Case-mix adjustment is generally considered indispensable for fair comparison of healthcare performance. Inaccurate results are also unfair to patients as they are ineffective for improving quality. However, little is known about what factors should be adjusted for. We reviewed case-mix factors included in adjustment models for key diabetes indicators, the rationale for their inclusion, and their impact on performance.

This systematic review included studies published up to June 2013 addressing case-mix factors for 6 key diabetes indicators: 2 outcomes and 2 process indicators for glycated hemoglobin (A1C), low-density lipoprotein cholesterol, and blood pressure. Factors were categorized as demographic, diabetes-related, comorbidity, generic health, geographic, or care-seeking, and were evaluated on the rationale for inclusion in the adjustment models, as well as their impact on indicator scores and ranking.

Thirteen studies were included, mainly addressing A1C value and measurement. Twenty-three different case-mix factors, mostly demographic and diabetes-related, were identified, and varied from 1 to 14 per adjustment model. Six studies provided selection motives for the inclusion of case-mix factors. Marital status and body mass index showed a significant impact on A1C value. For the other factors, either no or conflicting associations were reported, or too few studies (n ≤2) investigated this association.

Scientific knowledge about the relative importance of case-mix factors for diabetes indicators is emerging, especially for demographic and diabetes-related factors and indicators on A1C, but is still limited. Because arbitrary adjustment potentially results in inaccurate quality information, meaningful stratification that demonstrates inequity in care might be a better guide, as it can be a driver for quality improvement.

Diabetes is associated with elevated morbidity, comorbidities, complications, and early mortality rates. It has an increasingly high prevalence and correspondingly high healthcare costs. Benchmarking is expected to reflect differences in quality of care, but case-mix factors can affect these scores as well. This review provides an overview of case-mix factors that are ideally included for specific indicators, showing: 

A large variation in case-mix factors is used, and the factors’ impact varies greatly. 

Rationales for the selection of case-mix factors are rarely provided. 

More evidence-based knowledge on case-mix adjustment for diabetes quality indicators is necessary for fair quality-of-care comparisons.

There is an increasing quest for public information on the quality of healthcare, which increases the need for transparency of the performance of individual healthcare providers as well as hospitals. For this purpose, the use of quality indicators has widely been adopted. Among the disease entities with ample research on quality indicators is diabetes. Treating and monitoring patients with diabetes is challenging, as the disease is associated with elevated morbidity, comorbidities, complications, and early mortality rates. It has an increasingly high prevalence, and healthcare costs are correspondingly high.

Benchmarking performances in diabetes care is expected to reflect differences in quality of care delivered, but it is commonly known that population characteristics such as comorbidities or age can affect these scores as well. To facilitate fair comparison of quality of care, case-mix adjustment based on these population characteristics is suggested for adequate public reporting on healthcare provider performance.

Indeed, the results of several studies indicate that the ranking of hospitals and physicians might change after adjustment by case-mix factors.

 Other studies, however, do not report clear effects.

These conflicting results can be explained by the variety of factors used in the different risk-adjustment models.

 This variety may exist due to a lack of knowledge of, or consensus on, factors that need to be included in these models; nevertheless, knowledge of factors that could affect the quality of care does not axiomatically mean that we should adjust quality indicators for these factors. Ten years ago, Chin

 outlined the key elements of a diabetes case-mix adjustment tool based on the scientific knowledge so far at that time. One of these elements is validity, meaning that case-mix adjustment tools should do the required task of adjusting for factors affecting glycated hemoglobin (A1C) value beyond the provider’s reasonable control. For example, when female patients or patients from minority ethnic groups receive lower quality of care and the performance can be controlled by the healthcare provider, gender and ethnicity should not be included in a risk-adjustment model.

Instead, this gender- and ethnicity-specific information should be made explicit (eg, by calculating stratified quality indicator scores). This is just a single element of an appropriate case-mix adjustment tool. In addition, a case-mix tool developed for a single end point, such as A1C value, is not necessarily valid for other diabetes-related outcomes. Another recommended criterion is clinical face validity, meaning that the case-mix adjustment tool should comprehend all key factors. Also, Chin

 refers to a conceptualization of 6 fundamental domains (demographics, access to care, healthcare-seeking behavior, geographic location, duration and severity of diabetes, and comorbidity),

 which is a good start to ensure incorporation of relevant variables, and thus, clinical face validity. Other recommended characteristics of a case-mix adjustment tool are feasibility (ie, the tool should be simple to use and affordable) and applicability (ie, the tool should be applicable to a specific target population of patients with diabetes).

In order to assess the current scientific knowledge on diabetes case-mix adjustment and to advance an appropriate and valid case-mix adjustment, researchers and quality assessors may benefit from an overview of the evidence for inclusion of case-mix factors for specific indicators. This is also relevant for healthcare providers because only appropriate case-mix adjustment allows comparisons and benchmarking among providers. Furthermore, more insight into relevant case-mix factors may facilitate registration needs, as only context information that matters needs to be collected.

We therefore aimed to review the literature to: a) identify relevant case-mix factors for the 6 most commonly used quality indicators for diabetes (namely, the outcome and process indicators for A1C value, low-density lipoprotein cholesterol [LDL-C] value, and blood pressure)

; b) to explore the rationale for the selection of these factors; and c) to study the impact of these factors on performance (ie, diabetes quality indicator scores).

A systematic literature search was performed in the electronic databases PubMed, EMBASE, and the Cochrane Library on June 5, 2013, to identify scientific publications specifically addressing case-mix adjustment for diabetes quality indicators. The search was performed using the terms “diabetes mellitus,” “quality indicators,” “risk adjustment,” and “case-mix,” as well as combinations of these terms and related Mesh terms (

]). The search was performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines.

