Body Mass Index and Health Care Utilization in Diabetic and Nondiabetic Individuals

Background:Although controversial, most studies examining the relationship of body mass index (BMI) with mortality in diabetes suggest a paradox: the lowest risk category is above normal weight, versus normal weight in nondiabetic persons. One proposed explanation is greater morbidity of diabetes in normal weight persons. If this were so, it would suggest a health care utilization paradox in diabetes, paralleling the mortality paradox, yet no studies have examined this issue. Objective:To compare the relationship of BMI with health care utilization in diabetic versus nondiabetic persons. Design:Population-based cross-sectional study. Subjects:Adults in the 2000–2011 Medical Expenditures Panel Surveys (N=120,389). Measures:Total health care expenditures, hospital utilization (≥1 admission), and emergency department utilization (≥1 visit). BMI (kg/m2) categories were: <20 (underweight); 20 to <25 (normal); 25 to <30 (overweight); 30 to <35 (obese); and ≥35 (severely obese). Adjustors were age, sex, race/ethnicity, income, health insurance, education, smoking, co-morbidity, urbanicity, region, and year. Results:Among diabetic persons, adjusted mean total health care expenditures were significantly lower in obese versus normal weight persons ($1314, 95% confidence interval [CI], $513–$2115; P=0.001). By contrast, among nondiabetic persons, total expenditures were nonsignificantly higher in obese versus normal weight persons (−$229, 95% CI, −$460 to $2; P=0.052). Findings for hospital and emergency department utilization exhibited similar patterns. Conclusions:Normal weight diabetic persons used substantially more health care than their overweight and obese counterparts, a difference not observed in nondiabetic persons. These differences support the plausibility of a BMI mortality paradox related to greater morbidity of diabetes in normal weight than in heavier persons.

M ost 1-6 (but not all 7 ) studies examining mortality in diabetic individuals find the body mass index (BMI) category with lowest risk is overweight or obese, rather than normal weight as it is in nondiabetic persons. This counterintuitive pattern is known as the BMI mortality paradox in diabetes. 8 The concept of such a paradox remains controversial, [9][10][11] in part due to a lack of studies exploring potential explanations.
One proposed explanation for a mortality paradox is that above normal BMI is protective in diabetic persons. 9 However, this explanation lacks parsimony as it fails to explain the lack of a protective effect of above normal BMI in nondiabetic persons. Alternatively, given that diabetes (especially type 2, the most common form in adults 12,13 ) represents a spectrum of disease, persons who have diabetes despite being normal weight may have a more morbid form of the condition than heavier persons, for whom diabetes may primarily be a consequence of excess weight. 3,8,[14][15][16][17][18] If normal weight diabetic persons tend to have a more morbid form of the condition, they may consequently use more health care than heavier diabetic persons, resulting in a utilization paradox. However, whether a utilization paradox exists in diabetes is unstudied.
Longitudinal studies examining only persons with incident diabetes are often viewed as optimal in addressing such issues. 6,7 However, by excluding participants with existing diabetes, many of whom have longer-standing, earlier onset, more severe disease, such studies may produce biased findings. Cross-sectional studies, though limited in other ways, reduce the potential for such bias. Using cross-sectional data from the 2000-2011 national Medical Expenditures Panel Surveys (MEPS), we examined the relationship of BMI with total health care expenditures and hospital and emergency department utilization in diabetic and nondiabetic persons. the United States, employing an overlapping panel design. 19 The analytic sample for the current study included persons aged 18-90 years at entry. The study was exempted by the local Institutional Review Board.
In the MEPS, the Household Component includes information on respondent sociodemographics and health insurance, and a self-administered questionnaire includes items on smoking, health conditions, and height and weight. The full-year response rate varied from 70.5% to 59.4% for the 2000-2011 panels. 19 Measures

Health Care Utilization
The MEPS collects detailed information about health service use, including office and emergency department visits, inpatient hospitalizations, and prescription drug use. Self-reported health care use is validated and verified by standardized medical record abstraction among a subsample of respondents. The MEPS also ascertains from respondents and physicians the sum of insurance payments and out-ofpocket costs for services received. We used these data to specify total health care expenditures (in dollars), hospital admissions (Z1 admission), and emergency department utilization (Z1 visit).
BMI in kg/m 2 was constructed from self-reported height and weight. BMI categories employed were: <20 (underweight); 20 to <25 (normal); 25 to <30 (overweight); 30 to <35 (obese); and Z35 (severely obese). These categories correspond to those widely employed by clinicians, except for the underweight and normal weight categories, typically defined in clinical practice and most research before 2000 as <18.5 and 18.5 to <25. 20 A BMI of <20 was employed to distinguish underweight in the current analyses, since prior work indicates health status worsens sharply below that cut point, likely due to the effects of concurrent illnesses. 21 Classifying individuals with a BMI of 18.5 to <20 as normal weight would artificially increase the risk of health care utilization associated with normal weight and decrease the risk of utilization associated with overweight and obesity.

