Page 116 - Contributed Paper Session (CPS) - Volume 3

CPS1956 Areti B. et al.
National Statistics, 2016). Even though life expectancy is known to vary
substantially across space, very few studies have incorporated both temporal
and spatial information to investigate how the spatial patterns of mortality
evolve in time, and therefore to better understand the alarming behaviour
recently seen in England. In addition, most studies have traditionally used
standard statistical techniques, e.g. age-standardised mortality rates, thus not
accounting for the noise in the data. It is therefore hard to determine whether
spikes in deaths are true or data artefacts, in particular, when studying low
populated areas. In this paper, we analysed mortality counts in England at the
local authority level from 2001 to 2016 using Bayesian statistical methods,
which borrow strength from spatial and temporal neighbours to reduce the
high variability inherent to classical risk estimators, such as the crude mortality
rate. The main objective of this work was to investigate whether life expectancy
time trends are stable across England and highlight areas whose trends differ
to the national one over the last 5 years (2012 to 2016).

2. Methodology
We used mortality counts from 2001 to 2017 at the local authority level in
England. In total 324 local authorities were considered, after excluding the Isles
of Scilly and City of London. Information on age and sex was available for each
record. We used 19 5-year band age groups (0-4, 5-9, …, 90 plus) and we
analysed males and females separately as these are expected to have very
different mortality levels and trends. In this paper we present results for
females only. Population data by age and sex for each local authority for the
same time period were also used for the analysis. All data were provided by
Public Health England (PHE), originally held by Office for National Statistics
(ONS). Life expectancy tables were used to convert mortality rates to life
expectancy rates.
We developed a statistical model to analyse mortality counts by age
group, local authority, and year. The model was formulated within a Bayesian
hierarchical framework, which allowed to assign prior specification to the
unknown parameters through which we incorporate assumptions regarding
the structure of the data. The model is as follows:
,, ∼ ( ,, ,, )

where ,, and ,, are the mortality counts and population counts
respectively in area = 1, . . . ,324 , age group = 1, . . . ,19 and year
= 2001, . . . ,2017. Similarly, the parameter ,, represents the mortality rate
which we model on the log scale as ,, = + + + + + , + .
,
,

The overall intercept follows a flat prior. The spatial component s

assigned a convolution prior, widely known as the Besag York Mollie (BYM)
model (Besag et al. 1991). This is a Gaussian prior, ∼ ( , ) where is
2

105 | I S I W S C 2 0 1 9