Background: Claims databases might not capture or surrogate all the prognostic factors that influence health care delivery; thus, residual confounding might obscure causal effects. Screening sigmoidoscopy provides a relative reduction of 2% in the hazard of death. A reduction >10% should not be expected for screening colonoscopy.
Objectives: To estimate the amount of residual confounding in a scenario where causation of small magnitude, if any, is expected: screening colonoscopy and all‐cause mortality.
Methods: We emulated a target trial of one‐time screening colonoscopy in Medicare (2004‐2012) beneficiaries aged 70‐74. To adjust for potential baseline confounding, we used both a structural and a non‐structural approach. For the structural approach, we used subject‐matter knowledge to identify the potential confounders: age, race, sex, calendar year, reason for entitlement, use of preventive services, US Census Bureau division, comorbidity score, and history of 25 chronic conditions. We then used multivariate outcome regression, inverse probability of treatment weighting (IPW) and propensity score using deciles for adjustment. The non‐structural approach consisted on the use of a “high‐dimensional propensity score” (HDPS) to create indicators for adjustment using data dimensions (ICD‐9 diagnosis codes and CPT procedure codes from carrier claims, ICD‐9 diagnosis codes, ICD‐9 procedures codes and CPT procedure codes from outpatient and inpatient claims), which we added to the rest of confounders. The outcome model for all analyses was a pooled logistic regression model for monthly risk of death that included an indicator for screening strategy, a flexible function of follow‐up, and adjustment indicators, covariates or weights, as applicable.
Results: In the screening colonoscopy group, there were 46 872 individuals and 2592 deaths during a total of 174 230 follow‐up years. In the no screening group, there were 1 762 816 individuals and 162 298 deaths during a total of 6 362 550 follow‐up years. The unadjusted hazard ratio (HR) for all‐cause mortality was 0.58 (95% CI 0.55‐0.60). The adjusted HR was 0.69 (0.66‐0.72) using multivariate outcome regression, 0.70 (0.67‐0.73) using IPW, 0.69 (0.67‐0.72) using propensity score, and 0.72 (0.69‐0.74) using HDPS. The HDPS estimate did not change substantially under various choices of parameters for indicator selection.
Conclusions: Residual confounding can be intractable for the effect of preventive services on all‐cause mortality in Medicare.