CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: Treating ordinal outcomes as continuous quantities: When, why and how Authors:  Yilin Ning - National University of Singapore (Singapore)
Peh Joo Ho - National University of Singapore (Singapore)
Chuen Seng Tan - Saw Swee Hock School of Public Health (Singapore) [presenting]
Nathalie Stoer - Oslo University Hospital (Norway)
Ka Keat Lim - Duke-NUS Medical School (Singapore)
Hwee-Lin Wee - National University of Singapore (Singapore)
Mikael Hartman - National University of Singapore (Singapore)
Marie Reilly - Karolinska Institutet (Sweden)
Abstract: Ordinal variables are common in studies of patient care. Analysing such outcomes often utilizes the linear regression model to estimate the effect of an exposure or intervention of interest. The magnitude of the effect is quantified by the difference in mean ordinal scores of the two groups being compared, and this quantity is useful for the assessment of clinical significance. However, this approach may be inappropriate as it assumes the ordinal outcome is a proxy for the continuous scale but does not assess this assumption. We propose a new procedure using the cumulative link model to assess the proxy assumption and to estimate the difference in mean ordinal scores when appropriate. The procedure is applied to 5 subscales of fatigue measured using the Multidimensional Fatigue Inventory to investigate the effect of time since diagnosis on fatigue among breast cancer survivors. A statistically significant improvement over time since cancer diagnosis was found in the General Fatigue and Mental Fatigue scores, but only General Fatigue satisfied the proxy assumption. We can only draw conclusions on the magnitude of change in the General Fatigue score, which is expected to be 1-unit for every 6.5 additional years since diagnosis and clinical significance (i.e., a 2-unit difference) achieved at the 13-th year. The procedure offers a seamless way to assess both the statistical and clinical significance of an effect on ordinal outcomes when the proxy assumption is appropriate.