Methods of risk adjustment
Stratification of pediatric & adult populations
Two types of primary care practice measure sets and profiles were generated: adult (ages 18 years and older) and pediatric (ages 1–17 years). The adult profiles included members with commercial, Medicaid (ages 18–64 years), or Medicare (ages 18 years and older) as primary payer. The pediatric profiles included members ages 1–17 years with commercial payers or Medicaid as primary.
While many practices treat both adult and pediatric populations, others treat primarily only adults or children. Because the pediatric and adult populations have very different health, utilization, and cost distributions, segregating profiles by adult and pediatric populations provides a more accurate look at practice differences. An alternative — basing practice profiles on physician specialty — would be problematic since attribution is at the practice level and practice groups may have included practitioners with different primary care specialties.
Newborn infants under the age of 1 year were not included since they (1) have high cost compared to the remainder of the pediatric population, (2) have a small number of outlier cases requiring neonatal intensive care, and (3) are often reported as bundled newborn claims by payers resulting in incomplete reporting of expenditures.
Treatment of outliers
The method used in this study capped outliers in expenditure and utilization at the 99th percentile of patients for each measure. Capping was done at the state-level for each major payer type (i.e., commercial, Medicaid, Medicare), and capped values were used for practice-level analysis. For the 2014 study population, the dollars truncated by capping represented 7% for the adult population and 13% for the pediatric population.
Adjustment for demographics & health status
Demographic and health status information determined from the APCD data formed the basis for the risk-adjustment methods used for the Blueprint Practice Profiles. These factors included age, gender, presence of a Blueprint-selected chronic condition, health status as measured by 3M Clinical Risk Groups (CRGs), and (for adult profiles) the occurrence of a maternity diagnosis.
Adjustments were made for the partial length of enrollment in payer insurance reported for some members during the calendar year. Average members — i.e., cumulative member months divided by 12 — were reported for each practice.
For the purposes of risk adjustment and to facilitate interpretation of results, member age was grouped. Due to the potential for interaction effects of age and gender, the full model used age and gender groupings (e.g., males aged 18–34 years, females aged 18–34, etc.).
There are several systems for measuring health status being used in the United States, each with its different point of emphasis, yet no single system has emerged as the “gold standard.” In this study, the primary method of adjustment for each member’s health status was based on the application of 3M Clinical Risk Groups (CRGs) to the APCD data. CRGs used to measure health status are applicable to all ages, are updated regularly, are designed for use with claims data, are transparent in documentation, perform as well as other available systems, and represent a practical solution to meet the needs of the Blueprint project [
18]. The grouper classifies each member into a hierarchy of 1080 distinct clinical groups and nine major clinical CRG statuses based on the diagnoses reported. Due to small numbers, and to create an efficient model that was easily understood, these nine categories were further combined during the risk model development process into (1) Healthy (reference group), (2) Acute or Minor Chronic (e.g., acute ear, nose, or throat condition or minor join pain), (3) Moderate Chronic (e.g., diabetes or moderate chronic joint pain), (4) Significant Chronic (e.g., diabetes with other comorbid conditions such as congestive heart failure (CHF) or chronic obstructive pulmonary disease (COPD)), and (5) Cancer or Catastrophic (e.g., malignant breast cancer, HIV, cystic fibrosis, muscular dystrophy, quadriplegia).
The Blueprint program also targeted select chronic conditions: asthma, chronic obstructive pulmonary disorder (COPD), congestive heart failure (CHF), depression, diabetes, hypertension, ischemic heart disease, and attention deficit disorder (pediatric model only). A “chronic” (0/1) variable was created if a member was identified with any of these conditions. Since CRGs do not include pregnancy and child birth in clinical classification, a “maternity” (0/1) variable was created for members with pregnancy or delivery claims during the year.
Adjustment for practice’s Medicaid & Medicare population
The primary care practice profiles combine the populations from three different payer types (or “payers”) — commercial, Medicaid, and Medicare — that have significant differences in demographics, socioeconomic status, health status, provider reimbursement structures, and services covered and used. For the full model, Medicaid members were identified in the indicator variable as: Commercial = 0, Medicare = 0, Medicaid = 1.
Another distinguishing attribute of the Medicaid data was the inclusion of members who received “special Medicaid services” (SMS) uncommon in the commercial and Medicare populations. Members receiving SMS may have had a level of disability not adjusted for through the CRGs. Examples of SMS include members receiving day treatment, residential treatment, case management services, and special school services covered by the Department of Education. These types of services can contribute significantly to a member’s total expenditures. After evaluation of statistical distributions for these services, members with SMS expenditures over $500 during the 12-month study period were identified by a binomial (0/1) variable.
