Jay Bhattacharya, MD, PhD

 Low back pain (LBP) is a common and expensive clinical problem, resulting in tens of billions of dollars of direct medical expenditures in the United States each year. Although expensive imaging tests are commonly used, they do not improve outcomes when used in the initial management of idiopathic LBP. We estimated 1-year medical costs associated with early imaging of Medicare beneficiaries with idiopathic LBP.

 We used a 5% random sample of Medicare fee-for-service enrollees between 2006 and 2010 to determine 12-month costs following a diagnosis of idiopathic LBP. We analyzed costs of care and patient outcomes according to whether or not the patients had been referred for early imaging following their initial diagnosis.

 We employed an instrumental variables analysis using risk-adjusted physician-level propensity to order imaging for patients without LBP as an instrument for imaging use among patients with LBP. We selected this approach to adjust for confounding by indication when estimating the relative costs of early imaging of LBP compared with conservative management.

 Early imaging is strongly associated with increased costs of care in the first year following LBP diagnosis. Patients receiving an early magnetic resonance imaging scan accrued $2500 more in Medicare expenditures than conservatively managed patients, and patients who received computed tomography accrued $19,900 more.

 Medicare beneficiaries with low-risk LBP frequently receive early imaging studies. Early imaging was associated with greater long-term costs than a conservative diagnostic strategy; Medicare expenditures could be reduced by $362 million annually by managing newly diagnosed LBP in accordance with clinical guidelines.

Low back pain (LBP) is a common and expensive clinical problem, resulting in tens of billions of dollars of direct medical expenditures in the United States each year. Although expensive imaging tests are commonly used, they do not improve outcomes when used in the initial management of idiopathic LBP. Early imaging is strongly associated with increased costs of care in the first year following LBP diagnosis.

Patients receiving an early magnetic resonance imaging scan accrued $2500 more in Medicare expenditures than conservatively managed patients, and patients who received computed tomography accrued $19,900 more.

Medicare would save $362 million annually if all patients with newly diagnosed idiopathic LBP were managed in accordance with clinical guidelines.

Low back pain (LBP) is the most common type of pain reported by adults and represents the fifth most common reason for physician visits in the United States.

 One-fourth of adults in the United States report experiencing back pain lasting at least 1 day in the past 3 months,

 and more than 80% of adults will experience LBP at some point in their lives.

 The condition is costly: A 2008 study estimated that spine problems led to $85.9 billion in medical expenditures in 2005.

Although serious medical conditions, including neurological compromise, infectious and inflammatory processes, and malignancy, can cause LBP, most episodes of back pain are considered mechanical in nature and do not require specific therapy. Observational studies and randomized controlled trials have demonstrated that routine lumbar spine imaging (plain radiographic film followed by magnetic resonance imaging [MRI] or computed tomography [CT]) does not improve clinical decisions

 Further, such routine imaging leads to increased costs, unnecessary patient exposure to radiation and subsequent invasive treatments, and emotional turmoil.

 Given the robust evidence against routine imaging for LBP, medical specialty societies

 recommend against lumbar spine imaging for LBP within the first 6 weeks of diagnosis, unless patients present with “red flags” for underlying conditions such as malignancy, infection, or spinal fracture.

Thus, the standard of care is to defer imaging, relying instead on a focused history and physical exam in initial evaluation of patients with no red flags.

 Despite these clinical recommendations, however, diagnostic imaging has been employed in a steadily increasing proportion of initial evaluations of back pain among both privately insured patients

 and Medicare beneficiaries since the 1990s.

 An estimated 20% to 50% of patients undergo imaging studies for LBP within the first 6 weeks of diagnosis, often on the date of their initial physician visit. Further, a significant proportion of patients with LBP do not receive a physical exam when they first present to a physician.

Although a conservative approach to LBP has been shown to be cost-effective, it has not been adopted widely in clinical settings. According to the findings of one study, if guidelines for diagnosing LBP were fully implemented in practice, whether via cultural shifts in physician approaches to care or implementation of new incentive structures, annual cost savings in the United States could reach $300 million.

 However, that estimate is based on older prevalence and cost estimates collected through meta-analysis, rather than primary claims data; further, it did not account for regional and provider-specific variations in patterns of care, with resulting variability in costs.

We sought to estimate the cost implications to Medicare of full adoption of an initial conservative approach to LBP. To do so, we first estimated 1-year costs and the clinical sequelae associated with initial diagnostic strategy options. Because treatment assignment is not random, we used instrumental variable (IV) methods to account for confounding by indication. In this approach, a doctor’s risk-adjusted propensity to use imaging technologies for patients without back pain is treated as a quasi-randomizing variable. Finally, we estimated potential Medicare savings from switching those who receive initial imaging to conservative management.

We conducted an observational study evaluating clinical and cost outcomes at 1 year for Medicare patients with acute uncomplicated LBP who received either conservative management or imaging (MRI or CT) within 6 weeks of their initial diagnosis.

We analyzed a 5% sample of Medicare administrative claims data from 2006 to 2010 to perform this analysis. These data included patient claims, associated diagnosis and procedure codes, socio&shy;demographic information (eg, age, race, zip code of residence, and Medicaid eligibility status), and clinical characteristics (eg, diagnosis of LBP and the presence of comorbid medical conditions). Data were restricted to those of patients enrolled in fee-for-service Medicare and included data from Medicare Part A (inpatient) and Part B (outpatient).

In identifying patients with idiopathic LBP, we sought to ensure that their initial management would be uncorrelated with the etiology of their back pain. To construct the analytic sample, we first identified all patients who had at least 2 diagnoses of LBP (

International Classification of Diseases, Ninth Revision

] code 724.2) within 6 weeks in the outpatient setting between 2006 and 2010 and who did not have another diagnosis of LBP in the preceding year. We restricted our sample to those who (1) had at least 1 year of Medicare claims prior to and following the index LBP diagnosis and (2) were continuously enrolled in Medicare parts A and B from the date of diagnosis until death or the end of the study period. Because most beneficiaries become eligible for Medicare at 65 years, we limited our sample to patients 66 years and older to allow for 1 year of prior claims data.

