Susannah Higgins, MS; Ravi Chawla, MBA; Christine Colombo, MBA; Richard Snyder, MD; and Somesh Nigam, PhD
DxCG risk score refers to the concurrent medical risk score calculated using the commercial risk adjustment model developed by Verisk Health DxCG Risk Solutions, version 3.1 (Verisk Health, Cary, North Carolina).15
The DxCG models use linear, additive formulas obtained from ordinary least squares (OLS) regressions to combine expenses associated with clinical groupings (Condition Categories and Rx- Groups) and demographic factors (age, gender) to generate predictions. Hierarchies are imposed on Condition Categories before regression models are run to decrease their sensitivity to variations in coding and to strengthen the predictive power of the models across a wide range of benefit designs and pricing arrangements.15
The concurrent risk score in this analysis used each individual’s age, gender, and claims information from all medical encounters and enrollment data as inputs to predict annual total medical resource use for the patient.
Patients evaluated were enrolled in 17 PCMH practices, 15 of which were participating in the Pennsylvania Chronic Care Initiative pilot (cases), or 1 of 103 non-PCMH comparison practices in PA (controls). The high-risk patients were slightly more concentrated, appearing in only 14 PCMH practices (12 of which were part of the pilot) and 69 non-PCMH practices. Costs were assessed as per member per month (PMPM) costs and utilization was measured as encounters per 1000 patients per year.Inclusion/Exclusion
Criteria In order to qualify for inclusion, each member must have been enrolled with the same primary care physician for at least 3 months and have been associated consistently with a PCMH or non-PCMH practice for 6 months or more during each year of the study period, including baseline (2008). Only nonpediatric practices located in Pennsylvania were included in the analysis. Pediatric practices were excluded to reduce the variability in the age of members in the study and minimize the impact of any potential unobserved differences between pediatric practices and nonpediatric practices. Patients with end-stage renal disease or extremely high medical costs (>
$100,000 per year) were excluded to limit the impact of extreme outliers on the cost analysis.
The pool of eligible controls was limited to later adopters of the PCMH model in order to reduce potential practice self-selection issues. Patients enrolled in control practices that received NCQA recognition in 2011 were included in analyses for the first 2 program years, but were removed from the final year’s analysis to avoid contaminating the non-PCMH group with late adopters, along with their matched cases. This reduced the numbers included in program year 3 to 1141 cases and 1141 controls in the overall analysis, and 109 cases and 131 controls in the highrisk subgroup. While the guidelines remained largely similar from 2008 to 2011, some changes were made to the NCQA requirements, and a crosswalk table detailing the differences is available from the NCQA.16
An attrition diagram detailing all exclusion steps appears in Appendix A
Propensity score matching was used to select a sample of controls which were similar to the PCMH cases with respect to both practice- and patient-level demographics and characteristics. The propensity score included practice size, practice location, age, gender, DxCG risk score, chronic conditions (including asthma, diabetes, congestive heart failure, chronic obstructive pulmonary disease, and coronary artery disease), and median income for member’s zip code of residence. Matching successfully controlled for the significant differences at baseline between the group of cases and pool of potential controls, both for the overall comparison and among the cohort of highest risk patients (Table 2A
Differences in PMPM cost and utilization per 1000 patients between PCMH and non-PCMH practices for the 3 follow-up years were compared using regression analysis. This method provides an estimate of the differences between treatment and control groups before and after treatment. Since the members included in this study are the same in each time period (ie, panel data), the difference-in-differences model can be simplified and is more statistically powerful. The difference between baseline and each time period can be modeled as follows (separate model for each time period difference):
Y i,1 – Y i,0 = δ + βX i,1 + ε I
where Y i,1
– Y i,0
is the difference between the repeated outcome measure for each observation, δ is the effect of time on all units, X i,1
is the treatment indicator, and β is the treatment effect.17
Cost and utilization regressions controlled for concurrent risk in the comparison year and having a chronic condition in the baseline year, factors that could influence both cost and utilization in the comparison year. For example, when comparing 2010 with baseline, the 2010 risk score was added to the model to adjust for any new diagnoses a member might have acquired in 2010. Presence of a chronic condition at baseline was used as a covariate because chronic members often continue to incur costs, or incur even higher costs, over time. In order to test for the treatment effect for just the high-risk members, a high-risk indicator was introduced as covariate (ie, as a main effect and interaction term with the treatment indicator) into the model using the whole population. Additionally, a random intercept term was used in the model to account for any practice-level effects.RESULTS
PDF is available on the last page.