Changes in Healthcare Spending Attributable to Obesity: Payer- and Service-Specic Estimates

Background: National efforts to control US healthcare spending are potentially undermined by changes in patient characteristics, and in particular increases in rates of obesity. The objective of this study was to provide current estimates of the effect of obesity on healthcare spending overall, by service line and by payer using the NIH classications for BMI. Methods: We used a quasi-experimental design and analyzed the data using generalized linear models and two-part models to estimate obesity-attributable spending. Data was drawn from the 2006 and 2016 Medical Expenditures Panel Survey. We identied individuals in the different BMI classes based on self-reported height and weight. Results: Total medical costs attributable to obesity rose to $126 billion per year by 2016, although the marginal cost of obesity declined for all obesity classes. The overall spending increase was due to an increase in obesity prevalence and a population shift to higher obesity classes. Obesity related spending between 2006 and 2016 was relatively constant due to decreases in inpatient spending, which were only partially offset by increases in outpatient spending. Conclusions: Obesity related spending between 2006 and 2016 was relatively constant due to decreases in inpatient spending, which were only partially offset by increases in outpatient spending. Obesity class 2 and 3 were the main factor driving spending increases, suggesting that persons over BMI of 35 should be the focus for controlling spending.

Results: Total medical costs attributable to obesity rose to $126 billion per year by 2016, although the marginal cost of obesity declined for all obesity classes. The overall spending increase was due to an increase in obesity prevalence and a population shift to higher obesity classes. Obesity related spending between 2006 and 2016 was relatively constant due to decreases in inpatient spending, which were only partially offset by increases in outpatient spending.
Conclusions: Obesity related spending between 2006 and 2016 was relatively constant due to decreases in inpatient spending, which were only partially offset by increases in outpatient spending. Obesity class 2 and 3 were the main factor driving spending increases, suggesting that persons over BMI of 35 should be the focus for controlling spending.

Background
Obesity has been identi ed as one of the key drivers of increased healthcare spending and reduced life expectancy in the United States1-5 and worldwide6. Obesity has been linked to a multitude of health conditions, including coronary heart disease7, chronic renal failure8, many cancers, sleep apnea, gallbladder disease9, Type 2 Diabetes10 and other conditions. The link between obesity and chronic illness is the reason for the link between obesity and reduced life expectancy3, 4.
There has also been an extensive investigation of the impact of obesity on healthcare spending. Obesity was identi ed as one of the key drivers of increased healthcare spending during the 1996-2006 time period1, with the effect largely driven by increases in spending on chronic diseases caused by obesity5.
More recent work has found that the proportion of spending attributable to obesity increased by 29% from 2001 to 2015, from 6.1% to 7.9%11 with obese adults having higher inpatient and prescription drug spending, in particular12. The costs of obesity are higher in more obese individuals, both overall and for particular chronic illnesses, such as diabetes13. Interesting, there is some evidence that the effect of obesity on spending may have moderated in recent years, with a statistically insigni cant decrease in spending from 2010 to 2013, from $3,74814 to $3,429.
The more recent economic literature has begun measuring the effect of obesity by Body Mass Index (BMI) categories, mirroring the medical community. This is done using the National Institutes of Health This study makes a number of new contributions to the existing literature on the effect of obesity on healthcare spending. First, we measure the effect of obesity on spending by service line and payer using the NIH classi cations for obesity. Previous service line and payer speci c estimates used the more general obese / non-obese framework17, 18, which may miss important nuances if the effect of obesity is concentrated in the higher categories1. Second, reforms in the Affordable Care Act have shifted payer types, particularly through Medicaid expansions, which may have changed the distribution of payers from previous studies. Third, we examine the effect of different obesity classes on spending, by service line, to understand differences in how utilization occurs for different levels of obesity. Finally, we provide a careful examination of the suggestive evidence cited above that the effect of obesity may have moderated over more recent years. To do this, we analyze ten-year trends in obesity rates and obesityinduced spending and model the changes in spending for different BMI classes. In our empirical model, our dependent variables are healthcare expenditures, including total expenditures, inpatient, non-inpatient, and drugs expenditures. Non-inpatient is de ned as outpatient and o ce-based expenditures. The main explanatory variable is BMI categories. BMI was used to create dummy variables for four BMI categories, overweight (BMI 25-29.9), BMI obesity class 1 (30-34.9), BMI obesity class 2 (35-39.9) and BMI obesity class 3 (extreme) (above 40). BMI was calculated based on self-reported height and weight. The BMI class "normal" (18.5-24.9) was the reference group in all models. Individuals with a BMI less than 18.5 were coded as "underweight"; underweight is controlled for in the model but not reported in the tables. The models controlled for sociodemographic and health characteristics that are not in the causal pathway between obesity and spending.