Finally, to ensure that we were comparing patients who were similar at initial diagnosis, we restricted the final cohort to patients who did not have any red flag diagnoses in the 1 month prior to or 8 weeks following the index diagnosis and to patients surviving the first 6 weeks after the index diagnosis. Our red flag diagnoses are standard indications for early imaging in the context of LBP: spinal fracture, cancer, infection, disc herniation, spinal cord compression, aortic aneurysm, and nephrolithiasis. We also excluded patients with diagnoses in which imaging may have been contraindicated (patients with renal failure, pheochromocytoma, or hyperthyroidism; patients with implanted medical devices; patients relying on mechanical ventilation). After these exclusions, the remaining population largely consisted of patients whose LBP was idiopathic in nature.

We determined initial diagnostic strategies based on procedure codes within the Medicare claims data submitted within 6 weeks of the index diagnosis. We categorized patients as either receiving no lumbar spine imaging or receiving lumbar spine imaging (stratified as MRI or CT). Standard Current Procedural Terminology codes were used to define each lumbar spine imaging strategy.

When used to assess the relation of overall costs of care to early imaging of LBP, standard multivariable statistical methods can be compromised by confounding by indication: Physicians are more likely to order imaging studies early for patients with LBP who are in worse initial health.

To address this problem, we used risk-adjusted physician-level propensity to use imaging on patients without LBP over the previous 12 months (hereafter, physician propensity) as an IV. An IV adjusts for unobserved confounders by acting as a quasi-randomizer; it stratified our population using a variable that is unrelated to the unobserved confounder (in this case, severity of patient illness) but is associated with likelihood of receiving the treatment being studied (in this case, referral for early imaging). Here, we stratify by physician propensity to refer for imaging to remove the potential unobserved confounder of severity of the patient’s illness.

The IV method has been used in the medical literature to study a wide variety of topics for which unobserved confounders may introduce bias, including treatment of acute myocardial infarctions,

 management of bladder and pancreatic cancer,

 and interpretation of neuroimaging data.

 To the best of our knowledge, the IV approach has not yet been used in the medical literature to study management of LBP.

To adjust appropriately for confounding by indication, the IV method requires that physician propensity be (1) strongly correlated with the use of imaging in initial management of patients with LBP and (2) uncorrelated with the unobserved determinants of clinical management outcomes and costs among patients with LBP after the initial treatment period. Our statistical work strongly supported the first requirement, whereas the second requirement is true by construction: Physician propensity is measured for patients without LBP, so it cannot be correlated with the unobserved health of those with LBP. Regression analysis illustrating the strong correlation between physician propensity and the use of imaging in initial management of patients with LBP is available in the 

In addition to the IV approach, we adjusted for patient age, race, sex, Medicaid eligibility (as a proxy for socioeconomic status), inpatient visits during the previous 12 months (binary variables for 1 or 2 or more visits), skilled nursing facility (SNF) or hospice stays during the previous 12 months (binary variables for 0 or 1 or more visits in either setting), Elixhauser comorbid medical conditions (using 

 diagnosis codes during the year prior to index diagnosis),

 and geography (using a complete set of indicator variables for each hospital referral region).

To calculate cost outcomes, we identified all Medicare parts A and B claims and summed these to total Medicare costs by year from date of diagnosis for each patient. We inflated costs to 2010 US$ using the gross domestic product deflator. We accounted for the possibility of nonnormal distribution of costs by log-transforming costs and employing a standard statistical method that eliminates bias that might arise from the presence of patients with LBP with zero costs after the initial management phase.

To better illustrate the implications of our statistical analyses, we simulated the effects of switching all Medicare patients with LBP to a strategy of no imaging during the initial management phase (the first 6 weeks after diagnosis). The parameters and structure of our simulation matched the statistical models we used to analyze costs in the year after the initial visit. Using the simulation, we first calculated 1-year expenditures for patients with a particular set of demographic and health characteristics (as per the IV models described previously) if they had received no imaging instead of MRI or CT. We then compared this estimate with the original calculated cost for each diagnostic group and extrapolated the difference across the Medicare population.

Using a 5% sample of all Medicare beneficiaries from 2006 to 2010, we identified 744,262 patients with any diagnosis of LBP in the outpatient setting, of whom 250,771 (34%) qualified for the final study sample. A cohort diagram of the construction of the analytic sample, including all exclusion criteria, is available in the 

. Characteristics of the patients in the sample are presented by initial diagnostic strategy (conservative management, MRI, or CT) in 

. MRI was the most common form of early imaging, accounting for 21% of the study cohort. Only 3% of the cohort received CT, and 76% of patients received neither form of early imaging.

In general, patients undergoing MRI and patients receiving conservative management had similar characteristics in terms of age, race, socioeconomic status (as determined by Medicaid eligibility), history of recent hospitalization, recent medical costs, and observed comorbidities (Table 1). Thus, the observable characteristics of patients receiving MRI in our sample were similar to those of patients receiving conservative management, even before adjustment for relevant comorbidities and quasi-randomization with our IV. This implies that observed differences in medical costs for patients receiving MRI and patients receiving no imaging in our sample can reasonably be attributed to differences in the management of their LBP, rather than differences in the etiology of that LBP. However, patients in our study cohort who received CT imaging were slightly older at initial diagnosis and were more likely to be male, to be eligible for Medicaid, and to have had a hospitalization or received SNF or hospice care in the year prior to diagnosis. Patients undergoing CT also accrued higher medical costs in the year prior to their diagnosis and were more likely to have certain comorbidities, including diabetes, heart failure, and chronic obstructive pulmonary disease (Table 1). Patients receiving CT may have had contraindications to MRI that were associated with higher medical costs beyond those captured in our models.

