TWO UNIVERSITY BURSARIES TO SUPPORT PHD RESEARCH FOR 3 YEARS
ARE AVAILABLE IN THE DEPARTMENT OF MEDICAL STATISTICS
AT DE MONTFORT UNIVERSITY, LEICESTER, UK. THE TOPICS ARE:
1. THE DESIGN AND ANALYSIS OF SENSORY AND CONSUMER STUDIES.
2. THE ANALYSIS OF NON-NORMAL REPEATED MEASUREMENTS DATA.
Requests for application forms should be sent (by email, fax or post)
to:
Professor Byron Jones
Department of Medical Statistics
De Montfort University
The Gateway
Leicester
LE1 9BH
UK
Telephone: (+44) (0) 116 257 7463, Fax: (+44) (0) 116 250 6114 Email:
[log in to unmask]
--------------------------------------------------------------------------------
The details of each project are given below:
1. THE DESIGN AND ANALYSIS OF SENSORY AND CONSUMER STUDIES.
IN COLLABORATION WITH UNILEVER RESEARCH AND
THE INSTITUTE OF FOOD RESEARCH
A De Montfort University Bursary which covers tuition fees and
subsistence for three years study for a PhD will be available
beginning August 1st, 1998. The Bursary covers the fees at EEC levels.
In addition to the subsistence element of the Bursary (worth 5,500 GB
pounds per year), Unilever Research will give the student an
additional 1,500 GB pounds per year.
Unilever view this project as fully collaborative and will provide
access to a wealth of real and important problems, data from many
sensometric trials and access to research staff. As a consequence, the
student will be expected to spend some time each year at Unilever's
Colworth site near Sharnbrook Beds, at the company's expense. The,
student will also make visits to the Institute of Food Research Reading.
The project will be supervised by Professor Byron Jones (De Montfort),
Ian Wakeling (Institute of Food Research) and Graham Cleaver (Unilever
Research).
Applicants should have a good first degree in mathematics and/or
statistics or equivalent.
PROJECT DESCRIPTION
The sensory testing of foods by expert taste panels is a technique
vital to quality control and new product development within the food
industry. In an experiment a monadic sequence of food samples are
presented to each panel member and scores obtained for key sensory
attributes. The sensory profiles obtained are the principal means
of quantifying taste differences between samples and are therefore an
essential tool for food research scientists. In conjunction with
sensory experiments a parallel experiment is sometimes performed with
untrained consumers who evaluate their preference for the samples.
Sensory experiments raise challenging statistical design problems
which to date have not been fully addressed. Often there are more
samples to be profiled than the judges are able to taste in a single
sitting, so the immediate problem is which sample subset to present to
which judge in each session. But the ordering of the subset of
samples is also critical to the design as it can be used to balance
out the effects of position in the presentation order and also to
balance the residual effects that one sample may have on the next one
to be tasted. Designs for multi-session experiments should also take
account of any factorial structure among the samples and conform to
any practical limitations such as not having all samples available in
all sessions due to the limited resources required to prepare them.
During the course of the experiment other problems may arise, such as
non-attendance by some panel members, a change in performance, or
problems with sample preparation in a particular session. Therefore
designs which are robust to the removal of some assessors and some
sessions are therefore desirable.
The first aim of the project will be to investigate to what extent it
is possible to create designs that take some or all of the above into
account. It is clear that for many problems computer search algorithms
will be the only way forward and it is hoped to develop an algorithm
which would allow the experimenter to prespecify the most
important properties for the design to have. Many of these properties
are equally applicable to the design of consumer trials and the
project will also identify ways in which this class of experiment can
be improved.
Of equal importance to the design is the method of analysis that is
used to properly take account of the repeated measurements nature of
sensometric data. Sensory and consumer trials often lead to data which
have a hierarchical error structure which also incorporates some form
of autocorrelation. As linear and ordinal categorical attribute scales
are both used in sensometric research, the analysis of both continuous
and categorical repeated measurements will be of interest. In the
last few years there has been a tremendous growth in methods for
analysing repeated measurements data. Applying these methods, and
possibly developing more appropriate ones for sensometric data, will
form the second aim of the project.
References
----------
Schlich, P. (1993) Use of change-over designs and repeated
measurements in sensory and consumer studies. Food and Quality
Preference, Vol. 4, 223-235.
Wakeling, I.N. and Macfie, H.J.H. (1995) Designing consumer trials
balanced for first and higher orders of carry-over effect when only a
subset of k samples from t may be tested. Food and Quality
Preference, Vol. 6, 299-308.
Jones, B. and Kenward, M.G. (1989) Design and Analysis of Cross-over
Trials. Chapman and Hall: London.
Lindsey, J.K. (1993) Models for Repeated Measurements. Oxford
University Press: Oxford.
----------------------------------------------------------------------------------------------------------------------------------------------------------------
2. THE ANALYSIS OF NON-NORMAL REPEATED MEASUREMENTS DATA.
The Bursary is available to support a full-time PhD student
for three years. It will provide an annual subsistence
payment of 5,500 (UK pounds) and cover all University fees subject to status.
The project will be supervised by a panel headed by
Professor Byron Jones.
Applicants should have a First Class or Upper Second Class degree in
Mathematics or Statistics or equivalent.
Modelling repeated measurements is one of the most dynamic areas of
modern theoretical and applied statistics. With advances in
computer data management, such measurements are now routinely
collected for most types of medical research. Many models are
available for normally-distributed responses and the area is
well researched. However, in many fields of application, the
responses are not normally-distributed and existing models may not be
appropriate. For example, discrete and duration responses are commonly
recorded. Few multivariate distributions for such
responses, which will handle the dependence among the repetitions, are
currently available.
Two types of dependence can be expected: those due to clustering of
responses on each individual and those due to serial dependence over
time. The first can be modelled by introducing appropriate random
effects. These may require numerical integration for non-normal
responses. For serial
dependence, two types of model for normal data are commonly used. A
conditional or state dependence model has the current response
depending on the previous one. This can easily be extended to
non-normal responses. However, the more common normal autoregression
models have the current response depending on the previous residual or
`innovation'. Very little work has been done to extend such models to
non-normal responses. This project will involve development
of appropriate models for non-normal repeated measurement
(longitudinal) data involving both clustering of response on
individuals and in time.
Models developed will be applied to a variety of existing medical and
non-medical data sets. Various techniques may be necessary in order
to fit such models, once they are developed: numerical integration,
Gibbs sampling, EM algorithm, Kalman filtering. The student will
need to develop sufficient programming skills to write the
software needed to apply the models developed to the available data sets.
Key reference:
Lindsey, J.K. (1993) Models for Repeated Measurements. Oxford:
Oxford University Press.
----------------------------------------------------------------------
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|