Sliding Dichotomy. In the late 1990s researchers began questioning if the reason phase III clinical trials were failing to find successful treatments for ailments such as stroke and traumatic brain injury (TBI) was not due to the absence of successful treatments but rather to poorly defined or specified outcome measures resulting in under-powered trials [Barer, 1998, Xxxx et al., 1999]. The reasoning was the patient populations were so heterogenous, that not only were analyses often inadequately adjusting for these baseline differences but that outcome measures themselves needed to be adjusted to patient characteristics. In the TBI literature this resulted in the suggestion of a ‘sliding dichotomy’ by Xxxxxx et al [2005], which is an elaboration of Barer’s 1998 suggestion. The idea is to develop a prognostic score for individual patients, typically based on covariates such as age, medical history, and illness severity, and use this prognostic score to define new endpoints. Rather than going as far as patient-specific endpoints, Xxxxxx et al suggest dividing patients into ‘bands,’ each of which will have its own definition of a ‘good’ outcome (see figure 1.1). Traditionally, three ‘bands’ or groups are chosen based on tertiles of predicted probability of a favorable outcome. This approach is clinically appealing, as it reflects the genuine assessment of a patient that occurs in a hospital and is factored into treatment decisions. It also leads to the possibility of better powered (i.e., more efficient) clinical trials. Currently, the standard in TBI or stroke research is to design a phase III clinical trial such that it has 80% to 90% power to detect a treatment effect resulting in a 10 percentage point increase in good outcomes. However, this assumes that every patient enrolled in the trial has an equal probability of improving past some threshold to a ‘good’ outcome and this is simply clinically untrue [Xxxxxx et al., 2005]. There are two ways that the sliding dichotomy can be used to increase the power of a study to detect a significant treatment effect, and therefore reduce the number of patients that must be recruited. One option, suggested by Xxxxxxx et al [1999] is simply to only recruit patients with an ‘intermediate’ prognosis of good outcome, eliminating those at either extreme end of the spectrum. This certainly more closely resembles the equal probability of a good outcome implied by the standard design described above, but it also reduces the generalizability of ...
