Response to Letter by Miller and Palesch
Miller and Palesch question several aspects of the article by the Optimizing the Analysis of Stroke Trials (OAST) Collaboration.1 They suggest that the Friedman test (a nonparametric ANOVA) should have been used to assess differences between the statistical tests being compared; this is what we did when performing a 2-way ANOVA using ranked data rather than raw data, as stated in the Methods section (Comparison of Statistical tests) of the article.
Miller and Palesch also suggest that the level at which the modified Rankin Scale (mRS) is dichotomized should be chosen on the basis of expected risk and benefits. In reality, trialists tend to copy previous study designs so a cutpoint of 0 to 2/3 to 6 was fashionable several years ago until the National Institute of Neurological Disorders and Stroke (NINDS) rt-PA trial led a move to using 0,1/2 to 6.2 In reality, the ‘most efficient’ cutpoint is likely to be the one which splits the population roughly in equal halves so the choice should depend on stroke severity and prognosis. For example, trials of feeding and hemicraniotomy, which involve patients with very severe stroke, chose 0 to 3/4 to 6 and 0 to 4/5,6 respectively.3,4 However, as shown in the OAST article, ordinal approaches are generally more efficient than dichotomization.1
In regards of NXY-059 data, we cannot comment because the SAINT I Trial Steering Committee elected not to share their data with OAST (although they have with other groups such as VISTA). However, we note that the Cochrane-Mantel-Haenszel (CMH) test, as used in SAINT I, is not an ordinal test; rather, it compares 2 groups on a binary response, adjusting for control variables, and assuming a common odds ratio.
In regards of which approach to use, the statistical assumptions underlying the different tests must be considered. Importantly, the assumption of proportional odds is met in most stroke trials assessing functional outcome (with the notable exception of thrombolysis which increases both very good outcome and death5) so that ordinal regression is usually appropriate. The t test has been recommended for analyzing data from scales where there are 7 or more categories6 and is robust under a range of conditions.7 Finally, it is vital to consider which clinical assumptions are relevant to stroke trials; in general, we are interested in improving outcome irrespective of the underlying prognosis, and statistical tests which assess this will be more relevant than those which only see whether an intervention helps patients cross a particular boundary such as ‘poor outcome’ into ‘very good outcome’.
The Optimising Analysis of Stroke Trials (OAST) Collaboration. Can we improve the statistical analysis of stroke trials? Statistical re-analysis of functional outcomes in stroke trials. Stroke. 2007; 38: 1911–1915.
Vahedi K, Hofimeijer J, Juettler E, Vicaut E, George B, Algra A, Amelink GJ, Schmiedeck P, Schwab S, Rothwell PM, Bousser MG, van der Worp HB, Hacke W; DECIMAL, DESTINY, and HAMLET investigators. Early decompressive surgery in malignant infarction of the middle cerebral artery: a pooled analysis of three randomised controlled trials. Lancet Neurology. 2007; 6: 215–222.
Wardlaw JM, Zoppo G, Yamaguchi T, Berge E. “Thrombolysis for acute ischaemic stroke.” Cochrane Database Systematic Review. 2003; (3): CD000213.