Methods
The control variables are drawn from the MEPS data, and include gender, race/ethnicity, smoking status, marital status, region of the country, education and family income. Age was included and coded as a categorical variable for ages 18-34, 35-44, 45-54, 55-64, 65-74 and 75+. Expenditures were modelled using Generalized Linear Models (GLM) for total and non-inpatient expenditures; inpatient and drugs spending were modelled using two-part models (TPM)20, 21 For all the expenditures classes, we performed a Modi ed Park test to identify the distribution of the expenditure data and the coe cient of the conditional variance function. The test supported the choice for GLM with gamma family and log link for all models. We used the Hosmer-Lemeshow test for goodness of t. We calculated standard errors using bootstrap with 1000 iterations per model. Differences between coe cients were estimated using a standard t test. Observations with missing data for insurance (n=265 for 2006 and 256 for 2016) or BMI (n=741 and 701) were omitted from the analysis.
We also estimated the attributable fraction for obesity, which is equal to the ratio of the change in spending with and without obesity divided by total spending. The AF represents the proportion of spending attributable to the different BMI categories, controlling for other variables in the model. The estimated magnitude of the cost of obesity in previous work has varied considerably, perhaps driven by different study methodologies22. The advantage of using the AF methodology is that the estimates can be updated periodically to track the cost effect of BMI. This approach has been previously used in obesity as well smoking23 and falls in older adults24,25. Standard errors were calculated using a bootstrap method with 200 replications. We used STATA 15 for all analysis. Expenditure numbers from 2006 were adjusted to 2016 prices using the gross domestic product implicit price de ator (GDP de ator) from the Bureau of Economic Analysis26. The general price de ator was preferred to allow for differences in the social value of healthcare interventions.

Results
We rst estimated the marginal effect of obesity (in dollars), by BMI category, on overall healthcare spending (Table 1).  Being overweight had no effect on spending overall. Although 33% of the population was overweight in 2016, the marginal effect overall (Table 1) and by payer ( Meanwhile, there was a small decrease for Obese 1 in non-inpatient spending. Prescription drug spending was relatively at for Obese 3 and Obese 2, but declined from $643 to $379 for Obese 1. Table 3 suggests a shift in spending for obesity. For Obese 3, the most expensive spending category in 2006 was inpatient spending ($1,110), followed by prescription drugs ($1,031) and non-inpatient spending ($714). In contrast, the top expense in 2016 was for prescription drugs ($1,046), with inpatient spending third ($727). Obese 2 showed the same general pattern: a very slight decline in drug spending, an increase in non-inpatient spending and a decrease in inpatient spending. Overall, changes in the attributable fraction of healthcare spending varied depending on the service line and BMI category (Table 4). For inpatient care, the attributable fraction declined for Obese 2 (3.2% to 1.9%) and Obese 3 (3.9% to 2.5%), but increased for Obese 1 (2.4% to 3.5%). Non-inpatient and prescription drug spending had exactly the opposite pattern, with the attributable fraction decreasing for Obese 1 (5.6% to 4.0%) while increasing for Obese 2 (3.0% to 3.5%) and Obese 3 (2.2% to 2.5%). Finally, the attributable fraction for prescription drug spending decreased for both Obese 1 (7.6% to 5.3%) and Obese 2 (4.3% to 4.1%) and increased slightly for Obese 3 (4.2% to 4.3%). Attributable fraction (%) and Total spending ($ million). Standard error in parentheses.
[1] The Totals were calculated by adding up the columns. We also calculated AF and Total Spending independently and found similar results.
increase in prescription drug spending alone. Spending in Obese 2 had the largest overall increase, with an increase of nearly $4B in prescription drugs ($10.2B to $14.1B) and $6B in non-inpatient care. Spending for Obese 1 was largely at for prescription drugs and non-inpatient care. Inpatient spending declined for Obesity Class 2 ($3.2B) and Obesity Class 3 ($3.4B), but increased for Obese Class 1 ($4.1B).
The effect by payer varied (Table 5). Medicare experienced an increase in attributable fraction for Obese 1 (from 3.1% in 2006 to 3.6% in 2016) and Obese 2 (1.8% to 2.4%) and a decline for Obese 3 (from 4.0% to 2.3%). Medicaid also experienced an increase in attributable fraction for Obese 1 (from 2.7% to 3.5%), while Private Insurance saw a decrease for that same class (from 7.0% to 3.6%). Both Medicaid and Private Insurance saw deceases in the attributable fraction for Obese 2 (3.8% to 2.2% and 4.2% to 2.9%, respectively).

