Avoidable flaws in observational analyses: an application to statins and cancer

Abstract

The increasing availability of large healthcare databases is fueling an intense debate on whether real-world data should play a role in the assessment of the benefit-risk of medical treatments. In many observational studies, for example, statin users were found to have a substantially lower risk of cancer than in meta-analyses of randomized trials. Although such discrepancies are often attributed to a lack of randomization in the observational studies, they might be explained by flaws that can be avoided by explicitly emulating a target trial (the randomized trial that would answer the question of interest). Using the electronic health records of 733,804 UK adults, we emulated a target trial of statins and cancer and compared our estimates with those obtained using previously applied analytic approaches. Over the 10-yr follow-up, 28,408 individuals developed cancer. Under the target trial approach, estimated observational analogs of intention-to-treat and per-protocol 10-yr cancer-free survival differences were -0.5% (95% confidence interval (CI) -1.0%, 0.0%) and -0.3% (95% CI -1.5%, 0.5%), respectively. By contrast, previous analytic approaches yielded estimates that appeared to be strongly protective. Our findings highlight the importance of explicitly emulating a target trial to reduce bias in the effect estimates derived from observational analyses.

Figure 1.

Flowchart for selection of eligible individuals from CALIBER when emulating a target trial of statin therapy and cancer risk, 1999-2016. Numbers in parentheses represent unique individuals in each group. Counts of initiator and non-initiator individuals do not sum to the total number of eligible individuals because some eligible individuals contributed to both groups in different nested emulated trials.

Figure 2.

Standardized cancer-free survival curves comparing statin therapy with no statin therapy estimated by an observational analog to an (a) intention-to-treat analysis and (b) per-protocol analysis, CALIBER, 1999-2016.