Concern about the rising cost of healthcare is now a central issue in the American economy. Radiology procedures have come under increasing scrutiny since the Medicare Payment Advisory Commission identified imaging as showing the “highest cumulative growth in services per beneficiary” in recent years [1]. Noninvasive diagnostic imaging procedures are widely used procedures, and the accelerating development and use of high-technology imagery (e.g., CT, MRI, and PET) involves high costs. In an earlier study [2], we discussed the increasing use of noninvasive diagnostic imaging and high-cost techniques. The Deficit Reduction Act of 2005 [3] attempted to rein in radiology costs by reducing reimbursements to private offices for high-technology, high-cost procedures, such as MRI and CT. A Government Accountability Office report in 2008 [4] showed Medicare Part B imaging spending more than doubling, from $6.89 to $14.11 billion, between 2000 and 2006.

The New York Times reports intense interest on the part of the president and policy makers in “huge geographic variations in Medicare spending per beneficiary” [5]. A recent study by Iglehart [6] cites the Government Accountability Office's finding of nearly eight-fold variation among states in Medicare imaging expenditures, from $62 per capita in Vermont to $472 per capita in Florida.

The study of geographic variation in disease incidence was a foundation of modern epidemiology [7, 8]. Currently, the Dartmouth Atlas project [9] provides a wealth of information about geographic variation in health care utilization in the United States.

This study provides a comprehensive overview of geographic variation in noninvasive diagnostic imaging utilization in the Medicare population. To our knowledge, this type of overview has not been published before. There have been a number of more limited studies of geographic variation in noninvasive diagnostic imaging. Not surprisingly, the most frequently studied area has been breast cancer screening [10–15]. There have been several geographic studies of other cancer screening and staging examinations [16–19], spinal imaging [20], cardiac noninvasive diagnostic imaging [21, 22], and neurologic noninvasive diagnostic imaging [23]. Only one study [24] attempted a broad survey of geographic variation over the entire spectrum of imaging procedures. However, it reported data only to 2001, categorized noninvasive diagnostic imaging only according to technique, and reported statewide findings by percentile ranks, which are hard to interpret.

In this study, we are concerned with three questions: First, are utilization rates similar or dissimilar across geographic regions? Second, do trends in utilization from 1998 to 2007 differ among regions? Third, do rates and trends vary among procedure categories across regions?

The primary data sources for this study were the Centers for Medicare and Medicaid Services Part B Physician Supplier Procedure Summary Master Files for 1998–2007. These files aggregate Medicare Part B billing claims filed by physicians nationally for all procedures. The billing claims are classified by codes for the type of procedure, region, place of service, and specialty of the providing physician. Both the number of procedures performed and the allowed charges for each category are available. These databases are anonymous public use files and are exempt from review by institutional review boards.

The files cover all Medicare fee-for-service beneficiaries but do not include patients enrolled in managed care plans. Total Medicare enrollment ranged from 38.5 million beneficiaries in 1998 to 45.7 million in 2007. The percentage of healthcare maintenance organization enrollees was 17.2% in 1998, declined to 12.7% in 2004, and increased again to 19.2% in 2007. Therefore, the number of individuals covered by the data sets ranged from 31.9 million in 1998 to 36.9 million in 2007, with a peak of 37.8 million in 2004.

We evaluated all current procedural terminology codes related to noninvasive diagnostic imaging, which were selected from the fourth edition of Current Procedural Terminology for 1998–2007 [25]. These codes are designated by the American Medical Association and are referred to by Centers for Medicare and Medicaid Services as “CPT-4” or “level 1 Healthcare Common Procedure Coding System” codes. Noninvasive diagnostic imaging codes used for this evaluation included most codes in the 70,000 series (which are generally thought of as constituting radiology) and the cardiac and vascular sonography procedures listed in the 90,000 series. This analysis excluded surgical codes, the radiologic supervision and interpretation codes that accompany them, radiation oncology procedure codes, 3D reconstruction codes, and radioimmunoassay codes. Obstetric ultrasound was included in this analysis, although it is underrepresented in the Medicare population. We also included the imaging-related level 2 Healthcare Common Procedure Coding System codes [26], which are five-position alphanumeric codes approved and maintained jointly by the alphanumeric editorial panel (consisting of Centers for Medicare and Medicaid Services, the Health Insurance Association of America, and the Blue Cross and Blue Shield Association). These alphanumeric codes are sometimes used for gathering utilization data and setting reimbursement of procedures before the establishment of a current procedural terminology code by the American Medical Association.

