The socalled “obesity paradox” has been observed in a variety of settings [
Along with the emergence of the concept of metabolically obese normal weight, the “obesity paradox” phenomenon came into notion [
No other chronic disease is as known to be related to obesity as diabetes. The role of obesity in diabetes was so ingrained that some investigators suggested the term “
Using an inception cohort of diabetic patients, we conducted the current study aiming to reconcile the “obesity paradox” recently observed among diabetic patients. This is an attempt to examine the apparently contradictory statements regarding the obesitymortality association and to draw conclusions to either reconcile them or explain their existence. As such, we tested the hypotheses that (1) the contributions of anthropometric measures of obesity to mortality are nonlinear and (2) the confounding bias due to hip and waist circumference might have contributed to the “obesity paradox” among diabetic patients.
Detailed descriptions of the Tehran lipid and glucose study (TLGS) have been reported elsewhere [
Findings on covariate variables are expressed as means (SD) and frequency (%) for continuously and categorically distributed variables, respectively. We tested for trends across BMI tertiles by using the median in each tertile as a predictor. Statistical significance of trends across BMI tertiles was examined by implementing General Linear models. The flexible parametric model for survival analysis was used to plot mortality rates (hazard function) against followup time. We concentrated on continuous timedependent effect using restricted cubic spline functions to model the baseline cumulative hazard. We fitted nested models with and without spline functions of the baseline hazard and estimated the deviance (Dstatistics) for each model based on the maximum likelihood
We set the statistical significance level at a twotailed type I error of 0.05. All statistical analyses were performed using STATA version 12.0 (STATA, College Station, TX, USA).
We certify that all applicable institutional and governmental regulations concerning the ethical use of human volunteers were followed during this research. Informed written consent was obtained from all participants and the Ethical Committee of Research Institute for Endocrine Sciences approved this study. The investigations reported herein have been carried out in accordance with the principles of the Declaration of Helsinki as revised in 2000.
Among participants aged ≥30 years, we selected those not using glucose lowering agents but diagnosed to have new onset diabetes at the baseline examination and those who developed incident diabetes during any of the two consecutive followup examination cycles. Complete data on covariates were available for 1,322 of participants with the median followup time being 9.1 years.
Clinical and paraclinical characteristics of participants at the time when they were first diagnosed to have diabetes are stratified by tertiles of BMI in Table
Baseline characteristics of the participants with newonset diabetes mellitus stratified by tertile of body mass index.
Tertile 1  Tertile 2  Tertile 3 

Total  

Number of participants  441  441  440  —  1322 
Median body mass index (kg·m^{−2})  24.9  28.9  33.8  —  28.9 
Minimum body mass index (kg·m^{−2})  15.7  26.9  31.1  —  15.7 
Maximum body mass index (kg·m^{−2})  26.9  31.1  57.7  —  57.7 
Categorically distributed variables  
Male  259 (0.59)  213 (0.48)  118 (0.27)  <0.001  590 (0.45) 
Smoker  67 (0.15)  55 (0.13)  32 (0.07)  <0.001  154 (0.12) 
Blood pressure lowering drug usage  107 (0.24)  118 (0.27)  136 (0.31)  <0.001  361 (0.27) 
History of previous cardiovascular disease  76 (0.18)  90 (0.21)  80 (0.18)  0.957  246 (0.19) 
Assigned to life style modification intervention  160 (0.36)  170 (0.39)  166 (0.38)  0.549  496 (0.38) 
Continuously distributed variables  
Age (years)  55.22 (11.96)  53.48 (11.69)  52.25 (10.93)  <0.001  53.65 (11.59) 
Systolic blood pressure (mmHg)  127.40 (20.77)  130.02 (22.75)  132.52 (21.79)  <0.001  129.97 (21.86) 
Total cholesterol (mmol·l^{−1})  5.65 (1.23)  5.84 (1.32)  5.82 (1.21)  0.001  5.77 (1.26) 
Highdensity lipoprotein cholesterol (mmol·l^{−1})  1.03 (0.28)  1.01 (0.29)  1.04 (0.26)  0.002  1.03 (0.28) 
Waist circumference (cm)  87.87 (8.05)  97.31 (6.98)  106.03 (8.44)  <0.001  97.06 (10.87) 
Hip circumference (cm)  94.80 (5.03)  101.87 (4.56)  112.87 (5.58)  <0.001  103.16 (9.75) 
Outcome  
Allcause mortality  48 (10.9)  33 (7.5)  27 (6.1)  0.236  108 (8.2) 
Allcause mortality rate, per (10 000 personyear)  90.8 (65.8–125.3)  53.7 (35.0–82.3)  64.2 (44.0–93.5)  0.140  69.7 (56.3–86.2) 
Data are presented as mean (SD) or frequency (%) for continuously and categorically distributed variables, respectively.
Mortality is presented as per 10 000 personyears (95% CIs).
No consistent trend in mortality rates was observed across tertiles of BMI. In fact the first (
Contribution of different tertiles of the body mass index to allcause mortality.
Hazard Ratio (95% CIs)  Std. Err.  Wald 



