点击显示 收起
Magee-Women’s Research Institute
204 Craft Avenue
Pittsburgh, PA 15213
E-mail: bodnar{at}mwri.magee.edu
Department of Nutrition
University of North Carolina School of Public Health
Carolina Population Center
University of North Carolina
Chapel Hill, NC
E-mail: melissa_nelson{at}unc.edu
Dear Sir:
In a study recently published in the Journal, Mathews et al (1) examined the association between maternal nutritional status and the fetal growth of term infants in a well-nourished cohort in the United Kingdom. They concluded that maternal nutrient intake is unlikely to have an important effect on infant health. In an attempt to shed light on these contentious issues, such studies are of great value. However, we believe that the provocative findings of Mathews et al should be interpreted with extreme caution because of the potential biases introduced by their statistical modeling strategy.
The authors measured 17 biomarkers of nutritional status, which were included in multivariable linear regression models predicting birth or placental weight; however, the presentation of these analyses lacked a conceptual framework or distinct hypothesis that tested biologically plausible mechanisms. Additionally, the authors modeled all 17 biomarkers as continuous variables, seemingly without evaluating alternative transformations of each variable and therefore disregarding many conceivable structures that these relations may take on. In doing so, the authors may have biased their results by failing to recognize the often implausible assumptions that accompany such analyses (2-5). Modeling variables in their continuous form does not test whether the true relation between the biological marker and birth or placenta weight is linear; rather, it assumes that the relation is linear. It is unlikely that the effect of maternal nutritional status on fetal growth increases linearly across the distributions of each nutrient; more likely, these relations vary by nutrient concentration in a nonlinear fashion. Interestingly, Mathews et al acknowledge the previously documented U-shaped relation between hemoglobin and birth weight (6), yet there is no indication in their article that they explored this possibility. Furthermore, several of the biological markers they measured have not been studied extensively in relation to fetal or placental growth; thus, one can assume nothing about the dose-response relation. Although modeling a linear relation may be an appropriate first step in preliminary analyses, in most situations it is insufficient to describe the true dose-response relation (3), and further exploration is needed to gain a valid representation of the data. Unrealistic dose-response curves are usually fit with the authors’ approach, which leads to considerable bias (3).
Many simple alternatives to the authors’ method of modeling are available, which may show a more accurate representation of the true underlying structure of the data and, hence, the relation between maternal nutrition and fetal growth. Although categorical analysis is a viable alternative that avoids linearity assumptions, it performs poorly when the effects of the exposure are limited to the extremes of its distribution (4), which is often the case in the study of the nutrient status of well-nourished populations. Categorization also makes the unrealistic assumption of a constant effect within the category and a sudden jump in effect with the next category. Two simple alternatives that avoid such drawbacks are fractional polynomial regression (the inclusion of terms such as x2, x3, and x1/2 in the model) and spline regression. These modeling strategies not only avoid the power loss associated with categorical analysis (5), but they are flexible and make few assumptions about the dose-response relation (4); the data are more able to determine the most appropriate curve shape than are the investigators. Fractional polynomial and spline regression involve uncomplicated transformations of variables and can be easily implemented in any standard software package (2, 4). Greenland (4) provides an accessible review of such models.
Mathews et al conclude that "maternal intake of the nutrients measured is unlikely to be an important determinant of the long-term health of infants" (1). Yet, we argue that the potential biases in their results may explain their null findings. Further investigation of the relation between biomarkers of nutritional status and infant and placental birth weight is warranted.
REFERENCES