American Journal of Epidemiology Copyright 2005 by the Johns Hopkins Bloomberg School of Public Health All rights reserved

Vol. 161, No. 1 Printed in U.S.A. DOI: 10.1093/aje/kwi004

Tu et al. Respond to Barker Meets Simpson

Yu-Kang Tu1,2, George T. H. Ellison3, Robert West1, and Mark S. Gilthorpe1

1 2

Biostatistics Unit, Centre for Epidemiology and Biostatistics, University of Leeds, Leeds, United Kingdom. Leeds Dental Institute, University of Leeds, Leeds, United Kingdom. 3 St. Georges Hospital Medical School, London, United Kingdom. Received for publication September 22, 2004; accepted for publication September 28, 2004.

Weinberg (1) highlights the challenges and pitfalls faced when statistically analyzing data from observational (nonrandomized) studies to explore causal hypotheses. She reminds us that statistically significant associations between variables in such studies can be ambiguous. Extreme caution is required when statistical modeling of data from observational studies is used to infer causality from statistical associations between exposure and outcome variables. When knowledge regarding the relation between exposure(s) and potential confounder(s) is incomplete, no causal inference should be made based on such data alone. Indeed, under these circumstances, it is questionable whether one can invoke a statistical artifact (as we have done) without reference to an accepted causal model. This is why, when demonstrating a serious potential flaw in the statistical analyses of observational data believed to support the fetal origins of adult disease hypothesis, our article (2) invoked a widely accepted causal model linking adult blood pressure, birth weight, and current weight. Using our simple model, we demonstrated one reason why the statistical relation between birth weight and adult blood pressure should not be adjusted for current weight if this is not a true confounder (2). In Weinbergs more complex causal model (1), which she describes in the form of a directed acyclic graph and which includes prenatal environmental and genetic factors (set A factors) that influence both birth weight and current weight, adjustment for current weight can be justified provided there is no direct causal pathway from birth weight to current weight. However, if set B factors (which affect current weight and blood pressure) and/or set C factors (which affect current weight, blood pressure, and birth weight) are also present or there is a causal pathway from birth weight to current weight, adjustment for current weight remains inappropriate. These last two scenarios seem more likely than the first simply because, like any analytical variable, birth weight is both a correlate of and a proxy for a range of other variables. Like any biologic entity, birth weight is both a consequence and a cause of

other biologic phenomena; like any phenotype, birth weight is the product of the interaction between genetic and environmental factors. Thus, while many studies exploring the fetal origins of adult disease hypothesis use birth weight as a marker for prenatal factors responsible for size at birth, birth weight is also a variable that captures those other genetic and environmental phenomena with which it is associated, directly and indirectly, causally and by chance. The problem is therefore this: What is the correct causal graph to represent reality? Weinbergs directed acyclic graph (1) is one of many possible causal models linking prenatal factors, birth weight, current weight, and blood pressure. Although some causal models seem more plausible than others, it is crucial to acknowledge that several feasible models exist. For proponents of the fetal origins of adult disease hypothesis, current weight is a confounder because there is at least a possibility (in one or more potential causal models) that adjustment for current weight is justified. Until we have clearer evidence from experimental studies (3) and a greater understanding of the complex anatomic, physiologic, and biochemical processes linking birth weight, current weight, and blood pressure (4), all causal models will be subject to debate. Epidemiologic studies need to be more transparent and rigorous in their reports of the formulation of research questions, the underlining biologic mechanism(s), and the statistical testing of specific research hypotheses (5, 6). Whatever the preferred causal model, researchers into the fetal origins of adult disease need to be aware that adjustment for variables, such as current weight, is subject to the reversal (Simpson) paradox. Otherwise, old paradoxes never die.

Downloaded from http://aje.oxfordjournals.org/ by guest on July 4, 2012

REFERENCES 1. Weinberg C. Invited commentary: Barker meets Simpson. Am J Epidemiol 2005;161:3335.

Correspondence to Dr. Yu-Kang Tu, Biostatistics Unit, Centre for Epidemiology and Biostatistics, University of Leeds, 30/32 Hyde Terrace, Leeds, LS2 9LN, United Kingdom (e-mail: y.k.tu@leeds.ac.uk).

36

Am J Epidemiol 2005;161:3637

Tu et al. Respond to Barker Meets Simpson 37

2. Tu YK, West R, Ellison ETH, et al. Why evidence for the fetal origins of adult disease might be a statistical artifact: the reversal paradox for the relation between birth weight and blood pressure in later life. Am J Epidemiol 2005;161:2732. 3. Ceesay SM, Prentice AM, Cole TJ, et al. Effects on birth weight and perinatal mortality of maternal dietary supplements in rural Gambia: 5 year randomised controlled trial. BMJ 1997;315: 78690.

4. Gluckman PD, Hanson MA. The developmental origins of the metabolic syndrome. Trends Endocrinol Metab 2004;15:1837. 5. Huxley RR, Neil A, Collins R. Unravelling the fetal origins hypothesis: is there really an inverse association between birthweight and subsequent blood pressure? Lancet 2002;360:659 65. 6. Lucas A, Fewtrell M, Cole TJ. Fetal origins of adult disease the hypothesis revisited. BMJ 1999;319:2459.

Downloaded from http://aje.oxfordjournals.org/ by guest on July 4, 2012

Am J Epidemiol 2005;161:3637