Explaining Higher-than-Expected Medicare Costs in US Counties

Health care cost evaluation in the United States is a controversial and complex subject. John Wennberg, founder of the Dartmouth Atlas Institute, first began to examine the role of medical practice variation in the geographic variability of US health care in the 1970s. Wennberg and his colleagues found that physician preferences often drove health care cost and utilization but did not lead to better outcomes for patients. The Atlas researchers also thought they saw evidence that some areas of the country were superior at providing high-quality care at lower cost. For a time, policy makers seemed extremely focused on figuring out the formula for promoting this high-quality, lower-cost care everywhere in the nation. Then, in a 2013 report entitled “Variation in Health Care Spending: Target Decision Making, Not Geography”, the IOM (Institute of Medicine)’s Committee on Geographic Variation in Health Care Spending and Promotion of High-Value Care, headed by Joseph Newhouse and Alan Garber of Harvard, added an important new chapter to the conversation.

While the Dartmouth researchers had mostly focused their analyses on Medicare data, the IOM committee used a combination of Medicare and Medicaid data and private health insurance data, along with quality indicator and risk adjustment methods in their analysis. While the committee agreed with the Dartmouth researchers that there was substantial, pervasive geographical variation in health care delivery or practice patterns and that adjusting for age, sex, income, race and health status did not remove it (Fisher & Skinner, 2013), they also found that while practice patterns were (at least partially) responsible for variation in both utilization and cost, there was “no clear pattern suggesting that certain regions or providers uniformly deliver higher-value care than others” (Fineberg, H., cited in Newhouse et al., 2013, p. xii).

The report’s initially puzzling new finding that there was no regional association between commercial insurance spending and Medicare spending, led to a key insight: for the commercially insured, variations in price (rather than cost or utilization) explain most of the variation in health care spending. 

Many other important contributions to the conversation around US health care cost and quality fall under the umbrella of topics like health care access, risk-factor surveillance, health determinants, and the central issue of health equity.  An important project associated with these efforts is  the County Health Rankings initiative by the University of Wisconsin, Madison's Population Health Institute (2010). 

With funding from the Robert Wood Johnson Foundation (RWJF), the Health Rankings and Roadmaps program utilizes twenty sources of publicly available and well-validated data to create measures of the health of US counties that capture both the average length and quality of life (the Health Outcomes ranking), and the impacts of health determinants such as health behaviors, clinical care access and quality, social and economic factors, and the physical environment (the Health Factors ranking). 

The mix and alignment of the Health Rankings factors in identifying geographic regions at risk make it clear: Americans who are poor, not white, or live in areas with inadequate educational opportunities are likely to struggle with poor health. 

These additional barriers to good health exist on top of the other systemic problems already listed. They may also contribute to an understudied problem of underutilization of appropriate medical visits and services in some areas of the country. These areas include places with widespread cultural mistrust of the health care system and/or little geographic proximity to health care services.

With all these components to analyze when attempting to describe US health care cost and utilization, it is evident why researchers’ understanding of these trends is still evolving. Still, a picture of a system in need of reform is emerging. In the coming years, analysts, data scientists, and public and population health experts will continue to come together to test previous findings and to accelerate the process of developing overdue solutions that benefit all Americans.

This analysis made use of the Medicare DeSYNPuf data set available through CMS for the years 2008-2010 and chose inpatient and outpatient claims samples, Number 13. The DeSYNPuf data set is provided for data entrepreneurs, researchers, and data miners to allow for the development of novel research approaches and “innovations that may reveal unanticipated knowledge gains while preserving beneficiary privacy” (CMS, 2019).

The inpatient and outpatient claims were first left joined into the beneficiaries' data, creating a file with approximately one million rows (row count = 991,604). 

The age of each beneficiary was calculated for each year period that the beneficiary had representation in the beneficiaries’ table. 

Race and ethnicity variables for each beneficiary were combined and dummy coded, creating 4 binary variables: NH_WHITE, AA_BLACK, NW_HISPANIC and OTHER, for those who self-identified as “Another race”. Asian Americans / Pacific Islanders and Native Americans / Alaska Natives had no representation in this data set. 

