For the past several decades, much progress has been made in the assessment of prognosis in various diseases. Formerly, prognosis was often graded using stages or adjectives (e. g., grave or favorable). Now it is usually assessed quantitatively in terms of probability of event or time to event and these outcome measures are statistically related to treatment and other prognostic factors. Unfortunately, it was found that even the results of some randomized controlled trials (RCT), which had been defended by evidence-based medicine, did not agree with the long-term clinical outcomes (e. g., Acute Leukemia Group B trial evaluating 6MP and Dutch Gastric Cancer Group trial evaluating extended lymphadenectomy). Such discrepancies may have resulted from the inappropriate use of statistical models. The purpose of this report was to review commonly used statistical models from the clinician's point of view, and to study whether they agree with the long-term observations, and whether their parameters provide useful information for clinicians and patients.
The statistical models we checked included the logistic model, probit model, and accelerated failure time model, with special reference to the Boag lognormal model with cured fraction and its extensions, and the Cox proportional hazards model and its extensions as well as simulation of RCT using the Boag model and independent competing risk model.
The results of our studies show that the most commonly used proportional hazards model does not necessarily derive useful information for clinicians and patients and may occasionally misguide them into favoring a suboptimal therapy if follow-up is terminated prematurely.
In conclusion, as far as statistical models are used to estimate prognosis, error is unavoidable. In order to minimize error, the validation of the model should not be left to statisticians alone. Instead, clinicians should participate in this process with life-long follow-up of patients.
View full abstract