| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Full Review Prediction versus aetiology: common pitfalls and how to avoid them Merel van Diepen 1 , Chava L. Ramspek 1 , Kitty J. Jager 2 , Carmine Zoccali 3 and Friedo W. Dekker 1 1 Department of Clinical Epidemiology, Leiden University Medical Center, Leiden, The Netherlands, 2 ERA-EDTA Registry, Academic Medical Center, University of Amsterdam, Amsterdam, The Netherlands and 3 CNR-IBIM Clinical Epidemiology and Pathophysiology of Renal Diseases and Hypertension, Reggio Calabria, Italy Correspondence and offprint requests to: Merel van Diepen; E-mail: m.van_diepen@lumc.nl ABSTRACT Prediction research is a distinct field of epidemiologic research, which should be clearly separated from aetiological research. Both prediction and aetiology make use of multivariable model- ling, but the underlying research aim and interpretation of re- sults are very different. Aetiology aims at uncovering the causal effect of a specific risk factor on an outcome, adjusting for confounding factors that are selected based on pre-existing knowledge of causal relations. In contrast, prediction aims at ac- curately predicting the risk of an outcome using multiple pre- dictors collectively, where the final prediction model is usually based on statistically significant, but not necessarily causal, asso- ciations in the data at hand. In both scientific and clinical practice, however, the two are often confused, resulting in poor-quality publications with lim- ited interpretability and applicability. A major problem is the frequently encountered aetiological interpretation of prediction results, where individual variables in a prediction model are attributed causal meaning. This article stresses the differences in use and interpretation of aetiological and prediction studies, and gives examples of common pitfalls. Keywords: aetiological research, causality, multivariable mod- elling, prediction research, risk prediction INTRODUCTION Traditionally, the main focus of clinical epidemiology studies has been uncovering the underlying causes of disease, or aeti- ology. Pioneering examples are John Snow’s investigations into the cause of the 19th-century cholera outbreak [1] and the results of Richard Doll and Austin Bradford Hill, providing evi- dence for tobacco smoking as a cause for lung cancer in the 1950s [2]. Recently, however, more and more scientific effort has been devoted to a different line of epidemiologic research, prediction research. Prediction models provide individual risk estimates and many examples of their application in practice to complement clinical reasoning exist, such as the Apgar score for determining the prognosis of newborns, prenatal testing in pregnant women for assessing the risk of Down’s syndrome and the Framingham risk score for cardiovascular events. In line with this trend, prediction of a wide spectrum of chronic kidney disease (CKD) outcomes—from development of early-stage CKD to mortality and morbidity on dialysis [3–12]—has been the topic of an increasing number of articles. If properly de- veloped and implemented, models for accurate outcome predic- tion and risk stratification could be helpful tools in clinical decision-making [13, 14] and thus potentially benefit the health and quality of life of kidney disease patients. Unfortunately, as a relatively underexposed and developing field of epidemiology, prediction research has proven to be quite prone to error. Despite an extensive amount of literature on prediction methodology [13, 15–26], the methods used in many prediction articles are not up to standard, and the quality of re- porting of methods and results is often poor [27–31]. Such ob- servations have led to the development of the Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) guideline [32, 33] for report- ing prediction studies, which has been adopted by many leading medical journals. Adherence to this guideline allows journals and readers to adequately assess the quality and usefulness of a prediction study, thereby reducing research waste [34]. One aspect that is not commonly addressed in methodo- logical articles and guidelines for prediction studies, however, is | | | | | | | | | V C The Author 2017. Published by Oxford University Press on behalf of ERA-EDTA. All rights reserved. ii1 Nephrol Dial Transplant (2017) 32: ii1–ii5 doi: 10.1093/ndt/gfw459 Advance Access publication 27 February 2017 Downloaded from https://academic.oup.com/ndt/article/32/suppl_2/ii1/3056968 by guest on 27 January 2022