im symposium: vbcm doug thompson phd tom cavin asa, maaa august 2012

16
IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

Upload: tyler-weber

Post on 26-Mar-2015

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

IM Symposium: VBCM

Doug Thompson PhD

Tom Cavin ASA, MAAA

August 2012

Page 2: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• Delivering quality timely care by aligning member and provider incentives where focus is shifted from ‘quantity’ of service to ‘value’ of service

• Two models considered here– Intensive and Extended Medical Home

– Bridges to Excellence Physician Recognition

Value Based Care Models

Page 3: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

Patient Centered Medical Home

• Contractual arrangements with providers, designed to incent decreases in cost and increases in quality via improved coordination of care

• The models vary by the breadth of the managed member population

Page 4: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

BRIDGES TO EXCELLENCE

• Requires no contracting, only recognition through third party benchmarking of member biometric data

• Incentives for physicians of $100 per member per annum

• Specialized by condition ‘module’– Diabetes Care Recognition

– Cardiac Care Recognition

Bridges to Excellence

Page 5: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• EMH/IMH – IMH focuses on member identification and risk stratification techniques

• BTE– Focuses on finding a suitable group for a matched control

Analysis centered on applying a generalized linear model and setting classification level contingent on key explanatory variables

Considerations for propensity scoring and applicability in analysis

Two Models Two Approaches

Page 6: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

Population Identification: Treatment and Control

• Key determinants in measuring value center on finding the appropriate comparison population

– The number of parameters necessary to define the study and control populations are inversely proportional to the sample size

– Some input from outside sources such as the Care Continuum Alliance’s: Outcome Guidelines for certain exclusionary criteria

• Consistent definition of your population– Condition identification

– Age bounds

– Geography

– Benefit Design

• Exclusionary Criteria– High severity, low frequency exclusions

– Discontinuous Enrollment

Page 7: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• Is it appropriate in this setting? – Is your sample treatment a sufficiently randomized grouping or

concentrated in geography, morbidity or some other categorical variable?

– Are there unobserved covariates of ‘x’ not represented in your model?

• Propensity Scoring– A function created to describe the probability (between 0 and 1) of the

considered outcome

– Intuitively basic categories such as age/gender, similar geographies, and morbidities would eliminate much differentiation between groups

– Goal with propensity scoring is to isolate the non-intuitive relationships existing in the observed ‘x’ values and allow for a method by which each treatment member is matched to their nearest control counterpart

– Assists in eliminating the relative representation bias of ‘x’ dependencies for the control and treatment group

Population Matching

Page 8: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• The explanatory variables you choose should begin to ‘carve out’ their effects on the dependent variable of choice

– Looking for explanatory variables with low correlation– Avoid over-fitting your model

• Explanatory variables considered in BTE– Age/Gender– Geography– DCG Prospective Risk Score– BTE Recognition Flag

• Sampling for validity and fit– Stratified sampling applied for fitting and testing model – Measuring actual to residuals at a ‘class’ level

• Measuring Influence– Cook’s Distance– Fisher Scores

Regression Considerations

Page 9: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• Your model data set is larger and you are interested in bounding your model results without assuming a normal distribution

– Sample with replacement to match your original population counts

– Stratify on your treatment/control indicator

– Re-run regression model and predictor outputs for each bootstrapping sample created

– Percentile Bootstrap, take the desired percentiles (5th and 95th, for a 90% confidence intervals) as the bounds for your predictor variable

Bootstrapping to Find Confidence Intervals

Page 10: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• This is a key component of the IMH model (less central to EMH)– A goal of IMH is to focus management on the members who are expected

to benefit most (as opposed to a broader population)

Member Identification

Page 11: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• In the initial “tactical” phase, HCSC’s plans had license to use whatever member identification algorithm they deemed appropriate

• Approaches:1. Select members with high expected future costs based on Verisk models

2. Select members with high likelihood of benefitting from medical management based on internal algorithm

3. Select members with high prior-year costs

HCSC’s Approach to Member Identification: Tactical

Page 12: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• Next (Strategic) phase: Use a single, Enterprise-wide, optimized member identification approach

1. Predict which members are most likely to have high costs next year that could be impacted by PCPs

2. Consider a wide variety of leading indicators (“predictors”) of future costs

3. Use statistical modeling techniques to determine 1) what information is useful for prediction, 2) how to weight different pieces of information

HCSC’s Approach to Member Identification: Strategic

Page 13: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• Regression-based algorithm

• Target that the algorithm predicts: Next-year member spend, after excluding categories not expected to be impactible by PCPs (based on expert clinical judgment)

Identification Algorithm

Page 14: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• Inputs considered include:– Verisk models (e.g., 18, 26, 56, 51, 71)

– HCSC’s proprietary algorithms for selecting members for medical management programs

– Clinical Intelligence Rules

– Cost (prior year allowed amount)

– Selected utilization measures (admits, ER, selected procedures and Dxs)

– Rx data

– Behavioral health key measurements

Identification Algorithm Inputs

Page 15: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• This is similar to other prospective risk prediction models (e.g., CMS-HCC, CDPS, Verisk), except:

1. It has a different target (next-year PCP-impactible spend vs. total spend)

2. It considers a wider variety on inputs, enabling greater prediction accuracy

3. It is tailored to HCSC’s own member population

Identification Algorithm vs. Other Risk Prediction Models

Page 16: IM Symposium: VBCM Doug Thompson PhD Tom Cavin ASA, MAAA August 2012

CONFIDENTIAL AND PROPRIETARY; NOT FOR DISTRIBUTION; PROPERTY OF HEALTH CARE SERVICE CORPORATION

• HCSC has implemented several subtypes of Value Based Care Models (VBCMs), and now is in the process of evaluating and improving them

• HCSC has been using advanced analytical techniques to optimize its VBCM programs, as illustrated in this presentation

– Advanced analytics are essential for accurate program evaluation and optimal member identification

Key Takeaways