USING OBSERVATIONAL DATA TO EMULATE RANDOMIZED TRIALS OF SURVEILLANCE INTERVENTIONS: AN APPLICATION TO COLORECTAL CANCER

Purpose:   To show how modern methods for causal inference can be used to determine the effectiveness and optimal implementation interval of a screening or surveillance test. 

Method: During the NORCCAP randomized trial on the effects of sigmoidoscopy screening,  2197 patients with adenomas were identified.  We have observational follow-up data on these patients, which includes information on when they underwent surveillance colonoscopies, and what findings were made. Undergoing a surveillance colonoscopy may result in lower colon cancer mortality due to earlier detection, and also lower incidence due to removal of pre-malignant lesions.   The magnitude of any such effect is not known.

 In this data set, we fit three causal models, including two models to determine the overall effects of undergoing a single surveillance colonoscopy, and a dynamic marginal structural model to determine the optimal interval between surveillance colonoscopies.  We estimate the effect of undergoing colonoscopy both on colon cancer mortality, and on colon cancer incidence.  When using colon cancer incidence as the outcome measure,  the analysis subject to lead time bias.   We discuss the potential impact and direction of such bias. In these models, it is sometimes necessary to control for historical findings at previous colonoscopies, a group of covariates that are only seen in those who underwent the intervention.  The method allows control for such covariates.  

Results: The analysis demonstrates how to implement the methodology, but is not very informative due to the small sample size. We estimate that undergoing a surveillance colonoscopy 5 years after an adenoma was removed, results in an incidence hazard ratio of 0.73 (95% CI: 0.22- 1.54) in the following 6 years. We also estimate that a dynamic strategy under which high risk patients undergo colonoscopies every 3 years, results in incidence hazard ratios of 0.69 (95% CI:  0.12 – 4.44) when compared to strategies where they undergo colonoscopies every 11 years.   The analysis with colon cancer mortality as the outcome was not informative due to the small data set.

Conclusions:   When randomized trials are not available, emulation of randomized controlled trials from observational data can be used to guide decision makers in writing more rational clinical guidelines for setting the interval between surveillance tests. When large data sets are available, this method is an attractive alternative to decision tree models.