Body mass index, diabetes, and mortality in French women: explaining away a “paradox”

Abstract

Background

Obesity is associated with increased mortality in the general population but, paradoxically, with decreased mortality in individuals with diabetes.

Methods

Among 88,373 French women participating in the E3N-EPIC study who were free of diabetes in 1990, we estimated the mortality hazard ratios (HR) and 95% confidence intervals (CI) for body-mass index (BMI) levels by diabetes status.

Results

During an average 16.7 years of follow-up, 3,750 deaths and 2,421 incident diabetes cases occurred. In overweight/obese versus normal weight women, the mortality HR (95%CI) was 1.42 (1.32,1.53) in women without diabetes, and 0.69 (0.40,1.18) in women with incident diabetes. Mortality increased with BMI among women without diabetes and decreased as BMI increased in women with diabetes.

Conclusions

We found a direct association between BMI and mortality among women without diabetes but not among those with incident diabetes in the same population. Selection bias may be a simple explanation for this “paradox”.

Obesity is associated with increased mortality in the general population but with decreased mortality in individuals with chronic disease (e.g. diabetes).^2–4 This so-called obesity paradox has led some to suggest that patients with established chronic disease should avoid weight loss.⁵ However, the obesity “paradox” might just be a selection bias⁶ that arises from a misguided analysis.^7–9

Interestingly, despite the increasing interest in this topic,¹⁰ the obesity “paradox” has not been empirically described in a prospective study of individuals without chronic disease at baseline. Here we (i)provide such description, (ii)propose a likely explanation for the “paradox” and, (iii)discuss its practical implications.

i)Empirical illustration of the “paradox”

Our analysis included 88,373 French women in the E3N Study¹¹ followed through mailed questionnaires between 1990 (baseline) and 2007 who were free of diabetes and had a BMI ≥18.5 kg/m² at baseline (see eAppendix). We defined normal weight in 1990 as BMI 18.5–24.9 kg/m² and overweight/obesity as BMI ≥25 kg/m². Self-reported cases of diabetes were confirmed using supplementary questionnaires and a drug reimbursement database (eAppendix Figure 1 and eAppendix Table 1). Deaths were identified through the health insurance plan, postal service, and next-of-kin. We estimated unadjusted incidence rates and fit Cox regression models adjusted for baseline covariates (marital status, education, menopause, hormone therapy use, physical activity, smoking, hypertension, cardiovascular disease, cancer) to estimate mortality hazard ratios (HR) for overweight/obesity versus normal weight, and for BMI categories 18.5–22.4, 25.0–27.4, 27.5–29.9, and ≥30 versus 22.5–24.9 kg/m².

After an average 16.7 years of follow-up, 3,750 women died and 2,421 had incident diabetes. [see eAppendix Table 2 for age-adjusted characteristics by BMI group]. Overweight/obese women had higher mortality and diabetes incidence rates (38.3 and 56.0 per 10,000 person-years, respectively) than normal weight women (22.6 and 7.9 per 10,000 person-years, respectively). The adjusted HR (95% CI) for overweight/obesity versus normal weight was 6.10 (5.60, 6.64) for diabetes, and 1.33 (1.23, 1.43) for mortality (Table 1). Results did not materially change after excluding women with cancer/cardiovascular disease at baseline and smokers. Mortality increased with BMI (eAppendix Figure 2).

Table 1.

Hazard Ratios of Diabetes and All-cause mortality by overweight/obesity status at baseline, E3N study 1990–2007

		Cases	Person-years	Incidence rate per 10,000 person-years	Hazard Ratios (95%CI)
	Baseline BMI
Diabetes	<25	963	1,225,210	7.9	1.00
	≥25	1,458	258,926	56.3	6.10 (5.60–6.64)
Mortality	<25	2,740	1,214,797	22.3	1.00
	≥25	1,010	263,618	38.3	1.33 (1.23–1.43)

CI: Confidence interval; BMI: Body mass index.

Hazard ratios adjusted for marital status, highest level of educational level attainment, menopausal status and menopause hormone therapy use, physical activity, smoking status (never, past and current), treated hypertension, cardiovascular disease (stroke, myocardial infarction and angina) and cancer at baseline.

