Mortality Risk of High BMI in Life Insurance Applicants and the US Population

Body mass index (BMI), defined as weight (kg) divided by the square of height (m) is a commonly used measure of body composition and is frequently used as a measure of overweight and obesity. Elevated BMI has been shown, in numerous studies, to increase the risk of various adverse health outcomes, including diabetes, heart disease, stroke, various cancers, and all-cause mortality.^1-3 However, in the life insurance industry, many of these other health outcomes are already considered during underwriting, leaving the question of how the BMI itself affects mortality risk. This study seeks to determine the residual mortality risk associated with elevated BMI once these other health effects have been considered.

METHODS

This study makes use of 2 different data sets. The first is from Clinical Reference Laboratories, Inc, and includes laboratory and biometric variables collected from life insurance applicants at the time of underwriting. Variables include age, sex, height, weight, BMI, laboratory tests (lipids, liver function tests, renal function tests, hemoglobin A1c and others), and answers to simple yes/no questions about the presence of diabetes, hypertension and heart disease. Vital status is determined by a third-party death audit, which includes outcomes from the Social Security Death Master File as well as data from obituary listings, state databases, and other sources. The second data set is from the National Health and Nutrition Examination Survey (NHANES),⁴ which includes a representative sample of the US population. This data set includes similar variables, including height, weight, and BMI, as well as a variety of laboratory tests. Vital status assessment is determined via the paired mortality file, which is based on data from the National Death Index, a comprehensive source of vital status information.

Exclusion criteria included smoking (defined as answering “yes” to questions about current tobacco use or testing positive for urine cotinine at a level of 200 ng/dl), pregnancy (defined as answering “yes” to a pregnancy question or having a predicted probability of pregnancy > 0.5 on the basis of a previously determined logistic regression model), testing positive for hepatitis B, hepatitis C or HIV, having a missing value for cholesterol (a marker of whether blood testing was done), or having a BMI outside the study range of 22 to 80. Note that BMIs below 22 were excluded due to the previously observed inverse relationship between BMI and mortality in this range. The initial CRL data set included over 15 million applicants. After exclusions, there were approximately 7.2 million subjects, 39.6% female, mean(sd) age 44.9(12.9), with a mean follow-up time of 9.0 years, during which 157,951 deaths occurred.

In the NHANES data, participants of all ages were invited through a complex multistage probability sampling technique. This included oversampling of Mexican Americans, African Americans, low-income whites, and older people (aged 60 years and up) to create a representative sample of the general US population. Demographic data such as age and sex were collected through a questionnaire. Laboratory tests were also done in their survey which included cotinine level, lipids, liver function tests, renal function tests, hemoglobin A1c, and other biomarkers. Body measurements such as BMI and blood pressure were measured by trained health technicians. Diabetic status was determined by yes/no question. For consistency between NHANES and CRL variables, the presence of hypertension and heart disease were derived from multiple variables. Hypertension status was determined by the average of all blood pressure readings and the participant’s answer in the questionnaire (“Have you ever been told that you had high blood pressure?”). Heart disease status was determined by combining multiple yes/no questions: Have you ever been told that you had congestive heart failure, coronary heart disease, angina/angina pectoris, heart attack, or stroke? Similar exclusion criteria were also applied to the NHANES data. The initial data set included 70,190 participants. After the exclusion and removal of missing data, there were 23,486 participants, 48% female, mean (sd) age 52.8 (16.9). The mean follow-up time was 7.6 years, during which 2,735 deaths occurred.

To analyze the effect of BMI, each data set was split by sex and age group (<40, 40-59, and 60+ years). For each of these groups, 6 successive models were fit. The first included only age and BMI (adjusted so that the hazard ratio result is per 5-unit increase in BMI), the second adds disease question variables (hypertension, diabetes, and heart disease), the third adds in liver function studies (AST, ALT, GGT, alkaline phosphatase, and albumin), the fourth adds in renal function (creatinine level) and systolic blood pressure, the fifth adds in lipids (cholesterol:HDL ratio), and the final model adds in hemoglobin A1c. For each of these, the measure of interest is the hazard ratio for a 5-unit increase in BMI. The models were fit using Cox proportional hazards regression. For each group, models were constructed treating BMI as a continuous variable, and as a categorical variable with 5 categories (22-26.9, 27-29.9, 30-34.9, 35-39.9, and 40+). Note that when variables were heavily skewed (AST, GGT, alkaline phosphatase, creatinine, cholesterol:HDL ratio, and hemoglobin A1c), a log transformation was applied, and when univariate testing had demonstrated a non-linear relationship between a given variable and mortality (age, AST, and creatinine), restricted cubic splines with 4 knots were used to model the relationship.

Models were constructed using R version 4.4.3⁵ and packages including tidyverse, rms, Hmisc, survey, survminer, dplyr, haven, and survival.

RESULTS

In the CRL data, with continuous models, the hazard ratio associated with a 5-unit increase in BMI was 1.30 for men and 1.25 for women in the initial model controlled only for age (Table 3). By the fourth model, the hazard ratio reached a level that was not significantly different than the next 2 models (Figure 1). This reflects a certain “saturation effect” which may indicate that this final hazard ratio (1.14 for men and 1.10 for women) is the final residual effect of BMI. In both sexes, the largest change in the BMI hazard ratio occurred in the 3rd model, after adding liver function tests.