Our motivation for using an IV, physician propensity to refer for imaging, as a quasi-randomization tool in this analysis was that the health of patients who receive early imaging may differ from that of patients who do not receive imaging. This difference may be reflected in their observed health characteristics (Table 1) and addressed in part by including patient characteristics in a multivariable regression. However, such an approach does not account for unobserved health characteristics, which we address here by implementing IVs.

Another indication of unobserved differences in health between our patient groups would be differences in survival after their LBP diagnosis. We used Cox proportional hazard models to estimate the survival of patients who received MRI, CT, or no early imaging after their initial LBP diagnosis. We found no differences in survival across these groups, both in an unadjusted analysis (

) and in a survival analysis that adjusted for patients’ pre-existing conditions and for geographic variation in service use (

Although early use of imaging in LBP does not appear to be associated with improved clinical outcomes, our IV analyses of patient costs indicate that patients who receive early imaging have substantially higher average medical costs in their first year after diagnosis. Cumulative estimated Medicare expenditures in the 1 year post diagnosis are reported for each diagnostic strategy in 

 (see eAppendix for full regression output). The conservative management strategy was associated with the lowest annual Medicare expenditures, whereas those receiving an MRI accrued $2512 more and those receiving CT accrued $19,899 more in total Medicare expenditures in the year following diagnosis.

Table 2 also shows potential Medicare health expenditure savings from increased use of conservative diagnostic strategies for patients diagnosed with LBP. Shifting patients who receive CT to no imaging would save $203 million annually ($19,899 saved per patient per annual diagnosed cohort), and shifting those who receive MRI to no imaging would save $159 million annually ($2512 saved per patient per annual diagnosed cohort).

Our analysis confirms that imaging studies are overused in the initial management of low-risk patients with LBP

 and suggests that substantial Medicare savings could be realized from guideline-compliant care in this setting. We found that 24% of Medicare patients with uncomplicated acute LBP diagnosed from 2006 to 2010 received advanced imaging within 6 weeks of diagnosis. The proportion of patients receiving advanced imaging (MRI or CT) remained constant throughout this period, although clinical guidelines published in 2007 recommended an initial conservative diagnostic strategy (ie, no imaging within the first 6 weeks following diagnosis).

 Diagnostic strategies that included advanced imaging were associated with greater long-term costs than a conservative diagnostic strategy and were not associated with improved outcomes; patients receiving CT accrued the highest costs in the year following diagnosis. We find that $362 million could be saved annually within the Medicare context by treating patients with newly diagnosed LBP with a conservative initial diagnostic strategy, per clinical guidelines.

A fundamental premise of our IV approach is that physician practice patterns vary. One physician may adopt a practice style that favors less intervention while another may adopt a more aggressive attitude toward patient management. Early imaging may be a component of a broader management approach that some physicians take with patients with LBP. Furthermore, physician practice patterns change surprisingly little in response to new evidence and clinical guidelines.

 A key lesson that accompanied the Dartmouth Atlas’s observation of marked variation in physician practice was that these variations are driven more by a physician’s own impressions and experience than by the publication of new guidelines and research.

Among the limitations of this study is the absence of data on Medicare Part D costs, which prevented us from assessing the use of opioids and other medications to control LBP. Although our IV directly addresses the problem of confounding by indication, there may be residual confounding that the instrument has failed to account for. For example, we did not measure practice type or the physician’s financial incentives to order advanced imaging, both of which have been shown to influence the choice of LBP diagnostic techniques.

 We minimized such biases by using the IV method and by adjusting for a variety of patient demographic and health characteristics. We further limited the influence of high-cost outliers by using log-transformed 2-stage least squares regressions. Additionally, as with all analyses of a random 5% Medicare sample, our conclusions are drawn from a subset of the Medicare population rather than the entire population.

Finally, our research only addresses potential savings from adoption of a conservative diagnostic strategy for LBP. Clinical societies have been recommending conservative strategies for many years, with seemingly little success. Chou et al have hypothesized that financial incentives, including incentives linked to patient satisfaction and self-referral, along with defensive medicine considerations, promote overuse of advanced imaging techniques in this setting.

 Other studies have shown that patients at baseline may feel more satisfied if they receive advanced imaging for LBP; however, those who instead receive a 5-minute educational intervention on the risks associated with lumbar spine imaging and its minimal clinical usefulness feel equally satisfied with their care.

Insofar as payment moves from a per service basis to models closer to capitation, the culture of clinical practice will change: Physicians may be incentivized to adhere to more cost-effective and conservative diagnostic strategies through payment structures incorporating quality measurement, such as LBP-specific measures sponsored by the Agency for Healthcare Research and Quality National Quality Measures Clearinghouse.

 These shifts in physician preferences must come hand-in-hand with shared decision making, in which patients receive trustworthy information on the clinical usefulness of early imaging for LBP.

 Results from this study demonstrate that such shifts in clinical practice toward adoption of conservative diagnostic strategies for LBP, as supported by comparative effectiveness research, could lead to large health expenditure savings.&ensp;

This research was supported by the National Institutes of Health (NIH)/National Institute of Aging under grant R37 AG036791. The NIH had no role in the study design, conduct, or reporting of this study and the authors take full responsibility for the content of this research. The authors thank Tom MaCurdy and participants at the Stanford Medical School Research in Progress seminar for helpful feedback. This study was exempted from the Institutional Review Board of Stanford University School of Medicine.

Harvard University (AMG, MF), Boston, MA; National Bureau of Economic Research, Inc (AMG, MF, JB), Boston, MA; Stanford University School of Medicine (TDA, DV), Stanford, CA; Department of Anesthesia, University of California San Francisco (AD), San Francisco, CA; Acumen, LLC (ES), Burlingame, CA; Stanford Centers for Health Policy/Primary Care and Outcomes Research (JB), Stanford, CA.