We included original research explicitly addressing the impact of case-mix adjustment in diabetes care quality research (studies that both included adults and were written in English), with regard to 3 outcome indicators—A1C value (% or mmol/mol), LDL-C value (mg/dL), and blood pressure (mmHg)&mdash;and 3 process indicators&mdash;A1C measurements (% patients), LDL-C measurements (% patients), and blood pressure measurements (% patients).

Studies were excluded when they: a) addressed the quality of care provided to specific subgroups of patients with diabetes (eg, older patients) because these studies provide information on possible voids in provided quality of care to these subgroups rather than providing information on the relevance of including such factors in case-mix adjustment models; or b) applied case-mix factors as covariates to correct for variability between groups (intervention studies or trend studies).

The studies were first screened independently for eligibility based on title and abstract, by 2 reviewers (HC, GEV). In the case of disagreement, the publication was included in the first step of the selection process (see 

). Next, full-text copies were reviewed by the same 2 reviewers and a decision was made for final inclusion or exclusion. If the 2 reviewers disagreed on inclusion, a third reviewer (JCCB) was consulted. The list of included studies was completed by hand-searching the reference lists of the selected studies and the journals in which these studies were published.

From each included study, we extracted information on: aim; study design; study population; quality indicators; case-mix factors and their rationale for inclusion; the statistical association between case-mix factor and quality indicator; and, if available, the percentage, which explained variance and the effect of case-mix correction on ranking. Two reviewers (HC, JGMM) independently assessed the methodological quality of the studies and discussed their results for consensus. Using the Meta-Analysis of Observational Studies in Epidemiology guideline,

the methodological quality was evaluated using the following criteria (which were slightly rephrased to fit with the content of the current study): background and rationale reported, objectives and hypotheses stated, study design described, eligibility criteria specified, selection of sample described, selection method of case-mix factors described, and statistical methods described.

First, case-mix factors were categorized into the fundamental domains as described by Zhang et al

: demographic factors (eg, age, gender), diabetes-related factors (eg, duration of diabetes), comorbidity factors, geographic deprivation, and care-seeking factors (eg, free medical care). We added the domain “generic health” to classify the factors of health status and health status and quality of life, as used in several studies.

Then, we categorized the rationales for inclusion of case-mix factors according to the key elements of an ideal case-mix tool, as outlined by Chin,

 who stated that the tool: a) should be valid, implying that the tool adjusts for factors affecting the particular indicator beyond the provider’s reasonable control; b) needs to have clinical face validity, meaning that it does not lack key factors. With such validity, incorporation of relevant factors is ensured and the factor does not wrongly pardon a difference in the quality of care for which the provider should be accountable (eg, see gender example in introduction section); c) must be feasible, in terms of simplicity and affordability; and d) must be applicable to the target population; that is, the tool is developed on a similar set of patients.

Finally, we studied the impact of the case-mix variables on outcome. A factor was thought to have impact on outcome when an association between factor and indicator was shown in at least 3 studies, and no study reported otherwise. A similar reasoning was adopted for defining no impact. Impact was considered inconsistent when different studies reported both the presence and absence of an association between factor and indicator.

After applying the selection criteria, 13 studies were included in the review (Figure).

Characteristics and Quality Assessment of the Studies

 provides a comprehensive overview of the study characteristics. In summary (see also 

), the 13 studies were published between 1998 and 2012. Most of the studies were conducted in the Unites States (n = 10), 2 in the United Kingdom, and 1 in The Netherlands, and each of them had a retrospective observational design. The study samples ranged from 208 to 118,167 patients with diabetes.

 Six studies reported on both process and outcome indicators,

 with the outcome indicator for A1C value being the most frequently studied. Each study reported on more than 1 case-mix variable (see also next section). Effects of case-mix adjustment were studied across facilities

 One study reported effects on both levels.

 The criterion “selection method of case-mix variables described” could be applied to 6 studies,

 while the other quality criteria were satisfied by all studies (except the study design description

Case-Mix Factors Included in Adjustment Models and Rationales for Their Selection

 provides an overview of the factors included in case-mix adjustment models for quality indicator scores, stratified by domain and indicator. Altogether, 23 different case-mix factors were identified in the 13 studies. Demographic factors, especially age and gender, were most frequently included, followed by diabetes duration and comorbidity. The more extensive and diverse case-mix models were found for the outcome indicators, in particular for A1C value. The rationale for selection of case-mix factors was reported in 6 studies (eAppendix Table 3),

 mainly covering the elements of clinical face validity (eg, evidence from other studies) and feasibility (eg, data availability).

Impact of Case-Mix Factors: Outcome Indicators

Twelve studies investigated case-mix factors for this quality indicator. Impact on indicator scores (

) were found for marital status (demographic domain) and body mass index (BMI; diabetes-related domain); married patients with diabetes

 generally had lower A1C values compared with patients who were not married and those with higher BMI. For the factors of age,

inconsistent impact on outcome was reported. Other factors were rarely (n ≤2) studied. The percentages explained variance (eAppendix Table 2) induced by case-mix correction were generally small (3%-10.5%).

 Effects of case-mix correction on ranking particularly affected outliers and bottom- and top-performing hospitals in 3 studies.

 reported that after case-mix correction, clear differences among the 4 participating facilities emerged, while Berlowitz et al

 showed that case-mix correction reduced differences among facilities. Four studies were unable to demonstrate ranking effects after case-mix adjustment for A1C value.

Six studies investigated case-mix factors for this indicator

: inconsistent impact was found for the factors of age, gender, and comorbidity

; no association was found for diabetes duration.

 Other factors were rarely (n ≤2) studied and no data on explained variance was reported. Two studies showed a ranking effect of case-mix correction on facilities and healthcare providers,

 while 1 study showed no significant difference in ranking after case-mix adjustment.

