Measurement error and bias in real-world oncology endpoints when constructing external control arms.

While randomized controlled trials remain the reference standard for evaluating treatment efficacy, there is an increased interest in the use of external control arms (ECA), namely in oncology, using real-world data (RWD). Challenges related to measurement of real-world oncology endpoints, like progression-free survival (PFS), are one factor limiting the use and acceptance of ECAs as comparators to trial populations. Differences in how and when disease assessments occur in the real-world may introduce measurement error and limit the comparability of real-world PFS (rwPFS) to trial progression-free survival. While measurement error is a known challenge when conducting an externally-controlled trial with real-world data, there is limited literature describing key contributing factors, particularly in the context of multiple myeloma (MM). We distinguish between biases attributed to how endpoints are derived or ascertained (misclassification bias) and when outcomes are observed or assessed (surveillance bias). We further describe how misclassification of progression events (i.e., false positives, false negatives) and irregular assessment frequencies in multiple myeloma RWD can contribute to these biases, respectively. We conduct a simulation study to illustrate how these biases may behave, both individually and together. We observe in simulation that certain types of measurement error may have more substantial impacts on comparability between mismeasured median PFS (mPFS) and true mPFS than others. For instance, when the observed progression events are misclassified as either false positives or false negatives, mismeasured mPFS may be biased towards earlier (mPFS bias = -6.4 months) or later times (mPFS bias = 13 months), respectively. However, when events are correctly classified but assessment frequencies are irregular, mismeasured mPFS is more similar to the true mPFS (mPFS bias = 0.67 months). When misclassified progression events and irregular assessment times occur simultaneously, they may generate bias that is greater than the sum of their parts. Improved understanding of endpoint measurement error and how resulting biases manifest in RWD is important to the robust construction of ECAs in oncology and beyond. Simulations that quantify the impact of measurement error can help when planning for ECA studies and can contextualize results in the presence of endpoint measurement differences.