Consumer Behavior: Factors Affecting Member Attrition and Retention

Consumer Behavior: factors affecting member attrition and retention March 19, 2014 Prepared for:

Partners Summit, Las Vegas

Discussion objectives

•  Growing importance of consumer behavior and decision making in Healthcare

•  Discuss new approaches to identifying consumer trends –  Using more expansive data –  Applying new analytics approaches like machine learning

•  Review a case study –  Failure to recertify in 3 state study –  Engagement acceptance

2 COPYRIGHT © 2014 PREDILYTICS, INC. ALL RIGHTS RESERVED

Privacy and security of personal information is first and foremost Analytic insights must benefit the individual, governed by code of conduct and privacy

and security regulations

Computer science and big data Hype or a new way of business. . .


Consumer and Boomer revolution Impact on Healthcare Delivery

•  New generation of health care users entering the system, 77 Million Baby Boomers –  Transform industries as they emerge and

engage –  New behavior and purchasing patterns

•  Government policy shaping future of healthcare

•  Financial and funding constraints


•  45% annual growth in consumer and healthcare data

•  Explosion of healthcare mobility and telemetry solutions

•  95%1 of the “data wake” we all leave annually is not in the healthcare system

SOURCE: IDC; US Bureau of Labor Statistics; McKinsey Global Institute analysis, May 2011 Big data: The next frontier for innovation, competition, and productivity


Consumerization and realization of healthcare Impact on information

The ideas that drive new analytic approaches. . .

•  Use all available data to improve population and individual health –  Individual behavior is best predicted by socio-

economic and lifestyle characteristics and consumer activities, not typically found in EMR and Claims Data

•  Machine learning and advance computer science are required to convert massive amounts of data into actionable insights, by optimizing identification of targeted events at the actionable cohort

•  Identify individuals, predict engagement and deploy interventions with highest probability of success

•  Focus analytics efforts on the critical business and quality issues that drive organizational performance

6

Performance

Big Data

Advanced Analytics

Speed

Efficiency

Business Insights

Consumer Engagement

Results

COPYRIGHT © 2014 PREDILYTICS, INC. ALL RIGHTS RESERVED

Analytic solutions framework

7

Descriptive

Diagnostic

Predictive

Prescriptive

Hindsight Insight Foresight

Generates insight from big data to:

q  Improve quality and coordination of care q  Identify risk and asses opportunity q  Evaluate program impact


Big data approach How does it work and why is it different?

•  Big Data comes in the form of clinical, administrative claims, operating, demographic, workflow, purchasing, provider and consumer behaviors, etc. Examples include;

• Electronic Medical Record(e.g. Clinical values, notes) • Monitoring devices (e.g. wellness trackers, biometrics,

telemetry) • Consumer engagement (e.g. voting, financial, census,

Facebook, smartphones, portal/website utilization)

Big Data is the essence of collecting and storing data, both structured and unstructured, from as many different sources

as are readily available


Illustrative external data sources Public, Consumer, Financial, Social Media

Public Healthcare •  Medicare, Medicaid •  Population Stats •  Healthcare Providers, Cost, Quality •  AHRQ, NIH, CDC •  Health Outcomes

Consumer •  Consumer Behavior / Purchasing •  Ethnicity •  Social Security / Death Records •  Voter Registration •  Legal / Regulatory

Financial •  Consumer spending •  Credit risk •  Public records •  Real estate indicators

Social Media •  Facebook Activity •  Foursquare Check-in •  Twitter Activity •  Google Services, ETC.


Analysis approach and process

10

Customized files and reports with actionable insights •  Support operations •  Support business planning •  Reporting

Create predictive models and run client specific cohort(s) to generate insights

Predilytics supports implementation of analytic insights

Consumer Data

Client & Private Data


Background on Machine Learning How does it work and why is it different?

•  Predictive patterns in the data are discovered and retained •  The software builds on previous learnings and highly predictive equations evolve •  Genetic Algorithms (GAs) are a form of machine learning that are highly effective in

spotting subtle patterns in data sets. GA modeling technology and the output are transparent and more actionable

Software evaluates data and combinations of data sets millions of times

Machine learning is capable of exploring more data, faster and more thoroughly than traditional statistical techniques

•  Traditional modeling relies on statistical analyses of data, in particular various forms of regression, which carry with it certain limitations that are not found in iterative – based learning models

•  The results are more accurate predictive models


Machine learning is optimized for ‘Big Data’ predictive analytics


•  Linear Regression •  Logistic Regression • Time Series • Survival Analysis • Segmentation • Data Valuation • Variable Reduction

Machine Learning Optimize Prediction of X Start with “Random Walks” Learns Quickly & Transparently

Automation saves analyst time for more value-added tasks

Structure Predictive Modeling Task

X = f (A,B,C | D,E) + e

GA Enhances:

• Descriptive Summary

• Train / Test Samples • Univariate Graphs • Variable

Transformation • Missing Data

• Candidate Model Development

•  Lift Chart / ROC Curve

• Scoring Code

GA Automatic Features

Traditional Analytics

Genetic Algorithms (GA)

13

125 models per generation in 10 seconds

10,000 generations performed

1.25 Million equations evaluated with learning past to next generation

Low

Fitness Accuracy Scale

High

Model 7 Model 8 Model 9 Model 10 Model 11 Model 12 Generation Two

Model 13 Model 14 Model 15 Model 16 Model 17 Model 18 Generation (n)

Model (n)

Generation One Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 3 Model 4 Model 5 Model 6


The genetic algorithm advantage

•  Superior accuracy through the evaluation of far more data attributes and combinations of data attributes (often 15% to 20% improvement vs. traditional statistics approaches) o  Changing the economics of analytics – isolates the actionable

segment for intervention

•  Substantially improves the speed and segmentation of models: o  Decreasing modeling turnaround time o  Allowing for a proliferation of predictive models… breaks the analytic

bottleneck

•  Optimizes identification of targeted events at the actionable portion of the distribution, therefore optimizing the models predictive factors for the targeted event vs. trying to explain errors of the whole distribution

•  Clear, understandable results (No Black Box!)