 Other factors (as mentioned in Tables 1 and 2) were rarely (n ≤2) studied. No data on explained variance were reported. Although Greenfield et al

 demonstrated significant differences between unadjusted and adjusted indicator scores across healthcare providers, Gorter et al

Impact of Case-Mix Factors: Process Indicators

 Eight studies investigated case-mix factors for this indicator (Table 2).

 was inconsistent across studies, while no impact was found for gender.

 Other factors (as mentioned in Tables 1 and 2) were rarely (n ≤2) studied. In terms of explained variance, the difference in adjusted versus unadjusted results was 2.6% in 1 study on plan level.

An increase of comorbidity from 5% to 10% in patients with type 2 diabetes appeared to be associated with an increase of 1.3% in the screening rate of A1C. The regression model (using Healthcare Effectiveness Data and Information Set data) also included the total number of enrollees with diabetes, health plan type, and regions, overall explaining 24.2% of variation across health plans. No other data on explained variance or ranking were reported.

 Three studies investigated the impact of case-mix on this process indicator.

The impact of comorbidity was found to be absent.

 found that the overall regression model on LDL-C measurement explained 26.9% of variation across health plans. No other data on explained variance or ranking were reported.

 Two studies investigated the effect of case-mix correction on this process indicator.

 Therefore, all factors were rarely (n ≤2) studied. No data on ranking or explained variance of case-mix adjustment models were reported.

Twenty-three case-mix factors were identified; the most frequently considered factors were demographics (especially age and gender), diabetes duration, and comorbidity, and there was a predominant focus on outcome indicators (A1C in particular). The rationale for selection of case-mix factors in adjustment models was reported in 6 studies (42%): Face validity and feasibility of measurement were most commonly noted. There is evidence that marital status and BMI do affect outcome indicator scores for A1C. For the other case-mix factors, associations with indicators were inconsistent, absent, or marginally studied. Six studies showed impact on ranking,

 (2 of the 13 studies are not included here as ranking was not always studied).

To the best of our knowledge, this is the first systematic review aimed at identifying relevant case-mix factors for the 6 most commonly used quality indicators for diabetes care, which are outcome and process indicators for A1C, LDL-C, and blood pressure.

We found a consistent impact—ie, at least 3 studies reporting impact and no studies reporting otherwise&mdash;of marital status

 on the outcome indicator A1C. The demographic factor of marital status probably relates to a more regular lifestyle or getting support from a partner for self-management tasks, both resulting in better A1C values

 as an outcome indicator. Also BMI, a diabetes-related factor, appeared to be consistently associated with this indicator.

 Despite evidence of the association of BMI with hypertension and dyslipidemia,

 the relevance of this factor was rarely studied for the other diabetes indicators.

Inconsistent impact was found for age (on outcome and process indicators for A1C, outcome indicator LDL-C), gender (outcome indicators A1C and LDL-C), ethnicity, education, diabetes duration, health status (outcome indicator A1C), and comorbidity (on outcome and process indicators for A1C and LDL-C). Age and gender were among the factors studied most often, but their relevance was not clearly shown. It is worth highlighting that ethnicity has been considered a case-mix factor for outcome

 although its impact appeared inconsistent. Inclusion of this factor in case-mix adjustment models is controversial, as it suggests that ethnicity causes differences in outcome of healthcare which are beyond the control of the healthcare provider.

 Indeed, differences in blood-pressure levels exist across ethnicities; for example, black Africans generally have higher blood pressure, which is usually more difficult to treat,

 favoring case-mix adjustment on these grounds. However, it cannot be ruled out that ethnic minorities, due to prejudices, language barriers, or other factors, receive suboptimal medical care. Correction for ethnicity then masks poor quality of care.

The inconsistent results for diabetes duration may be explained by different measurement methods.

 One could question the relevance of diabetes-related factors, such as creatinine clearance, when case-mix adjustment is used for profiling reasons, as kidney function is subject to many influences.

 Although nephropathy is directly associated with diabetes management, the relative contribution of this specific variable to the quality of diabetes care needs more research. The inconsistent results for comorbidity regarding A1C indicators might be explained by variability in operationalization. While some studies used the Charlson comorbidity index,

 others used the US National Committee for Quality Assurance’s definition

Finally, it is worth noting that for the case-mix factors of diabetes duration (for the outcome indicators for LDL-C and blood pressure), gender (outcome indicator blood pressure and process indicator A1C), and age and education (outcome indicator blood pressure) no impact was found in at least 3 studies. Researchers and quality assessors may consider omitting these variables from their specific case-mix adjustment models, which will be beneficial for the statistical power of their analysis. Also, stakeholders or users of quality information should be aware of the value of case-mix and case-mix factors, that case-mix models may vary across different quality indicators, and that it may not be self-evident that background characteristics such as age and gender need to be included.

Because the case-mix adjustment models varied widely, it is hard to draw any conclusions on the final impact of ranking. Researchers are encouraged to further explore this issue.

The review was subject to some limitations. The included studies were cross-sectional, which makes it impossible to address causality. We judged the impact of a case-mix factor as robust when at least 3 studies showed the same (statistically significant) impact, or no impact, without other studies showing opposite results. This cut-off was arbitrarily set and we realize that robustness may require more studies showing similar results. In addition, in the case of 1 or more studies showing opposite results, this was classified as inconsistent results. Some of these inconsistent results could possibly be attributed to the variety in the study populations included—that is, primary versus secondary care, which may be related to the severity of the diabetes and might have affected the impact of case-mix factors. Again, the number of studies included in this review is too small to draw firm conclusions. Finally, in reviewing the scientific literature, we did not seek to limit the relevance of the results for research aims, but we did seek to expand the relevance, when possible, for use in real-world quality assessment and benchmarking.