Discussion
In this paper we nd that spending associated with obesity has changed in some important ways over the past ten years. First, we show that the spending on obesity is increasingly focused in individuals in Obesity Class 3 (Extreme). These individuals are 5% of the total population and only about one in six obese persons fall into this class. Yet more than a quarter of obesity related costs (26.1%) are concentrated in this group. And this is the group that is proportionately growing the fastest, with a 32% increase over the past decade.
For other obesity classes, spending has been more effectively controlled and total spending has been relatively at. The models separating the effect of changes in obesity prevalence and the relationship of obesity and spending indicate that the latter is the reason for the moderation in effect. This is largely due to a shift from inpatient care to outpatient care coupled with slight reductions in prescription drug spending. Also, despite the coverage expansions in the Affordable Care Act, the majority of spending remains paid for by private insurance (67B), rather than Medicare (43B) or Medicaid (16B). Spending for overweight persons is insigni cantly different from normal weight spending, which may suggest a lost opportunity to intervene.
There are a number of limitations to this study. First, the analysis is based on MEPS data. Other data sources may have different spending numbers, particularly due to the inclusion or exclusion of long-term care spending. The advantage of MEPS is its widespread usage as a measure of healthcare spending, which allows comparisons to other studies. Second, our data is based on self-report height and weight as there are no nationally representative data set that includes both measured height/weight and annual medical spending2. Previous research concluded that reporting error in weight can lead to bias in estimates of the healthcare consequences of obesity and the extent of underreporting increases with measured weight27.Our results could be interpreted as an upperbound of the true effect, however, interpreting the annual coe cients as an upper bound seems reasonable. However, as we look at changes, the bias may cancel out. Endogeneity is also possible in this study if there are unobserved characteristics associated with both insurance and obesity. Comparisons across different studies should be done with caution because of differences in model design, sampling frame and control variables. Cautions shoudl be used in interpreting differences between payers, given that patients with different arrangements face very different prices, which may lead to differences in demand or access for a given level of health. We do not include the uninsured in our study because our focus is on changes by payer.
Future research should examine how obesity affects the uninsured.
We stratify the analysis by age and insurance status to be able to compare results to comparable work done with earlier years, so time trends could be established. It would have been interesting to stratify by gender as well, but it is di cult to treat the gender issue carefully without providing signi cant additional context and result tables.

Conclusions
Overall, we nd that the obesity attributable fraction of healthcare spending has actually declined over the past decade, despite increased obesity prevalence. Our result suggest this success is due to the shift from inpatient to outpatient settings for care. These ndings suggests there are two potential conclusions that may be drawn from this. First, obesity has not been the key driver of increases in healthcare spending over the past decade. Obesity related spending has increased, but other spending (the denominator) has increased more quickly. To understand why costs have increased over the past decade, analysts need to look for other culprits.
Second, obesity may be a more important cost driver in the next decade. The proportion of the population which is obese is increasing. Over the past decade, the increased prevalence was offset with changes in the pattern of spending -inpatient to non-inpatient -which moderated the increase. Without further reductions in per capita spending, the effect of increases in the proportion of the population which is obese may have a larger effect on healthcare spending. This is particularly true because of the increase in extreme obesity. Future efforts to control obesity-related spending are likely to be most impactful if they concentrate on individuals with BMI over 40 as well as preventing individuals from progressing to high levels of obesity.