 This research was supported by the National Institutes of Health/National Institute of Aging under grant R37 AG036791.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

 Concept and design (AMG, TDA, AD, MF, ES, DV, JB); acquisition of data (AMG, JB); analysis and interpretation of data (AMG, DV, JB); drafting of the manuscript (AMG, TDA, AD, MF, ES, DV, JB); critical revision of the manuscript for important intellectual content (AMG, TDA, AD, MF, ES, DV, JB); and statistical analysis (AMG, DV, JB).

Articles

This instrumental variables analysis estimates that Medicare would realize $362 million in annual savings if all patients with newly diagnosed low back pain were managed conservatively.

Medicare Savings From Conservative Management of Low Back Pain

 To determine whether electronic health record (EHR) access influences the number of laboratory and imaging tests ordered, which is a frequently cited mechanism for EHR-enabled cost savings.

 We analyzed data on non—federally employed office-based physicians from the 2008 to 2012 Electronic Health Medical Records Survey, a supplement to the National Ambulatory Medical Care Survey.

 We estimated logistic regressions to determine the relationship between EHR utilization and the volume of laboratory and imaging tests ordered in our study population, controlling for age, sex, race, clinic type, payer type, health status, comorbidities, and new patients.

 Physicians who actively used an EHR system ordered more complete blood count (CBC) tests than physicians who did not (odds ratio [OR], 1.34; 

 <.001), even after adjusting for patient demographics, health status, and case mix. EHR-using physicians also ordered more computerized tomography scans (OR, 1.41; 

 <.001); the difference for magnetic resonance imaging scans was not significant (OR, 1.08; 

 = .449). Subgroup analysis highlighted differences in ordering among various patient cohorts.

 Using the most recent available nationally representative data, excluding federal and Veterans Affairs' hospitals, we found that physicians with EHR access ordered more tests than their non-EHR counterparts, thus contradicting a common rationale for EHR implementation. We argue that EHR use may actually increase healthcare expenditures by facilitating the ease of ordering tests. Whether these extra tests carry clinical utility requires further analysis.

	Using the most recently available national data, physicians with electronic health record (EHR) access ordered more tests than their non-EHR counterparts, thus contradicting a common rationale for EHR implementation. 

		Against a backdrop of policies suggesting cost savings for EHR use, these results call for a reassessment of the unqualified expectation that EHRs will reduce medical expenditures and increase clinical efficiency. 

		Adopting EHRs is not enough; providers must also foster the organizational and delivery processes required to realize systemwide efficiencies. 

		Implementing EHR systems may become cost-effective only when complemented by models of care that emphasize quality, value, and efficiency.

Over the last few years, the federal government has enacted seminal legislation to promote the adoption and use of health information technology (IT) among physicians. Most notably, in 2009, Congress passed the Health Information Technology for Economic and Clinical Health (HITECH) Act, which established incentive payments both for individual providers and hospital systems that demonstrate “Meaningful Use” of electronic health record (EHR) systems through 2015.

 An estimated $30 billion has been dedicated to this effort, with the goal of improving healthcare quality, cost-effectiveness, and efficiency.

 Much of this investment was predicated on an influential study by the RAND Corporation, a nonprofit global policy think tank. In 2005, RAND predicted that widespread EHR adoption would save $81 billion annually.

 A 2012 reassessment, however, altered this estimate: rather than decreasing, healthcare costs have climbed by over $800 billion since the first report was issued.

 Although EHRs are still thought to yield cost and quality benefits, recent studies have shown this effect to be inconsistent at best.

The global market for EHRs has grown substantially since HITECH’s passage. In the United States alone, it is expected to expand further from $9.6 to $12.2 billion by 2021.

 Recently, however, physicians adopting EHRs have voiced their concerns about their limitations and cost-effectiveness.

 Given the role of health IT in health reform, it has become increasingly urgent that we determine the financial and clinical impact of the EHR rollout: do EHRs save money while improving quality and efficiency?

In this study, we examined whether the introduction of EHRs in ambulatory medical practices has reduced the ordering of diagnostic and imaging tests. Although some agree that EHRs make data more accessible and reduce the likelihood of duplicate orders, other studies show that the accessibility of EHRs makes tests easier to order thus increasing the number.

study analyzed data from the 2008 National Ambulatory Medical Care Survey (NAMCS)—before the passage of the HITECH Act&mdash;and found an association between electronic access to results and increased ordering of imaging tests.

 Nevertheless, because EHR systems have evolved since 2008, an updated evaluation of their effect on test ordering is warranted. We extended this previous analysis to include years 2008 through 2012, the most recent publicly available NAMCS data. We also conducted an extensive subgroup analysis as the impact of EHRs can vary by patient risk and demographics. 

We analyzed data from the 2008 to 2012 Electronic Health Medical Records Survey, a mail survey supplement to the NAMCS. We included the years 2008 to 2012 to capture patient records both before and after the HITECH Act was implemented in 2009. For 2008 and 2009, the EHR supplement used the same sampling scheme as the NAMCS. In 2010, the NAMCS expanded the sampling scheme to better represent national EHR usage, and in 2012, it implemented 2 sampling designs: one for national, regional, and division-level analysis, and a second to enable state-based analyses for the 34 most populous states in the country. State-based analyses were not incorporated into our study because of their absence from the data prior to 2012. In aggregate, the NAMCS presents data on patient demographics, EHR usage, and laboratory tests for a random sample of office-based physicians with direct patient interaction, excluding pathologists, anesthesiologists, and radiologists.

The basic sampling unit of the NAMCS is 1 physician—patient interaction with a non–federally employed physician. From 2008 to 2011, a multi-stage sampling design was used. First, the primary sampling unit (PSU) comprised counties and county equivalents; second, physicians were selected within PSUs according to their specialty; and third, specific patient encounters were selected within physician offices (eAppendix Figure A [

]). In 2012, this design was simplified to 2 stages: the first stage stratified physicians based on primary care status and the second stage selected individual patient encounters (eAppendix Figure B).

