Final published version
Licence: CC BY-NC-ND: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License
Research output: Contribution to Journal/Magazine › Journal article › peer-review
A probabilistic predictive Bayesian approach for determining the representativeness of health and demographic surveillance networks. / Utazi, Chigozie Edson; Sahu, Sujit K.; Atkinson, Peter Michael et al.
In: Spatial Statistics, Vol. 17, 08.2016, p. 161-178.Research output: Contribution to Journal/Magazine › Journal article › peer-review
}
TY - JOUR
T1 - A probabilistic predictive Bayesian approach for determining the representativeness of health and demographic surveillance networks
AU - Utazi, Chigozie Edson
AU - Sahu, Sujit K.
AU - Atkinson, Peter Michael
AU - Tejedor, Natalia
AU - Tatem, Andrew J.
PY - 2016/8
Y1 - 2016/8
N2 - Health and demographic surveillance systems, formed into networks of sites, are increasingly being established to circumvent unreliable national civil registration systems for estimates of mortality and its determinants in low income countries. Health outcomes, as measured by morbidity and mortality, generally correlate strongly with socioeconomic and environmental characteristics. Therefore, to enable comparison between sites, understand which sites can be grouped and where additional sites would aid understanding of rates and determinants, determining the environmental and socioeconomic representativeness of networks becomes important. This paper proposes a full Bayesian methodology for assessing current representativeness and consequently, identification of future sites, focusing on the INDEPTH network in sub-Saharan Africa as an example. Using socioeconomic and environmental data from the current network of 39 sites, we develop a multi-dimensional finite Gaussian mixture model for clustering the existing sites. Using the fitted model we obtain the posterior predictive probability distribution for cluster membership of each 1×11×1 km grid cell in Africa. The maximum of the posterior predictive probability distribution for each grid cell is proposed as the criterion for representativeness of the network for that particular grid cell. We demonstrate the conceptual superiority and practical appeal of the proposed Bayesian probabilistic method over previously applied deterministic clustering methods. As an example of the potential utility and application of the method, we also suggest optimal site selection methods for possible additions to the network.
AB - Health and demographic surveillance systems, formed into networks of sites, are increasingly being established to circumvent unreliable national civil registration systems for estimates of mortality and its determinants in low income countries. Health outcomes, as measured by morbidity and mortality, generally correlate strongly with socioeconomic and environmental characteristics. Therefore, to enable comparison between sites, understand which sites can be grouped and where additional sites would aid understanding of rates and determinants, determining the environmental and socioeconomic representativeness of networks becomes important. This paper proposes a full Bayesian methodology for assessing current representativeness and consequently, identification of future sites, focusing on the INDEPTH network in sub-Saharan Africa as an example. Using socioeconomic and environmental data from the current network of 39 sites, we develop a multi-dimensional finite Gaussian mixture model for clustering the existing sites. Using the fitted model we obtain the posterior predictive probability distribution for cluster membership of each 1×11×1 km grid cell in Africa. The maximum of the posterior predictive probability distribution for each grid cell is proposed as the criterion for representativeness of the network for that particular grid cell. We demonstrate the conceptual superiority and practical appeal of the proposed Bayesian probabilistic method over previously applied deterministic clustering methods. As an example of the potential utility and application of the method, we also suggest optimal site selection methods for possible additions to the network.
KW - Bayesian inference
KW - BIC
KW - Central clustering
KW - Finite Gaussian mixture model
KW - Gibbs sampling
KW - Predictive clustering
U2 - 10.1016/j.spasta.2016.05.006
DO - 10.1016/j.spasta.2016.05.006
M3 - Journal article
VL - 17
SP - 161
EP - 178
JO - Spatial Statistics
JF - Spatial Statistics
SN - 2211-6753
ER -