Case Example

23

Overview

3 State study of Medicaid Recertification Identify health plan members likely to: •  Lose Medicaid eligibility by not recertifying (e.g. Dual Eligibles)

–  Identify those who fail to recertify, but are still eligible for Medicaid Optimizing these goals provides enhanced business performance •  Improve intervention targeting to increase reimbursement and drive

increased value for Altegra’s customers •  Improve recertification rates, reach and engagement rates and member

retention

16


Data sources

17

Altegra – derived time series data, member recertification and disenrollment, date of birth & age, race, gender

Predilytics-household level demographics including measures of affluence, household composition, length of residence, age, ethnicity, gender of head of household, home values, financial stress predictors (from unemployment stats)

US Census – zip code level data including distributions related to affluence, heritage, race, age of household members, languages spoken, educational achievements, employment, and population density, gender mix, veterans, disabilities, mobility


Analysis cohort

18

Popula'on

Total Members

Members who were enrolled as of August 2012, Medicaid cer:fied, and with ac:ve plans across 3 states (Georgia, Florida, Texas) 78,707

Number of Unique Members in Household Data 13,686729

Successful Match to Household Data 51,170

Match Rate 65%

Members who failed to recer2fy between September 2012 and August 2013 19,538

Recer2fica2on failure rate (Failed recer2fica2on members / total enrolled members as of August 2012) 38%

* An active plan was defined as any plan with members enrolled in September of 2013 Analysis cohort


0 25 50 75

100 125 150 175 200 225 250

1 2 3 4 5

210

133

78

52

27

Consolidated Failure to Recertify Model Lift

19

Model performance

Average

Members Projected

Rate of Failed Recertification

All 38%

Top 10% 87%

Top 20% 80%

Bottom 20% 10%

Rates indicates how likely a member is of not recertifying for Medicaid

Model Population Training Population 35,822 Validation Population 15,353

Top 20% of members are 2x times more likely to fail to recertify

1) Three State Model is combination of FL, GA and TX data, August 2012 to August 2013

Quintile

Lift


Descriptive analytics: Recertification failure by county

20

For geographic areas with at least 100 members.

Florida Texas

Georgia


Systematic issues, County Office performance •  Addressed by

Altegra’s Government Affairs Outreach

Model predictors Consumer variables

•  Charitable giving – areas where 75% or more of individuals contribute to charities are 35% less likely to fail to recertify

•  Party affiliation – individuals who are unaffiliated with a political party are 2 times more-likely to fail to recertify.

•  Foreign Made Car ownership – individuals who own foreign made cars are nearly 2 times more likely to fail to recertify than those own domestic built cares

•  Employment Patterns – (% engaged in Manufacturing) More manufacturing, lower probability of recertification failure, indicating lower skill or blue collar job stability


41% 37%

29% 27%

0%

10%

20%

30%

40%

50%

0 to 25% 25 to 50% 50 to 75% 75 to 100% Percent of Population (ZIP) That Have Made Charitable

Contributions

41% 40%

25% 22% 20% 20% 15%

0%

10%

20%

30%

40%

50%

Unknown Unaffiliated Other Republican Democrat Green Libertarian Registered Parties

29% 33% 32%

39% 43%

0%

10%

20%

30%

40%

50%

10 to 9 8 to 7 6 to 5 4 to 3 2 to 1 Likelihood of Owning a Domestic Sedan

(1: Most Likely, 10: Least Likely )

Three state recertification failure model validation Excellent validation observed

22

0

20

40

60

80

100

120

140

160

180

200

220

240

1 2 3 4 5 6 7 8 9 10

226

195

155

111

85 71

59 44

33 21

229

195

156

111

85 74

57 42

29 22

Recertification Model Validation Lift by Decile2

Training

Validation

Average

LIft

1)  Three State Model is combination of FL, GA and TX data, August 2012 to August 2013 2)  Population study cohort size of 19,538, or 1,954 per decile, split 70% training and 30% validation

Decile


23

Current Served Populations •  Historical experience indicates 1/3 of

population at risk of not recertifying

•  With predictive analytics “at-risk” individuals can be identified increase probability of failure to recertify to 90% likelihood

•  Improve business performance by appropriately allocating resources to targeted cohort

Failure to recertify risk

Applying analytics to allocate resources


New Consumers / Exchange Populations •  Integration of consumer behavior, social

claiming can identify risk in unknown populations

•  Health exchanges •  Assigned

capitated populations

COPYRIGHT © 2014 PREDILYTICS, INC. ALL RIGHTS RESERVED 24

Big Data

Healthcare Analytics

Machine Learning

Delivering machine learning healthcare data analytics to generate meaningful insight to solve healthcare industry challenges

Discussion

Machining learning modeling performance Accepted assessment model validation – Intervention engagement

25

•  3,677 members were selected for assessments in 2012 who were in the randomly selected member validation group (not used to create the model equation)

•  To verify the model’s predictive power, the model equation was applied to this group as they appeared on the file in June 2012

0%

10%

20%

30%

40%

50%

60%

70%

1 2 3 4 5 6 7 8 9 10

Engagemen

t Accep

tance Ra

te

Decile

Model Projec:on Actual 2012 Result

The model projection tracks closely with the actual 2012 results


Consumer Behavior: Factors Affecting Member Attrition and Retention

Health & Medicine

machine learning

big data

targeted events

rights reserved

recertification failure

data attributes

tx data

state model