OUTCOME MEASURES IN STROKE
The Modified Rankin Scale and Ordinal Logistic Regression
Author: Else Charlotte Sandset, Department of Neurology, Oslo University Hospital, Norway
Since stroke is the second leading cause of disability worldwide, level of disability or dependence is naturally of interest in stroke research. There are many scales measuring different aspects of disability or dependence, however the most widely used to measure functional outcome is the modified Rankin Scale (mRS). The mRS is an ordinal scale that ranges from 0 (no symptoms at all) to 6 (death). Reliability to determine the patient’s mRS category score is satisfactory both when assessed by telephone interview and by consultation.1
Good outcome following stroke is commonly defined as scores 0 – 2 and poor outcome scores of 3 – 6. This allows for conventional statistics such as binary logistic regression where the results are presented intuitively as an odds ratio representing the risk of good versus poor outcome. Still, treatment or intervention where a patient goes from being bedridden (mRS score of 5) to being able to walk without assistance (mRS score of 3) is highly clinically relevant, and this information is lost when dividing the scale in two. Also, utilising the full scale of the mRS increases statistical sensitivity, so by using statistical methods taking the whole scale into consideration (ordinal methods) we are more likely to show that a treatment or intervention is effective.2
So what is an ordinal statistical method? Several methods are now increasingly used in stroke trials, one being ordinal logistic regression. In ordinal logistic regression analysis the results are presented as a common odds ratio which represents a shift in scores on the mRS. This means patients with mRS score of 0 is compared to patients with scores 1 – 6, 0 – 1 to 2 – 6, 0 – 3 to 4 – 6, 0 – 4 to 5 – 6 and 0 – 5 to 6 and get an overall point estimate – the common odds ratio. To simplify it, let us take a trial with common odds ratio of 0.80 for a treatment or intervention, this means if you have an mRS of 4, the treatment increases your chance of getting a mRS score of 0-3 by 20%.
Although many benefits, there are some issues, the main one being the lack of intuitive interpretation of the common odds ratio. Also, the proportion odds model assumes that treatment effect is consistent across all categories of the mRS, meaning that the OR is the same for mRS 0 – 1 vs 2 – 6 as for mRS 0 – 4 versus 5 – 6. This assumption can be tested using a “goodness-to-fit” test, and a report from «the Optimising Analysis of Stroke Trials (OAST) Collaboration» found that 85% of datasets examined did not significantly violate this assumption.3
The European Stroke Organisation Outcomes Working Group strongly encourages the use of ordinal methods both for calculating sample size for clinical trials and for analysing the results. 2
1 Wilson, J. T. et al. Improving the assessment of outcomes in stroke: use of a structured interview to assign grades on the modified Rankin Scale. Stroke 33, 2243-2246 (2002).
2 Bath, P. M. et al. Statistical analysis of the primary outcome in acute stroke trials. Stroke 43, 1171-1178, doi:STROKEAHA.111.641456 ;10.1161/STROKEAHA.111.641456 (2012).
3 Bath, P. M., Gray, L. J., Collier, T., Pocock, S. & Carpenter, J. Can we improve the statistical analysis of stroke trials? Statistical reanalysis of functional outcomes in stroke trials. Stroke 38, 1911-1915, doi:STROKEAHA.106.474080 ;10.1161/STROKEAHA.106.474080 (2007).