Chronic condition markers were converted to binary variables, including B_ALZHDMTA (Alzheimer’s disease or another dementia), B_CHF (congestive heart failure), B_CNCR (any cancer), B_COPD (chronic obstructive pulmonary disease), B_DEPRESSN (depression), B_DIABETES (type I or type II diabetes), B_ISCHMCHT (ischemic heart disease), B_OSTEOPRS (osteoporosis), B_CHRNKIDN (chronic kidney disease), B_RA_OA (rheumatoid or osteoarthritis), B_STRKETIA (stroke or TIA) and B_ESRD (end-stage renal disease).

The Elixhauser Co-morbidity Classification system, developed by Elixhauser et al. (1998) is a method of categorizing co-morbidities of patients using International Classification of Diagnosis (ICD) codes found in administrative claims data. It is widely used to measure burden of disease or case mix within a healthcare system as well as to risk adjust for severity of illness when comparing health care outcomes. The “MCHP Elixhauser Co-morbidity Index SAS Macros for Hospital Data to be used with ICD-9CM Diagnosis Codes” was used to calculate the total number of disease categories present within each individual claim (Quan et al., 2006). 

An overall Elixhauser Score (single numeric index) for each claim was then calculated by applying weights obtained from the algorithm developed by Van Walraven et al. (2009) based on the association of co-morbidities and death to each category of co-morbidity present in that beneficiary’s claim. 

The number of codes for each claim was also recorded as CODES_COUNT, and the number of HCPCS type 1 (procedure / CPT) codes for each claim was recorded as HCPCS_COUNT. CLAIM_TYPE was recorded for each claim as “inpatient” for the inpatient claims, and “outpatient” for outpatient claims. CLM_UTLZTN_DAY_CNT was recorded as the length of stay for inpatient claims.

Beneficiary Aggregation

Aggregation was performed for each beneficiary for each year. Inpatient and outpatient visit totals for each year were counted as IP_VISITS and OP_VISITS. 

The number of unique providers for each beneficiary for each year was recorded as NUM_PRVDRS, by counting the number of unique NPI numbers the patient had for each year. 

Costs and plan coverage months for each type of Medicare plan were averaged across the years for each beneficiary. Care was taken to indicate chronic conditions across years, but to indicate each condition only once, and to keep track of the count of new conditions for the three-year period for each beneficiary (NEW_CHRONIC).

Once these values were calculated, the year-level summaries created by taking means of the claims information for each beneficiary was aggregated to the individual beneficiary level. To do this, mean values were taken for all the calculated ages, Elixhauser scores, inpatient lengths of stay, and the numbers of codes and procedures. 

Visit counts and costs were summed across the years, to give a total number of visits and total cost for each beneficiary for the three-year period. The total number of beneficiaries represented in this data set was 114,530 for the years 2008-2010. There were 85,524 beneficiaries with one or more visits (users), and 61,000 users with > 3 visits. Just under 12,000 users had between 9 and 15 visits during the three-year period. This aggregated data represented summaries at the beneficiary / user level of a total of 842,288 visits, 776,259 outpatient visits and 66,029 inpatient visits. 

Prior to aggregation at the county level, 350 beneficiaries were removed for having an outlying >40 visits each during the three-year period.

Pivot, County Aggregation

With one row for each beneficiary, the data was pivoted to aggregate across beneficiaries for each county within each state. 

The total number of counties in the contiguous US states, Alaska, Hawai’i, and Washington DC (not including the territories) is 3,143, but only 3,010 were initially represented in the data set. 

FIPS codes were used to identify both states and counties. 

Some counties had only one beneficiary (n=121). 

There were 693 counties with 5 beneficiaries or fewer, and 1,267 counties with 10 beneficiaries or fewer. 

At this point, the UW Madison / RWJF County Health Rankings were joined into the data for all the counties represented in the data set.  Some of the counties in the data set did not have rankings, so they were dropped from the data (n=63). 

Finally, outlier counties where the mean Elixhauser score for all of the users in the county was > 6 were removed after it was determined that these counties had small numbers of beneficiaries (n=2). 

The final number of counties in the aggregated county-level data set was then 2,947.  Finally, for all the analyses (but not for the descriptive statistics), an additional 38 counties were dropped, because while they had beneficiaries, they had zero users during the three-year analysis period. 

This brought the final total # of counties analyzed via regression and factor analysis to 2,909 out of 3143, or 92.6% of US counties.

