Home Teaching Research Vita Offices
My
current research focuses on three distinct but interrelated topics.
1.
High-Dimensional
Inference: My research in this area is centered on developing
parametric and nonparametric methods for analyzing high-dimensional data,
especially for comparison of treatments or groups. High dimensional
refers to both the sample size and dimension being large without any
restriction on the rate of growth between the two. Recent
high-dimensional methods make strong assumptions about the dependence between
the variables, which may not be satisfied, for example, in elliptically
contoured models or other dependence models. Our high-dimensional
approach focuses on asymptotic methods that are applicable under weaker
dependence conditions. The weaker conditions, among other things,
make it possible to derive results for rank-based high-dimensional inference.
Representative
Publications:
a)
Harrar, S. W.
and Hossler, J. Z. (2016), Methods for
High-Dimensional Multivariate Repeated Measures Data under General Conditions. Statistics: A Journal of Theoretical and
Applied Statistics 50(5), 1056--1074.
b)
Harrar, S. W.
and Kong X. (2016), High-Dimensional Repeated Measures Analysis with Unequal
Covariance Matrices, Journal of
Multivariate Analysis 145(1),
1--21.
2.
Nonparametric
Analysis of Clustered Data: Clustered data naturally arise when
subjects or units of analysis are grouped in some way such that there is
considerable amount of dependence within cluster. Methods
for clustered data have been widely investigated from GEE, mixed model or
Bayesian points of views. Not much attention has been given to the
nonparametric analysis of clustered data. In this context, my
research pertains to inference on the so-called nonparametric effects
(Mann-Whitney effects) which quantity the tendency of one sample to be larger
(smaller) than the other while accounting for the intra-cluster dependence.
This approach, besides not making assumptions about data distributions not
their moments, does not imply equal distribution under the null
hypothesis. This phenomenon is referred to as the nonparametric
Behrens-Fisher problem.
Representative
Publications:
a)
Roy, A., Harrar, S. W. and Konietschke,
F., The nonparametric Behrens-Fisher Problem with Dependent Replicates.
b)
Feyasa, M. B., Harrar,
S. W. and Wencheko, E., Nonparametric Procedures for Partially Paired Data in Two Groups, (submitted).
3.
Inference
under Diagnostic Misclassification: In clinical trials or
other comparative studies, diagnostic or screening devices often determine the status
of patients for having the condition or trait of interest. Treatment
modalities are allocated to groups based on the results of the diagnostic
(screening) devices. Typical statistical analysis makes the assumption that the
devices are 100% accurate but often times, these devices are prone to errors of
unknown rate. Needless to say, ignoring these errors leads to
incorrect inference and optimistic sample size estimates. In our
researches in this direction, we develop methods to estimate and account for
diagnostic errors while concurrently make inference about treatment
effects.
Representative
Publications:
a)
Harrar, S. W.,
Amatya, A. and Kalachev, L.
(2016), Assessing Treatment Efficacy in the Presence of Diagnostic Errors, Statistics in Medicine 35(29), 5338--5355.