From the past (the E5 xoma trial):
Statistical Analysis (as explained in the full article of the E5 xoma sepsis trial, between parentheses are my comments).
Statistical analysis of the study followed a plan prepared during study design , which specified the methodology to be used and groups to be analyzed, including division of patients into shock and nonshock populations (meningo trial is all shock, a more difficult group).
All tests were two tailed. (like my recent examples).
Efficacy measures were survival over the 30 day study period and resolution of organ failures (meningo includes follow up to 90 days, again more difficult, but looking for different sequelae).
Survival was computed using the method of Kaplan-Meir and included all patient deaths . Survival differences between the treatment groups were tested by the Cox proportional hazards model, adjusting for prognostic variables identified in the analysis plan: age and the presence of organ failures like ARDS, ARF, DIC at entry. (here is probably easier for the analysis of the meningo trial due to lack of preconditions that could cause death aside from meningococcemia, since all are young children).
Relative risk was calculated using established techniques.(references #31. Cox, Regression models and life tables... and #32. Breslow, Covariance analysis of censored survival analysis...)
Mortality at days 14 an 30 is provided solely for descriptive purposes to facilitate comparison with previous studies but was not used for analysis.
An interim analysis, detailed in the analysis plan, was conducted. The level of significance for the stopping rule on reduction in mortality was P less than 0.005, which was not achieved.(reference for the "stopping rule" #33 from Miller, RG Jr. Simultaneous Statistical Inference. 2nd ed. New York, NY John Wiley & Sons Inc; 1981:8
Anyway, I have no time to look for that book, but looking at the
P less than 0.005, look well there are two zeros there, not one, it seems to be a tough cookie to chew, no wonder the study was not stopped).
Applying Bon-ferroni's rule, the requisite value for significance was
P=0.045. A sample size of 454 patients was calculated based on 80% power (like my examples) at this significance level for detection of a 50% reduction in mortality in patients with gram-negative sepsis.
Clinical symptoms were analyzed by evaluating investigator assessment of findings as normal or abnormal. Symptom data were analyzed using shift tables of changes from baseline and frequencies of values within, above and below normal ranges.
Shift tables of frequencies were analyzed by Chi Square analyses for individual days, and summary statistics were generated using Cochrane-Mantel-Haenzel methodology, controlling for baseline (now I really need more teeth to chew this cookie). The means for each group were compared using Student's t test.
(all for entertainment purposes, the poster does not claim full understanding of the above, no disclaimers apply, nothing else to do but to wait for news). |
