Qian Liu1, Abigail R Smith1, Laura H Mariani1,2, Viji Nair2, Jarcy Zee3. 1. Arbor Research Collaborative for Health, Ann Arbor, Michigan; and. 2. Michigan Medicine, University of Michigan, Ann Arbor, Michigan. 3. Arbor Research Collaborative for Health, Ann Arbor, Michigan; and Jarcy.Zee@ArborResearch.org.
Abstract
BACKGROUND AND OBJECTIVES: Identifying novel biomarkers is critical to advancing diagnosis and treatment of CKD, but relies heavily on the statistical methods used. Inappropriate methods can lead to both false positive and false negative associations between biomarkers and outcomes. This study assessed accuracy of methods using computer simulations and compared biomarker association estimates in the NEPhrotic syndrome sTUdy NEtwork (NEPTUNE), a prospective cohort study of patients with glomerular disease. DESIGN, SETTING, PARTICIPANTS, & MEASUREMENTS: We compared three methods for analyzing repeatedly measured biomarkers in proportional hazards models: (1) time-invariant average, that averages values over all follow-up and uses the average as a baseline covariate, (2) time-varying last observation carried forward (LOCF), that assumes the covariate is unchanged until the next observed value, and (3) time-varying cumulative average, that updates the average using values at or before each measurement. RESULTS: Under both true mechanisms of LOCF and cumulative average, simulation results showed the time-invariant average method often gave extremely inaccurate results. When LOCF was the true association mechanism, the cumulative average method often gave overestimated association estimates that were further away from the null. When cumulative average was the true mechanism, LOCF always underestimated the associations, i.e., closer to the null. In NEPTUNE, compared with the LOCF or cumulative average methods, hazard ratios estimated from the time-invariant average method were always higher. CONCLUSIONS: Different analytic methods resulted in markedly different results. Using the time-invariant average produces inaccurate association estimates, whereas other methods can estimate additive (cumulative average) or instantaneous (LOCF) associations depending on the hypothesized underlying association mechanism and research question.
BACKGROUND AND OBJECTIVES: Identifying novel biomarkers is critical to advancing diagnosis and treatment of CKD, but relies heavily on the statistical methods used. Inappropriate methods can lead to both false positive and false negative associations between biomarkers and outcomes. This study assessed accuracy of methods using computer simulations and compared biomarker association estimates in the NEPhrotic syndrome sTUdy NEtwork (NEPTUNE), a prospective cohort study of patients with glomerular disease. DESIGN, SETTING, PARTICIPANTS, & MEASUREMENTS: We compared three methods for analyzing repeatedly measured biomarkers in proportional hazards models: (1) time-invariant average, that averages values over all follow-up and uses the average as a baseline covariate, (2) time-varying last observation carried forward (LOCF), that assumes the covariate is unchanged until the next observed value, and (3) time-varying cumulative average, that updates the average using values at or before each measurement. RESULTS: Under both true mechanisms of LOCF and cumulative average, simulation results showed the time-invariant average method often gave extremely inaccurate results. When LOCF was the true association mechanism, the cumulative average method often gave overestimated association estimates that were further away from the null. When cumulative average was the true mechanism, LOCF always underestimated the associations, i.e., closer to the null. In NEPTUNE, compared with the LOCF or cumulative average methods, hazard ratios estimated from the time-invariant average method were always higher. CONCLUSIONS: Different analytic methods resulted in markedly different results. Using the time-invariant average produces inaccurate association estimates, whereas other methods can estimate additive (cumulative average) or instantaneous (LOCF) associations depending on the hypothesized underlying association mechanism and research question.
Authors: Wenjun Ju; Viji Nair; Shahaan Smith; Li Zhu; Kerby Shedden; Peter X K Song; Laura H Mariani; Felix H Eichinger; Celine C Berthier; Ann Randolph; Jennifer Yi-Chun Lai; Yan Zhou; Jennifer J Hawkins; Markus Bitzer; Matthew G Sampson; Martina Thier; Corinne Solier; Gonzalo C Duran-Pacheco; Guillemette Duchateau-Nguyen; Laurent Essioux; Brigitte Schott; Ivan Formentini; Maria C Magnone; Maria Bobadilla; Clemens D Cohen; Serena M Bagnasco; Laura Barisoni; Jicheng Lv; Hong Zhang; Hai-Yan Wang; Frank C Brosius; Crystal A Gadegbeku; Matthias Kretzler Journal: Sci Transl Med Date: 2015-12-02 Impact factor: 17.956
Authors: Jacek Rysz; Anna Gluba-Brzózka; Beata Franczyk; Zbigniew Jabłonowski; Aleksandra Ciałkowska-Rysz Journal: Int J Mol Sci Date: 2017-08-04 Impact factor: 5.923
Authors: Crystal A Gadegbeku; Debbie S Gipson; Lawrence B Holzman; Akinlolu O Ojo; Peter X K Song; Laura Barisoni; Matthew G Sampson; Jeffrey B Kopp; Kevin V Lemley; Peter J Nelson; Chrysta C Lienczewski; Sharon G Adler; Gerald B Appel; Daniel C Cattran; Michael J Choi; Gabriel Contreras; Katherine M Dell; Fernando C Fervenza; Keisha L Gibson; Larry A Greenbaum; Joel D Hernandez; Stephen M Hewitt; Sangeeta R Hingorani; Michelle Hladunewich; Marie C Hogan; Susan L Hogan; Frederick J Kaskel; John C Lieske; Kevin E C Meyers; Patrick H Nachman; Cynthia C Nast; Alicia M Neu; Heather N Reich; John R Sedor; Christine B Sethna; Howard Trachtman; Katherine R Tuttle; Olga Zhdanova; Gastòn E Zilleruelo; Matthias Kretzler Journal: Kidney Int Date: 2013-01-16 Impact factor: 10.612
Authors: Conxita Jacobs-Cachá; Ander Vergara; Clara García-Carro; Irene Agraz; Nestor Toapanta-Gaibor; Gema Ariceta; Francesc Moreso; Daniel Serón; Joan López-Hellín; Maria José Soler Journal: Clin Kidney J Date: 2020-08-11