During model development, it was determined that a practice’s percentage of total members covered by Medicaid was a statistically significant predictor of higher total expenditures. Practices in Vermont varied considerably in their percentage of members who were covered by Medicaid. Therefore, the full risk-adjustment model included a practice’s percent Medicaid for each Medicaid enrollee in the practice.
Given widely observed healthcare disparities, women covered by Medicaid may be at higher risk for poor maternity and neonatal outcomes than women covered by commercial plans [
19‐
22]. To account for these differences, an interaction term was added between Medicaid and maternity.
As was done for Medicaid, the full risk-adjustment model identified Medicare-eligible beneficiaries through the indicator variable: Commercial = 0, Medicaid = 0, Medicare = 1. The model also included a variable for “practice’s percent Medicare” for members contributing to the practice’s percent Medicare. Using Medicare-specific eligibility elements, “disability” (0/1) and “end-stage renal disease” (0/1) variables also were created. Pediatric members covered by Medicare were excluded from the pediatric profiling due to small numbers.
Full model & the computation of risk-adjusted rates
The risk-adjustment methods used for reporting used SAS (Version 9.3) regression methods (SAS GENMOD procedure). The full model included age/gender stratification groups, Blueprint-selected chronic conditions, CRG classification, maternity status, and the additional Medicaid and Medicare adjustors. Adjusted rates were produced by summing the differences between each member’s actual value and the member’s predicted measurement from the model. Rates were weighted for partial lengths of enrollment. Detailed descriptions of the model’s computation of risk-adjusted rates and 95% confidence intervals for the adult and pediatric populations are provided below.
To calculate the adjusted rate, adjusted values were computed for each member by adding model residuals (e) to the population grand mean \( \left(\overline{y}\right) \). To report the overall adjusted rate for each practice, the mean of the adjusted values for the members in each practice \( \left({\overline{y}}_{\mathrm{practice}}\right) \), in each HSA (\( \overline{y} \)
hsa), and statewide (\( \overline{y} \)
statewide) were computed. The following equations represent the models for the adult and pediatric practice profiles.
Adult model
$$ \begin{array}{l}y = \alpha + \left(F\_ AGE1834\right){\beta}_1 + \left(F\_ AGE3544\right){\beta}_2 + \left(F\_ AGE4554\right){\beta}_3 + \left(F\_ AGE5564\right){\beta}_4 + \\ {}\left(F\_ AGE6574\right){\beta}_5+\left(F\_ AGE7584\right){\beta}_6+\left(F\_ AGE85 PLUS\right){\beta}_7 + \left(M\_ AGE3544\right){\beta}_8 + \\ {}\left(M\_ AGE4554\right){\beta}_9 + \left(M\_ AGE5564\right){\beta}_{10} + \left(M\_ AGE6574\right){\beta}_{11}+\left(M\_ AGE7584\right){\beta}_{12}+\\ {}\left(M\_ AGE85 PLUS\right){\beta}_{13} + (MEDICAID){\beta}_{14} + (MEDICARE){\beta}_{15} + \left( DUAL\ ELIGIBILITY\right){\beta}_{16}+\\ {}\ (SMS){\beta}_{17} + \left( PRACTICE\_ PERCENT\_ MEDI\right){\beta}_{18}+\left( PRACTICE\_ PERCENT\_ MCARE\right){\beta}_{19} + \\ {}(DISABLED){\beta}_{20} + (ESRD){\beta}_{21}+(CHRONIC){\beta}_{22} + \left(CRG\_ ACUTE\_ MINOR\right){\beta}_{23} + \\ {}\left(CRG\_ CHRONIC\right){\beta}_{24} + \left(CRG\_ SIGNIFICANT\_ CHRONIC\right){\beta}_{25} + \\ {}\left(CRG\_ CANCER\_ CATASTROPHIC\right){\beta}_{26} + (MATERNITY){\beta}_{27} + \left( MATERNITY*\ MEDICAID\right){\beta}_{28} + \varepsilon \end{array} $$
Pediatric model
$$ \begin{array}{l}y = \alpha + \left(F\_ AGE0104\right){\beta}_1 + \left(M\_ AGE0511\right){\beta}_2 + \left(F\_ AGE0511\right){\beta}_3 + \left(F\_ AGE1217\right){\beta}_4 + \\ {}\left(M\_ AGE1217\right){\beta}_5 + (MEDICAID){\beta}_6 + (SMS){\beta}_7 + \left( PRACTICE\_ PERCENT\_ MEDI\right){\beta}_8+\\ {}\left( CHRONIC\_ PED\right){\beta}_9 + \left(CRG\_ ACUTE\_ MINOR\right){\beta}_{10} + \left(CRG\_ CHRONIC\right){\beta}_{11} + \\ {}\left(CRG\_ SIGNIFICANT\_ CHRONIC\right){\beta}_{12} + \left(CRG\_ CANCER\_ CATASTROPHIC\right){\beta}_{13} + \varepsilon \end{array} $$
$$ \overline{y} = \left(\frac{{\displaystyle \sum }{y}_i}{MMA}\right) $$
$$ {y}_{\mathrm{adj}}=\overline{y}+e $$
$$ {\overline{y}}_{\mathrm{practice}} = \left(\frac{{\displaystyle \sum }{y}_{ad{j}_i}}{{\displaystyle \sum }MM{A}_i}\right)\;\mathrm{f}\mathrm{o}\mathrm{r}\ \mathrm{each}\ \mathrm{practice} $$