For the RAM, the outcome variable used was Total Cost per Beneficiary for each county and the predictor variables were County Mean Age, County Sex Proportion (the mean proportion of males to females), Diverse Proportion (the mean proportion of non-white to white), County Mean Chronic Conditions (the mean number of chronic conditions), the Health Factors Proportional Rank (the County Health Ranking for each county within its state, represented as a proportion where 0 represents the lowest-ranked county and 1 represents the highest-ranked) and the severity of illness expressed as the Total Elixhauser Score per User for each county.

A Generalized Linear Model (GLM) with log link and gamma distribution was used to create the RAM.  The GLM does not assume homoscedastic error nor normality.  It is free from retransformation bias, and works well with positively skewed and zero-heavy distributions (Barnett, 2017; Davidian M, 2007). 

The deviance value/DF=0.585, is low and indicative of good fit at 1572.3 when compared with its asymptotic chi square with 2687 degrees of freedom, obtains a p-value of >0.99 (SAS Institute, 2013).  Of the six predictor variables, 4 variables- Diversity Proportion / Race (βRACE = 0.282), Mean Number of Chronic Conditions (βCC = 0.348), Severity of Illness (βSEV = 0.369), and Health Factor Ranking of the county (βHFR =0.133), significantly contributed to the explanation of the variance in the cost variable.

When the link function is natural log, then β represents the percent change in y for a unit change in x. The interpretation of the beta coefficients was as follows:

1. For every unit increase in the mean number of chronic conditions, there was a 35% increase in the cost per beneficiary. 

2. For every unit increase in the total Elixhauser score per user, there was a 37% increase in cost per beneficiary. 

3. For every unit (0.1) increase in the Health Factor proportionate ranking, there was a 13% increase in cost per beneficiary. 

4. For a every unit (0.1) increase in diversity proportion, there was a 28% increase in cost per beneficiary.

c) Comparing the accuracy of predictions between an OLS model with a Square Root transformed outcome variable, and GLM Log link Gamma Regression model.

The RMSE for GLM log-linked gamma regression model was 35791.84. 

GLM RAM was applied to the dataset. 1,543 out of 2,694 counties had O/P ratio < 1 indicating that the observed costs were lower than the predicted costs. These counties were categorized as having "lower-than-expected" costs per beneficiary.

The remaining 1,151 counties had O/P ratio > 1, indicating that the observed costs were higher than the predicted costs. These counties were categorized as having "higher-than-expected" costs per beneficiary.

Between-group differences in demographic and illness factors were assessed using the nonparametric Mann Whitney U test. There was no significant difference in Mean Age, Sex Proportion, Mean Number of Chronic Conditions and Total Elixhauser Scores per User between the two groups of counties. Group 1 (the higher-than-expected cost group) had significantly more counties with greater racial diversity compared to Group 2 (the lower-than-expected cost group) (p<0.01).

The variables representing healthcare utilization were Total Number of Visits per Beneficiary (Inpatient and Outpatient), Proportion of Inpatient Visits, Mean Length of Inpatient Stay, Average Number of Providers per User and Average Number of Procedures per User. The distribution of these variables was non-normal, and therefore between group differences in healthcare utilization variables was assessed using the Mann Whitney U test. 

As compared to the lower-than-expected cost counties (Group 2), the higher-than-expected cost counties (Group 1) had a significantly higher Number of Visits per Beneficiary (Mean visits-7.89 v/s 6.5), Proportion of Inpatient Visits (9% v/s 7%), Mean Length of Inpatient Stay (1.88 days v/s 1.45 days), Average Number of Providers per User (2.13 v/s 1.75) and Average Number of Procedures per User (12.4 v/s 9.8). 

To attempt to determine what was behind the co-variability of cost with the other significant variables, an exploratory factor analysis (EFA) was conducted utilizing the variables associated with the cost outcome.  This technique is used “to examine the underlying patterns or relationships for a large number of variables to determine whether the information can be condensed or summarized in a smaller set of factors or components (Hair et al., 2019). 

Sixty-seven counties with costs > $200,000 per beneficiary were first removed in order to improve the reliability of the variance in the cost and utilization data.