Health Conditions and Sociodemographics
Diabetes was self-reported, as were 8 other chronic conditions: cancer, hypertension, coronary heart disease, myocardial infarction, cerebrovascular disease, asthma, emphysema, and arthritis. Agreement between respondent-reported and clinician-reported health conditions is high. 22 Respondents also self-reported whether they used insulin and, in 2008-2011 only, the year of their diabetes diagnosis. Self

Data Analysis
Data were analyzed using Stata 13.1 (Stata Corporation, College Station, TX), adjusting for the complex survey design of MEPS. Data were analyzed using longitudinal strata and primary sampling unit identifiers and survey weights, to derive estimates representative of the US civilian, noninstitutionalized adult population.
The primary analyses examined how diabetes moderated the association of BMI category with total health care expenditures, hospitalizations (Z1 vs. none), and emergency department utilization (Z1 visit vs. none) (the dependent variables). Hospitalization and emergency department use were modeled using logistic regression. We employed a 2stage modeling approach to examine expenditures because they are highly skewed. 23 Logistic regression was used to model predicted expenditures versus no expenditures; among those with any expenditure, a generalized linear model was implemented using a gamma distribution and a log link. This distribution and link were found appropriate using the modified Park test. 23 The results from the 2 stages were combined by multiplying together the predicted probability of any utilization (from the logistic regression) and the amount of utilization (from the generalized linear model), and confidence intervals (CIs) were derived using the delta method. 24 As the optimal modeling of expenditures is uncertain we also examined a single-stage Poisson model. Results from the Poisson model were consistent with those of the 2-stage model and are not presented here. The key independent variables included in all models were diabetes status (present vs. absent), BMI category, as well as a diabetes status Â BMI category interaction term. All models adjusted for sociodemographic characteristics ( To facilitate interpretation of the net adjusted associations between BMI category and health care utilization, the findings of all models are presented as adjusted predicted marginal effects. 25 We examine the difference between the marginal utilization for the BMI category associated with the lowest utilization in diabetic persons with that found among normal weight diabetic persons. We contrast that difference with the difference observed between the same BMI categories among nondiabetic persons. We refer to the resulting difference as a "difference in difference." As a measure of effect size, we also report the difference in difference as the percentage of the adjusted utilization in normal weight nondiabetic persons, the modal group. We conducted several sensitivity analyses. Because MEPS data do not permit determination of whether diabetic individuals had type 1 or type 2 diabetes, we conducted expenditure analyses: (a) limited to persons aged Z65 (in whom the relative prevalence of type 1 vs. type 2 diabetes is low 26,27 ); and (b) excluding respondents who reported taking insulin.
Additional sensitivity analyses excluded smokers and persons with cancer, to explore potential reverse causation (ie, smoking or cancer causing both low BMI and increased health care use). Finally, for diabetic persons in the 2008-2011 surveys, we examined the effects of additionally adjusting for duration of diabetes, a possible confounder in a cross-sectional analysis.

