Re: Randomized Controlled Study of Stroke Unit Versus Stroke Team Care in Different Stroke Subtypes
To the Editor:
In a recent article in Stroke, Evans et al examined the effect of stroke unit care by subtype and concluded that stroke units improve the outcome in patients with large-vessel infarcts but not in those with small-vessel disease.1 However, this conclusion appears to be based on a misinterpretation of the analysis and lack of essential details in the presentation of results.
The authors conclude that in small-vessel occlusions, “mortality and institutionalization … were influenced by age and stroke severity but not by management strategy.” This is based on an odds ratio for stroke unit care of 4.9 (95% CI 0.9 to 25.0) and probability value of 0.06. In contrast, for large-vessel infarcts, the conclusion is “the odds for dying or being institutionalized were also increased nearly 3-fold in stroke-team patients.” This is based on an odds ratio for stroke unit care of 2.8 (95% CI 1.3 to 6.2) and probability value of 0.01. The actual effect size for stroke unit care at 12-month follow-up is larger for patients with small-vessel disease, and yet because the probability value is larger than the arbitrary cut-off of 0.05, the authors conclude that the effect of stroke unit care in this subgroup is not clinically important.
The authors’ conclusion depends heavily on the arbitrary distinction between “significant” and “nonsignificant” results (based on whether P<0.05 or P>0.05). As recommended recently, the description of differences based on “statistical significance” is not acceptable—estimates (with 95% CIs) and exact probability values should be given for all comparisons.2 This would include all those in Table 3 in their article, where the odds ratio and 95% CI are omitted and the probability value merely described as “NS.”
The correct analysis is to use the complete data set (including both subtypes) to examine the strength of the evidence for interaction, ie, to examine whether the effect of stroke unit really differs between the 2 subgroups. The evidence for this interaction should be presented in the article, and only if this evidence is strong should claims be made about different effects of stroke unit care for those with large- and small-vessel disease.3
Given the clinical importance of the conclusion that stroke unit care offers no “significant” advantages for those with small-vessel disease, it is vital that these conclusions are based on good scientific practice. It is also crucial that adequate information be provided in every published article for the reader to be able to judge the clinical significance of the results and to assess the validity of the conclusions drawn.
The letter by Tilling and Wolfe raises some important issues on the subgroup analysis and interpretation of data as well as on the delicate balance between presenting details of statistical analysis versus presenting a clear clinical message. We believe that relevant details have been presented in the results and the interpretation of data are accurate and appropriate.
We agree that statistical significance should not be assessed on the grounds of a probability value alone, and more so in the case of a borderline probability value like that of 0.06. The subgroup analyses (small-vessel and large-vessel) presented in this article followed appropriate tests of interactions between strategy and vessel subtype in the complete data set as recommended.1 These interactions were significant, with the strength of significance decreasing across time. For example, strategy for the small-vessel group was not significant at 3 months, with no deaths on the stroke unit and only 1 death in stroke team (Fisher’s exact test P>0.40). The odds ratios for the strategy (stroke team versus stroke unit) by vessel-subtype interaction were 6.5 at 6 months and above 2.0 at 12 months (95% CI 0.7 to 61; P=0.10). (Considering the low power of the test of interaction, the threshold value for its significance is around 10%.) Hence the conclusion that stroke units may have limited effectiveness in patients with small-vessel disease was based not on isolated subgroup comparisons but on preliminary tests for interaction between strategy and subtype in the whole data set. Such tests should always precede subgroup analysis, as was done in this article.2
The style of writing the article reflects the aim to present relevant data and clear analyses in a familiar and “user friendly” format, which enables readers (mostly stroke practitioners) to judge the clinical significance of results, rather than to overburden them with precise details of all statistical analyses undertaken, many of which would be of little interest to clinicians. Hence, corresponding pair-wise comparisons of data and important comparisons with both CIs and probability values were presented in the article. We disagree with Tilling and Wolfe that enough information has not been given in Table 3 of the article. With the exception of the 12-month mortality, all estimates (with 95% CIs) and precise probability values have been given for all variables that were relevant to the model presented. The odds ratio and CIs for mortality at 12 months were 7.2 (95% CI 0.8 to 60; P=0.06), similar to those for mortality or institutionalization. The term nonsignificant has been used only when the probability approached unity.
We believe it is important that significant clinical messages are clearly presented and do not get lost in a morass of statistical computations, which detract from the main message and are of little interest to most clinicians. Tilling and Wolfe’s letter also appears to imply a concern that the findings of the article may deprive patients of specialist stroke care. They apparently have missed the fact that comparisons were undertaken between a stroke unit and a specialist stroke team, hence both groups were receiving specialist stroke care. This may be one of the reasons for the lack of difference in outcome in patients with small-vessel disease, in whom mortality and need for rehabilitation are limited, as discussed in the article. Furthermore, it is important to recognize the limitations of proven interventions to ensure that they are used appropriately and in the most cost-effective way. No one will argue that stroke unit care in the acute phase reduces mortality and specialist stroke care facilitates recovery.3 However, an unquestioned belief that stroke units may be the only method of providing specialist stroke care in the postacute phase to patients with low probability of death and limited disability may be misplaced and prove detrimental to the development of high-quality and cost-effective services for stroke patients.