Pitt Logo LinkContact Us

Multiple Imputation and Quantile Regression Methods for Biomarker Data subject to Detection Limits

Lee, MinJae (2011) Multiple Imputation and Quantile Regression Methods for Biomarker Data subject to Detection Limits. Doctoral Dissertation, University of Pittsburgh.

[img] PDF - Primary Text
Restricted to University of Pittsburgh users only until 31 January 2016.

Download (600Kb) | Request a copy

    Abstract

    Biomarkers are increasingly used in biomedical studies to better understand the natural history and development of a disease, identify the patients at high-risk and guide the therapeutic strategies for intervention. However, the measurement of these markers is often limited by the sensitivity of the given assay, resulting in data that are censored either at the lower limit or upper limit of detection. Ignoring censoring issue in any analysis may lead to the biased results. For a regression analysis where multiple censored biomarkers are included as predictors, we develop multiple imputation methods based on Gibbs sampling approach. The simulation study shows that our method significantly reduces the estimation bias as compared to the other simple imputation methods when the correlation between markers is high or the censoring proportion is high. The likelihood based mean regression for repeatedly measured biomarkers often assume a multivariate normal distribution that may not hold for biomarker data even after transformations. We consider a robust alternative, median regression, for censored longitudinal data. We develop an estimating equation approach that can incorporate the serial correlations between repeated measurements. We conduct simulation studies to evaluate the proposed estimators and compare median regression model with the mixed models under various specifications of distributions and covariance structures. Missing data is a common problem with longitudinal study. Under the assumptions that the missing pattern is monotonic and the missingness may only depend on the observed data, we propose a weighted estimating equation approach for the censored quantile regression models. The contribution of each individual to the estimating equation is weighted by the inverse probability of dropout at the given occasion. The resultant regression estimators are consistent when the dropout process is correctly specified. The performance of our estimating procedure is evaluated via simulation study. We illustrate all the proposed methods using the biomarker data of the Genetic and Inflammatory Markers of Sepsis (GenIMS) study. Appropriate handling of censored data in biomarker analysis is of public health importance because it will improve the understanding of the biological mechanisms of the underlying disease and aid in the successful development of future effective treatments.


    Share

    Citation/Export:
    Social Networking:

    Details

    Item Type: University of Pittsburgh ETD
    Creators/Authors:
    CreatorsEmailORCID
    Lee, MinJaeleem2@mwri.magee.edu
    ETD Committee:
    ETD Committee TypeCommittee MemberEmailORCID
    Committee ChairKong, Lanlkong@pitt.edu
    Committee MemberJeong, Jong-Hyeonjeong@nsabp.pitt.edu
    Committee MemberWeissfeld, Lisalweis@pitt.edu
    Committee MemberYende, Sachinyendes@upmc.edu
    Title: Multiple Imputation and Quantile Regression Methods for Biomarker Data subject to Detection Limits
    Status: Unpublished
    Abstract: Biomarkers are increasingly used in biomedical studies to better understand the natural history and development of a disease, identify the patients at high-risk and guide the therapeutic strategies for intervention. However, the measurement of these markers is often limited by the sensitivity of the given assay, resulting in data that are censored either at the lower limit or upper limit of detection. Ignoring censoring issue in any analysis may lead to the biased results. For a regression analysis where multiple censored biomarkers are included as predictors, we develop multiple imputation methods based on Gibbs sampling approach. The simulation study shows that our method significantly reduces the estimation bias as compared to the other simple imputation methods when the correlation between markers is high or the censoring proportion is high. The likelihood based mean regression for repeatedly measured biomarkers often assume a multivariate normal distribution that may not hold for biomarker data even after transformations. We consider a robust alternative, median regression, for censored longitudinal data. We develop an estimating equation approach that can incorporate the serial correlations between repeated measurements. We conduct simulation studies to evaluate the proposed estimators and compare median regression model with the mixed models under various specifications of distributions and covariance structures. Missing data is a common problem with longitudinal study. Under the assumptions that the missing pattern is monotonic and the missingness may only depend on the observed data, we propose a weighted estimating equation approach for the censored quantile regression models. The contribution of each individual to the estimating equation is weighted by the inverse probability of dropout at the given occasion. The resultant regression estimators are consistent when the dropout process is correctly specified. The performance of our estimating procedure is evaluated via simulation study. We illustrate all the proposed methods using the biomarker data of the Genetic and Inflammatory Markers of Sepsis (GenIMS) study. Appropriate handling of censored data in biomarker analysis is of public health importance because it will improve the understanding of the biological mechanisms of the underlying disease and aid in the successful development of future effective treatments.
    Date: 31 January 2011
    Date Type: Completion
    Defense Date: 31 August 2010
    Approval Date: 31 January 2011
    Submission Date: 15 September 2010
    Access Restriction: 5 year -- Restrict access to University of Pittsburgh for a period of 5 years.
    Patent pending: No
    Institution: University of Pittsburgh
    Thesis Type: Doctoral Dissertation
    Refereed: Yes
    Degree: PhD - Doctor of Philosophy
    URN: etd-09152010-104021
    Uncontrolled Keywords: Gibbs sampler; Median regression; Multiple imputation; Detection limits; Longitudinal data; Quantile regression; Drop-out; Left-censored data
    Schools and Programs: Graduate School of Public Health > Biostatistics
    Date Deposited: 10 Nov 2011 15:02
    Last Modified: 02 May 2012 11:31
    Other ID: http://etd.library.pitt.edu/ETD/available/etd-09152010-104021/, etd-09152010-104021

    Actions (login required)

    View Item

    Document Downloads