Analysis of compliance factors for colorectal cancer screening using a Bayesian network

Abstract

Background: Compared to standard explanatory analyses based on multivariate regressions, Bayesian network analyses enable multiple hypotheses and clear graphical representations of complex interactions. They provide visual descriptions of causal pathways to distinguish between direct/indirect factors. We compared multivariate regression and a Bayesian network to assess factors associated with colorectal cancer (CRC) screening. Methods: The 5th French observational survey, EDIFICE 5, was conducted (Nov 22- Dec 7, 2016) by phone interviews of a representative sample of 1501 individuals (age, 50-75 y). The present analysis focuses on 1299 individuals with no history of cancer (50-74 y). Bayesian analysis was performed with the bnlearn R Package. Parameters of the Bayesian analysis were based on the literature and our own data (logistic regression). "Blacklist/whitelist"-type restrictions were used to reset current understanding of the correlations between variables. We also analyzed the network topology. Results: In our sample, 36% (N=469) declared never having undergone CRC screening (colonoscopy, fecal occult blood test) in their lifetime. The Bayesian model revealed 5 direct correlating factors: age, smoking status, social vulnerability, psychological reassurance in the screening test (PRST), and confidence in the efficacy of the test. The latter 2 account for 43% of the observed sum of the mutual informations. Other relevant factors typically seen in the literature and regression analysis had an indirect impact: level of education, self-perception of own risk of CRC, gender, temporal perspective, confidence in their physician and fear of the disease. Multiple regression analysis identified PRST (OR=0.84, 95% CI 0.80-0.88, P<0.01) and fear of the disease (OR=0.90, 95% CI 0.84-0.96, P<0.01) as the two main criteria. Conclusions: We showed that Bayesian network analysis provides a novel representation of factors associated with CRC screening, and may explain why interventions focusing on indirect factors might be ineffective if the next step of the causal pathway remains unchanged. We suggest that Bayesian networks should be used more often to drive timely interventions (short term vs long term).