For the final stage of both sampling designs, physicians were each assigned a calendar week, and surveys were completed for a systematic random sample of patients that were seen during the assigned week.

 The CDC compiled survey data into the NAMCS database and, in 2012, they incorporated physicians practicing at community health clinics (CHCs) in a separate survey. Therefore, we excluded all CHCs from our study population.

We generated national estimates from the NAMCS, accounting for the complex sampling design. Demographic data were compiled for the total study population. Typical variables included, but were not limited to, age, sex, payer type, and clinic type. Median income was analyzed for individuals within the 2008 to 2011 dataset, but was then no longer collected by NAMCS in 2012. Clinic type was defined as private practice or nonprivate practice in the NAMCS.

We analyzed several individual test-ordering practices for our primary outcomes, including complete blood count (CBC) and radiographic studies, because they are common tests, expensive in aggregate, and readily available in EHR systems. For each test, we compared the probability of ordering by physicians using EHRs with that of physicians not using EHRs. We used a Bonferroni correction to adjust for multiple comparisons.  

We estimated a patient-level multivariate logistic regression to determine the relationship between the ability to order laboratory tests electronically and the probability of ordering each test. We then computed an adjusted estimate for EHR and non-EHR physicians in which confounders were held at mean values. As covariates, all multivariate logistic regressions included, as available: age, sex, race, clinic type, payer type, health status, major 

International Classification of Diseases, Ninth Revision, Clinical Modification 

) diagnosis categories, and whether the physician had seen the patient before. The 

 categories included were: infectious/parasitic, neoplasms, endocrine/metabolic/immune disease, blood disease, mental disorders, nervous disease, sense organs disease, circulatory disease, respiratory disease, genitourinary disease, pregnancy/childbirth complications, skin disease, musculoskeletal disease, congenital anomalies, perinatal conditions, ill-defined conditions, injury/poisoning, and external injury.

To assess the overall effect of EHRs on imaging orders, 1 variable combined all major imaging modalities, such as, but not limited to, x-ray, computed tomography (CT), magnetic resonance imaging (MRI), electrocardiogram, ultrasound, and bone mineral density. We also analyzed 3 modalities individually: x-ray, CT, and MRI. For imaging, EHR usage was defined as the presence and usage of an electronic system for viewing imaging results, encompassing both the imaging study and its interpretation. Such computerized systems might be either part or independent of a full EHR, which is assumed to contain patient medical histories, previous laboratory and imaging results, and point-of-decision support. For those occasional patients seen at facilities where the EHR system was “turned off” or “not used routinely,” we categorized observations as occurring at facilities without an EHR system.

The study population (n = 183,519), as described in 

, includes a higher proportion of female (58.4%) and white patients (84.5%). The modal patient was between 45 and 64 years old (29.6%). Approximately half of the study population had private insurance (53.4%), one-third had a median family income above $52,387 (32.5%), and most were seen in private practice office settings (92.1%).

 illustrates the health status of our study population, with just over a quarter of patients diagnosed with hypertension (27.4%). Other notable chronic health conditions included hyperlipidemia (16.5%), arthritis (13.4%), diabetes (11.6%), and depression (9.2%). The most common primary diagnoses were “external causes of injury” (19.9%), “diseases of the respiratory system” (9.8%), and “musculoskeletal disease” (9%).

For physician practices with and without EHRs, 

 shows the probabilities of CBC test ordering. Superscripted notes indicate that the difference between these groups was significant at the 95% confidence interval (CI) after the Bonferroni correction. In subgroup analysis, we found that availability of EHRs was associated with higher rates of CBC testing regardless of sex, race, median household income, or new patient status, and for all individuals over 25 years of age. This correlation was additionally seen in those receiving either Medicare or private insurance and in private or freestanding clinic settings. Furthermore, several comorbid conditions and primary 

 diagnoses—which are further detailed in Table 3&mdash;were associated with higher rates of CBC testing. The availability of an EHR did not statistically significantly reduce the probability of ordering a CBC test in any subgroup.

At the 95% CI using the Bonferroni correction, the probability that a provider would order imaging studies was significantly higher with an EHR system, again, regardless of sex, race, median household income, new patient status, or even age, as shown in 

. It was higher for those seen in private, health maintenance organization (HMO), and faculty practice plan settings, as well as for all insurance types except worker’s compensation and “other,” for which the data were not statistically significant. Several comorbid conditions and primary 

  diagnoses (further detailed in Table 4) were associated with higher rates of CBC testing. In no single subgroup did the availability of EHRs statistically significantly reduce the probability of ordering imaging studies. 

Multivariate logistic regressions demonstrated that the probability of ordering a CBC test is significantly greater for physicians with EHR access (odds ratio [OR], 1.34; 

 <.001)—a 30% increase in the adjusted likelihood of ordering a CBC test compared with physicians not using an EHR after adjusting for patient demographic information and a detailed set of clinical risk-adjusters (

). Although the difference for MRIs is not significant, physicians using EHRs have a greater probability of ordering CT scans (OR, 1.41; 

 <.001), at a 41% and 37% increase, respectively. Considering imaging in aggregate, physicians using EHRs have a greater probability (OR, 1.26) and a 23% increased adjusted likelihood of ordering imaging than physicians not using EHRs (

<.001). A multivariate logistic subgroup analysis shows statistically significant differences among various subgroups (eAppendix Table).

The availability of an EHR system is associated with a measurable increase in the ordering of CBC and imaging tests in outpatient settings, even after adjusting for an extensive set of demographic and case-mix variables. Physicians with EHR access exhibited a higher probability of ordering an imaging study (OR, 1.26; 

 <.001). This difference is particularly pronounced for Medicare patients and patients with private insurance. These findings contradict one of the most common arguments for EHR implementation: that EHRs reduce excessive testing and, subsequently, unnecessary costs.

