Research

The added value of ordinal analysis in clinical trials: an example in traumatic brain injury

Bob Roozenbeek^1,2*, Hester F Lingsma², Pablo Perel³, Phil Edwards³, Ian Roberts³, Gordon D Murray⁴, Andrew IR Maas¹, Ewout W Steyerberg² and the IMPACT (International Mission on Prognosis and Clinical Trial Design in Traumatic Brain Injury) Study Group and the CRASH (Corticosteroid Randomisation After Significant Head Injury) Trial Collaborators

• * Corresponding author: Bob Roozenbeek b.roozenbeek@erasmusmc.nl

Author Affiliations

¹ Department of Neurosurgery, Antwerp University Hospital, Wilrijkstraat 10, 2650 Edegem, Belgium

² Department of Public Health, Erasmus MC, P.O. Box 2040, 3000 CA Rotterdam, The Netherlands

³ Epidemiology and Population Health Department, London School of Hygiene & Tropical Medicine, Keppel Street, London, WC1E 7HT, UK

⁴ Centre for Population Health Sciences, University of Edinburgh, Teviot Place, Edinburgh, EH8 9AG, UK

For all author emails, please log on.

Critical Care 2011, 15:R127 doi:10.1186/cc10240

Published: 17 May 2011

Abstract

Introduction

In clinical trials, ordinal outcome measures are often dichotomized into two categories. In traumatic brain injury (TBI) the 5-point Glasgow outcome scale (GOS) is collapsed into unfavourable versus favourable outcome. Simulation studies have shown that exploiting the ordinal nature of the GOS increases chances of detecting treatment effects. The objective of this study is to quantify the benefits of ordinal analysis in the real-life situation of a large TBI trial.

Methods

We used data from the CRASH trial that investigated the efficacy of corticosteroids in TBI patients (n = 9,554). We applied two techniques for ordinal analysis: proportional odds analysis and the sliding dichotomy approach, where the GOS is dichotomized at different cut-offs according to baseline prognostic risk. These approaches were compared to dichotomous analysis. The information density in each analysis was indicated by a Wald statistic. All analyses were adjusted for baseline characteristics.

Results

Dichotomous analysis of the six-month GOS showed a non-significant treatment effect (OR = 1.09, 95% CI 0.98 to 1.21, P = 0.096). Ordinal analysis with proportional odds regression or sliding dichotomy showed highly statistically significant treatment effects (OR 1.15, 95% CI 1.06 to 1.25, P = 0.0007 and 1.19, 95% CI 1.08 to 1.30, P = 0.0002), with 2.05-fold and 2.56-fold higher information density compared to the dichotomous approach respectively.

Conclusions

Analysis of the CRASH trial data confirmed that ordinal analysis of outcome substantially increases statistical power. We expect these results to hold for other fields of critical care medicine that use ordinal outcome measures and recommend that future trials adopt ordinal analyses. This will permit detection of smaller treatment effects.