Despite the compelling call for public transparency on quality of healthcare—for which case-mix adjustment on quality indicators is generally considered essential for fair comparison among healthcare providers&mdash;there still are large interstices in this field of research, as was shown by the current review. There appears to be a tendency for inclusion of factors based on face validity, often without a clear rationale, and scientific evidence on relevance and impact of these factors is still limited. From the available studies, we only found evidence for 2 factors&mdash;marital status and BMI&mdash;to be used in case-mix models for the outcome indicator A1C. Simple advice to the developers of diabetes quality indicators is therefore not possible. 

Radboud University Medical Center, Radboud Institute for Health Sciences, Scientific Institute for Quality of Healthcare (IQ healthcare) (HC, GEV, JCCB), Nijmegen, The Netherlands; Rijnstate Hospital (JGMM), Arnhem, The Netherlands. 

 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 (HC, JGMM, GEV, JCCB); acquisition of data (HC, GEV); analysis and interpretation of data (HC, JGMM, GEV, JCCB); drafting of the manuscript (HC, JGMM, GEV, JCCB); critical revision of the manuscript for important intellectual content (HC, JGMM, JCCB); provision of patients or study materials (JGMM); administrative, technical, or logistic support (JGMM); and supervision (JCCB). Note: HC and JGMM share first authorship.

Radboud University Medical Center, IQ Healthcare, Geert Grooteplein 21, 114 IQ Healthcare, 6525 EZ Nijmegen, The Netherlands. E-mail: h.calsbeek@radboudumc.nl.

1. Renders CM, Valk GD, Griffin SJ, Wagner EH, Eijk Van JT, Assendelft WJ. Interventions to improve the management of diabetes in primary care, outpatient, and community settings: a systematic review. 

Articles

WEB EXCLUSIVE

A systematic review of the impact and rationale for the selection of adjustment factors (case-mix factors) used to describe performance in diabetes care.

Case-Mix Adjustment for Diabetes Indicators: A Systematic Review

To compare the rates of mammographic screening and reasons for not receiving mammography between Medicare Advantage (MA) and Medicare fee-for-service (FFS) beneficiaries.

 Unadjusted rates of mammographic screening and reasons for not receiving mammography were compared between MA and FFS beneficiaries in total and within subgroups. Probabilities of mammographic screening were also compared between the MA and FFS beneficiaries after adjusting for age, education, race/ethnicity, health status, marriage status, and region of residence using modified Poisson regression models.

We found significantly higher unadjusted rates of mammographic screening in the MA beneficiary population compared with the FFS population (50.3% vs 44%, respectively; 

 <.0001). MA beneficiaries were more likely to receive a mammogram than FFS beneficiaries (unadjusted relative risk [RR], 1.13; 95% CI, 1.07-1.19; and adjusted RR, 1.06; 95% CI, 1.01-1.12). Furthermore, the top 6 reasons why MA and FFS beneficiaries did not receive mammograms were: “not needed,” “doctor did not recommend,” “forgot,” “not recommended annually,” “did not like mammograms,” and “appointment due soon.” Among non-Hispanic blacks and Hispanics specifically, FFS beneficiaries were more likely to report doctors not recommending mammograms than MA beneficiaries.

MA beneficiaries were 1.06 times more likely to receive mammography screening compared with the FFS population. Additionally, providers were less likely to recommend mammographic screening to non-Hispanic black and Hispanic FFS beneficiaries compared with non-Hispanic black and Hispanic MA beneficiaries.

		Medicare Advantage (MA) beneficiaries were more likely to receive mammography screening compared with fee-for-service (FFS) beneficiaries. 

		Three of the top 6 reasons for not receiving mammography involved whether mammography was needed. Although discrepancies in clinical guidelines remain, minimal mammographic screening recommendations should be disseminated to both providers and patients.

		While clinical guidelines were being debated, providers were less likely to recommend mammographic screening to non-Hispanic black and Hispanic FFS beneficiaries compared with non-Hispanic black and Hispanic MA beneficiaries.

Breast cancer is the most common type of cancer among women and is the second-leading cause of cancer-related death among women.

 The decline in breast cancer mortality has been attributed, in part, to the early detection of breast cancer with mammographic screening.

Although the value of mammographic screening has been demonstrated, the frequency of screening and age of the target populations remains a topic of debate. In 2009, the US Preventive Services Task Force (USPSTF) recommended that women aged 50 to 74 years should receive screening mammography every 2 years.

 The American Cancer Society (ACS) disagreed and continued to support the longstanding recommendation that women 40 years or older should receive screening mammography every year.

One important determinant of the use of mammographic screening is health insurance status.

 The 2 types of insurance compared in this study are Medicare Advantage (MA) and fee-for-service (FFS). MA plans are private health plans that receive a capitated amount per beneficiary from Medicare to provide all Part A and B benefits. MA prescription drug plans receive a separate payment from Medicare to provide Part D benefits.

MA plans started as Medicare managed care in the 1970s, and then were called “Medicare + Choice” after the 1997 Balanced Budget Act. It was not until the Medicare Modernization Act of 2003 that MA plans were called “Medicare Advantage.”

 When Medicare managed care plans first began in the 1970s, they were distinguished from traditional Medicare (or FFS) not only because they were private health plans operating under capitated payments, but also because they were mostly health maintenance organizations (HMOs).

MA enrollment has increased from 13% of all Medicare beneficiaries in 2004 to 31%, or 16.8 million beneficiaries, today.

 Sixty-four percent of MA beneficiaries are enrolled in HMOs and 31% are enrolled in preferred provider organizations.

Another key difference between MA and FFS is that MA plans are incentivized by star ratings whereas traditional FFS is not. More specifically, MA plans are given a 1 to 5 star rating annually, which is shared with the public

; consumers can then use this information when selecting plans. Additionally, MA plans with higher star ratings receive bonus payments. Providing preventive services, such as breast cancer screenings, is one of many measures used to calculate star ratings, and thus, MA plans have more of an incentive to ensure that their beneficiaries are receiving mammography.