Among women without diabetes, mortality was higher in the overweight/obese than in normal-weight (38.8 vs. 22.5 per 10,000 person-years). The mortality HR was 1.42 (1.32, 1.53) and did not change after exclusion of smokers and women with cancer or cardiovascular disease (Table 2). Mortality increased with BMI (Figure 1). Conversely, among women with diabetes, mortality was lower in overweight/obese individuals than in normal-weight individuals (26.2 vs. 43.3 per 10,000 person-years). The mortality HR was 0.69 (0.40, 1.18) (Table 2). After excluding women with cancer/cardiovascular disease and smokers the HR was 0.41 (0.18, 0.92). Mortality decreased as BMI increased (Figure 2). These findings illustrate the “paradox” via a direct comparison between women with and without diabetes from the same population.

Table 2.

Hazard Ratios of All-Cause Mortality by Overweight/Obesity Status at Baseline, Stratified by Diabetes, E3N study 1990–2007

	No diabetes				Incident Diabetes
Baseline BMI	Deaths	Person-years	Mortality rate per 10,000 person-years	Hazard Ratios (95%CI)	Deaths	Person-years	Mortality rate per 10,000 person-years	Hazard Ratios (95%CI)
All (n=83,373)
<25	2,714	1,208,798	22.5	1.00	26	5,998	43.3	1.00
≥25	982	252,922	38.8	1.42 (1.32–1.53)	28	10,695	26.2	0.69 (0.40–1.18)
Excluding women with chronic disease (n=83,288)
<25	2,192	1,146,877	19.1	1.00	23	5,617	40.9	1.00
≥25	755	235,470	32.1	1.40 (1.28–1.52)	19	9,850	19.3	0.54 (0.29–1.01)
Excluding women with chronic disease and smokers (n=45,195)
<25	1,206	617,685	19.5	1.00	15	3083	48.7	1.00
≥25	445	132,064	33.7	1.43 (1.28–1.60)	10	5381	18.6	0.41 (0.18–0.92)

CI: Confidence interval; BMI: Body mass index in kg/m² at baseline in 1990.

Hazard ratios adjusted for marital status, highest level of educational attainment, menopausal status and menopause hormone therapy use, physical activity, smoking status (never, past and current), treated hypertension, cardiovascular disease (stroke, myocardial infarction and angina) and cancer at baseline.

Figure 1. Hazard Ratios of All-Cause Mortality by Body Mass Index among Women without Diabetes (Panel A), excluding Women with Chronic Disease (Panel B), and additionally excluding Women Who Had Ever Smoked (Panel C).

Hazard ratios were adjusted for age, marital status, highest level of educational level attainment, menopausal status and menopause hormone therapy use, physical activity, smoking status (never, past and current), cardiovascular disease (stroke, myocardial infarction and angina), cancer and treated hypertension.

I bars denote 95 percent confidence intervals. P for trend: Panel A <0.0001; Panel B <0.0001; Panel C <0.0001.

Figure 2. Hazard Ratios of All-Cause Mortality by Body Mass Index among Women with Diabetes (Panel A), excluding Women with Chronic Disease (Panel B), and additionally excluding Women Who Had Ever Smoked (Panel C).

Hazard ratios were adjusted for age, marital status, highest level of educational level attainment, menopausal status and menopause hormone therapy use, physical activity, smoking status (never, past and current), cardiovascular disease (stroke, myocardial infarction and angina), cancer and treated hypertension.

I bars denote 95 percent confidence intervals. P for trend: Panel A 0.20; Panel B 0.08; Panel C 0.02.

In sensitivity analyses, we found similar estimates when we i)used the BMI just before diabetes diagnosis [eAppendix Table 3]; ii)used BMI as a time-varying exposure; iii)repeated the analyses starting follow-up in 1993 (when dietary information was first available) and adjusted for coffee, fruits and vegetables, and processed red meat intake [eAppendix Table 4]; iv)replaced BMI by waist circumference and started the follow-up in 1994[eAppendix Table 5]; and v)censored women after two missed questionnaires.

ii)Explanation of the “paradox”

Biological explanations for the “paradox” rely on potential benefits of obesity, ^12,13,14 or differences² (perhaps genetic¹⁵) between normal-weight and overweight/obese individuals with diabetes that put normal-weight individuals at a higher mortality risk. However, there is an explanation that does not require to posit any benefits of elevated BMI: selection bias due to conditioning on a variable affected by exposure.^7–9

An analysis restricted to individuals with diabetes is conditioned on a variable (diabetes) that is affected by exposure (BMI) and unmeasured risk factors for mortality such as lifestyle and genes. According to the simplified causal diagram in Figure 3a, stratifying the analysis on diabetes (a collider) will generally induce an association between BMI and mortality through other common causes U, even if no association existed in the unstratified analysis. Similar causal diagrams have been proposed to explain the birth weight “paradox” ¹⁶ and some components of Simpson’s “paradox”.¹⁷

Figure 3. Simplified Causal Diagram to Represent the Association between BMI and Mortality when Conditioning on Diabetes.