 Although our results do not differentiate between clinically indicated and redundant tests, rates of both expensive and inexpensive tests are higher in practices with EHRs.

EHR systems have been federally subsidized since 2009, when the HITECH Act earmarked billions of dollars in reimbursement to early adopters. Proponents argued that EHR use would improve care coordination, increase efficiency, expose duplicate testing, and, thereby, reduce costs. Preliminary evidence upheld this potential; however, these studies typically examined health technology systems developed in-house in highly controlled single-clinic or emergency department environments.

Evidence on the quality and cost-effectiveness of EHRs beyond these benchmark hospitals has been mixed.

 One study found hospitals across the country with advanced EHRs had a 9.66% lower cost per admission than those without advanced EHRs.

 By contrast, another study found that inpatient cases cost 7% more in hospitals with advanced EHRs than in those without.

 A third, analyzing Medicare claims data from 1998 to 2005, found an initial 1.3% increase in billed charges with no evidence of cost savings—even 5 years after adoption.

 The initial promises of EHRs to “cut waste…reduce the need to repeat expensive medical tests” and “save billions of dollars,” have yet to be achieved.

We propose 2 possible interpretations of the observed correlation between EHR access and test ordering: one in which computerized access simplifies the ordering process leading to more frequent ordering, and another in which the same physicians who readily adopt EHRs also order more tests for their patients. Our results support the former interpretation. First, we found striking increases in both test and imaging orders for EHR-equipped physicians across nearly every subgroup; no variable, from patient demographics to insurance type to comorbidities, eliminated this effect. Second, from 2008 to 2011—when this data was available&mdash;we found that the largest effect of EHRs on test ordering was in large practice settings, such as HMOs, in which individual physicians are least likely to influence institutional IT decisions. In those settings, the argument that doctors who are most likely to adopt EHRs are the same doctors who are most likely to order excessive tests bears less relevance. If the selection bias interpretation was correct, we would have expected a larger difference in test ordering between EHR and non-EHR doctors in small practice settings.

 Because we observed the opposite, selection bias is a less likely interpretation of our results.

Our study has a few limitations. Although we used CBC ordering as our single measure for all laboratory testing, other laboratory tests may be affected differently by EHRs. Nonetheless, as CBC is typically among the first-ordered laboratory tests in many clinical situations, it arguably reflects the overall trend for laboratory tests.

 Our measure for EHR implementation represents, at minimum, the capacity to order and view patient diagnostic information, not necessarily advanced clinical decision support—helpful for filtering vast quantities of patient information&mdash;because such distinctions were not available from our data source. Still, our results are suggestive of the broader impact of EHRs as it is reasonable to conclude that ordering of laboratory tests and imaging are basic functions of all EHR systems.

Another limitation stems from our lack of data beyond 2012. In 2011, the federal government implemented Meaningful Use regulations, which tied federal incentive payments to specific care delivery improvements enabled by EHRs.

 Because we reviewed years 2008 through 2012 only, we cannot be certain whether additional functionalities developed in the last 3 years might have reduced the quantity of laboratory and imaging tests ordered. Still, cost data for evolved functionalities like clinical decision support, one of the most publicized of Meaningful Use, remain conflicted to modest at best.

 Moreover, at a time when less than one-third of office-based providers are meeting Stage 2 Meaningful Use requirements, perhaps it is the EHR programs studied—however rudimentary&mdash;that most accurately reflect the current usage and usability of EHRs nationwide.

 It remains to future studies to evaluate EHR systems as they continue to evolve.

Finally we were limited by the constraints of our data source. Given that our basic sampling unit was a single patient encounter and not the patient, long-term outcome variables, such as mortality and complications, could not be included. Moreover, our study does not cleanly distinguish between clinically necessary and unnecessary tests. We can infer clinical utility for some subsets of patients: those with a primary diagnosis of cancer, for example, for whom imaging was 47% (

 <.001) more likely to be ordered if EHRs were available. Nonetheless, from our analysis, it also appears that EHRs may simply promote excessive testing more generally. It is interesting, for instance, that this effect holds true—across both imaging and laboratory testing&mdash;even for patients seen primarily for depression and mental disorders, diagnoses typically not associated with CBC or imaging requirements. Furthermore, those diagnoses that would almost necessitate CBC testing&mdash;specifically, infection and blood diseases&mdash;saw no significant difference in ordering frequency between EHR and non-EHR practices. This suggests that physicians will order critical diagnostic tests and imaging regardless of EHR status.

Our results demonstrate a positive relationship between EHR implementation and the volume of laboratory and imaging tests that physicians order. Against a backdrop of policies suggesting cost savings for EHRs, these results call for reassessment of the hope that EHRs can reduce medical expenditures and increase clinical efficiency. Adopting EHRs is not enough: providers must also foster the organizational and delivery processes required to realize systemwide efficiencies. Implementing EHR systems may become cost-effective only when complemented by models of care that emphasize quality, value, and efficiency.

Ibrahim Hakim, BBA; Sejal Hathi, BS; Archana Nair, MS; and Trishna Narula, MPH, contributed equally and are all joint first authors on the paper. The authors alone are responsible for the statements in the paper and for any errors; all had equal access to the data. 

 School of Medicine (IH, SH, AN, TN), and Institute for Economic Policy Research (JB), and Graduate School of Business (SH), Stanford University, Stanford, CA; National Bureau of Economic Research (JB)

 Dr Bhattacharya acknowledges funding from the National Institute on Aging (NIA) for his work on this project (R37 AG036791 and P30 AG17253).

 Concept and design (IH, SH, AN, TN, JB); acquisition of data (SH, AN, TN); analysis and interpretation of data (IH, SH, AN, TN, JB); drafting of the manuscript (IH, SH, AN, TN, JB); critical revision of the manuscript for important intellectual content (IH, SH, AN, JB); statistical analysis (IH, SH, AN, TN, JB); provision of patients or study materials (TN); obtaining funding (JB); administrative, technical, or logistic support (SH); and supervision (JB).