RESULTS
There were 138,944 adults aged 18-90 entering the MEPS cohorts between 2000 and 2011; 120,389 (88.3%, population weighted) had complete data and were included in the current analyses. Compared with participants who did not report diabetes, those reporting diabetes were older and more likely to be Hispanic (any race) or non-Hispanic black, have low household income and education, have public insurance, reside in the South and in nonurban areas, have more comorbid chronic health conditions, be nonsmoking, and be obese or severely obese (Table 1). Participants reporting diabetes also had higher total health care expenditures and were more likely to have hospital and emergency department utilization compared with those not reporting diabetes (Table 1).
In unadjusted utilization analyses (Table 2), across the range of study BMI categories, total expenditures and hospital and emergency department utilization were higher among respondents with versus without diabetes. Among diabetic persons, all 3 unadjusted utilization measures were lowest among those in the overweight category. By contrast, for nondiabetic persons, unadjusted mean total expenditures were lowest in the normal weight category, and unadjusted hospital and emergency department utilization were lowest among (and did not significantly differ between) normal weight and overweight persons.
In adjusted analyses, among diabetic persons the mean total health care expenditures were significantly lower in the obese group than in the normal weight group ($1314, 95% CI, $513-$2115; P = 0.001) ( Table 3 and Fig. 1). By contrast, among nondiabetic persons the adjusted mean total expenditures were nonsignificantly higher in the obese group than in the normal weight group (À $229, 95% CI, À $460 to $2; P = 0.052) ( Table 3 and Fig. 1). The adjusted difference between nondiabetic and diabetic persons in the difference in total expenditures for the normal weight versus obese category was significant ($1543, 95% CI, $708-$2378; P < 0.0001). This difference represents 40.4% (95% CI, 18.4%-62.4%; P < 0.0001) of the adjusted total expenditures in normal weight nondiabetic persons. The overall BMI category Â diabetes status 2-way interaction term also was significant (F 4445 = 3.97; P = 0.003).
Regarding adjusted hospital utilization (Table 3 and Fig. 2A), among nondiabetic persons the utilization was not significantly different in the normal weight group than in the overweight group (0.2%, 95% CI, À 0.3% to 0.6%; P = 0.52). Contrasting with this finding, in diabetic persons adjusted hospital utilization was significantly higher in the normal weight group than in the overweight group (3.9%, 95% CI, 1.8%-6.0%; P < 0.001). The adjusted difference between nondiabetic and diabetic persons in the difference in hospital admissions for the normal weight versus overweight category was significant (3.7%, 95% CI, 1.6%-5.8%; P = 0.007). This difference represents 47.1% (95% CI, 20.1%-74.1%; P = 0.001) of the adjusted hospital admission percentage in normal weight nondiabetic persons. The overall BMI category Â diabetes status 2-way interaction term also was significant (F 4445 = 5.29; P = 0.001).
Regarding adjusted emergency department utilization (Table 3 and Fig. 2B), among nondiabetic persons the utilization was not significantly different in the normal weight group than in the overweight group (À 0.2%, 95% CI, À 0.8% to 0.4%; P = 0.55). By contrast, in diabetic persons adjusted emergency department utilization was significantly higher in the normal weight group than in the overweight group (3.5%, 95% CI, 0.9%-6.1%; P = 0.01). The adjusted difference between nondiabetic and diabetic persons in the difference in emergency department use for the normal weight versus overweight category was significant (3.6%, 95% CI, 1.0%-6.3%; P = 0.007). This difference represents 27.7% (95% CI, 7.4%-48.0%; P = 0.008) of the adjusted emergency department use in normal weight nondiabetic persons. The overall BMI category* diabetes status 2-way interaction term also was significant (F 4445 = 5.00; P < 0.001).
In a sensitivity analysis of adjusted total expenditures limited to those aged Z65 years (N = 19,952 without diabetes and N = 4673 with diabetes), the adjusted difference between nondiabetic and diabetic persons in the difference in total expenditures for the normal weight category versus the overweight category (the nadir for diabetic persons) was significant ($2581, 95% CI, $640-$4522; P = 0.009). This difference represents 36.6% (95% CI, 9.0%-64.1%; P = 0.009) of the adjusted total expenditures in normal weight nondiabetic persons (Appendix Figure 1, Supplemental Digital Content 1, http://links.lww.com/MLR/A898). Similar patterns were observed in analyses excluding those taking insulin (Appendix Figure 2, Supplemental Digital Content 2, http://links.lww.com/MLR/A899) and when smokers and individuals with cancer were excluded (Appendix Figure 3, Supplemental Digital Content 3, http:// links.lww.com/MLR/A900). Finally, the expenditure pattern observed for diabetic persons was not substantially changed in an analysis additionally adjusting for duration of diabetes (data not shown). Details of the sensitivity analyses are available from the authors.