Previous studies have found that women with Medicare HMO insurance had an earlier-stage breast cancer diagnosis than women with Medicare FFS insurance, which may lead to increased survival.

 Among breast cancer survivors, another study found that mammography was more often reported by those with Medicare plus private than those with only public insurance.

 Regarding the Medicare population as a whole, a study using 1996 data showed that Medicare HMO beneficiaries were more likely to receive mammograms than FFS beneficiaries.

 Updating these figures and better understanding why patients decline mammograms is needed. This study has 2 aims: 1) compare mammographic screening rates between Medicare FFS and MA beneficiaries in 2009, and 2) compare reasons for not receiving mammographic screening between Medicare FFS and MA beneficiaries in 2009.

This study used 2009 Medicare Current Beneficiary Survey (MCBS) data, collected via computer-assisted personal interviews.

 The MCBS is a longitudinal panel survey of a representative sample of the Medicare population

 that captures many beneficiary characteristics, including age, highest degree of education, race/ethnicity, self-reported health status, marriage status, region of residence, and MA enrollment. Additionally, the MCBS asks beneficiaries about mammography utilization and their reasons for not receiving mammography.

This study was approved by the University of Maryland Baltimore Institutional Review Board.

 shows how the sample was selected. Only females without a history of breast cancer were included. Since mammogram type (diagnostic or screening) was not specified, females with a history of breast cancer were excluded so calculated mammogram receipt rates would be focused on screening mammograms.

Of 14,695 Medicare beneficiaries, 7460 (51%) were females with no history of breast cancer. Of those, 5568 (75%) were FFS beneficiaries and 1892 (25%) were MA beneficiaries. Of the FFS and MA beneficiaries, 5417 (97%) and 1875 (99%) responded to whether they received a mammogram, respectively.

Next, 3033 (54%) and 932 (49%) FFS and MA beneficiaries did not receive mammograms, respectively. Of the FFS and MA beneficiaries who did not receive mammograms, 2175 (72%) and 755 (81%) reported why they did not receive a mammogram, respectively (not shown in the Figure).

All measures except age were self-reported from the survey. Age came from administrative records maintained by CMS. Receipt of screening mammogram was based on females with no history of breast cancer reporting whether they had received a mammogram in the last year. MA enrollment was based on whether patients reported current coverage by a Medicare HMO. However, this Medicare HMO was more similar to an MA plan because it included not only traditional risk HMOs, but also cost HMOs and other healthcare prepayment plans.

For preliminary analyses, the FFS and MA beneficiary populations were compared using χ

 tests of homogeneity to determine if differences in characteristics between the 2 populations were statistically significant.

For the first aim, unadjusted mammographic screening rates were calculated and compared between the FFS and MA beneficiary populations. Next, mammographic screening rates were compared between the 2 populations within strata of potential confounders, including age, education, race/ethnicity, health status, marriage status, and region of residence.

Lastly, a series of modified Poisson regression models were constructed to assess the effect of MA enrollment on the receipt of screening mammography. Poisson regression with robust error variance has been previously used to estimate relative risk for binary outcomes.

 Relative risks were reported instead of odds ratios to avoid overestimation of risk if events are common, due to misinterpretation of odds ratios as a measure of risk.

 The final adjusted model contained all 6 potential confounders mentioned above: age, education, race/ethnicity, health status, marriage status, and region of residence. The previous study using data from 1996 also controlled for age, education, race/ethnicity, and health status.

 Marriage status and region of residence were also controlled for in the study because these characteristics were statistically significantly different between the FFS and MA populations in preliminary analyses and were considered potential confounders of the relationship between health plan type and mammogram use.

For the second aim, unadjusted reported rates for reasons for not receiving a mammogram were calculated and compared between the FFS and MA beneficiary populations. Next, these rates were compared within strata of the same 6 potential confounders. Chi-squared and Fisher’s exact tests were used throughout to determine if differences were statistically significant at 

shows the characteristics of FFS versus MA beneficiaries. In general, these results indicate that there are greater proportions of MA beneficiaries, compared with FFS beneficiaries, who are older, racial minorities, married, healthier, and living in western regions, mountainous regions, or Puerto Rico.

 compares unadjusted mammographic screening rates in Medicare FFS versus MA beneficiaries in the entire sample and then within strata of the 6 potential confounders: age, race/ethnicity, education, self-reported health status, marriage status, and region of residence.

A statistically significantly higher proportion of MA beneficiaries than FFS beneficiaries received a mammogram in 2009 (Table 2) (50.3% vs 44%, respectively; 

 <.0001). Also, statistically significantly higher proportions of MA beneficiaries received a mammogram in 2009 compared with FFS beneficiaries within specific strata: aged under 65 years, non-Hispanic white race/ethnicity, non-Hispanic black race/ethnicity, other races/ethnicities, high school graduate or lower education, good or fair health status, single marriage status, northern regions, eastern regions, and Puerto Rico.

MA enrollees were more likely to receive a mammogram compared with Medicare FFS enrollees before and after adjusting for confounders (Table 2) (unadjusted relative risk [RR], 1.13; 95% CI, 1.07-1.19; and adjusted RR, 1.06; 95% CI, 1.01-1.12).

 shows the top 6 reasons why beneficiaries did not receive mammograms in 2009. These reasons, from most to least reported, were that the beneficiary did not think it was needed, the doctor did not recommend it, the beneficiary had forgotten, the beneficiary did not think the mammogram was recommended annually, the beneficiary did not like mammograms, and the beneficiary had an appointment due soon. The proportions of beneficiaries reporting these reasons in the FFS and MA beneficiary populations were not statistically significantly different.

