Pronk NP, Goodman MJ, O'Connor PJ, Martinson BC. Relationship Between Modifiable Health Risks and Short-term Health Care Charges. JAMA. 1999;282(23):2235-2239. doi:10.1001/jama.282.23.2235
Author Affiliation: HealthPartners Center for Health Promotion and HealthPartners Research Foundation, Minneapolis, Minn.
Context If physical inactivity, obesity, and smoking status prove to contribute
significantly to increased health care charges within a short period of time,
health plans and payers may wish to invest in strategies to modify these risk
factors. However, few data are available to guide such resource allocation
Objective To examine the relationship of modifiable health risks to subsequent
health care charges after controlling for age, race, sex, and chronic conditions.
Design, Setting, and Participants Cohort study of a stratified random sample of 5689 adults (75.5% of
total sample of 7535) aged 40 years or older who were enrolled in a Minnesota
health plan and completed a 60-item questionnaire.
Main Outcome Measure Resource use as measured by billed health care charges from July 1,
1995, to December 31, 1996, compared by health risk (physical activity, body
mass index [BMI], and smoking status).
Results The mean annual per patient charge in the total study population was
$3570 (median, $600), and 15% of patients had no charges during the study
period. After adjustment for age, race, sex, and chronic disease status, physical
activity (4.7% lower health care charges per active day per week), BMI (1.9%
higher charges per BMI unit), current smoking status (18% higher charges),
and history of tobacco use (25.8% higher charges) were prospectively related
to health care charges over 18 months. Never-smokers with a BMI of 25 kg/m2 and who participated in physical activity 3 days per week had mean
annual health care charges that were approximately 49% lower than physically
inactive smokers with a BMI of 27.5 kg/m2.
Conclusions Our data suggest that adverse health risks translate into significantly
higher health care charges within 18 months. Health plans or payers seeking
to minimize health care charges may wish to consider strategic investments
in interventions that effectively modify adverse health risks.
Sedentary lifestyle, obesity, and tobacco use are strongly related to
a variety of long-term adverse health outcomes, and over a lifetime, have
significant costs to society.1,2
In formulating clinical policy and making resource allocation decisions, health
care delivery organizations need to know the short-term relationship of modifiable
health risks to health care charges. Managed care organizations (MCOs) attempt
to maximize the value obtained from each health care dollar spent. Because
adverse, modifiable health risks are detrimental to health and contribute
to higher costs of health care, many MCOs may be interested in devoting resources
to initiatives that favorably affect the health-related behaviors of their
members. This is particularly true in light of recent reports documenting
the efficacy of behavior change intervention strategies,3
lower health care charges with risk reduction,4
and a good return on investments in health-related behaviors.5
However, little financial data are available to guide resource allocation
decisions, or to estimate the potential financial impact of programs that
may modify health risks. Moreover, to our knowledge, no report has adequately
controlled in the analysis for chronic conditions, which may confound the
relationship of health risks to charges.
In this study we analyze the short-term cost to health plans of modifiable
health risks. These data provide an estimate of excess costs incurred by health
plans that do nothing to influence modifiable health risks, and may assist
health plans in deciding whether strategic investments to modify certain health
risks in members is a wise use of scarce resources.
The study was approved by the HealthPartners Institutional Review Board
and conducted at HealthPartners, a Minnesota health plan with both owned and
contracted clinics. All patients who were aged 40 years or older and enrolled
on December 15, 1994, were potential subjects for the study. These individuals
were classified using the International Classification of
Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) and pharmacotherapy databases as having 1 of 4 chronic conditions.
