Endpoints in intensive care unit based randomized clinical trials

With few exceptions, intensive care unit (ICU)-based randomized clinical trials (RCTs) have failed to demonstrate hypothesized treatment effects. Undoubtedly, some of these failures are attributable to interventions that truly do not provide hoped-for benefits. However, this dissertation pursues the thesis that many null findings represent “false negatives” that are due not to ineffective therapies but to flawed study designs or analytic approaches. We examine the design and statistical methods traditionally employed in ICU-based RCTs, and their potential impacts on the efficient measurement and interpretation of treatment effects. Paper one presents a systematic review of 146 contemporary ICU-based RCTs in which we find that most trials were underpowered to detect small but potentially important mortality differences between treatment arms. We also find that the majority of RCTs (73%) specified primary outcomes other than mortality, that trials employing nonmortal primary outcomes more frequently identified significant treatment effects, and that both mortal and nonmortal endpoints were heterogeneously defined, measured and analyzed across RCTs. Thus, papers two and three focus on nonmortal endpoints, using ICU length of stay (LOS) as a case study to evaluate how best to measure and analyze duration-based nonmortal endpoints. In paper two, we conduct a statistical simulation study, demonstrating that nonmortal endpoints are interlinked with and confounded by mortality, and that the manner in which investigators choose to account for deaths in LOS analyses may influence their conclusions. In paper three, we examine another potential source of error in LOS analyses, namely the measurement error attributable to the additional ICU time that patients commonly accrue after they are clinically ready for ICU discharge. Using simulated data informed by our own ICU-based RCT, we demonstrate that this “immutable time” (which cannot plausibly be altered by the interventions under study) combines with clinically necessary ICU time to produce overall LOS distributions that may either mask true treatment effects or suggest false treatment effects. Our work provides evidence of the potential benefits and pitfalls when employing nonmortal outcomes in ICU-based RCTs, and also identifies a clear need for standardized methods for defining and analyzing such outcomes