$$ {\overline{y}}_{\mathrm{hsa}} = \left(\frac{{\displaystyle \sum }{y}_{ad{j}_i}}{{\displaystyle \sum }MM{A}_i}\right)\;\mathrm{f}\mathrm{o}\mathrm{r}\ \mathrm{the}\ \mathrm{practices}\ \mathrm{in}\ \mathrm{each}\ \mathrm{H}\mathrm{S}\mathrm{A} $$
$$ {\overline{y}}_{\mathrm{statewide}} = \left(\frac{{\displaystyle \sum }{y}_{ad{j}_i}}{{\displaystyle \sum }MM{A}_i}\right)\;\mathrm{f}\mathrm{o}\mathrm{r}\ \mathrm{all}\ \mathrm{members}\ \left(\mathrm{equals}\ \mathrm{the}\ \mathrm{grand}\ \mathrm{mean}\right) $$
where:
-
α is the intercept
-
ε is the error term
-
ŷ is the predicted value from the regression model for each member
-
e is the residual
-
MMA is the average enrollment for each participant (i.e., the cumulative member months of enrollment during the year divided by 12)
-
Subscript
i
indicates a value for an individual member
For practice-level reporting, 95% confidence intervals were generated based on the standard error of the mean. For utilization measures (e.g., inpatient hospitalizations), the Poisson distribution was utilized. The outlier capping and risk-adjustment models were run separately for each individual expenditure and utilization measure reported to practices.
Outcome measures
Blueprint practice profiles reported included 27 expenditure, 15 utilization, 10 HEDIS, and 5 additional National Quality Forum (NQF) endorsed measures. This study focused on four measures for the adult and pediatric practice reporting: (1) total expenditures per capita, (2) total expenditures per capita excluding SMS, (3) total Resource Use Index excluding SMS, and (4) a quality composite measure constructed from HEDIS measures (described below). Total expenditures were based on the allowed amount on claims, which includes both the plan payments and the member’s out-of-pocket payments (i.e., deductible, coinsurance, and copayment). For enhanced expenditure parity across payers, an additional total expenditures measure — this time excluding SMS costs — was also examined. Within the Medicaid population, SMS represented 26% of adult Medicaid population expenditures and 61% of pediatric Medicaid population expenditures.
Because pricing and reimbursement can vary, total expenditure measures do not provide a measure of actual consumption of resources (i.e., the actual frequency and intensity of all services used). Therefore, Blueprint used a measure of overall resource use: the total Resource Use Index (RUI), which is based on the NQF-endorsed measure (NQF #1598) Total Care Relative Resource Values (TCRRVs). The measure was implemented by applying HealthPartners’ Total Cost of Care (TCOC) software to the claims data [
23]. The resulting standardized relative RUI was included to measure aggregate resource consumption across all components of care (i.e., inpatient, outpatient facility, professional, and pharmacy). The RUI for each practice was calculated by dividing the adjusted TCRRV rate by the statewide TCRRV rate. Since the TCRRVs do not include many of the special Medicaid services, these services were excluded from the RUI.
Effective and preventive care measures were produced by applying NCQA HEDIS specifications to the APCD data. A composite adult measure was constructed at the practice level using the practice average for three HEDIS measures: comprehensive diabetes care hemoglobin A1c testing, breast cancer screening, and imaging studies for low back pain. A composite pediatric measure was constructed using the practice average for three HEDIS measures: well-child visits, appropriate testing for pharyngitis, and appropriate treatment for upper respiratory infection. These measures were a subset of six adult and four pediatric HEDIS measures available in the profile data and were selected based on sufficient sample size at the practice level for statistical reliability and to ensure that no focus area was overweighted (e.g., limiting cancer screening measures to a single measure). Rates for HEDIS measures were not risk adjusted, and NCQA provides no recommendations for risk adjustment.
Comparing risk-adjustment models
The following risk-adjustment models were compared:
-
No adjustment
-
Age and gender (no interaction)
-
Age and gender (no interaction) and CRGs
-
Full model (includes age, gender, CRGs, maternity, and payer-specific variables)
At the patient level, the percentage of variance explained by each model was evaluated using the regression r2. The relative difference between models was evaluated using the log-likelihood ratio test. At the practice level, each model’s results were evaluated using the coefficient of variation (CV).