Eleven other reasons for not undergoing mammography were asked in the survey. These were: “cost,” “test useless,” “beneficiary not at risk for breast cancer,” “doctor recommending against it,” “inconvenience,” “beneficiary afraid of results,” “radiation causing cancer,” “beneficiary had a mastectomy,” “beneficiary never heard of a mammogram,” “beneficiary too ill,” and “other.” These reasons were not included in Table 3 because they were reported by less than 5% of both the FFS and MA beneficiaries. Additionally, only 8 differences were found to be statistically significant when comparing proportions of FFS and MA beneficiaries’ reporting reasons for not receiving mammography within strata of age, education, race/ethnicity, health status, marriage status, and region (see 

 Among beneficiaries aged under 65 years, a higher proportion of FFS versus MA beneficiaries reported not receiving mammograms because the doctor did not recommend it (Table 4) (30.11% vs 12.70%, respectively; 

 <.05). Among beneficiaries 85 years or older, a higher proportion of MA versus FFS beneficiaries reported not receiving mammograms because it was not recommended annually (Table 4) (21.54% vs 14.98%, respectively) or they had forgotten (Table 4) (14.36% vs 6.42%, respectively; 

 Among beneficiaries who graduated high school at the most, a higher proportion of MA versus FFS beneficiaries reported not receiving mammograms because the appointment was due soon (Table 4) (12.94% vs 8.28%, respectively; 

 Among non-Hispanic blacks, a higher proportion of FFS versus MA beneficiaries reported not receiving mammograms because the doctor did not recommend it (Table 4) (30.29% vs 16.87%, respectively; 

 <.05). In the same group, a higher proportion of MA vs FFS beneficiaries reported not receiving mammograms because they did not like mammograms (Table 4) (15.66% vs 6.22%, respectively; 

 <.05). Among Hispanics, a higher proportion of FFS versus MA beneficiaries also reported not receiving mammograms because the doctor did not recommend it (Table 4) (34.15% vs 9.09%, respectively; 

 Among beneficiaries reporting excellent or very good health status, a higher proportion of MA versus FFS beneficiaries reported not receiving mammograms because they were not recommended annually (Table 4) (18.45% vs 12.98%, respectively; 

MA beneficiaries are slightly more likely to receive mammographic screening compared with FFS beneficiaries. Reasons for not receiving mammography differ between MA and FFS beneficiaries within specific subgroups of age, race/ethnicity, education level, and health status.

Carrasquillo et al (2001) used 1996 data and found that Medicare HMO beneficiaries tended to be slightly younger minorities with similar education and perceived health as FFS beneficiaries.

 This study similarly finds that MA beneficiaries tend to be racial minorities and have similar education compared with FFS beneficiaries. However, this study also finds that compared with FFS beneficiaries, more MA beneficiaries are older, healthier, married, and living in western regions, mountainous regions, or Puerto Rico. The increased MA enrollment over the past decade may have contributed to these changes in demographic characteristics between the MA and FFS populations.

Medicare Advantage beneficiaries were 1.06 times more likely to receive mammography compared with fee-for-service (FFS) beneficiaries. FFS providers were less likely to recommend mammograms to racial/ethnic groups.

The Effect of Medicare Advantage Enrollment on Mammographic Screening

 To analyze inappropriate use of magnetic resonance imaging (MRI) for patients with low back pain in a healthcare system with no financial incentives for overuse.

 We used administrative data to assess the appropriateness of lumbar spine (LS) MRI in the Veterans Health Administration.

All veterans who received LS MRI in the outpatient setting in fiscal year 2012 were included. We based our assessments of appropriateness on CMS criteria, which have been endorsed by the National Quality Forum. Generalized estimating equations were used to evaluate characteristics of inappropriate scans.

Of the 110,661 LS MRIs performed, 31% were classified as inappropriate. Most scans that were considered appropriate were characterized as such because they were preceded by conservative therapy (53%). "Red flag" conditions were responsible for a much smaller percentage of scans being considered appropriate; 13% of scans were preceded by conservative therapy and were performed in patients with a red flag condition, while only 4% of scans were considered appropriate because of red flag conditions only. Scans ordered in the emergency department and in urgent care, primary care, and internal medicine clinics were most likely to be classified as inappropriate. Resident physicians were significantly less likely than other provider types to order inappropriate LS MRIs (odds ratio, 0.80; 

 <.0001). Approximately 24% of providers ordered 74% of inappropriate scans.

We found that 31% of LS MRIs were inappropriate in a healthcare system largely absent of financial and other incentives for ordering. The problem of inappropriate ordering of LS MRI is concentrated in a small number of providers; any provider-facing interventions to reduce inappropriate order should therefore be targeted, rather than aimed at all providers who order LS MRI.

	Using CMS criteria, we found that 31% of lumbar spine magnetic resonance imaging (LS MRI) performed in fiscal year 2012 was inappropriate. This high level of inappropriate ordering is particularly striking in a system largely free of financial and other incentives for ordering, suggesting that nonpecuniary factors play a role in these inappropriate ordering choices. We found a high concentration of inappropriate ordering among few providers. Results indicate that efforts to reduce inappropriate use of LS MRI should involve cultural change, as well as patient and provider education, rather than financial incentives alone, and that provider-facing efforts should be targeted instead of universal.

Most Americans will experience low back pain (LBP) at some point in their adult lives.

 Many patients with LBP receive imaging to explore the cause of their pain

; advances in imaging technology have revealed unprecedented anatomical detail for LBP patients. The underlying difficulty is that, for many patients, anatomical and functional abnormalities do not correlate well, and thus imaging can fail to guide treatment of the primary complaint.

 In one study of individuals who underwent magnetic resonance imaging (MRI), approximately 90% were detected to have a degenerated or bulging disc, 36% had a herniated disc, and 21% had spinal stenosis—but all patients were asymptomatic.

