Effect of Time Measurement Error on Assessing Treatments with Time Dependent Effect

Abstract

In the emergency medicine setting, it is often difficult to accurately record the time of injury without measurement error. If the treatment effect varies depending on the time from injury to treatment initiation, this type of measurement error may affect our analysis of the treatment effect. A Phase III trial for Tranexamic acid (TXA) in trauma patients with significant hemorrhage, CRASH-2, suggested a significant association between the time from injury to treatment initiation and treatment effect in that the benefit of TXA treatment reduces as the time to treatment initiation gets longer. TXA is a drug used to prevent fibrinolysis and reduce surgical blood loss for patients with major trauma, with traumatic brain injury (TBI) or after surgery. In this work, I study the effect measurement error might have on a combined analysis of two ongoing clinical trials of TXA in TBI. One of these trials is a Phase II clinical trial conducted by the Resuscitation Outcomes Consortium (ROC) in the United States and Canada. It is a double blind, randomized placebo controlled clinical trial evaluating the efficacy of two dosing regimens of TXA in patients with TBI in a pre-hospital setting. The trial plans to enroll about 1,000 patients. The second trial, CRASH-3, is a Phase III double blind, randomized, placebo controlled multi-national trial of TXA in patients with TBI. This trial plans to enroll about 13,000 patients. For the purpose of this study, I define a time dependent effect as a treatment effect that varies as times from injury to treatment initiation vary. Any observed treatment effect may be smaller or larger than the true effect due to time measurement error and the time to treatment initiation may be underestimated or overestimated. For treatments with potential time dependent effect like TXA, I investigate in this study whether and how much their estimated treatment effect may be affected by measurement error using simulations. I use logistic regression to fit models of interest with and without measurement error, with and without interaction term and compare power, (for some model comparisons) bias and Type I error rates. I first investigated the effect of measurement error under the setting resembling the ROC TXA trial, and then under other more general settings. Finally, I also investigated the potential effect of measurement error for the dataset combining the simulated CRASH-3 data and simulated ROC TXA data using meta-analysis. The results were consistent with our expectation that the measurement error could reduce power for detecting treatment and interaction effect and increase estimation bias for treatment effect at time zero and interaction term and this impact of measurement error was only associated with the strength of absolute measurement error. However, if our assumptions are appropriate, an average 0.5 hour absolute measurement error in the ROC TXA trial does not meaningfully impact our analysis results and an average 1 hour absolute measurement error in the meta-analysis (combining simulated CRASH-3 and ROC TXA data) can still generate a power above 80% to detect both treatment and interaction effect.