Identifying Confounders Using Bayesian Networks and Estimating Treatment Effect in Prostate Cancer With Observational...

Randomized controlled trials are considered the golden standard for estimating treatment effect but are costly to perform and not always possible. Observational data, although readily available, is sensitive to biases such as confounding by indication. Structure learning algorithms for Bayesian Networks (BNs) can be used to discover the underlying model from data. This enables identification of confounders through graph analysis, although the model might contain noncausal edges. We propose using a blacklist to aid structure learning in finding causal relationships. This is illustrated by an analysis into the effect of active treatment (v observation) in localized prostate cancer.

In total, 4,121 prostate cancer records were obtained from the Netherlands Cancer Registry. Subsequently, we developed a (causal) BN using structure learning while precluding noncausal relations. Additionally, we created several Cox proportional hazards models, each correcting for a different set of potential confounders (including propensity scores). Model predictions for overall survival were compared with expected survival on the basis of the general population using data from Statistics Netherlands (Centraal Bureau voor de Statistiek).

Structure learning precluding noncausal relations resulted in a causal graph but did not identify significant edges toward treatment; they were added manually. Graph analysis identified year of diagnosis and age as confounders. The BN predicted a treatment effect of 1 percentage point at 10 years. Chi-squared analysis found significant associations between year of diagnosis, age, stage, and treatment. Propensity score correction was successful. Adjusted Cox models predicted significant treatment effect around 3 percentage points at 10 years.

A blacklist in conjunction with structure learning can result in a causal BN that can be used for confounder identification. Treatment effect found here is close to the 5 percentage point found in the literature.