 For most patients, LBP resolves on its own

; even sophisticated imaging of the lumbar spine (LS) for nonchronic pain does not improve outcomes.

Overuse of LS MRI is suspected to be a large problem; multiple societies have prioritized it as a quality-of-care measure or a measure of prudent stewardship of resources. The American College of Physicians and the American Association of Neurological Surgeons recommend against the use of LS MRI for nonspecific or nonpersistent LBP in their Choosing Wisely campaign.

 The National Quality Forum (NQF), a nonprofit entity that sets national priorities for quality measurement, has endorsed a CMS measure regarding inappropriate use of imaging in LBP.

The reasons for overuse of imaging are not well understood. Fee-for-service financial incentives have been suggested as a potential reason for overuse of lumbar spine imaging,

 In this study, we explored inappropriate use of LS MRI in a US healthcare system—one largely free of financial incentives for overuse, and in which physicians are largely insulated from malpractice concerns&mdash;to understand whether inappropriate ordering occurs in the absence of these financial or legal influences. The federal government accepts liability for any instances of medical negligence or wrongdoing on the part of the Department of Veterans Affairs (VA) providers; concerns over malpractice or incentives to practice defensive medicine are therefore much lower for VA providers compared with providers in the larger US healthcare environment. It is also not well understood whether inappropriate ordering is equally distributed across providers, or concentrated among relatively few high-ordering providers; this information carries important implications for quality improvement activities. Here, we use NQF-endorsed CMS criteria to evaluate inappropriate ordering of LS MRI across the entire VA, including the patient-, provider-, and system-level factors associated with the practice.

To study the appropriateness of this imaging modality, we identified the population of veterans who received an outpatient LS MRI in VA in fiscal year (FY) 2012. If patients received more than 1 LS MRI, we retained information regarding the first MRI. Additionally, we searched VA inpatient and outpatient data from FY 2010 to FY 2012 to evaluate the care and condition of these veterans prior to receipt of LS MRI. CMS criteria pertain to LS MRIs with a diagnosis of LBP on the MRI claim; because VA MRIs do not have associated diagnosis codes, we evaluated all LS MRIs. We considered the clinic visit with the ordering provider that immediately preceded the scan to be the ordering visit.

We evaluated the appropriateness of LS MRI according to the NQF-endorsed CMS measure #0514 (“MRI Lumbar Spine for Low Back Pain”),

). This measure classifies an LS MRI as appropriate if it was: 1) preceded by conservative therapy, or 2) conducted in a patient who has a “red flag” condition. CMS considers physical therapy or chiropractic care in the 60 days prior to the scan, or an evaluation and management visit occurring between 28 to 60 days prior to the scan, to be conservative therapy. Physical therapy, chiropractic care, and evaluation and management visits were identified by Current Procedural Terminology codes specified by the NQF-endorsed CMS measure.

 We classified an evaluation and management (E&M) visit as conservative therapy only if the scan was not ordered during that visit.

We made minor modifications to the CMS criteria regarding red flag conditions. The CMS measure excludes LS MRIs occurring in patients with red flag conditions from the denominator; here, we included them but considered them to be appropriate in order to present a complete view of all LS MRIs in VA. CMS red flag conditions include trauma in the 45 days preceding the scan; LS surgery in the 90 days preceding the scan; or intravenous drug abuse, neurologic impairment, cancer, HIV diagnosis, or diagnosis of unspecified immune deficiency in the 365 days preceding the scan. The measure includes all cancer; we excluded primary skin and primary prostate cancer as these types of neoplasms would not cause lumbar compression and rarely have an impact on spine pain. Additionally, there are concerns regarding a high false-positive rate for prostate cancer due to overuse of prostate-specific antigen (PSA) screening. To the CMS red flag conditions, we added intraspinal abscess or compression fracture in the 30 days prior to the scan, and osteomyelitis in the previous 365 days. All red flag conditions were identified through

 International Classification of Diseases, Ninth Revision, Clinical Modification 

) codes in inpatient and outpatient data sets.

From these data, we identified which LS MRI scans were considered inappropriate. LS MRIs were considered inappropriate in patients who did not have a red flag condition in the time frame of interest and those who did not receive appropriate conservative therapy.

We also investigated patient-, provider- and facility-level characteristics associated with inappropriate scans, as well as any regional differences in proportion of inappropriate scans. Patient-level characteristics included age, gender, and race. Provider-level variables included ordering provider type (attending physician, resident, physician’s assistant, and advanced practice nurse/nurse practitioner) and the clinic in which the scan was ordered (internal medicine/primary care, physical medicine and rehabilitation; neurology, neurosurgery, orthopedics, pain medicine, “other” medicine services, miscellaneous other medicine services, chiropractic care, and emergency department [ED] and urgent care). Other medicine services refer to various primary and specialty care clinics separate from the other categories of care (eg, geriatrics, oncology, hospice, dementia clinic).  Miscellaneous other services refer to all ancillary and support services (eg, cast clinic, social work, telephone triage).  The facility-level variable included complexity, defined in 5 levels (1a, 1b, 1c, 2, and 3; from most to least complex).

Characteristics of inappropriate scans were evaluated using generalized estimating equations to account for clustering of patients within facilities. Our generalized estimating equations employed a logit link, a binomial distribution family, and an independent correlation structure. We ran sensitivity analyses in which we evaluated facility-level physical therapy penetration—defined as physical therapy visits per 1000 pro-rated patients&mdash;and in which, we evaluated alternative definitions of conservative therapy. Specifically, we required the conservative therapy visits to have an accompanying diagnosis of low back pain. Post estimation commands were used to compare outcomes among different levels of each categorical level; for example, to compare residents with each other provider type, physician’s assistants to each other provider type, etc. All analyses were conducted in SAS version 9.2 (SAS Institute, Inc, Cary, North Carolina). 