 Candidate Variables

The candidate variables for this analysis were taken from a list of variables that had significance again the outcome of cost at the 90% confidence interval, including:

·        The proportion of users to all beneficiaries in the county (CNTY_USR_PROP)

·        The county prevalence of end-stage renal disease (TOT_ESRD_PREV)

·        The average number of Medicare (any type) coverage months per beneficiary per year (TOT_CVRG_MOS_PB_PY)

·        The average number of Part D / prescription drug coverage months per beneficiary per year (TOT_RXCVRG_MOS_PB_PY)

·        The total number of inpatient visits per user (TOT_IP_VSTS_PU)

·        The total number of outpatient visits per user (TOT_OP_VSTS_PU)

·        The total number of ICD-9 codes per user (CODES_PU)

·        The total number of HCPCS Level I procedure codes per user (PROCS_PU)

·        The total number of providers per user (PRVDRS_PU)

·        The total number of inpatient days (length of stay) per user, as the square root of that # (SR_TOT_IPDAYS)

·        The total number of chronic conditions per beneficiary / chronic condition prevalence (TOT_CC_PREV)

·        The total number of new chronic conditions per beneficiary for the three-year period / chronic condition incidence (TOT_CC_INCD)

·        The total Elixhauser Score per User (TOT_ELIX_PU)

·        The total cost per beneficiary, as the square root of that # (SR_TOT_COST_PB)

·        The Health Outcomes proportionate rank (HO_PROP_RANK)

·        The Health Factors proportionate rank (HF_PROP_RANK)

·        The proportion of beneficiaries in the county self-identified as not white (DIVERSE_PROP)

EFA must only be performed on metric variables, in order to attempt to reveal a “factor structure”. The sample size has more observations than variables, and there are enough variables that may reveal underlying factors to create the possibility of their discovery.  In factor analysis, some degree of multicollinearity is desirable, because the objective is to determine which variables may have a common, possibly explanatory or classificatory, factor in common.

The Bartlett Test for Sphericity

The candidate variables were subjected to The Bartlett Test for Sphericity, a test for the overall significance of all the correlations within a correlation matrix.  A significant p-value in the Bartlett Test indicates that the variables may be fit for factor analysis.  The Chi Square score for the Bartlett Test for the variables chosen was 10155.95, with a p-value of 0.0.

Variable-Specific Measures of Intercorrelation

In addition to conducting the Bartlett Test and examining the correlation matrix of the variables, the measurements of sampling adequacy (MSA) for all the variables were calculated using the Kaiser-Meyer-Olkin (KMO) test.  The variables that were determined to have inadequate MSA (< 0.60), were dropped from the variables list one at a time, and the KMO test was re-evaluated in between each elimination to assess possible changes in the adequacy of the other variables.  The final list of variables to be included in the analysis and their individual and overall KMO / MSA scores are listed in the table below.

Since there are only three possible factors with Eigenvalues / latent roots greater than one, three factors will be retained. The Eigenvalues / latent roots of the three retained factors were: Factor 1, 4.23, Factor 2, 2.59, and Factor 3, 1.59.  Note that the magnitude of the Eigenvalues does not indicate the conceptual importance of the factors, only the amount of variance captured by them.

Variable Loadings Analysis

Rules of thumb for interpreting the practical significance of factor loadings are:

·       Factor loadings <+/- .10 can be considered equivalent to zero for a structure assessment.

·       Factor loadings between +/- .30 and +/- .40 meet the minimal level of interpretation.

·       Loadings +/- .50 are considered significant.

·       Loadings exceeding +/- .70 are considered indicative of “well-defined structure” (meet the goal).

These loadings are summarized as follows:

Factor One has the highest loadings for: Total Outpatient Visits (TOTAL_OP_VSTS), Number of CODES, Number of Procedures (PROCS) and Number of Providers (PRVDRS). 

It has low loadings on Chronic Condition Prevalence (CC_PREV), Chronic Condition Incidence (CC_INCD) and Elixhauser per User (ELIX_PU). It has very low or negative loadings for User Proportion (USR_PROP), End-Stage-Renal-Disease Prevalence (ESRD_PREV), # of Coverage Months (CVRG_MOS), # of Drug Coverage Months (RXCVRG_MOS), and # of Inpatient Days (IPDAYS). It has the highest relationship to DIVERSE_PROP at 0.126 and is associated with 0.243 of COST variation.