The total number of codes studied in each year varied as a result of the creation of new codes and discontinuation of outdated codes. There were 412 noninvasive diagnostic imaging codes in 1998 and 575 such codes in 2007. For the analyses in this article, we aggregated these codes into 28 categories, reflecting both technique and anatomic location. The number of codes in each category ranged from one for breast ultrasound to 95 for skeletal radiography. Three of the categories—CT cardiac, CT vascular, and PET head, did not exist in all the years of the study.

We used Centers for Medicare and Medicaid Services region codes for the geographic analysis. Ten regional codes (Fig. 1) are named for the city in which the Centers for Medicare and Medicaid Services regional office is located, as follows: Boston, New York, Philadelphia, Atlanta, Chicago, Dallas, Kansas City (Missouri), Denver, San Francisco, and Seattle. These 10 regions include all 50 states and the territories of the United States. One additional code, “Traveler's Railroad,” which identifies a small percentage of Medicare recipients by miscellaneous categories (e.g., railroad workers) rather than by geographic location, was excluded from the analysis of geographic data.

The regional utilization rate was calculated by determining the number of eligible fee-for-service Medicare beneficiaries for each year in each region. These were obtained by summing the county-level data in the Centers for Medicare and Medicaid Services Medicare Managed Care Market Penetration for All Medicare Plan Contractors—Quarterly State/County Data Files, which list total beneficiaries and health maintenance organization enrollees in each county. Two rate statistics were used: per capita and per 100,000 beneficiciary population.

We calculated overall utilization in each of the 10 Centers for Medicare and Medicaid Services regions for each of the years of our study, as well as the percentage change from 1998 to 2001, 2001 to 2004, 2004 to 2007, and 1998 to 2007. We also calculated the overall utilization rate and percentage change across all regions. Pearson's correlation coefficients (r) were computed for the relationship between the 1998 and 2007 rates to the 10-year rate of change for each region. For each of the 28 procedure categories, we calculated the rate per 100,000 population, determined the region with the highest rate, the region with the lowest rate, and the ratio of the highest to lowest region, which is a relative risk statistic. We also calculated allowed charges and relative value units (a measure of work involved in the procedure) per capita for each region and the high-to-low ratio. We used the total professional component relative value unit from the 2007 files published by the American Medical Association. Relative value unit values for a given procedure may change small amounts from year to year, as the amount of work involved is reevaluated; using the same relative value unit across all years eliminates changes in the relative value units per capita that we calculated that are due to this reevaluation process.

The Medicare Part B database represents the total fee-for-service Medicare population. Because these are complete population counts, no inferential statistics are required, as would be the case if we were trying to infer population statistics from sample data.

Underlying regional variation in imaging utilization are concerns about cost and quality. Wide geographic variation in utilization raises questions about underutilization and quality on the low side and overutilization and costs on the high side. There are many factors involved in regional variation in utilization, including morbidity and risk factors in the population, access-to-care factors (e.g., the distribution of providers, physical barriers, and insurance issues), and general population factors (e.g., age, sex, race, education, and socioeconomic status). However, the most rational practice of medicine should presumably show as little variation as possible, declining to a floor level of variation caused by these factors.