In FY 2012, 110,661 LS MRIs were ordered in the outpatient setting by 16,273 unique providers. Most patients received 1 LS MRI; only 3.7% of patients received more than 1. Each provider ordered a median of 3 LS MRIs and a mean of 6.8 LS MRIs. Of the 110,661 total scans, 33,998 (30.7%) were classified as inappropriate; these patients had no red flag conditions coded and did not receive conservative therapy in a VA facility prior to their scan (

). The majority of scans classified as appropriate were done so because the patient was deemed to have received prior conservative therapy (65.3%). Another 16.6% of LS MRIs were classified as appropriate because the patient had a red flag condition. (Note that the numbers add up to more than 100% because patients could have had both a red flag condition and have received conservative therapy.)

describes the demographics of the population, which on average, was male, white, and middle-aged. The greatest proportion of scans (42%) was ordered for patients seen at the most-complex (level 1a) facilities. However, the greatest proportion of inappropriate scans (36%) was ordered for patients seen at the least-complex facilities (level 3) (Table 2). The number of scans ordered varied greatly by clinic setting, with the greatest percentage of scans ordered by the primary care/internal medicine clinic (62%), and the fewest percentage ordered by neurosurgery and orthopedics (6% combined). The greatest percentage of total scans (69%) was ordered by medical doctors or doctors of osteopathy, and the lowest was ordered by physicians’ assistants (8%). We did not find meaningful regional variation in proportion to scans considered inappropriate.

Results from generalized estimating equations revealed that patients aged under 35 years were more likely to have an inappropriate scan than those of all other ages (

). Black patients were less likely to have an inappropriate scan compared with white patients (odds ratio [OR], 0.80; 

 <.0001). Scans ordered at the lowest-complexity facility (level 3) were more likely to be inappropriate compared with those ordered at the highest-complexity facilities (OR, 1.27; 

 = .0096). These multivariate results also show the ED or urgent care clinic to be the settings most strongly associated with an inappropriate scan (OR, 1.42; 

 <.0001, with reference category of primary care/internal medicine clinic). In fact, ED/urgent care providers were more likely than providers practicing in any other clinic to order inappropriate scans (Table 3). Primary care/internal medicine clinics were also associated with a high likelihood of ordering inappropriate scans; all other clinic settings were less likely to order inappropriate scans than the primary care/internal medicine clinic (Table 3). Neurology, neurosurgery, and orthopedic clinics all practiced similarly; after adjusting for covariates, each was equally likely to order an inappropriate scan compared with the others (Table 3). Medical doctors (MDs), doctors of osteopathy (DOs), nurses, and physician’s assistants were also equally likely to order an inappropriate scan; however, residents were less likely to order an inappropriate LS MRI scan compared with MD/DOs (OR, 0.80; 

Results also show that 10.9% of providers ordered 50.1% of inappropriate scans, and 23.6% of providers ordered 74% of inappropriate scans (

). There appeared to be a weak, volume-percent inappropriate relationship; that is, there was little evidence of a relationship between the volume of scans ordered by a provider and the percentage of those scans deemed inappropriate.

We ran sensitivity analyses in which we modified the categorization of E&M codes. CMS considers any E&M visit between 28 and 60 days prior to the MRI to be conservative therapy, as long as the MRI was not ordered during that visit. Using this definition in our base-case, 63% of scans were conducted in patients who had a qualifying E&M visit. In sensitivity analyses, we restricted the E&M visit to have a diagnosis of LBP. Using this more restrictive definition, we found only 34% of LS MRI scans had a qualifying E&M visit, and thus, 53% of LS MRIs were considered inappropriate.

We also ran other sensitivity analyses. We tested a model with a covariate of facility-level physical therapy penetration. This variable was not a significant predictor of inappropriate LS MRI. Additionally, we ran a model on the subset of scans ordered by providers who ordered 5 or more scans per year, to avoid drawing conclusions based on a large number of providers each ordering few scans. Results did not change with respect to direction or significance. Third, we tested the inclusion of all prostate cancer in the red flag conditions, thus considering all LS MRI imaging in patients with a prostate cancer diagnosis in the previous 365 days to be appropriate. This resulted in slightly fewer (28.9%) LS MRIs being considered inappropriate.

We estimated the cost of providing each LS MRI in VA using the Decision Support System, the activity-based cost allocation system used by VA. We estimated that VA spent $44.4 million on LS MRIs in FY 2012, including $13.6 million on inappropriate scans.

The extent of overuse of LS MRI has been the subject of much debate. Here, we find that, despite a strong body of evidence regarding the poor utility of LS MRI in nonchronic LBP, and the identification of LS MRI as a low-value service by multiple specialty societies,

 31% of LS MRIs conducted in VA in FY 2012 were inappropriate.

This article analyzes use of lumbar spine magnetic resonance imaging in a national sample of patients with low back pain.

February 2016

Case-Mix Adjustment for Diabetes Indicators: A Systematic Review

The Effect of Medicare Advantage Enrollment on Mammographic Screening

Inappropriate Ordering of Lumbar Spine Magnetic Resonance Imaging: Are Providers Choosing Wisely?

Molecular Testing Patterns in Metastatic Non—Small Cell Lung Cancer

USPSTF Colorectal Cancer Screening Guidelines: An Extended Look at Multi-Year Interval Testing

The Value of Value-Based Insurance Design: Savings From Eliminating Drug Co-Payments

The Promise and Perils of Big Data in Healthcare

Continuity of Care and Changes in Medication Adherence Among Patients With Newly Diagnosed Diabetes

Characterizing Health Plan Price Estimator Tools: Findings From a National Survey

Longitudinal Adherence to Colorectal Cancer Screening Guidelines