Factor Two  has the highest loadings on County User Proportion (CNTY_USR_PROP), Chronic Condition Prevalence (CC_PREV) and high loadings on End-Stage-Renal-Disease Prevalence (ESRD_PREV), # of Coverage Months (CVRG_MOS) and # of Prescription Coverage Months (RX_CVRG_MOS) as well as Chronic Condition Incidence (CC_INCD). 

It has very low or negative loadings on Total Outpatient Visits (TOT_OP_VSTS), # of CODES, # of Procedures (PROCS), # of Providers (PRVDRS), and Elixhauser per User (ELIX_PU). It is associated with 0.39 of COST variance.

Factor Three has the highest loadings on # of Inpatient Days (IPDAYS) and Elixhauser per User (ELIX_PU). It has moderately high loadings on End-Stage-Renal-Disease Prevalence (ESRD_PREV), # of CODES, Chronic Condition Prevalence (CC_PREV), and Chronic Condition Incidence (CC_INCD).

 It has low loadings on # of Procedures (PROCS) and very low or negative loadings on User Proportion (USR_PROP), # of Coverage Months (CVRG_MOS), # of Prescription Coverage Months (RXCVRG_MOS), # of Outpatient Visits (OP_VSTS), # of Providers (PRVDRS) and DIVERSE_PROP.

Communalities & Uniquenesses

The communalities are the amounts of these variables' variance explained by their loading on the factors.  The uniquenesses are the opposite of the communalities; they are variances associated with only a specific variable. Many times, demographic variables, or variables with many different causal factors will show up with high uniquenesses, suggesting their commonalities may be difficult to determine or – in the case of demographic variables – they simply “are what they are”.  Based on its uniqueness, 0.448 of cost variance is unaccounted for by the factor analysis.

Factor Variance

Sum of squared loadings for each factor (total factor variance): 2.75, 2.66, 1.87.

Proportional variance of each factor vs. the other factors (approximately equal between Factor one and Factor two): 0.196, 0.190, 0.133.

Cumulative variance of all factors (from left to right): 0.196, 0.386, 0.519.

So, the First Factor is the most significant, the Second Factor is also very significant, and the Third Factor is somewhat less significant in explaining the total variance in the variables. 

The three factors together account for approximately 52% of the total variance in the variable set.

Describing the Latent Factors

A basic assumption of factor analysis is that some underlying structure does exist in the set of selected variables. The presence of correlated variables and the subsequent definition of factors do not guarantee relevance, even if they meet the statistical requirements. Based on domain knowledge, the following labels have been proposed for the three factors extracted from the covariance among these variables, significant for the outcome of cost:

Factor one is our High Utilization Factor. 

It explains / groups counties by total visits, number of codes, number of procedures, and number of providers. It explains some of the cost for the higher-than-predicted counties, but a moderate amount. Can we call this "overutilization"? Maybe, but it might be an over-generalization to do so. There's utilization in counties dominant in other factors also, but because of their loadings we may be able to conclude that it is more likely that the need for those procedures and hospitalizations was higher.

Factor two is our Illness / Primary Care Factor. 

It explains / groups counties by sickness: ESRD prevalence, chronic conditions prevalence, and incidence, but not by severity of illness / Elixhauser score. It is associated with the proportion of users, the number of coverage months, and a moderate amount of cost. This looks like diagnosed people who are seeing a provider.

Factor three is our Hospitalization Factor. 

It explains / groups counties by the length of inpatient stays, has a moderately high association with ESRD prevalence (sick people who often need hospitalization), the number of codes per user, and a very high association with the Elixhauser score (again, sicker people). It has moderate association with chronic conditions prevalence and incidence, but low/mod association with procedures. It has the highest factor loading for cost (0.58). Regardless of reason, hospitalizations are expensive. It also has very low or negative associations for coverage months and # of providers. 

This outcome suggests the power of preventative care.  If people are covered and seeing the provider (Factor Two), we know that hospitalizations become less frequent. It also suggests there may be areas where there are few PCPs but there are still hospitals. In these areas (which might have low Health Factors or Outcomes Ranks) people sometimes do not have regular primary care but do go to the hospital once a crisis occurs.

Improvements in the Factor Analysis

Later, I revisited and improved this analysis, revealing a possible 4th factor.  The updated Jupyter Notebook can be found here: 

County-level Aggregated Medicare Data - New Factor Analysis.ipynb

Calculating and Evaluating the Dominant Factors of High-Cost Counties

Once the three factors were described, the value for each factor for each county was calculated, and the dominant factor determined through simple comparisons. 

