2015TrialsOpen access

Identification and evaluation of novel methods for the analysis of ordinal data in acute stroke trials in comparison to ordinal logistic regression

Aimie Nunn, Laura J. Gray

Open full text 0 citations

Abstract

Ordinal outcomes are common; however, selecting an appropriate analysis is often problematic. Acute stroke trials routinely measure dependency using an ordinal scale as the primary outcome. Historically trials have dichotomised ordinal scales to compare the proportion dependent across groups, which limits statistical power to detect an effect. The OAST Collaboration (2007) showed ordinal logistic regression increased statistical power. Alternative methodologies for ordinal data proposed since 2007 were compared to ordinal logistic regression (OLR). PubMed and conference proceedings were searched for novel methods for analysis of ordinal outcome scales in stroke. Identified statistical methods were applied to data from the International Stroke Trial which randomised 19,435 patients to receive no treatment, aspirin, heparin or aspirin and heparin. All identified methods were compared to OLR in detection of the effect of aspirin versus no aspirin. Two regression based techniques were described for the analysis of ordinal scales: partial proportional odds and adjacent categories models. Two methods were proposed using prognosis based cut points on ordinal scales: sliding dichotomy and trichotomy. Three absolute measures of treatment effect were identified: permutation test, number needed to treat (NNT) for ordinal data and win ratio. None of the methods assessed were observed to increase the power to find a treatment effect compared to OLR. None of the recently proposed methods appear to offer statistical advantage over OLR. However, reporting an absolute measure of effect such as the NNT alongside the results from OLR aids interpretability of the common odds ratio.

Open-access reader

About this research paper

What this paper is about

Ordinal outcomes are common; however, selecting an appropriate analysis is often problematic. Acute stroke trials routinely measure dependency using an ordinal scale as the primary outcome. Historically trials have dichotomised ordinal scales to compare the proportion dependent across groups, which limits statistical power to detect an effect. The OAST Collaboration (2007) showed ordinal logistic regression increased statistical power. Alternative methodologies for ordinal data proposed since 2007 were compared to ordinal logistic regression (OLR). PubMed and conference proceedings were searched for novel methods for analysis of ordinal outcome scales in stroke. Identified statistical methods were applied to data from the International Stroke Trial which randomised 19,435 patients to receive no treatment, aspirin, heparin or aspirin and heparin. All identified methods were compared to OLR in detection of the effect of aspirin versus no aspirin. Two regression based techniques were described for the analysis of ordinal scales: partial proportional odds and adjacent categories models. Two methods were proposed using prognosis based cut points on ordinal scales: sliding dichotomy and trichotomy. Three absolute measures of treatment effect were identified: permutation test, number needed to treat (NNT) for ordinal data and win ratio. None of the methods assessed were observed to increase the power to find a treatment effect compared to OLR. None of the recently proposed methods appear to offer statistical advantage over OLR. However, reporting an absolute measure of effect such as the NNT alongside the results from OLR aids interpretability of the common odds ratio.

Why it matters

A significance statement is not available in the OpenAlex record.

Key contribution

A contribution statement is not available in the OpenAlex record.

Method / approach

Method details are not available in the OpenAlex metadata.

Main findings

Findings are not separately available in the OpenAlex metadata.

Limitations

Limitations are not available in the OpenAlex metadata.

Applications

Application details are not available in the OpenAlex metadata.

Available abstract

Ordinal outcomes are common; however, selecting an appropriate analysis is often problematic. Acute stroke trials routinely measure dependency using an ordinal scale as the primary outcome. Historically trials have dichotomised ordinal scales to compare the proportion dependent across groups, which limits statistical power to detect an effect. The OAST Collaboration (2007) showed ordinal logistic regression increased statistical power. Alternative methodologies for ordinal data proposed since 2007 were compared to ordinal logistic regression (OLR). PubMed and conference proceedings were searched for novel methods for analysis of ordinal outcome scales in stroke. Identified statistical methods were applied to data from the International Stroke Trial which randomised 19,435 patients to receive no treatment, aspirin, heparin or aspirin and heparin. All identified methods were compared to OLR in detection of the effect of aspirin versus no aspirin. Two regression based techniques were described for the analysis of ordinal scales: partial proportional odds and adjacent categories models. Two methods were proposed using prognosis based cut points on ordinal scales: sliding dichotomy and trichotomy. Three absolute measures of treatment effect were identified: permutation test, number needed to treat (NNT) for ordinal data and win ratio. None of the methods assessed were observed to increase the power to find a treatment effect compared to OLR. None of the recently proposed methods appear to offer statistical advantage over OLR. However, reporting an absolute measure of effect such as the NNT alongside the results from OLR aids interpretability of the common odds ratio.

Key concepts: Ordinal regression, Ordinal data, Ordered logit, Ordinal Scale, Logistic regression, Statistics, Medicine, Mathematics

Related papers

Back to paper searchBrowse research topicsOriginal source
Identification and evaluation of novel methods for the analysis of ordinal data in acute stroke trials in comparison to ordinal logistic regression — Research Paper | ScholarLens