De-Mystifying the Clone-Censor-Weight Method for Causal Research Using Observational Data: A Primer for Cancer Researchers

Abstract

Background: Regulators and oncology healthcare providers are increasingly interested in using observational studies of real-world data (RWD) to complement clinical evidence from randomized controlled trials for informed decision-making. To generate valid evidence, RWD studies must be carefully designed to avoid systematic biases. The clone-censor-weight (CCW) method has been proposed to address immortal time and other time-related biases.

Methods: The objective of this manuscript is to de-mystify the CCW method for cancer researchers by describing and presenting its core components in an accessible and digestible format, using visualizations and examples from cancer-relevant studies. The CCW method has been applied in diverse settings, including investigations of the effects of surgery within a certain time after cancer diagnosis, the continuation of annual screening mammography, and chemotherapy duration on survival.

Results: The method handles complex data wherein the treatment group to which an individual belongs is unknown at the start of follow-up. The three steps of the CCW method involve cloning or duplicating the patient population and assigning one clone to each treatment strategy, artificially censoring the clones when their observed data are inconsistent with the assigned strategy and weighting the cloned and censored population to address selection bias created by the artificial censoring.

Conclusions: The CCW method is a powerful tool for designing RWD studies in cancer that are free from time-related biases and successfully, to the extent possible, emulate features of a randomized clinical trial.

Keywords: cancer screening; cancer treatment; causal inference; observational study; target trial emulation.

FIGURE 1

Visual depiction (A) and associated dataset (B) for the motivating example. In this example, older adults diagnosed with early‐stage lung cancer are followed for up to 365 days to assess the outcome of overall survival. The two treatment groups are surgery within 182 days of diagnosis and no surgery within 182 days of diagnosis. Patients 1 and 2 underwent surgery at some point during follow‐up; Patients 3–5 did not undergo surgery. Patients 1, 3, and 5 died within 365 days of diagnosis; Patients 2 and 4 were followed up until the end of the study (365 days) without an event.

FIGURE 2

Visual depiction of the cloning and censoring components of the CCW method. When using the CCW method, patient data are duplicated, and clones are assigned to receive either treatment A (surgery within 182 days of diagnosis) or treatment B (no surgery within 182 days) (A). Data are reviewed and patients are then artificially censored when they deviate from their assigned treatment strategy (B). For example, the treatment B clone for Patient 1 is artificially censored when Patient 1 undergoes surgery within 182 days. The treatment A clone for Patient 2 is censored at the end of 182 days because it does not undergo surgery during this time window; however, the Patient 2 treatment B clone is followed until the end of the study period (even though they underwent surgery after 182 days, that is consistent with treatment B).

FIGURE 3

Visualization of the third component of the CCW method after hypothetical inverse probability weighting is applied (A) and the final analytic dataset (B). Panel A shows that for each treatment arm, compliant individuals (i.e., those who remain under follow‐up and not artificially censored) are upweighted to represent individuals in their treatment group who are artificially censored. Weighting can be implemented via inverse probability of treatment weighting in the original data (before duplication) or via inverse probability of censoring weighting in the duplicated data. For example, Patient 1 assigned to treatment A (surgery) is upweighted after Patients 2, 4, and 5 are artificially censored for not undergoing surgery within 182 days. Patients 2–5 assigned to treatment B (no surgery) are upweighted after Patient 1 undergoes surgery and deviates from the assigned treatment strategy. Panel B shows the final analytic dataset structure after applying the weights, with a new person‐time row for an individual clone anytime their weight changes. In this example, the impact of surgery within 182 days of diagnosis on 1‐year survival is being assessed. All individuals are followed for up to 1 year after diagnosis. Appendix 2 provides more detail about the weight calculations for a mock patient.

FIGURE 4

Study flow diagram reporting the core components of the CCQ method for the motivating example by Maringe et al.