View Figure

• Download as Powerpoint
• Download Figure

View Figure

• Download as Powerpoint
• Download Figure

In categorical models of the CRL data, the effect of various levels of BMI was significant at the 0.001 level in most age/sex categories. However, a BMI of 27-30 was non-significant in models 3-6 in young men, models 4-6 in older men, models 2-6 in young women, models 3-6 in middle-aged women, and models 3-6 in older women. In young women, only the hazard ratio associated with the highest BMI category was significant in models 3-5, and no category was significant in the final model - though this may be due to a marked reduction in the number of deaths due to missing hemoglobin A1c data (Table 3). Figure 3 illustrates the diminishment of the BMI hazard ratios in the categorical models along with the clustering of the hazard ratios in models 3 through 6.

View Figure

• Download as Powerpoint
• Download Figure

The NHANES data were divided into 2 age groups (<60 and ≥60 years), unlike the CRL data, which used three age groups (<40, 40-59, and ≥60 years). This was due to the lower number of deaths among the younger individuals in NHANES.

Using NHANES data with continuous models, the hazard ratios for both continuous and categorical were lower than those in CRL. For men, the hazard ratio was 1.22 and it was 1.16 for women in the continuous model controlled only for age (Table 4). The confidence interval was larger in the NHANES model (Figure 3) compared to CRL model (Figure 1), reflecting higher uncertainty in the estimated hazard ratios.

By Model 3, the hazard ratios in NHANES reached a plateau, with no further significant changes observed through Models 4 to 6 (Figure 3). The final hazard ratios were 1.10 for men and 0.98 for women, and they represent the residual association between BMI and mortality after full adjustment of the covariates. While CRL showed the largest change at Model 3, NHANES demonstrated comparable drops at Models 2 and 3, after adding disease question variables and liver function tests, respectively.

Compared to CRL, NHANES results showed limited significance at the 0.001 level. Table 4 shows the hazard ratios and their significance, both in continuous and categorical models. The continuous models hazard ratios were only significant (p < 0.001) for the first few models (Models 1-2 for men, and Model 2 for women). Models 4-6 were significant at p < 0.05 for males, but not in females.

In categorical models, high significance (p < 0.001) appeared only in younger men with BMI 40+ in Model 1 and in older men with BMI 40+ in Model 1-2. In females, the hazard ratios were only significant at 0.001 in the older age group, for BMI of 40+ in Model 1, BMI of 27-30 in Models 5-6, and BMI of 30-35 in Models 4-6.

Figure 4 illustrates the hazard ratios for men and women in NHANES, revealing a slightly different pattern compared to CRL, particularly among females under age 60.

View Figure

• Download as Powerpoint
• Download Figure

DISCUSSION

Overall, there was a good compatibility between the insurance population (CRL) and the general US population. However, the results from NHANES are generally lower and lack significance compared to the CRL data.

The lack of statistical significance from the NHANES models likely reflects the smaller sample sizes within BMI categories and the overall lower number of deaths in the NHANES dataset. In Figure 4, the pattern of younger females’ hazard ratio is likely influenced by the limited number of deaths in this subgroup (n = 173; see Table 2), which may contribute to greater variability and less stable estimates.

Table 1.Basic Characteristics of the CRL Study Population

View Fullscreen

Table 2.Basic Characteristics of the NHANES Study Population

View Fullscreen

Table 3.Categorical Models, CRL data

View Fullscreen

Table 4.Categorical Models, NHANES data

View Fullscreen

Other studies have demonstrated a similar mediation effect of the BMI–mortality connection. Indeed, some have shown a mortality benefit of being overweight or obese when other disease states are present and accounted for. This is termed the “obesity paradox.” Osadnik et al studied consecutive primary care patients in the LIPIDOGRAM program in Poland⁵ and found that the mortality effect of elevated BMI was reduced after the inclusion of sex, education, diabetes, hypertension, and dyslipidemia in multivariable adjusted models. Indeed the “overweight” category (BMI 25-29.9) had the lowest mortality hazard. Afzal et al., in a large study of the Danish population,⁶ showed that in models controlled for cholesterol, activity, income, and smoking, the lowest risk BMI was 27.0 in the most recent (2003-2013 cohort), and that this risk nadir had moved upward substantially since the earliest cohort (1976-1978), when it was 23.7. The present study does show a similar effect in the NHANES data, with the 27-30 BMI group and, less commonly, the 30-35 BMI group having a lower risk than the reference, normal weight, group in some models. This effect was largely not present in the CRL data, and this may have to do with the screening effect of the underwriting process and other differences between the two populations.

This study demonstrates that BMI is just one part in a cluster of factors identifying mortality risk. The relationship between BMI and mortality is multifactorial, and the expression of that risk is demonstrated in the other criteria of the metabolic syndrome (also known as Syndrome X). Dyslipidemia, hypertension, metabolic dysfunction-associated steatotic liver disease (MASLD), coronary disease and type 2 diabetes are all recognized to be associated with mortality and high BMI. Thus, it is not surprising that, as variables related to these disorders are added to Cox models examining the relationship between BMI and mortality, that the coefficients would shrink. It is perhaps surprising that the CRL-based models converged such that only individuals in the highest BMI category (40+) had hazard ratios around 1.5 (a bit higher for middle-aged men). For the insurance industry, these findings suggest that, when there is a thorough assessment of other risk factors, the mortality risk of BMI is substantially blunted but still elevated to a significant degree in the heaviest applicants. So, higher ratings for applicants with high BMI are often warranted, though consideration should be given to lowering the additional debits when ratings are already being given for diseases or conditions that are, in effect, the manifestation of the risks associated with high BMI.

CONCLUSION

The mortality effect of high BMI is substantially mediated by the inclusion of variables related to the diseases and conditions promoted by overweight and obesity.