A diagnosis of diabetes mellitus was assigned if the patients had 2 or more ICD-9 codes 250.xx, or a filled prescription for a diabetes-specific
drug such as insulin, a sulfonylurea, or a biguanide in 1994. Heart disease
was assigned if the patients had 1 or more ICD-9
codes 412, 413.9, 429.2, or 428.0 in 1994. Hypertension was assigned if the
patients had 1 or more ICD-9 codes 401, 401.1, or
401.9 in 1994. Dyslipidemia was assigned if the patients had an ICD-9 code of 272.4 in 1994. A more detailed description of the identification
of patients with specific conditions and the sensitivity, specificity, and
positive predictive value of this method has been published.6
From 158,415 patients with none of the 4 conditions, a random sample
of 3000 members (1.89%) was selected. From 34,159 patients who had 1 chronic
condition, a random sample of 2500 patients (7.3%) was selected. From 7571
patients who had 2 or more of the chronic conditions, a random sample of 2500
(33%) was selected. Hence, the total study population included a stratified
random sample of 8000 individuals aged 40 years or older. In August 1995,
a mailed survey was administered and 533 subjects who were unable to complete
it due to death, disenrollment, or language problems were excluded. In addition,
159 proxy respondents were excluded from all analyses. Following postcard
reminders at 7 days, a second mailing to nonrespondents at day 21, and telephone
follow-up, data of a total of 5977 respondents (representing 79% of the total
sample [5977/7535]) were available for analysis.
The 60-question survey instrument included items on demographics, health
status, use of preventive services, modifiable health risks, and readiness
to change modifiable health risks. The core of the survey items was adapted
from the Centers for Disease Control and Prevention's Behavioral Risk Factor
Surveillance System, which has reliability coefficients for behavioral risk
factors above 0.70.7
Health care charges billed from July 1, 1995, to December 31, 1996,
were used to measure relative resource use, and were gathered from the HealthPartners
claims system. Each encounter in either MCO-owned or MCO-contracted clinics
generated such a claim.
Important independent variables included age, race, sex, and chronic
disease status. Prior research has shown that health care charges are associated
with these variables before and after adjustment for functional health status
and other factors.8,9 Age and
sex were obtained from MCO administrative databases. Age was calculated in
years from date of birth to the date of the initial survey. Chronic disease
status was determined based on 1994 data.
Body mass index (BMI) was calculated as self-reported body weight in
kilograms divided by self-reported height in meters squared, and was centered
on its mean value. Physical activity was assessed via self-report and quantified
in relation to recommended guidelines.10 In
addition, physical fitness was predicted using self-reported variables according
to procedures outlined by Ainsworth et al.11
Respondents who reported ever having smoked at least 100 cigarettes and who
indicated they currently smoked were defined as current smokers. Those who
reported ever having smoked at least 100 cigarettes but denied currently smoking
were defined as former smokers.
We used a conditional probability (2-part model) approach to study charges.12 This approach recognizes that charges are the product
of the probability of utilization multiplied by the expected level of utilization
conditional on positive utilization.
The statistical properties of each model were systematically assessed
in nested fashion. First, the significance of each individual partial regression
coefficient was assessed using a t test. This tested
the hypothesis of whether an individual variable was related to variation
in health care charges. Collinearity was assessed using variance inflation
factors. In the semilogarithmic models, none of the variance inflation factors
were greater than 1.20, suggesting that collinearity was not a significant
problem in these models. Second, we analyzed residuals to be sure the distributional
assumptions of the semilogarithmic linear regression were met. Third, overall
goodness of fit was measured with the coefficient of determination, the adjusted r2. Since the sample size was fixed for each
model, the denominator used to calculate r2 was fixed, allowing direct comparison of semilogarithmic models. Overall
logistic model fit was assessed using the Hosmer-Lemeshow χ2
test.13 In the logistic regression, the performance
of an individual coefficient was tested using a χ2 test.
The sampling structure of the survey required that we weight each observation
based on its sampling probability to obtain population estimates. This was
done in all multivariate analyses using standard methods.13
We further tested a generalized linear model assuming a gamma distribution
and log link, which produced essentially identical results. Analyses were
restricted to individuals with complete responses on all analysis variables.
Results from the semilog model were retransformed using standard methods to
present estimated effects in terms of mean charges in dollar units.14
Of the 5977 subjects (79%) who responded to the 1995 survey, 5689 had
complete data on all study variables, and provide the basis of this report.
