Including Age-Sex and Intercept Terms and Explanatory Power of MassHealth SDH Total Cost of Care Risk Adjustment Models 1, 2, and 3 in CY2017

This cross-sectional study examines administrative data from the Massachusetts Medicaid program MassHealth and develops models to account for caring for members with greater complexity and allocate payments more equitably.


eMethods: Modeling Approach
Our modeling approach for social determinants of health (SDH) Model 3 was to begin with Model 2, which was currently in use by MassHealth, and to then examine incremental changes motivated by good modeling practices, newly available data and variables, and policy considerations.Simply refitting Model 2, developed on 2015 data, to the more recent 2017 data produced a small increase (0.3%) in explanatory power.This refitted version of Model 2 (R 2 =51.5%) provided a baseline against which the following subsequent changes to the model could be compared.
As summarized in the text of the manuscript and in Supplemental Table 1, most of the gains in explanatory power associated with the transition from Model 2 to Model 3 were obtained with the first few modeling revisions including substituting an updated DxCG medical morbidity summary score, adding a pharmacy based medical morbidity score (the RxCG), and adding terms for serious mental illness (SMI) and substance use disorder (SUD).
Subsequent modeling revisions were made primarily to address concerns regarding the underpricing of specific subgroups of members that were identified by MassHealth as potentially vulnerable, costly, and distributed unevenly across MassHealth ACOs, therein presenting substantial concerns regarding equity if mispriced.These subgroups included youth with medical complexity, members with behavioral health conditions including comorbid SMI and SUD, and members with comorbid medical, behavioral and social risk factors (e.g., homelessness, unstable housing, high neighborhood stress scores).
In Model 3, in line with MassHealth's commitment to payments that support delivery system reforms integrating physical and behavioral health care, we examined more deeply the various relationships among behavioral health conditions and costs.We worried that making distinctions among finer categories of SMI (beyond those differences captured by the summary risk scores) would be overly susceptible to coding manipulation ("gaming").Thus, we focused on SUD, finding that opioid use disorders (OUD), but not alcohol use disorders, were associated with increased costs.We also sought to improve accuracy for high acuity members, especially youth.Although all our models are constrained to accurately model mean cost within each age-sex category, we suspected that overpayments for healthy children were being balanced against underpayments for sick ones.This concern was acute because of MassHealth's recent launch of 17 Medicaid ACOs with risk-based contracts, and medically complex youth were unevenly distributed across ACOs.Therefore, we explored interaction terms between pediatric age groups and the DxCG risk score.
We were concerned that individuals with both medical and social risk factors used more resources than those factors individually account for, focusing on interactions of social risk factors with medical morbidity, specifically, the DxCG score.For example, we expect higher costs for a person with mental health issues and no stable place to live, but even more excess costs in the presence of multiple medical problems.Similarly, neighborhood disadvantage can make chronic conditions more difficult to manage, for example, when pollution and low-quality housing exacerbate asthma in children, making hospitalizations more likely.Thus, in Model 3, payments for homelessness or unstable housing, when there is also a behavioral health problem, are indexed to (that is, interacted with) the DxCG score, as are payments associated with neighborhood stress.These interactions generate larger payments for members with both social risk factors and high medical acuity than for these problems individually.
Incremental model adjustments were mostly guided by examination of observed-to-expected ratios (i.e., the ratio of observed costs to model expected costs) for subgroups defined by demographic (age, sex, race/ethnicity, rurality), medical (e.g., percentiles of the DxCG score), behavioral (i.e., SMI, SUD, and component diagnoses), and social characteristics.Other changes were made in pursuit of simplicity and parsimony, and in the case of the age-sex category indicators, the initial decision to remove them was reversed in light of subsequent evidence of mispricing of certain age-sex categories without the indicators.
Policy considerations also informed modeling decisions.Members with dual SMI-SUD comprised only 3% of the population, while those with any behavioral health condition made up nearly 15% of the population.Recognizing the challenge of caring for individuals with social and medical risk factors is made more difficult by the presence of any behavioral health condition, and since differences in performance were not meaningfully different between models, we chose the model with a three-way interaction of housing problems and DxCG score with any behavioral health condition rather than only those with dual SMI-SUD.Finally, due to the real-world problems with retaining negative coefficients in payment models, whereby health systems would be incentivized to under-code an individual's characteristics, model grooming was performed to remove negative coefficients.

Variable Description and Codes
Other Substance Use Disorder d BH = behavioral health SMI = serious mental illness, SUD = substance use disorder, OUD = opioid use disorder, AUD = alcohol use disorder.e The DxCG score for Models 1 and 2 is the DxCG v4.2 Cotiviti, Inc., Model 312 diagnosis-based concurrent risk score normalized to have mean = 1 in the full population and the DxCG score for Model 3 is the DxCG v5.3 (Cotiviti, Inc., Model 88), normalized to have mean = 1 in the full population.
Disability StatusDisability status is a hierarchy.At the top is enrollment as client of the Department of Mental Health (DMH); else, as a client of the Department of Developmental Services (DDS); else, as entitled to Medicaid due to disability (all other disabled).

Coefficients Including Age-Sex and Intercept Terms and Explanatory Power of MassHealth SDH Total Cost of Care Risk Adjustment Models 1, 2, and 3 in CY2017
Notes: Analyses are weighted (WGT = fraction of the year during which the member was eligible) and include only those enrolled for at least 183 days in 2017.Cost is calculated using "rate book" prices from the actuaries and excludes long-term services and supports (LTSS) spending.Cost is top-coded at $200,000 and annualized = minimum of dollars spent/WGT or $200k.NA = not applicable.a SMI = serious mental illness b OUD = opioid use disorder c RxCG = pharmacy-based concurrent risk score (Cotiviti, Inc., Model 86) normalized to have mean = 1 in the MassHealth population.d The DxCG score for Models 1 and 2 is the DxCG v4.2 Cotiviti, Inc., Model 312 diagnosis-based concurrent risk score normalized to have mean = 1 in the full population and the DxCG score for Model 3 is the DxCG v5.3 (Cotiviti, Inc., Model 88), normalized to have mean = 1 in the full population. e

Observed-to-Expected Ratios for MassHealth SDH Total Cost of Care Risk Adjustment Models 1, 2, and 3 in 2016 a
The mean cost in CY16 was $4942 (top-coded at 200,000 and annualized).Expected values were made from SDH Models 1-3.Those expectations were top and bottom coded ($60-200,000) and then standardized to the CY16 mean costs to calculate the observed to expected (O:E) ratios.b Disability status indicates MassHealth eligibility as client of the Department of Mental Health (DMH), Department of Developmental Services (DDS), or entitled to Medicaid due to disability ("All other disabled").c NSS = Neighborhood Stress Score, normalized to the MassHealth population; NSS+ equals the Neighborhood Stress Score when it is positive and 0 otherwise.
© 2023 Alcusky MJ et al.JAMA Network Open.a