Practical methods for ordinal data meta-analysis in stroke

When randomised trials are conducted to see whether treatments work, these frequently have outcome measures that are ordinal scales (such as high, moderate, low or no pain), rather than binary measures (such as dead or alive) or continuous measures (such as an exact measure of blood pressure). There are too many trials performed for anyone to read all of the reports of their results, so the results of groups of trials are summarised using systematic reviews. In these reviews, ordinal scales are often analysed as if they were binary ? so high, medium and low levels of pain might be combined together and the pain scale analysed as any pain versus no pain. This is equivalent to throwing away 30% of the data. However, there are better ways of analysing ordinal scales. We wish to review this methodology. We then wish to investigate how practical each of the methods is by assessing how often sufficient data are presented to use the method, how often the available data fulfil any statistical rules that are needed for the methods to work, how easy to understand the results are, and how much detail they show of the way the treatment effect operates. If we can utilise the extra statistical power held in the ordinal scales, then guideline authors, clinicians and healthcare users will benefit, as they will be able to learn whether treatments are beneficial more quickly, using fewer patients, fewer trials, and less money.
We aim to do this work using stroke as an example. Ordinal outcome scales are very common in stroke trials, and there are several scales that are consistently used in many trials. Thus there will be a substantial amount of data for us to use, on outcomes that it makes clinical sense to combine. In addition, the analysis of ordinal scales in stroke trials provides some interesting examples of treatment effects that work in different ways (?kill or cure? treatments as compared with treatments that improve all patients? outcomes systematically), and we wish to see whether ordinal methods can highlight the differences between these. We have easy access to all the original reports of trials in stroke from a local trials register. This register is incredibly comprehensive, and structured so that we can easily find all the information relating to any one trial. Although we will use stroke data, our findings will be usable in other disease areas.

Randomised trials frequently have ordinal outcome measures, yet currently, in systematic reviews, these are mainly dichotomised and analysed as binary data. This is equivalent to throwing away 30% of the data. In addition, individuals who fall close to, but on different sides of the cut-point used to dichotomise, will be assumed by the analysis to be very different, yet they are likely to be very similar. Finally, if data are not dichotomised, then any the relationships between variables can be fully explored, for example by exploring non-linearity. We wish to review the available methodology for the meta-analysis of ordinal data. We then wish to investigate how each of the methods could be used in practice by assessing how often sufficient data are presented (or can be obtained), how often the available data fulfil any distributional assumptions (and indeed whether sufficient data are presented to check these assumptions), how easy to understand the results are, and how much detail they show of the way the treatment effect operates. If we can utilise the extra statistical power held in the ordinal scales, then guideline authors, clinicians and healthcare users will benefit, as they will be able to learn whether treatments are beneficial more quickly, using fewer patients, fewer trials, and less money. We will communicate the results of the above work by developing a workshop on ordinal methods to present at a Cochrane Colloquium.
We aim to do this work using stroke as an example. Ordinal outcome scales are very common in stroke trials, and there are several scales that are consistently used in many trials. Thus there will be a substantial amount of data for us to use, on outcomes that it makes clinical sense to combine. In addition, the analysis of ordinal scales in stroke trials provides some interesting examples of treatment effects that work in different ways (?kill or cure? treatments as compared with treatments that improve all patients? outcomes systematically), and we wish to see whether ordinal methods can highlight the differences between these. We have easy access to all the original reports of trials included in Stroke Group reviews. The stroke group trials register is incredibly comprehensive, and structured so that all references in it are linked to the study that they were based on. Our findings will be translatable to other therapeutic areas.