[Paper Review] A Spatio-Temporal Multivariate Shared Component Model with an Application in Iran Cancer Data

Behzad Mahaki, Yadollah Mehrabi+2 more|PubMed|Jul 19, 2017

Health Systems, Economic Evaluations, Quality of Life29 references19 citations

TL;DR

This paper proposes a spatio-temporal multivariate shared component model that jointly analyzes incidence rates of seven cancers in Iran from 2005 to 2009, integrating spatial and temporal trends across shared latent risk components. The model identifies distinct geographical and temporal patterns, revealing strong spatial clustering for esophagus and stomach cancers, and highlights shared risk factors across disease pairs like breast and prostate cancer.

ABSTRACT

Background: Among the proposals for joint disease mapping, the shared component model has become more popular. Another advance to strengthen inference of disease data is the extension of purely spatial models to include time aspect. We aim to combine the idea of multivariate shared components with spatio-temporal modelling in a joint disease mapping model and apply it for incidence rates of seven prevalent cancers in Iran which together account for approximately 50% of all cancers. Methods: In the proposed model, each component is shared by different subsets of diseases, spatial and temporal trends are considered for each component, and the relative weight of these trends for each component for each relevant disease can be estimated. Results: For esophagus and stomach cancers the Northern provinces was the area of high risk. For colorectal cancer Gilan, Semnan, Fars, Isfahan, Yazd and East-Azerbaijan were the highest risk provinces. For bladder and lung cancer, the northwest were the highest risk area. For prostate and breast cancers, Isfahan, Yazd, Fars, Tehran, Semnan, Mazandaran and Khorasane-Razavi were the highest risk part. The smoking component, shared by esophagus, stomach, bladder and lung, had more effect in Gilan, Mazandaran, Chaharmahal and Bakhtiari, Kohgilouyeh and Boyerahmad, Ardebil and Tehran provinces, in turn. For overweight and obesity component, shared by esophagus, colorectal, prostate and breast cancers the largest effect was found for Tehran, Khorasane-Razavi, Semnan, Yazd, Isfahan, Fars, Mazandaran and Gilan, in turn. For low physical activity component, shared by colorectal and breast cancers North-Khorasan, Ardebil, Golestan, Ilam, Khorasane-Razavi and South-Khorasan had the largest effects, in turn. The smoking component is significantly more important for stomach than for esophagus, bladder and lung. The overweight and obesity had significantly more effect for colorectal than of esophagus cancer. Conclusions: The presented model is a valuable model to model geographical and temporal variation among diseases and has some interesting potential features and benefits over other joint models.

Motivation & Objective

• To develop a joint modeling framework that integrates spatio-temporal dynamics with multivariate disease mapping for improved inference on cancer incidence.
• To identify shared latent risk components that explain spatial and temporal variation across multiple cancer types in Iran.
• To enhance estimation precision for less prevalent cancers by borrowing strength across related diseases through shared components.
• To explore geographical and temporal patterns in cancer incidence beyond univariate or purely spatial models.
• To provide a flexible Bayesian hierarchical model capable of estimating the relative contribution of each shared component to individual cancers over time and space.

Proposed method

• Proposes a Bayesian hierarchical model where the relative risk for each cancer is decomposed into shared spatial and temporal components.
• Each shared component is associated with a subset of cancers and captures common spatial and temporal trends across those diseases.
• Spatial and temporal random effects are modeled using conditional autoregressive (CAR) and random walk priors, respectively, to account for spatial correlation and temporal dependence.
• The model estimates the relative weight (scaling parameter) of each shared component for each relevant cancer, allowing interpretation of component importance.
• Prior distributions are assigned to shared components and their variances, with full Bayesian inference via Markov Chain Monte Carlo (MCMC) sampling.
• Model comparison is performed using Deviance Information Criterion (DIC), and model fit is assessed through posterior predictive checks and residual analysis.

Experimental results

Research questions

• RQ1 How do spatial and temporal patterns of cancer incidence vary across seven prevalent cancer types in Iran from 2005 to 2009?
• RQ2 Which shared latent components explain common geographical and temporal trends across multiple cancers?
• RQ3 What is the relative contribution of each shared component to the incidence risk of individual cancers over time and space?
• RQ4 How does the joint modeling approach improve estimation precision and hotspot detection compared to univariate models?
• RQ5 What are the geographical clusters of high and low risk for each cancer, and how do they compare across disease pairs?

Key findings

• Sistan and Baluchestan province consistently showed the lowest relative risk for all seven cancers across all time periods.
• High-risk provinces included Razavi Khorasan, Semnan, Gilan, Mazandaran, Yazd, Isfahan, East Azerbaijan, Fars, West Azerbaijan, Kurdistan, Tehran, Ardebil, and Golestan.
• Esophagus and stomach cancers exhibited highly similar spatial patterns, as did bladder and colorectal cancers, and breast and prostate cancers.
• Lung cancer showed a distinct spatial pattern that diverged from all other cancers, indicating unique risk factors.
• Temporal effects of shared components were relatively stable over the five-year period, suggesting limited short-term change in underlying risk trends.
• The model demonstrated improved precision for less common cancers by borrowing information across related diseases through shared components.