Assigning Treatment Strategy Using Real-World Data

A Challenge of Real World Data: How to Assign Individuals to a Treatment Strategy When Their Data Are Consistent With Several Treatment Strategies at Baseline

Garcia-Albeniz X, Hsu J, Hernan MA


When using observational data, quantifying the effect of treatment strategies that are sustained over time is not straightforward, because only people who live for a long time can receive treatment for a long time. 

A way to appropriately manage this potential for immortal time bias consists of, first, cloning your observational data set into as many exposure groups as you want to study, second, censoring those clones when they deviate from their assigned strategy and, third, using inverse probability weighting to adjust for the potential of selection introduced by that artificial censoring. This is a well-established approach, which has been used for over a decade in fields like HIV, cardiovascular, and cancer research. 

In this study we explored an alternative to cloning, based on randomly allocating people to either strategy. Censoring and adjusting for bias remained unchanged. We used the emulation of a target trial of breast cancer screening using Medicare claims data: we estimated, in more than a million women over 70 years old who were just screened for breast cancer, what was the effect of stopping screening vs. continuing screening for 8 more years on breast cancer-specific mortality.

We found that random assignment can be a valid alternative to cloning at baseline. The tradeoff is more variability of the point estimate and wider confidence intervals. 

Read more about the research at this link.