Inter-individual variability in the improvement of physiological risk factors for disease: gene polymorphisms or simply regression to the mean?

In an informative and balanced review, Mori et al. (2009) explored the perplexing question of why people differ in their individual physiological responses to exercise training. We agree fully with Mori et al. (2009) that gene polymorphisms could account for inter-individual differences in the exercise-modulated response of physiological risk factors of disease, and that a better understanding of associated issues should lead to more effective exercise prescription. In this respect, we wonder whether a statistical artifact, originally identified in the 19th century, should also be considered carefully by researchers on this topic. Mori et al. (2009) observed that ‘prominent improvement’ was apparent amongst the individuals with the unhealthiest initial values of phenotypes measured at baseline. In an accompanying editorial, Joyner & Nose (2009; p. 5525) summarized the findings reported in another symposium review (Nose et al. 2009) by stating that ‘high-intensity exercise can have a profound effect on physiological regulation in older humans, especially in the least fit’. Presumably, it was found by Mori et al. (2009) that the individuals with the smallest values of some phenotypes (e.g. peak oxygen uptake) measured at baseline increased the most in response to exercise, and that the individuals with the greatest values of other phenotypes (e.g. body mass index, BMI) measured at baseline decreased the most at follow-up. Such observations are commonly made in exercise science and, while we agree that they could be valid, we are also aware of warnings that have been published about their interpretation (e.g. Shephard, 2003). For example, in their classic review, Bouchard & Rankinen (2001) used data from the HERITAGE Family Study to show that age, sex and race have little impact on inter-individual differences in training responses. Rather, these authors reported that the initial level of a phenotype is a major determinant of training response for some traits, such as sub-maximal exercise heart rate and blood pressure (BP). We maintain that such observations could be just as consistent with the effects of the regression-to-the-mean statistical artifact as they are with any real biological explanation (e.g. gene polymorphisms) of inter-individual differences in response. As illustrated recently in the context of the BP-lowering effects of exercise (Taylor et al. 2010), regression-to-the-mean can compromise the interpretation of individual differences in physiological responses to exercise, or indeed any intervention. Because of inherent measurement error in phenotypes such as BP, individuals who are initially found to lie at the extreme ends of a population distribution are likely to record, at follow-up, values that are closer to the population mean. Therefore, individuals with initially low values of a phenotype will be naturally higher on a follow-up test and vice versa (Atkinson et al. 2001). This observation might be nothing at all to do with any real differential physiological effects of the intervention that is studied; it is simply explained by the often insidious effects of regression-to-the-mean. In their review, Mori et al. (2009) went on to explore whether the individual magnitudes of change in their outcomes of interest were related to gene polymorphisms, and this exploration was made only in the obese subjects who completed the exercise intervention. The fraction of the variability in responses explained by polymorphisms was reported to be small (<5%). Nevertheless, we wonder whether the regression-to-the-mean artifact might have ‘swamped’ this analysis, as it did in our analysis of other factors influencing the BP response to exercise (Taylor et al. 2010). It is interesting that the individuals with the A/A genotype showed the largest reduction in BMI, yet these subjects also showed the highest mean initial BMI. Nevertheless, the G/A genotype subjects showed the lowest initial BMI yet also showed a relatively large reduction in BMI. It would be especially interesting if the explained variance due to polymorphism might just appear small because the effects of an uncontrolled regression-to-the-mean artifact might be so large? Methods are available to control for the regression-to-the-mean artifact when examining individual differences in the response to interventions (Bland & Altman, 1994; Yudkin & Stratton, 1996). For example, the relationship between gene polymorphisms and individual changes in phenotypes due to an exercise intervention can be examined alongside the quantification of the regression-to-the-mean artifact present in a randomized comparator group of subjects who do not complete the intervention. It would be interesting to examine whether the fraction of variability in individual response that is explained by gene polymorphisms increases if this design-based approach and/or other types of statistics-based controls for regression to the mean (e.g. analysis of covariance) are incorporated into the data analyses of studies on this important topic.