Characteristics of survey respondents and nonrespondents are shown in Table 1. Health care charges were highly
skewed, with a small proportion of the population accounting for a large proportion
of expense. The upper quintile of subjects accounted for 86% of total charges
and the upper decile accounted for 71% of total charges. The median annual
charge in our total study population was $600 (interquartile range, $151-$2080)
compared with a mean (SD) of $3570 ($12,823). Also, a significant proportion
(15%) of study subjects had no medical encounters or charges during the 18-month
The independent variable with the most missing data was BMI. Missing
data precluded the calculation of BMI for 229 study subjects (198 users).
Among those with health care charges, those with missing BMI had higher charges
when measured on the log scale (P<.001), were
older (P = .006), more likely to be men (P<.001), and more likely to be hospitalized (P = .005). The likelihood of health care charges did not differ between
those with and without missing data for BMI.
Table 2 shows characteristics
of study subjects with and without health care charges during the 18-month
study period. Those with no health care charges were significantly younger,
more likely to be men, more likely to be current smokers, and had significantly
less chronic disease.
In models predicting probability of medical charges, each 10-year increase
in age resulted in a 32% (P<.001) increase in
the odds of nonzero medical charges. Men were 44% (P<.001)
less likely to use medical services than women. Other predictors examined
in this model, including race, diagnoses of diabetes or heart disease, BMI,
physical activity, and smoking status, were not predictive of the probability
of having health care charges.
In all models, we introduced blocks of variables in the following order:
demographic characteristics, chronic diseases, modifiable health risks (physical
activity, obesity, and smoking), and interactions. Acceptable measures for
alcohol intake and nutrition were not available for inclusion in this analysis.
We assessed a limited set of 2-way interaction terms, and none were found
to be significant. Measurements of hypertension and lipid disorders were omitted
since models including these measurements consistently indicated that they
did not predict medical charges and did not improve model fit.
Table 3 shows the combined
results of a 2-part model predicting the presence of charges in the total
sample, and the log of medical charges for those with positive, nonzero log
charges. For the modifiable health risks, medical care charges for study subjects
with no days of physical activity were approximately 4.7% higher than charges
for those who were physically active 1 day a week. Each 1-unit increase in
BMI was associated with a 1.9% increase in medical charges. Current smokers
had medical care charges 18% higher than never-smokers and former smokers
had medical care charges 25.8% higher than never-smokers did. The final column
in Table 3 demonstrates the relationship
of these and other risk factors with annualized median health care charges.
The changes in mean dollars implied by our models vary depending on
the specific characteristics of individuals. Table 4 presents descriptive estimates of mean annual charges for
hypothetical patients with high-risk vs low-risk profiles of modifiable health
risks. These estimates are based on least-squares regression results, estimated
only for subjects with nonzero charges. The high-risk profile was associated
with a fixed-percentage increase of 49% mean annual charges higher than the
low-risk profile. However, across the range of individuals with characteristics
as presented in column 1, the absolute increase in mean charges ranged from
approximately $1500 to more than $2500.
In this study, we found that modifiable health risks were significantly
related to higher health care charges, even after controlling for age, sex,
race, diabetes, and heart disease. Excess medical care charges related to
physical inactivity, obesity, and smoking, were substantial. The findings
suggest that health plans that do not systematically support members' efforts
to improve health-related behaviors may be incurring significant short-term
health care charges that may be at least partly preventable.
Goetzel et al15 described the relationship
between modifiable health risks and health care costs and found that 7 of
10 health risks were significantly related to higher health care expenditures.
In that report, the estimated per-capita annual impact of poor exercise habits
and current smoking amounted to $172 and $227, respectively, in an employed
population that completed a voluntary health-risk appraisal. However, previous
studies have failed to control for diagnosed chronic conditions such as heart
disease or diabetes.15,19 In a
similar analysis, we have presented evidence that employed health plan members
who have increased behavioral risk factors, including obesity and low physical
fitness incur significantly higher annual excess health care expenses.16 The magnitude was $135 per member per year for obesity
and $176 per member per year for low fitness. The present study establishes
the relationship of modifiable health risks in a population-based, random
sample, while also controlling for the presence of major chronic conditions.
Despite differences in the study sample demographics and characteristics,
the results of these studies all support the hypothesis that behavioral risk
factors significantly influence short-term health care expenditures.