 Sejal Hathi, BS, Stanford School of Medicine, Stanford Graduate School of Business, 291 Campus Dr, Stanford, CA 94305. E-mail: shathi@stanford.edu. 

1. Ferris N. Health IT gold rush under way. 

. 2010;29(4):583-584. doi: 10.1377/hlthaff.2010.0217.

2. EHR incentive programs. CMS website. https://www.cms.gov/Regulations-and-Guidance/Legislation/EHRIncentivePrograms. Updated April 22, 2015. Accessed May 5, 2015.

3. Walker J, Pan E, Johnston D, Adler-Milstein J, Bates DW, Middleton B. The value of health care information exchange and interoperability. 

. 2005;24(1):W5-10-18. doi: 10.1377/hlthaff.w5.10.

Using the most recently available national data, physicians with electronic health record (EHR) access ordered more tests than their non-EHR counterparts, thus contradicting a common rationale for EHR implementation.

Electronic Health Records and the Frequency of Diagnostic Test Orders

 Unwarranted geographic variation in spending has received intense scrutiny in the United States. However, few studies have compared variation in spending and surgical outcomes between the United States healthcare system and those of other nations. In this study, we compare the geographic variation in postsurgical outcomes and cost between the United States and Japan.

 This retrospective cohort study uses Medicare Part A data from the United States (2010-2011) and similar inpatient data from Japan (2012). Patients 65 years or older undergoing 1 of 5 surgeries (coronary artery bypass graft, abdominal aortic aneurysm repair, colectomy, pancreatectomy, or gastrectomy) were selected in the United States and Japan.

 Reliability- and case-mix—adjusted coefficient of variation (COV) values were calculated using hierarchical modeling and empirical Bayes techniques for the following 5 outcomes: postoperative mortality, the development of a complication, death after complication (failure to rescue), length of stay, and the cost of the hospitalization. Sensitivity analyses were also performed by calculating patient demographic-and case-mix–adjusted COV values for each outcome using weighted age- and sex-standardized values.

 The variability of the postsurgical outcomes was uniformly lower in the United States compared with Japan. Cost variation was consistently higher in the United States for all surgeries.

 Although the US healthcare system may be more inefficient regarding costs, the presence of higher geographic variation in postoperative care in Japan, relative to the United States, suggests that the observed geographic variation in the United States—both for health expenditures and outcomes&mdash;is not a unique manifestation of its structural shortcomings.

This investigation compares the amount of observed geographic variation in the United States and Japan for the surgical outcomes and costs of 5 common, higher-risk surgeries. Key findings include:

		The United States has less variation in surgical outcomes compared with Japan, but it has higher geographic cost variation.

		The surgical outcome variation in the United States is not unique; it exists even in a country that consistently ranks highly on public health quality measures. This variation is likely the result of unique structural inefficiencies in each healthcare system.

		There is an opportunity to reduce wasteful spending by standardizing the cost of care in the United States.

Countries around the world are attempting to contain increasing healthcare costs without sacrificing quality of care. Relative to other developed countries, the United States fares poorly on measures of both healthcare expenditure and health outcomes. Although the United States spent $7929 per person (16.4% of gross domestic product [GDP]) in 2010—the most of any Organisation for Economic Co-operation and Development (OECD) country&mdash;health outcomes, such as life expectancy, are among the worst in the developed world.

Much research has focused on small-area geographic variation in healthcare spending as evidence of inefficient healthcare in the United States.

 A recent report released by the Institute of Medicine (IOM) has affirmed that higher healthcare expenditures within a region are not related to better outcomes.

 In principle, the lack of relationship between spending and outcomes suggests that any unwarranted variation is potentially wasteful, and, if eliminated, might decrease costs while maintaining quality. Variations in healthcare expenditure have been identified as a potential area for up to a 30% reduction in waste (approximately $650 billion) in the US healthcare system.

Apart from issues of economic inefficiency, geographic variations in treatments and outcomes have important clinical and health policy implications. Outcome differences across regions are evidence for the presence of inequalities in care across regions, which represent a distinct and important problem in itself. Regional differences in quality of care may indicate opportunities to improve care for patients in places with worse outcomes. Identifying centers that have better outcomes than others and promoting the practice styles of those centers may be a practical way to improve patient outcomes.

Researchers have proposed several reasons for the observed geographic variation in the United States. Studies cite fragmentation of care,

 racial and socioeconomic barriers to care, and the diversity of financing and coverage options as reasons that the United States has unwarranted variation.

 Other studies have also investigated physician peer effects and the ways that social influences can affect medical decisions.

One premise underlying the interpretation of these findings is that the American healthcare system is unique in its inefficient stewardship of healthcare dollars. This naturally leads to a question: If the United States is so singularly ineffective at controlling costs and ensuring quality relative to other countries, then does it also experience singularly high geographic variation in costs and outcomes relative to other countries?

In this study, we compared the variability in postoperative outcomes and cost of hospitalization for 5 major surgeries in the United States and Japan. There are several key similarities and differences between the 2 countries. Japan is a modern OECD country that has access to technologies and treatments similar to those used in the United States; however, in 2010, Japan spent only $3204 per person (9.5% of GDP) on healthcare while attaining the highest average life expectancy at birth of any OECD country. Its system also provides universal access to care, and it is consistently well-graded from a public health standpoint.

 In other words, it does not have many of the shortcomings of the US healthcare system that potentially lead to unwarranted variation.

Although the prevalence of various surgical procedures has been studied, to our knowledge, there have been only few and limited comparisons of the variability in surgical outcomes among countries.