This overview of geographic variation in radiology utilization is not reassuring. First, geographic variation is substantial. Over the last 10 years, the ratio of utilization in high and low regions has ranged from 1.47 to 1.56, with the last 5 years higher than the previous five. A relative risk of 1.5 is substantial; it means that one population has an excess of 50% in utilization over another. Generally, in epidemiology, relative risks > 1.2 draw researchers' attention. Differences in utilization between regions remain relatively stable, with no sign of lower utilization regions increasing relative to higher utilization regions.

Although a relative risk of 1.5 merits attention, it is quite different from the eight-fold difference in imaging costs between Vermont and Florida cited by Iglehart [6]. Which figure is more accurate to describe geographic variation in imaging practice? We calculated the relative risk on the basis of allowed charges and relative value units for procedures performed as well as utilization. The relative risk for relative value units, a measure of the complexity of the procedure performed, was 1.52 in 2007, almost exactly the same as the utilization relative risk of 1.54. The relative risk based on allowed charges is much higher (2.20). Regional comparisons of Medicare dollars spent on imaging are subject to a distortion based on the complex billing of radiology procedures. Both professional and technical charges are billed for radiology procedures, and the allowed technical charges (i.e., the reimbursement for the cost of equipment) dwarf the allowed professional charges (i.e., the reimbursement for doctor services). All allowed technical charges are reported in the Medicare Part B Physician Supplier Procedure Summary Master Files datasets for offices, but not for hospital outpatient or inpatient procedures. Thus, a different mix of office and hospital providers in different regions can yield very different relative risks for charges, compared with procedure counts, even when procedure counts are weighted by the complexity of the imaging performed. When the regions compared are very small (e.g., individual states), the variation between office and hospital providers, on top of differences in the underlying factors mentioned at the beginning of this discussion, can produce huge variations. It may be convenient for policy makers to pick out enormous differences to suggest that huge savings are available if one could just impose order on the chaos, but they are apt to be disappointed, because the differences, although large, are not huge. Furthermore, they can just as easily suggest underutilization on the low end as overutilization on the high end.

Bhargavan and Sunshine [24] painted a reassuring picture of regional differences and rates of change. They reported strongly negative relationships between regional utilization in 1998 and regional change from 1998 to 2001 (r = –0.951) and from 1995 to 2001 (r = –0.947), suggesting that higher utilization regions were leveling out. We certainly do not find this trend in the period 1998–2007. Relative risks are increasing yearly, there are almost no changes in relative position of regions over the period, and we find positive, though nonsignificant, correlations between either the starting or ending period and the rate of change.

Although a number of high-technology, high-cost imaging techniques have decreased in regional variation over the past 10 years, the number of categories of imaging techniques with relative risks > 2.0 have increased. We can discount some of the recently introduced, currently low-utilization categories, such as PET: newly introduced techniques will always show high variation because they are influenced by just a few early adopters in a region. The other high relative risks are composed entirely of advanced imaging categories, particularly cardiovascular advanced imaging. This finding is disturbing because cardiovascular imaging is performed predominantly in offices by cardiologists and is subject to self-referral issues [27]. Also troubling is the increase in variability of general ultrasound, a low-cost, high-technology imaging technique. Ultrasound is often described as an “operator-dependent” technique, and the increasing regional variation in utilization may suggest variation in distribution of highly trained personnel.

This study attempts to provide a high-level overview of regional variation in imaging utilization and has a number of limitations. It is based only on the fee-for-service Medicare population and may be limited in generalizability. Comparing utilization of more than 500 radiology procedures over 10 years and 10 regions forces very great analytic compression. It does not explore factors such as place of service or specialty of performing physician or the effect of the Deficit Reduction Act of 2005. Restricting the data source to the Physician Supplier Procedure Summary Master Files does not allow us to explore confounding factors, such as morbidity, demographics, or access. However, it seems very clear that there is substantial regional variation in imaging utilization, that this variation is not going away, and that it should be further explored.

Address correspondence to L. Parker ([email protected]).

This work was funded in part by the American College of Radiology.