Search

Yilin Ning,1,2 Peh Joo Ho,3,4 Nathalie C Støer,5,6 Ka Keat Lim,7,8 Hwee-Lin Wee,3,9 Mikael Hartman,1– 3,10 Marie Reilly,11 Chuen Seng Tan3 1NUS Graduate School for Integrative Sciences and Engineering, National University of Singapore, Singapore, Singapore; 2Yong Loo Lin School of Medicine, Department of Surgery, National University of Singapore and National University Health System, Singapore, Singapore; 3Saw Swee Hock School of Public Health, National University of Singapore and National University Health System, Singapore, Singapore; 4Genome Institute of Singapore, Singapore, Singapore; 5Norwegian National Advisory Unit on Women’s Health, Oslo University Hospital, Oslo, Norway; 6Department of Research, Cancer Registry of Norway, Oslo, Norway; 7Department of Population Health Sciences, School of Population Health & Environmental Sciences (SPHES), Faculty of Life Sciences & Medicine, King’s College London, London, UK; 8Programme in Health Services & Systems Research, Duke-NUS Medical School, Singapore, Singapore; 9Department of Pharmacy, Faculty of Science, National University of Singapore, Singapore, Singapore; 10Department of Surgery, National University Hospital, Singapore, Singapore; 11Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Stockholm, SwedenCorrespondence: Chuen Seng TanSaw Swee Hock School of Public Health, National University of Singapore and National University Health System, Tahir Foundation Building, 12 Science Drive 2, #10-01, Singapore, 117549, SingaporeTel +65-66013206Email ephtcs@nus.edu.sgPurpose: Assessing the clinical importance of an exposure effect on a quality of life (QoL) score often requires quantifying the effect in terms of a difference in scores. Using the linear regression model (LRM) for this purpose assumes the ordinal score is a proxy for an underlying continuous variable, but the analysis offers no assessment for the validity of the assumption. We propose an approach that assesses the proxy assumption and estimates the exposure effect by using the cumulative link model (CLM).Patients and methods: CLM is a well-established regression model that assumes an ordinal score is an ordered category generated from applying thresholds to a latent continuous variable. Our approach assesses the proxy assumption by testing whether these thresholds are equidistant. We compared the performance of CLM and LRM using simulated ordinal data and illustrated their application to the effect of time since diagnosis on five subscales of fatigue among breast cancer survivors measured using the Multidimensional Fatigue Inventory.Results: CLM had good performance in estimating the difference in means with simulated ordinal data satisfying the proxy assumption, even when the outcome had only a few categories. When the proxy assumption was inadequate, both the CLM and LRM had biased estimates with poor coverage. The proxy assumption was appropriate for four of the five subscales in our real data application to fatigue scores, which highlighted the importance of assessing the proxy assumption to avoid reporting invalid estimates in terms of the difference in scores.Conclusion: The proxy assumption is critical to the interpretation of the exposure effect on the difference in mean QoL scores. CLM offers a valid test for the presence of an association, a method for assessing the proxy assumption, and when the assumption is adequate, an assessment for clinical significance using the difference in means.Keywords: cumulative link model, ordered probit model, ordinal outcome, ordinal regression, probit link, quality of life