 Surgical outcomes as a quality metric are ideal to examine, given their well-established definitions and close association between the surgical process and the postoperative outcome. If the observed variability in Japan is low, the implication would be that the United States is truly unique. If, however, the variability in Japan is high, that would suggest that variability could be an inherent feature of healthcare systems. Our main aim is to study whether the United States is truly uniquely inefficient or whether geographic variation is part of the natural fabric of modern healthcare delivery.

Our analysis uses 2 data sources: a large Japanese administrative database and the Medicare 5% sample standard analytical file. The Japanese administrative database contains anonymous patient and encounter-level information collected over the 9-month period between April 1, 2012, and December 31, 2012. The data were originally collected from hospital forms that were submitted to the Japanese government as part of an ongoing effort to reform the healthcare reimbursement system in Japan.

 Hospitals that participate in this program are designated as Diagnosis Procedure Combination (DPC) hospitals, and together, they account for more than half of all the acute care beds in Japan. The DPC hospitals represent approximately 20% of all hospitals in Japan; they include all major academic medical centers and most of the nation’s larger hospitals. Our sample encompasses 100% of the patients from roughly half of these DPC hospitals. Please see eAppendix Figure A (

) for a comparison of our sample with the full set of DPC hospitals. The data contain information on patient status at admission and discharge; cost of admission; and details on surgeries, procedures, and drugs administered during the patient’s stay at the hospital. Medical clerks and licensed information managers generally oversee the collection and reporting of DPC data at individual hospitals to optimize the accuracy of the data. Studies correlating data from electronic health records and DPC data have shown very high compliance.

 In total, our data encompass information for approximately 5 million patients in more than 600 hospitals located across every region of Japan.

Data on patients in the United States were obtained from the Medicare 5% sample from 2010 to 2011. This is a random sampling of 5% of Medicare beneficiaries across all hospitals in the United States and has been used extensively to analyze different outcome measures.

 In our study, we analyze all inpatient claims by hospitals for this 5% sample. The information about each hospitalization includes dates of hospital admission and discharge, discharge status, disease(s) diagnosed, and cost of hospitalization.

Patients undergoing any of 5 relatively common, higher-risk surgeries (abdominal aortic aneurysm repair, coronary artery bypass grafting, colectomy, gastrectomy, or pancreatectomy) were included in the analysis. We follow the definitions outlined in the US Agency for Healthcare Research Quality’s Inpatient Quality Indicator guidelines when available.

 These surgeries have been previously used in related investigations studying the variability of surgical quality and outcomes.

 Japanese patients were selected using a Japan-specific surgery classification scheme (K-codes). K-codes were individually selected and verified to correspond to their 

International Classification of Diseases, Ninth Revision, Clinical Modification

) counterparts for this study; they have similarly been used in other studies to create comparable cohorts.

 Comorbidities were defined using Elixhauser classification definitions for both the US and Japanese cohorts. Prior validation studies have shown that both 

 coding algorithms produce similar estimates of comorbidity prevalence in administrative data.

 Patients were restricted to those 65 years or older.

We investigated 5 outcomes: in-hospital mortality, postsurgical complications, failure to rescue, length of stay, and cost of hospitalization. Postsurgical complications are defined using 

 codes in the US Medicare dataset, and corresponding 

 codes for the Japanese dataset. Complications included pulmonary failure, pneumonia, myocardial infarction, deep vein thrombosis/pulmonary embolism, acute renal failure, postoperative hemorrhage, surgical site infection, and gastrointestinal bleeding. The specific 

 codes for these complications have been previously validated as having reasonable agreement when comparing administrative data with the medical record.

 Failure to rescue is defined as death after any of the previously listed complications. Variations in cost in the United States were calculated using the Medicare Part A reimbursement to hospitals.

We subtracted disproportionate share payments, indirect medical education payments, and payments classified under total prospective payment system capital from United States cost totals, because these represent intended systemic cost variation and would inflate our variation estimates for the United States. Regions of comparison were defined for each country so that they were similar; variability was compared among states in the United States and among prefectures in Japan, graphically depicted in eAppendix Figure B.

In this study, we calculated outcome-specific coefficients of variation (COVs), defined as the sample standard deviation divided by the mean. The COV is a unitless measure that allows for a comparison of variation between a random variable with different means from 2 samples. For example, when used to measure the consistency of an elite athlete’s performance across multiple races, the COV will usually be between 1% and 2%.

Potential confounders, such as patient characteristics (age and gender for both countries’ cohorts, and additionally race for the US cohort), comorbidities, and urgent versus scheduled admissions, were included in all regression models. We additionally used hierarchical modeling techniques to adjust our COV estimates for reliability.

 Using random effects logistic regression models, we generated empirical Bayes predictions of the outcomes that were used to calculate our final COV values. This technique minimizes the impact of statistical outliers such as those from small-volume hospitals whose estimates can be considered less reliable than those of larger-volume hospitals. By variably shrinking the contribution of small-volume hospital estimates toward the average mortality rate across all hospitals, we are able to better estimate systematic, nonrandom variation as opposed to variation due to chance.

We additionally performed a sensitivity analysis by calculating age- and sex-standardized COV values for both countries to the 2000 US population, weighting all estimates by the number of procedures performed in each state and prefecture. These demographic- and case-mix—adjusted COVs were evaluated by performing multivariate regressions for each outcome of interest. We then calculated the adjusted COV (adjusted for age, gender, race, comorbidities, and urgency of admission) based on these regression estimates.

We assessed the geographic variation of in-hospital mortality, complication rate, failure to rescue rate, length of stay, and the cost of hospitalization for 5 major surgical procedures in the United States and Japan.

Compared with Japan, the United States has substantially less geographic variation in surgical outcomes, but it has higher variation in cost.

Jay Bhattacharya, MD, PhD

Articles

Medicare Savings From Conservative Management of Low Back Pain

Electronic Health Records and the Frequency of Diagnostic Test Orders

Geographic Variation in Surgical Outcomes and Cost Between the United States and Japan

Effects of Physician Payment Reform on Provision of Home Dialysis