Bayesian hierarchical models with applications to cervical, oesophageal and lung cancers in Kenya's Counties

Abstract

Cancer is an event associated with space and time. Counties relative risks esti- mates can be obtained using Bayesian hierarchical models. The general objective of the research was to obtain county based estimates through Bayesian hierarchi- cal modeling of cervical, oesophageal and lung cancers in Kenya's counties from 2015 to 2016, period which complete data was available. Speci c objectives were: to model over-dispersion and conduct spatial correlations tests in order to model three cancer cases distribution in Kenya' counties; to model cervical cancer cases using Poisson-Gamma and spatial-temporal models; to model the e ects of co- variates on spatial-temporal distribution of oesophageal and lung cancer cases in Kenya's counties. The data was obtained from National Cancer Registry (NCR) which carried a 2 year retrospective study in ten counties. Cervical cancer cases were 1064, oesophageal cancer cases 1599 while lung cancer cases were 256. A simple Poisson log-linear model dispersion parameter for cervical was 31.202, oe- sophageal was 49.241 and lung cancer cases was 6.134 which were greater than 1 indicating over dispersion. Spatial correlation tests conducted for the three cancers revealed that there was no spatial auto correlations of the residuals since for cervical cancer p-value=0.2104>0.05, oesophageal p-value= 0.4155>0.05 while for lung cancer p-value=0.4120>0.05. The model revealed that the highest cervi- cal cancer relative risk was in Embu=7.92 and lowest in Bomet which was 1.53. The smoking and alcohol use interaction oesophageal cancer model revealed that Bomet=11.16 had the highest risk while Kiambu had the lowest relative risk 0.6. Smoking and alcohol use were signi cant risk factors of oesophageal cancer. The multiplicative e ect of smoking was 1.012, thus 1.2 % higher to smokers compared to non-smokers. Alcohol use was 1.0346 thus 3.5 % higher to alcohol users. The interaction model revealed that oesophageal cancer was 16.88 % higher to alcohol users while it was 4.60 % higher to smokers. The interaction model for lung cancer revealed that in Nairobi=5.97 had highest risk while lowest in Kakamega=0.1. In the lung cancer model the multiplicative e ect of smoking was 1.4021, indicating 40.21 % higher to smokers as compared to non-smokers, 1.3689 for alcohol use variable that is 36.89 % higher to alcohol users. In interaction model the e ect was 7.86 times higher for smokers. In conclusion, simple Poisson regression mod- els were not appropriate to model the three cancers due to over dispersion nature of the data sets. The spatial correlation tests revealed that there was no spatial auto correlation for the three types of cancer. Application of Bayesian hierarchical models enabled generation of relative risks and identi cation of the risk patterns of various counties, a major milestone since previous studies focused on speci c regions. We recommend that, since all counties had cervical cancer relative risk greater than 1, step up screening and avail vaccines to the appropriate groups. To mitigate oesophageal cancer, counties should create awareness on e ects of smok- ing and alcohol use. In case of lung cancer, counties with relative risks greater than 1 should disseminate information elaborating the e ects of smoking and alcohol use