DISCUSSION
In a national sample including both diabetic and nondiabetic persons, we found evidence of a BMI health care utilization paradox in diabetes. Diabetic persons had higher total health care expenditures and hospital and emergency department utilization than nondiabetic persons across all BMI categories. However, for all three outcomes, the utilization nadir for diabetic persons was significantly shifted to the right (ie, to a higher BMI category-see Figs. 1 and 2) as compared with nondiabetic persons. The observed effects (differences in differences) were substantial, varying from 28% to 47% of the respective utilizations of normal weight nondiabetic persons (the modal group).
To our knowledge, only one prior study has examined health care utilization by BMI category in diabetes, finding expenditures were highest among normal weight diabetic persons. 28 However, the study was limited to Veterans Health Administration clinic users aged 65 and older, and lacked a concurrent nondiabetic control group. The absence of a nondiabetic control group precluded examination of a utilization paradox in that sample. Our finding of a BMI health care utilization paradox in diabetes, based on comparisons between concurrent nationally representative samples of diabetic and nondiabetic persons, has not previously been reported.
Although the observational nature of our analyses precludes causal inference, the study findings are consistent with the notion that normal weight diabetic persons may have a more severe or morbid variant of diabetes than their heavier counterparts, perhaps due to genetic and physiological differences. 3,8,[14][15][16][17][18] The hypothesis that above normal BMI may somehow be protective in diabetes seems less plausible due to lack of parsimony, given no evidence of a protective effect of above normal BMI in our concurrent nondiabetic sample. Our findings are consistent with diabetes in normal weight persons being a more morbid form of the disease than that observed in higher weight persons and do not suggest that diabetic individuals with a normal BMI should gain weight, or that overweight or obese diabetic persons should avoid losing weight. Moreover, there is evidence of potential benefits of weight loss in diabetes. 29 Despite the fact that nearly all of the prior studies examining mortality risk by BMI category in diabetes found overweight or obese diabetic persons had lower mortality risk than their normal weight counterparts, 1-6 the notion of a BMI mortality paradox remains controversial. 9-11 The controversy was newly fueled by the study of Tobias et al, 7 in which normal weight was the BMI category of lowest mortality risk in a cohort of persons with incident diabetes. However, the study was limited to examining a selected sample of nurses and physicians aged 30 and older with incident diabetes (mean age at diagnosis Z60 y), meaning people with earlier onset and potentially more severe diabetes were excluded. Such an approach may account for the differing findings of the study conducted by Tobias and colleagues as compared with the other studies in this realm. All of the other studies examined more broadly representative samples in terms of age, other sociodemographics, comorbidity, and duration of diabetes.
Also, the study of Tobias and colleagues and several other studies in this realm lacked a concurrent nondiabetic comparison group. 2,[4][5][6] Both of the studies that compared mortality risk by BMI category in concurrent diabetic and nondiabetic samples found evidence of a BMI mortality paradox in diabetes. 1,3 The current study findings, stemming from analyses of concurrent samples with and without diabetes, support the notion of greater morbidity (and mortality risk) in normal weight versus heavier diabetic persons, manifested in higher health care utilization in normal weight persons.
MEPS data do not permit determination of whether diabetic individuals in our sample had type 1 or type 2 diabetes. However, more than 90% of adults with diabetes have type 2 diabetes, 12,13 suggesting that most of the diabetic individuals in our sample had type 2 diabetes. Over 85% of persons with type 2 diabetes are overweight or obese. 30 Thus, our findings are likely to reflect the net impact of overweight and obesity on health care utilization in type 2 diabetes. Further supporting this conclusion is that the BMI utilization paradox was similar in 2 different sensitivity analyses of total expenditures aimed at reducing the potential influence of type 1 diabetes on the findings, one limited to persons aged Z65 years (among whom the incidence and prevalence of type 1 diabetes are low 26,27 ) and one excluding all persons taking insulin. Nonetheless, analyses of data sets that permit separate analyses for type 1 and type 2 diabetic persons will be required to further explore their relative contributions to the BMI utilization paradox. We also observed similar results to those of our main analyses in 2 further sensitivity analyses of total expenditures, one excluding smokers and persons with cancer and the other (limited to diabetic persons) adjusting for duration of diabetes. These findings do not exclude the possibility that reverse causation and duration of diabetes, respectively, affected our main analyses. However, they do suggest that the findings of the main analyses are unlikely to primarily reflect the effects of reverse causation or diabetes duration. As others have emphasized, a more robust examination of reverse causation will require studies in which subjects are followed throughout their life-course, with weight and utilization measured repeatedly. 31 A strong point of our study was the use of national data collected within the past 15 years from concurrent, broadly representative samples of diabetic and nondiabetic persons. Our study also had limitations. As noted previously, the study was observational, so causal associations cannot be inferred, and unmeasured confounding could have contributed to the findings. While our cross-sectional analysis had some advantages over prospective studies of incident diabetes-such as the inclusion of individuals with relatively early onset and more severe diabetes-a disadvantage was the lack of information regarding the durations of comorbid conditions. Still, the effects of unmeasured comorbidity and duration of comorbidity would have to be substantial, and to operate differentially on persons with versus without diabetes, to have meaningfully impacted on the findings. BMI was derived from self-reported height and weight. Prior studies suggest a complex relationship between selfreported and objectively measured BMI, with differences in BMI category misclassification resulting from self-reports based on sociodemographics (eg, country of residence, sex, race/ethnicity) and BMI category (eg, tendency to underestimate BMI among heavier persons vs. overestimate BMI among less heavy individuals). [32][33][34][35][36] Research also indicates that people who perceive themselves as being normal weight are less likely to report impaired health status than those who perceive they are overweight, regardless of actual BMI 37 ; as such they may perceive less need for health care. The net effects of such relationships on the associations we observed are uncertain. Studies employing objectively measured BMI and diabetes will be required to explore these issues.
In conclusion, in comparing total health care expenditures as well as hospital and emergency department utilization by BMI category in concurrent national samples of diabetic and nondiabetic persons, we found evidence of a health care utilization paradox in diabetes. For all 3 measures, utilization was higher in normal weight than heavier diabetic persons, differences not observed in nondiabetic persons. Although the concept of a BMI mortality paradox in diabetes remains controversial, the current findings are consistent with notion that the paradox may be due to a more morbid form of diabetes in normal weight versus heavier persons.