Then, the top twenty highest cost counties were examined based on their factor values and dominance. Of these top twenty counties (with the sixty-seven even higher-cost outliers previously removed), seven out of twenty had Factor one dominance, eight out of twenty had Factor two dominance, and five had Factor three dominance. However, most of these counties had roughly equal dominance across two or even all three factors, and several had low or negative values for all the factors, suggesting that there are other things still accounting for cost that remain unmeasured – it is never as simple as just three factors!

As noted in the introduction, the Dartmouth Atlas Project, through their ongoing analysis of Medicare data, has noted the geographic variations and inequalities of health care in the United States, and exposed the ways in which these geographic variations are not simply related to variations in illness in different parts of the nation.  

The IOM’s Committee on Geographic Variation in Health Care Spending and Promotion of High-Value Care augmented this view by adding information about the commercially insured and showing that while practice patterns vary, there’s no obvious way to determine where they will be of higher quality based on geographic region.  

This analysis contributes to the conversation by characterizing the sources of variability in the variables related to Medicare cost and utilization, and by demonstrating the mix of these sources across counties in the United States with higher-than-expected Medicare outlays.  No two counties’ mix of Utilization, Illness and Hospitalization factors is exactly the same, and an additional 48% of the variance in spending remains unaccounted for. 

After performing risk adjustment with the RAM, some specific key findings from our analysis of utilization behaviors follow:

1) The higher-than-expected cost counties had greater racial diversity as compared to the lower-than-expected cost counties. This points to systemic variation in the US health care system with respect to race. These differences could be attributed to variations in cultural belief systems with respect to health, or the availability of resources to access quality health care in a timely manner.  The RWJF 2020 County Health Rankings key finding report (Givens et al., 2020) found a decade of data demonstrating that racial opportunity gaps persist. 

2) The higher-than-expected cost counties have more total visits per beneficiary, more providers per user and more procedures per user.  This finding reflects the differences in medical practice patterns between higher-than-expected cost and lower-than-expected cost counties. These findings are similar to those of a prospective cohort study conducted by Fischer et al. (2003) to evaluate whether regions with higher Medicare spending provided better care. They found that after adjusting for regional differences in illness and price, the higher spending regions received 60%  more care including more frequent physician visits, greater use of medical subspecialists and more frequent diagnostic tests and minor procedures. In addition, they found that the quality of care was no better in these higher spending regions and was worse for several preventive care measures. Their study sample included patients who were hospitalized for hip fracture, colorectal cancer, or acute myocardial infarction.

One explanation for higher medical costs when multiple providers are involved in healthcare delivery per user could be the lack of communication among providers leading to repeated testing/procedures. Front line healthcare delivery in the US is filled with care silos – moments where care is not coordinated amongst multiple providers and thus important information about the care delivered to each patient stays with the doctor and not the patient. These care linkage deficiencies (CLD) have been known to contribute to unnecessary hospitalizations. 

3) The average lengths of inpatient stays were higher in the higher-than-expected cost counties. One explanation for longer inpatient stays could be the occurrence of hospital-related adverse events. Fry et al. (2009) and Zhan et al. (2003) both used the excessive length of stay as an indicator of real, but unreported, medical errors.

Another important explanation of higher-than-expected cost could be the practice patterns of  providers in a particular geographic areas. In addition to the Dartmouth Atlas findings, Cutler et al. (2019), used "strategic" survey questions of physicians and patients to identify physician incentives and beliefs, as well as patient preferences and their respective contribution to variations in healthcare spending. They found that patient preferences contributed minimally, whereas physician treatment intensity combined with the frequency with which physicians recommend that their patients return for routine office visits explained 60% of the variance in spending. In this study, the physicians were characterized as either "cowboys" or "comforters". The former were those physicians who consistently and unambiguously recommended intensive care over and above the established clinical guidelines, and the latter were those who recommended palliative care for the severely ill.  Upon examining the factors associated with being a "cowboy" or a "comforter", they found that, to a large degree, physicians have differing views on how to treat the same patients. 

Copyright © 2021

Citation: Kapoor, D. & Rasku, K. (2021).  "Geographical Variations in the United States Health Care System: A Multivariate Analysis of Cost-Related Factors in Aggregated Medicare Claims Data";  UC Davis CPE.