Page 235 - Special Topic Session (STS) - Volume 2

STS489 Glory A. et al.
population. We considered a condition at cut-point of blood pressure (BP) ≥
140/90 mmHg or self-reported diagnosis or on medication as captured in
the NiDS for the four consecutive surveys.
Exposure variables: A key exposure variable investigated in this study was
the effect of geographic location of respondent (at the time of the survey) on
the risk of hypertension given that closer districts (neighbors) are more likely
to have similar disease patterns. In addition, we controlled for known
individual risk factor variables such as the age of respondent (as a continuous
covariate), gender, race (Black African/Coloured/Indian-Asian/White), and
educational attainment. Others include lifestyle factors such as exercise,
alcohol and binary indicators for smoking, diabetes, fever and arthritis.
Statistical Analysis: We considered the class of Bayesian generalized geo-
additive mixed regression models in which the probability of hypertension in
individual ( = 1, . . . . , ) in district ( = 1, . . . . ) at time ( = 1, … . . ),

follows a Bernoulli distribution with mean = ( |, ) . Using

appropriate prior specification, we modelled the likelihood of hypertension
by replacing the linear predictor with a more flexible structured additive
predictor with a logit link specification. The flexibility of this class of models
allows us to account for nonlinear effects of continuous covariates, spatial
heterogeneity and spatial dependency structure between neighbouring
districts as well as temporal dependence in the observed data within a
unified framework. Full Bayesian inference was implemented using Markov
Chain Monte Carlo simulation method. Model evaluation and comparison
were carried out using the Deviance Information Criterion (Spiegelhalter et
al. 2002).
Model 1: Spatial model regression framework
~( )

= ( ) = 0 + + ( ) + ()

1

() = () + ()

Model 2: Spatio-temporal model regression framework
~( )

= ( ) = 0 + ( ) + ( ) + ( )

1
2
( ) = ( ) + ( )

Where and are the structured additive predictor with a logit link
function, 1 ( ) and 2 ( ) are the nonlinear effect of age and year
modelled as nonparametric smooth function using Bayesian P-Spline with a
second order random walk. In addition, are the vectors of covariate
values with their unknown regression parameters assigned a vague prior
distribution. The spatial effect of district () is decomposed into
224 | I S I W S C 2 0 1 9