Abstract
Background and purpose: Arguments persist as to whether parametric or nonparametric methods should be used to analyse ordinal data in trials. This paper aims to assess methods used for presenting and analysing an ordinal scale, the Barthel Index, in trials of poststroke interventions.
Methods: All randomized controlled trials (RCTs) of poststroke interventions published from 1995 to 2004 in two journals (Strokeand Clinical Rehabilitation) were scrutinized for methods used to present and analyse Barthel scores. Computer simulations were used to compare the type I errors and the statistical power of different statistical methods under a range of assumed circumstances.
Results: One hundred and fifty-six RCTs were identified within the two journals. The central tendency of Barthel scores was measured by the median in 47 trials and by the mean in 35 trials. Non-parametric analyses of Barthel scores were conducted in 47 trials and parametric methods used in 18 trials. The results of computer simulations demonstrate that the t-test has a similar type I error rate and statistical power when compared with the rank sum test. However, when a zero final Barthel score is assigned to patients who have died, the statistical power of the t-test is much reduced. The possible maximal statistical power of dichotomization and ordinal regression is usually much lower than that of the rank sum test.
Conclusions: To facilitate comparison and meta-analysis, we recommend that mean values (with standard deviations or standard errors) of Barthel scores should be routinely reported in trials of poststroke interventions. The rank sum test appears the most powerful inferential technique for detecting differences in Barthel scores.
Methods: All randomized controlled trials (RCTs) of poststroke interventions published from 1995 to 2004 in two journals (Strokeand Clinical Rehabilitation) were scrutinized for methods used to present and analyse Barthel scores. Computer simulations were used to compare the type I errors and the statistical power of different statistical methods under a range of assumed circumstances.
Results: One hundred and fifty-six RCTs were identified within the two journals. The central tendency of Barthel scores was measured by the median in 47 trials and by the mean in 35 trials. Non-parametric analyses of Barthel scores were conducted in 47 trials and parametric methods used in 18 trials. The results of computer simulations demonstrate that the t-test has a similar type I error rate and statistical power when compared with the rank sum test. However, when a zero final Barthel score is assigned to patients who have died, the statistical power of the t-test is much reduced. The possible maximal statistical power of dichotomization and ordinal regression is usually much lower than that of the rank sum test.
Conclusions: To facilitate comparison and meta-analysis, we recommend that mean values (with standard deviations or standard errors) of Barthel scores should be routinely reported in trials of poststroke interventions. The rank sum test appears the most powerful inferential technique for detecting differences in Barthel scores.
Original language | English |
---|---|
Pages (from-to) | 347-356 |
Number of pages | 10 |
Journal | Clinical Rehabilitation |
Volume | 20 |
Issue number | 4 |
DOIs | |
Publication status | Published - 2006 |