Domanski et al. find body mass index (BMI) to be an independent-risk factor for major adverse coronary events (MACE) in men, but not in women. These results are first based on a dichotomization of BMI at 30 kg/m2 and then on a further categorization of BMI into five pre-specified groups. These groupings, the first of which is based on guidelines presented by NIH1 while the second closely mimics the categorization presented by the World Health Organization,2 were developed for identifying subclasses of risk for general health concerns associated with obesity and not for any predictive ability to accurately model the relationship between BMI and the risk of the specified cardiac events. If weight categories of BMI are to be used, and there is an extensive literature suggesting that they should not,36 the categorization employed for the statistical analysis should reflect the nature of the association between the exposure BMI and the outcome MACE. The approach presented in Domanski et al., since it is based on a pre-determined categorization, does not allow for an unrestricted assessment of the relationship between BMI and MACE. Furthermore, any categorization may find non-significant results due to low power induced by small counts in certain BMI groups. For example, the authors find a significant increase between the ‘obese’ and ‘normal’ groups of men (HR=1.26, P<0.01), while a much larger effect (HR=0.40) between ‘morbidly obese’ and ‘normal’ women is found to be non-significant (P=0.12). Investigation of the relative widths of the hazard ratio confidence intervals for these two comparisons in the authors' figures reveals striking differences in the sample sizes included in the relevant gender by BMI groupings. While the authors do eventually consider BMI as a continuous variable, they curiously restrict it to having a linear effect (results again show a significant increase risk for men, but not for women). This is particularly puzzling as they previously claim a J-shaped association for men (it should be kept in mind, however, that this association is dependent on the categorization selected). It would seem that a smoothing technique such as cubic splining,4 which requires neither categorization nor an assumption of a linear effect, would have been more appropriate for modeling BMI in the analyses presented. Future research studies should avoid grouping continuous variables into categories, especially if it cannot be demonstrated that such a categorization accurately reflects the relationship between this variable and the specific outcome of interest. In addition, continuous variables should not be assumed to have a linear effect unless the assumption of linearity can be justified.

References

1
Clinical guidelines on the identification, evaluation, and treatment of overweight and obesity in adults. The Evidence Report. NIH Publication No.98-4083, September
1998
The National Institutes of Health
2
World Health Organization
Chronic disease information sheets: Obesity and overweight
(29 September 2005)
3
Royston
P
Altman
DG
Sauerbrei
W
Dichotomizing continuous predictors in multiple regression: a bad idea
Stat Med
 
2006
25
127
141
4
Harrell
FE
Jr
Regression Modeling Strategies: with Application to Linear Models, Logistic Regression, and Survival Analysis
 
2001
New York
Springer-Verlag
5
Taylor
JMG
Yu
M
Bias and efficiency loss due to categorizing an explanatory variable
J Mult Anal
 
2002
83
248
263
6
Ragland
DR
Dichotomizing continuous outcome variables: dependence of the magnitude of association and statistical power on the cutpoint
Epidemiology
 
1992
3
40
43

Comments

0 Comments