11/18/2009 — Estimating the ROC curve of a continuous biomarker with ignorable verification bias

Danping Liu

The receiver operating characteristic (ROC) curve is a powerful tool to evaluate the classification accuracy of a biomarker. In many large-scale screening tests, the gold standard is subject to missingness due to high cost or harmfulness to the patient. A complete-case analysis would yield biased results, known as “verification bias”. Under missing at random (MAR) assumption, several methods are proposed to correct for the verification bias based on imputation, inverse probability weighting, or both. The model framework is built on a location-scale model for the marker value, with unspecified residual distribution. Bias-corrected estimation of three different types of ROC curves are discussed: (i) the ordinary ROC curve as a plot of sensitivity and specificity, (ii) the covariate-specific ROC curve when the marker value is associated with individual covariates, and (iii) the adjusted ROC (AROC) curve as an “average” of covariate-specific ROC curves. The proposed method is applied to the data from Alzheimer’s disease research.

The receiver operating characteristic (ROC) curve is a
powerful tool to evaluate the classification accuracy of a biomarker. In
many large-scale screening tests, the gold standard is subject to
missingness due to high cost or harmfulness to the patient. A
complete-case analysis would yield biased results, known as
“verification bias”. Under missing at random (MAR) assumption, several
methods are proposed to correct for the verification bias based on
imputation, inverse probability weighting, or both. The model framework
is built on a location-scale model for the marker value, with
unspecified residual distribution. Bias-corrected estimation of three
different types of ROC curves are discussed: (i) the ordinary ROC curve
as a plot of sensitivity and specificity, (ii) the covariate-specific
ROC curve when the marker value is associated with individual
covariates, and (iii) the adjusted ROC (AROC) curve as an “average” of
covariate-specific ROC curves. The proposed method is applied to the
data from Alzheimer’s disease research.

11/04/2009 – HIV Vaccine Trails: The first, the Step and The Thai

Erin Gabriel

There have been more than 100 HIV vaccine trials since 1988. There have been 3 main types of trials and not one has shown a substantial reduction in acquisition until the Thai trial that combined two previously non-efficacious vaccine types. The two types of vaccines were tested previously in the STEP trial and the VaxGen trials. The STEP trial was stopped early due to lack of efficacy along with some evidence of harm. The long term follow-up has shown that actual harm is unlikely and I will discuss the different approaches to analyzing this challenging data. The VaxGen trial found no effect of the purely antibody vaccine. The Thai trial has caused a stir with it’s greater than 30% estimated effect. I will discuss the results and why it should give us hope for the future.

10/28/2009 – Estimation and Inference for Measures of Accuracy of a Medical Test in the Presence of Verification Bias.

Michael Sachs

Often, the accuracy of a medical test for diagnosing or predicting a binary event is measured against some definitive assessment of that binary event. Sometimes, for ethical or practical reasons, a definitive assessment of the event is not possible on all the subjects in a study sample. In such studies where not all subjects are definitively diagnosed, naive estimates of measures of accuracy may be subject to bias. This type of bias is called verification bias. The most common measure of accuracy of a medical test is the Receiver Operating Characteristic (ROC) Curve. Other measures that can also be scientifically relevant include Positive and Negative Predictive Values, covariate-specific ROC curves, covariate-adjusted ROC curves, the Predictiveness Curve, the Proportion of Explained Variation, and Total Gain. In his investigation, we propose to review these measures and consider their estimation in the presence of verification bias. Existing estimation methods are based on missing data procedures such as reweighting, imputation, and “doubly-robust” procedures that are hybrids of reweighting and imputation. We also consider a newly proposed method based on the theory of semi-parametric biased sampling models. Preliminary study suggests that the doubly-robust procedures offer the best balance between efficiency and robustness to model misspecification. Finally, we propose to investigate study design considerations and make recommendations for sample-size and whether to do a single phase or a two-phase study in situations where one would like to estimate a particular measure with a specified degree of precision.

10/21/2009 — The Stepped Wedge Design: Outstanding Issues

Tanya Granston

The stepped wedge design of cluster-randomized trials has been growing in popularity and becomes increasingly relevant due to the need to efficiently evaluate the rollout of interventions that are individually efficacious in a community setting.  The design becomes especially practical for HIV/AIDS prevention and intervention trials, as the areas in most need are in resource limited settings where it is very likely not feasible to introduce interventions all at once and where there are substantial clusters/communities/groups to provide answers regarding intervention effects.  However, there remain limitations and outstanding issues to this design, which could potentially limit its usefulness where it’s most needed, and there are interesting extensions to the design that need to be explored.

10/14/2009 — A Quasi-Likelihood Approach to Genome Wide Association Studies

Matthew Bryan

A common issue in Genome Wide Association Studies (GWAS) is relatedness between subjects due to restricting sampling to a confined homogenous population.  As a result, the genetic data can be correlated across subject which may not be accounted for by common genetic testing methods.  This presentation will discuss a method suggested by Timothy Thornton and Mary McPeek that is designed to account for relatedness between subjects for genetic data.  The method uses a quasi-likelihood model approach to generate a quasi-likelihood score statistic.  This statistic can be shown to be locally most powerful for testing genetic associations with a phenotype.  Thus the method is expected to provide higher power for detecting small genetic effects.