Study
NLST
(Learn more about this study)
Project ID
NLST-1459
Initial CDAS Request Approval
Jul 22, 2025
Title
Estimating the effect of cancer screening on those who would be diagnosed if screened: an application of the Treatment Reactive Average Causal Effect
Summary
Researchers are often interested in the effect of treatment on an outcome among a particular subgroup of the population defined by a post-treatment variable of interest. For example, in estimating effects of early screening for cancer in RCT settings, we would most likely want to know the effect of screening on mortality among the group of people who would be diagnosed with cancer. Simply conditioning on cancer diagnosis is problematic as it introduces statistical biases. Hazlett et. al. (2025) propose an estimand for these settings, coined the Treatment Reactive Average Causal Effect (TRACE), and propose a partial identification approach for this estimand. By reasoning about the effect among the group that would not be diagnosed if screened, we can identify and estimate the range of plausible values for the TRACE. We detail the use of this approach in the cancer screening setting, to estimate the effect of screening on mortality among those who, if screened, would be diagnosed with cancer. We compare the TRACE estimand with other possible estimands in this setting, and discuss the benefits and drawbacks of our proposed estimand.
Aims
- Detail the use of the Treatment Reactive Average Causal Effect (TRACE) and the partial identification approach in the cancer screening setting, focusing on the different assumptions that can be made about the non-reactive group
- Estimate the effect of cancer screening on mortality among those who, if screened, would be diagnosed with cancer
- Show the bias that occurs when estimating the effect of cancer screening on mortality among those who were diagnosed (conditioning on observed diagnosis)
- Highlight the use of the TRACE in this particular setting, and the importance of focusing on this particular quantity rather than conditioning on observed diagnosis
- Compare the TRACE estimand with other possible estimands in this setting
Collaborators