However, the present study does not prove that changing modifiable health
risks can reduce health care charges. The fraction of such charges that can
be reversed is not yet known. To reduce charges, health plans would have to
invest resources in effective intervention strategies to reduce the burden
of modifiable health risks in members. Such behavioral interventions are sometimes
viewed with skepticism for several reasons. First, the cost of the interventions
may offset potential savings in health care charges. Second, interventions
that target modifiable health risks may have limited effectiveness. Third,
for some modifiable health risks, there may be substantial lag time between
change in the risk factor and subsequent changes in health care costs. Finally,
if health plan disenrollment rates are high, savings associated with health-risk
reduction may not be fully realized by the health plan that invested in behavior
While health-risk reduction is not a panacea for controlling health
care costs, we believe that payers often undervalue its merits. Other reports
document better clinical outcomes,17,18
and suggest lower subsequent costs,3,5,19,20
when modifiable health risks are improved. While some behavior change interventions
are expensive, others are less costly, yet effective, and the return on investment
may be favorable.5 Moreover, we have shown
in previous work that readiness to change and willingness to participate in
health improvement activities tend to be greatest in those at highest risk
of poor outcomes.16,21,22
Further, this report documents a short lag time between the measurement
of modifiable health risks and their impact on health care charges. This short
lag time, only 18 months in this study, is plausible from the biological point
of view for physical activity and weight management.23,24
Moreover, while disenrollment rates may vary widely from one health plan to
another,25,26 there is evidence
in recent reports that older, sicker health plan members have lower rates
of disenrollment than do younger, healthier members.27
Because the percentage change in health care charges is relatively constant
across age groups, potential short-term savings may be highest in older, sicker
However, an additional problem relates to the notion that when behavior
change interventions target only high-risk individuals, the observed shift
from high to average risk is offset by individuals who move from average risk
to high-risk status. To achieve lasting health benefit in defined populations,
health plans and their community partners would be well-advised to address
the needs of both average-risk and high-risk members, through intervention
at both the individual and community levels.28
The relationship of smoking status to health care charges is complex.
The higher charges among former smokers than current smokers are likely to
reflect the occurrence of health problems that caused smokers to stop, for
example, after a myocardial infarction.29 Our
data corroborate previously published reports30
suggesting that primary prevention of smoking is an important strategy to
improve population health and reduce charges.
This study has several limitations. Our data were limited to a large
group of insured patients enrolled in a single health plan. The magnitude
of these patients' health care charges may differ from those with indemnity
insurance, although the expected impact would be to make our estimates of
charges conservative. Further, the relatively high educational level and demographic
homogeneity of the MCO membership studied here limit the generalizability
of our results to other groups, especially minority populations. While we
did not include data on alcohol intake and nutrition in our analyses, both
appear to be related in a complex fashion to health care charges.15 Our measurements of BMI and physical activity are
based on self-report, but reporting bias would likely make our estimates of
the effect of modifiable risks on charges conservative.
These limitations should be balanced against the strengths of this study.
The population we studied was large enough to ensure sufficient power in the
analysis, and is comparable with populations enrolled in many other health
plans. The data on health care charges, while not perfect, are derived from
a single source with uniform databases and charges, and are virtually complete.6 The ability to control for important comorbidities
and the use of appropriate analytic models are also strengths of this study.
Also, few other studies have analyzed the relationship of modifiable health
risks to direct health care charges using empiric, individual level data and
appropriate multivariate statistical models.31
Our results suggest that adverse, modifiable health risks contribute
substantially to health care charges. While the reversibility of these charges
with changes in health risks remains uncertain, a recent report4
suggests that reduction in risks may lead to reduced charges, and other work5 indicates a high return on investment to payers with
interventions that reduce health risks. Other data32,33
indicate that risk reduction leads to reduced mortality, which provides a
plausible biological pathway between risks and charges. Health plans and self-insured
employers seeking to maximize health return on each dollar spent for medical
care may wish to consider strategic investments in interventions that effectively
improve modifiable health risks. From a behavioral perspective, primary prevention
of smoking and increased physical activity appear to have substantial potential
to reduce health care charges.