Two further seminars are planned as part of the Edinburgh/Napier series.
Both will be held at 3.30pm at Napier's Merchiston campus, Room B4 (just
near entrance to the library from the foyer).
Tea and biscuits will be available in B4 from around 3pm, and informal
discussions over a drink usually follow the seminar.
Friday 7th June
PETER M FAYERS (University of Aberdeen)
Measuring health related quality of life
Friday 21st June
Chris Robertson (University of Starthclyde)
Modelling regional variations in disease rates
ABSTRACTS
FAYERS
This talk will cover some of the material that Peter presented (with David
Hand) as a read paper to the RSS last December.
Development and validation of multi-item measurement scales is usually based
on psychometric theory. This may be inappropriate for quality-of-life and
other clinical measurements,
where the fundamental distinction between indicator and causal variables
suggests a need for
both clinimetric and psychometric approaches in scale development.
ROBERTSON
An increasing number of Epidemiological investigations are concerned with
the comparison of disease rates in a number of areas. Examples include
cancer maps, investigation of regional variation in cervical cancer
screening. As well as regional variation there is also a temporal
dimension when disease rates are compared over a number of time periods.
Often the analysis of disease rates over time is accomplished within and
Age Period Cohort Model in which there is a well known identifiability
problem associated with the linear temporal trends. In this paper we
present a methodology for assessing regional variation and modeling
temporal changes in this variation.
One of the problems is in assessing if there is regional variation in
excess of what one would expect over and above sampling variation. We
present a diagnostic graph which can be used to compare regional variation
with anticipated sampling variation assuming that the numbers of cases in
an age group by time period cell arises from an (over-dispersed) Poisson
distribution.). The regional variance in each age group by time period
cell is plotted against the average of the inverse of the number of cases
in these cells. If there is no regional variation then the points are
expected to lie along a line of slope 1, passing through the origin. The
presence of regional variation is characterized by an intercept greater
than zero.
We further show that the data can be modeled by imbedding an identifiable
Age Period Cohort model, based upon local curvatures, within a two or
three level Poisson multilevel model. We also discuss the use of a normal
approximation which can be used when there are a large number of cases.
We show that there is a logical difference between the longitudinal Age
Period Cohort model with a random effect for cohort and the longitudinal
Age Period Cohort model with a random effect for period, as a different
covariance structure is implied by the different random effects. The
identifiability problem carries over to the random effects and we
illustrate how to detect non identifiable models and to overcome this by
the use of appropriate constraints.
We investigate regional variation in trends in the incidence of cancer of
the tongue in 14 regions of England plus Scotland and Wales using data
from 1971 to 1996 among men and women aged 35 to 84. The graphical
analysis shows the presence of modest regional variation in the incidence
rates, especially among men but less so for women. When fitting the
multilevel the three level model with age group nested within period
which was nested within region was more flexible than the two level model
with age and period nested within region. For men, there was evidence
that regional variation has increased in recent periods. Regional
variation was also greater among younger men.
|