the statistical methodology utilized should consider this possibility. As such, area under the curve (AUC) can be implemented as a summary measure for estimating change in biomarker levels. Parametric statistical models for repeated measures analysis are useful for characterizing the nature of that change over time, particularly as they easily accommodate both truncated and missing data. In EG studies, left truncation results when a biomarker level falls below the lower limit of detection. We propose two parametric approaches to provide direct estimation of the trends in biomarkers over time while implementing adjustments for left truncation. The focus is on estimation and hypothesis testing for AUC. The first paper derives a piecewise linear random-effects regression model fit to 3 biomarkers representing varying degrees of missingness due to lower detection limits using 2 ad hoc (naive) approaches for handling non-detect values and a likelihood approach accounting for left censoring (Lyles, Lyles and Taylor, 2000). These naive approaches replace non-detect biomarker values by the limit of detection and half that limit, which may result in bias, while the maximum likelihood method gives valid results when dropouts are missing at random. The second paper outlines AUC methodology for repeated measures biomarker data by using a nonlinear "Gamma Curve" mixed model with adjustment for left truncation based on a maximum likelihood approach in comparison to the ad hoc approaches outlined in the first paper. The third paper presents a simulation study that includes methods from the first two papers as well as Wilcoxon Sign Rank test methods from Preisser, Sen, and Offenbacher (2011). The simulation design, motivated by EG studies, focuses on properties of hypothesis tests (size and power) in the presence of left truncation and/or missing data to evaluate whether the parametric methods are reliable for small sample sizes or whether larger samples are needed to reliably use the methods. Evidence for recommending certain sample sizes for EG studies and an evaluation of whether the nonparametric method is robust to left truncation and crude single imputation methods are also provided. The proposed methodology is illustrated using longitudinal data from an EG study whereby the benefits of toothbrushing are temporarily withheld, then restored (Offenbacher et al, 2010).
Date of publication
Keyword
Subject
DOI
Identifier
Resource type
Rights statement
  • In Copyright
Advisor
  • Edwards, Lloyd
  • Preisser, John
  • Sen, Pranab Kumar
  • Ivanova, Anastasia
  • Offenbacher, Steven
Degree
  • Doctor of Public Health
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2014
Language
Publisher
Place of publication
  • Chapel Hill, NC
Access right
  • There are no restrictions to this item.
Date uploaded
  • April 23, 2015

Relations

Parents:

This work has no parents.

Items

Thumbnail Title Date Uploaded Visibility Actions
file details Wekheye_unc_0153D_14864.pdf 2019-04-09 Public Download