U, other unmeasured risk factors. The graph is conditional on measured confounders and, for simplicity, assumes a null direct “effect” of BMI on mortality.

To see how the bias arises, first suppose that the only causes of diabetes are obesity and a genetic factor that, independently, increases the risk of mortality. In this oversimplified scenario, all individuals with diabetes are either obese or have the genetic factor (or both), and all normal-weight individuals with diabetes necessarily have the genetic factor that results in higher mortality. In other words, an inverse association between obesity and mortality---the “paradox”---is expected among individuals with diabetes. This inverse association is also expected in the more realistic setting with multiple causes of diabetes other than obesity and the genetic factor (e.g., lifestyle). In this case, normal-weight individuals with diabetes are more likely to have other risk factors for mortality.

The above explanation can also accommodate the possibility that diabetes may be a collection of similar diseases with different etiologies and with different effects on mortality. Suppose that Diabetes – Type A is a disorder caused by high BMI, that Diabetes – Type B is a condition caused by other causes U, and that there are no common causes of Types A and B. Also suppose that Diabetes –Type B, but not Type A, increases mortality. This scenario is depicted in Figure 3b, which is an elaboration of Figure 3a. Under this scenario an analysis restricted to patients with Diabetes –Type A would not introduce selection bias (Diabetes –Type A is not a collider). However, because the type of diabetes is unknown in practice, any analysis restricted to individuals with diabetes (without specifying the type) will be conditioned on a collider and may therefore introduce selection bias. The bias has a structure similar to that of the birth weight paradox¹⁶, some components of Simpson’s paradox¹⁷, Berkson’s fallacy¹⁸, adjustment for time-varying confounders affected by prior treatment⁶, and including prevalent users in drug safety studies^19,20.

Our analysis has several strengths that help rule out alternative explanations---other than conditioning on a collider--- to the “paradox”. These strengths include measurement of BMI before the diagnosis of diabetes, a long-term follow-up, use of incident cases of confirmed diabetes, no differential access to health care among participants, and a detailed assessment of lifestyle factors and clinical diagnoses. However, compared with other cohorts,² E3N participants are leaner, which limits our ability to study high levels of BMI.

iii)Practical implications of the “paradox”

If the “paradox” is indeed an example of selection bias, then the lower mortality risk in overweight diabetics should not be the basis for weight management recommendations to individuals with diabetes. Similarly, the lower mortality in low birth weight babies of smokers (the birth weight “paradox”¹⁶) should not be the basis to recommend maternal smoking, and the lower risk of heart disease in prevalent users of estrogen plus progestin hormone therapy should not be the basis to recommend preventive hormone therapy²¹. The argument that conditioning on diabetes is necessary to estimate the effect of BMI in diabetics and non-diabetics separately is invalid for the same reason that one cannot generally estimate effects of a randomized treatment within levels of a post-randomization variable.

The implications of the obesity paradox are harder to describe than those of similar paradoxes because “the causal effect of BMI” is an ill-defined concept.^22,23 Even in the absence of confounding and other biases, it is unclear what causal effect, if any, is estimated when BMI is the exposure. For the same reason, conditioning on diabetes after baseline does not guarantee that the estimates can be interpreted as the direct effects of BMI on mortality, i.e., effects not mediated through diabetes. [If the effect of BMI were well defined, one would still need to condition on all common causes of diabetes and mortality (e.g., genetic and lifestyle factors),^24,25 which is generally impossible, for an unbiased estimation of direct effects.]

Leaving aside causal inference considerations, suppose the goal is to estimate the association between baseline BMI and mortality in a population of patients without chronic disease at baseline. Then the “paradox” resolves itself by simply not conditioning on post-baseline disease. Studies restricted to patients with prevalent disease are conditioned on post-baseline disease by design and thus are not appropriate to estimate the association between BMI and mortality. One can attempt to reconstruct this association using external data and strong statistical assumptions,⁹ but a safer method is to refrain from conducting these analyses in studies restricted to patients with prevalent disease and to focus our research efforts on unrestricted studies. Of course, the same problem arises in studies with patients without chronic disease at baseline when the analysis is restricted to patients with disease.

In summary, the inverse association observed between BMI and mortality among patients with diabetes, and perhaps other chronic diseases, could be explained by selection bias due to stratification on a variable affected by the exposure. Being explicit about the causal question of interest and the causal assumptions helps clarify and interpret observations that may seem paradoxical at first sight.