Model 1  
Body mass index (kg·m^{−2})  
First tertile  1.69 (0.99–2.88)  0.46  1.91  0.056 
Second tertile  1 

Third tertile  1.17 (0.66–2.07)  0.34  0.54  0.588 
Model 2  
Body mass index (kg·m^{−2})  
First tertile  2.36 (1.30–4.30)  0.72  2.81  0.005 
Second tertile  1 

Third tertile  0.96 (0.52–1.77)  0.30  −0.12  0.904 
Logwaist circumference [ 
20.35 (1.62–254.79)  26.24  2.34  0.019 
Model 3  
Body mass index (kg·m^{−2})  
First tertile  1.54 (0.83–2.86)  0.49  1.38  0.168 
Second tertile  1 

Third tertile  1.49 (0.80–2.76)  0.47  1.25  0.211 
Loggeneral cardiovascular risk  2.81 (2.11–3.75)  0.41  7.02  <0.001 
Logwaist circumference [ 
0.65 (0.04–10.50)  0.93  −0.30  0.763 
As shown in Figure
Inverse relationship of hip and waist circumference with body mass index.
Figure
(a) Nonlinear contribution of body mass index to allcause mortality. (b) Nonlinear contribution of body mass index to allcause mortality, allowing for waist circumference. (c) Nonlinear contribution of body mass index to allcause mortality, allowing for hip circumference. (d) Nonlinear contribution of body mass index to allcause mortality, allowing for both waist and hip circumference.
As panel (b) depicts, when we added waist circumference to the BMI in the multivariateadjusted model the steepness of BMImortality association curve slope for values below 27 kg·m^{−2} increased, whereas the steepness of BMImortality association curve slope for values above this threshold decreased. To examine if the increasing levels of BMI associated with decreasing trend in mortality up to BMI of 27 kg·m^{−2} were due to negative confounding effect of increasing hip circumference, we further adjusted the model for hip circumference and observed that the steepness of the slope of the association curve was observed to decrease considerably so that it was no longer statistically significant (panel (d)). As shown in panel (c), when we introduced hip circumference into the multivariable model already incorporating BMI, the steepness of BMImortality association curve slope for values below 27 kg·m^{−2} decreased, whereas, above this threshold, the steepness of the BMImortality association curve increased towards positive values.
Figure
(a) Nonlinear contribution of waist circumference to allcause mortality. (b) Linear contribution of hip circumference to allcause mortality.
Using an inception cohort of adults with diabetes, we investigated the nonlinear contribution of the anthropometric indices of obesity to allcause mortality allowing for potential confounding bias due to established CVD risk factors. We observed that BMI, waist, and hip circumference were all associated with allcause mortality in a curvilinear fashion.
The finding of interest was that even across values generally considered as normal, further decreases in the BMI was associated with increased mortality. Some observations regarding this finding were as follows.
When the effect of waist circumference was taken into account, for values of BMI below 27 kg·m^{−2}, the steepness of the slope increased. This implies that if waist circumference remains constant decreasing BMI would confer an excess mortality. Meanwhile, higherthannormal values of BMI were not associated with increased mortality, except for extremely high values. That is, except for extremely high BMI values, increasing levels of BMI are safe as long as waist circumference remains constant.
Conversely, when we controlled our analysis for hip circumference, the opposite was true. That is, when we introduced hip circumference into the multivariable model already incorporating BMI, the steepness of BMImortality association curve slope for values below 27 kg·m^{−2} decreased, whereas, above this threshold, the steepness of the BMImortality association curve increased. This implies that if hip circumference remains constant, decreasing levels of BMI will not result in excess mortality. In other words, BMIassociated mortality observed in the lower extreme is merely an artifact resulting from uncontrolled confounding bias due to hip circumference.
An inverse Ushaped association was observed between the waisttohip ratio (WHpR) and BMI; that is, the ratio increased with increasing levels of BMI up to about 27 kg·m^{−2} and the association was reversed thereafter. In other words, the magnitude of increase in the waist circumference exceeded the increase in hip circumference up to
To conclude, If we had not considered the effect of waist circumference, we would have overestimated the mortality associated with increasing values of BMI above normal values. Also, had we not taken the effect of hip into account, we would have overestimated the mortality associated with decreasing levels on BMI below normal values.
Some studies have suggested that waist circumference, either alone or in combination with BMI, may have a stronger relation to some health outcomes than BMI alone [
We used an inception cohort of diabetic patients which is now a standard for prognostic studies. Our patients were identified at an early and uniform point (inception) in the course of their disease. The cohort came from a largepopulationbased study of both sexes, with accurate and valid data on risk factors at baseline and continuous surveillance of mortality and CVD events based on standard criteria. In an attempt to reduce variability in the duration of newonset diabetes, we restricted our analysis to participants who were not taking glucose lowering agents in the first examination. We used appropriate advanced statistical methods to capture nonlinear association of BMI, waist, and hip circumference with mortality. Both FPG and the standard 2 hPCPG levels were available in the TLGS for identifying incident diabetes cases. Controlling the analyses for established CVD risk factors made the estimates unlikely to be affected by the bias stemming from potential confounding effects of these risk factors. Some limitations of our study merit mentioning. First, the small number of incident events precluded stratification of analyses by sex. Second, large confidence intervals on the rightsided tail of the estimated curves imply lack of statistical power and our estimates might not have been stable at high BMI, hip, or waist circumference quantities. Third, despite our best efforts to control for bias from preexisting disease, it is likely that we did not eliminate such bias completely. Thus, the increased risk of death from specific causes associated with leanness may reflect preexisting, but unrecognized, disease processes. Fourth, due to our sample, we were not able to investigate association of measures of obesity with specific causes of death. Finally, the population studied was of Persian ancestry and, thus, cannot be readily extrapolated to other populations.
Despite extensive use of BMI in research and clinical practice, there are very few studies testing its diagnostic accuracy and except for that of RomeroCorral et al. no study has done this in a large, multiethnic adult population representing men and women of many age strata [
In a pooled analysis of five longitudinal cohort studies, Carnethon et al. observed that adults who were of normal weight at the time of diagnosing diabetes had higher mortality than their overweight/obese counterparts [
The results of studies demonstrating the obesityparadox have consistently ignited a question of “should we start advising people to become more obese?”. We demonstrated, herein, that what previous studies have referred to as “obesityparadox” is de facto “BMIparadox.” Unfortunately, an implicit assumption made in many epidemiologic analyses is that BMI alone is a sufficient measure of obesity effects in regression analyses. Michels et al. have argued that this is not necessarily true and that whether BMI alone adequately captures the effect of anthropometric variables on health outcomes depends on many factors [
We have previously extensively studied different indices of obesity with respect to health risks in general population. These indices were either directly or inversely associated with mortality [
Previous studies showed that adults who were of normal weight at the time when they were first diagnosed to have diabetes had higher mortality than their overweight/obese counterparts. Using arbitrary definition for normalweight, they were not able to discover Ushaped associations.
Among newly diagnosed diabetic patients, BMI, in its extremes of both leanness and obesity was associated with increased allcause mortality, independent from established CVD risk factors. However, the association faded when we controlled our multivariate analyses for the confounding bias originating from hip and waist circumference. Our findings indicate that body mass as measured by BMI does not contribute to allcause mortality independent from central obesity. Rather, BMI harbors intermixed positive and negative confounding effects on waist and hip circumferencerelated mortality. In particular, failing to control for the confounding effect of hip circumference may stymie unbiased hazard estimation. What previous studies have referred to as the “obesityparadox” could be reconciled by the fact that BMI is not a perfect measure of obesity. Although not necessarily in the same direction, both body mass and fatness contribute to mortality. Such confounding bias originating from each part should be accounted for while studying obesityrelated health risks, since these parameters exhibit covariation in many settings.
The authors declare that there is no conflict of interests regarding the publication of this paper.
The authors express their appreciation to the participants of district 13 of Tehran for their enthusiastic support in this study. The authors also express their gratitude to Dr. Henry S Kahn (Medical Epidemiologist, Division of Diabetes Translation, Centers for Disease Control & Prevention, USA) for his thoughtful comments. Mohammadreza Bozorgmanesh designed the study, performed the statistical analysis, interpreted the analyses, and drafted the paper. Farzad Hadaegh interpreted the analyses and revised the paper critically for important intellectual content. Banafsheh Arshi and Farhad Sheikholeslami contributed in the interpretation of the analyses and drafting the paper. Fereidoun Azizi revised the paper critically for important intellectual content. Fereidoun Azizi is the guarantor and takes full responsibility for the work as a whole, including the study design, access to data, and the decision to submit and publish the paper. All authors read and approved the final paper. This study was supported by Grant no. 121 from the National Research Council of the Islamic Republic of Iran.