Sensitivity of series system reliability estimation to the non-informative masking assumption

Maximum likelihood estimation for series systems with masked failure data relies on the non-informative masking assumption (Condition C2): that the candidate set reported for a failed component is independent of the true cause of failure. This assumption is rarely verifiable and frequently suspect. We characterize the bias that arises when C2 is violated but assumed to hold, proving that the MLE converges to a pseudo-true parameter whose individual component parameters are biased while the system hazard remains consistent. Using a Bernoulli masking model that generates controlled C2 violations of varying severity, we conduct a systematic simulation sweep on a 5-component Weibull series system. The results establish a practical breakdown boundary: system-level reliability estimates are robust to C2 violations across the full severity range, but individual component estimates degrade with severity, with scale parameters showing substantial bias (up to 40 percent relative error) under strong informative masking, compounded by occasional extreme outlier estimates driven by near-singular Fisher information. Since the masking mechanism and component parameters are jointly non-identifiable, we recommend sensitivity analysis over model elaboration. Software is available in the mdrelax R package.