Garcia-Albeniz X, Hsu J, Hernan MA. 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. Presented at the 35th ICPE International Conference on Pharmacoepidemiology & Therapeutic Risk Management; August 27, 2019. Philadelphia, PA. [abstract] Pharmacoepidemiol Drug Saf. 2019 Aug 20; 28(S2):476. doi: 10.1002/pds.4864

BACKGROUND: When using real world data to emulate a hypothetical target trial, eligible individuals are assigned to a treatment strategy at time zero (baseline). However, when the exposure strategy involves timing, the information available at baseline may be insufficient to determine an individual’s treatment strategy, that is, for some individuals, the baseline data could be consistent with more than one strategy. For example, in emulating a target trial of stopping vs. continuing annual mammograms among elderly women, many subjects have baseline data that could be consistent with either strategy.

OBJECTIVES: To compare two approaches for the assignment of individuals to treatment strategies at baseline: random assignment and cloning.

METHODS: The simplest way to allocate women to a screening strategy when their data are consistent with both strategies is to randomly choose at baseline one of the two strategies: “stop” or “continue” screening. A second option is cloning the eligible women and assigning one clone to each strategy. While both approaches are unbiased, cloning is expected to be more efficient. However, the relative efficiency of the two approaches has not been studied in real data. Under both approaches, we censored women when they stopped following the assigned strategy (i.e., when receiving a screening mammogram if assigned to “stop” and at month 14 after the last mammogram if assigned to “continue”). We estimated the breast cancer-specific mortality hazard ratio for “continue” vs “stop” screening using a pooled logistic regression that included the screening strategy, time and baseline covariates. To adjust for the potential selection bias due to censoring, the model was inverse probability weighted using weights that depended on baseline and time-varying variables.

RESULTS: 1,235,459 eligible women aged 70-74 years received a screening mammogram at baseline. Using the random assignment approach, we randomly assigned half of the women to each strategy: continue screening for 8 years vs. stop screening. The hazard ratio (95% CI) of breast cancer mortality was 0.74 (0.57-0.96) for continuing vs. stopping screening. When we repeated the procedure 100 times, the hazard ratios varied from 0.62 (0.48-0.79) to 0.94 (0.72-1.22), with an average hazard ratio of 0.79. Using the cloning approach, we created and assigned 1,235,459 clones to each strategy and the corresponding hazard ratio was 0.78 (0.64-0.96).

CONCLUSIONS: Cloning individuals is more efficient than randomly assigning them to strategies that have overlapping initial follow-up.

Share on: