Transcript Investigations and Saves - University of Illinois at Urbana

Predictive Modeling Project
Stephen P. D’Arcy
Professor of Finance
University of Illinois at Urbana-Champaign
ORMIR Presentation
October 26, 2005
Motivation - To Advance the Science of
Predictive Modeling by:
• Applying predictive modeling to a key aspect
of insurance operations
• Sharing the results of this research fully so that
other researchers can replicate the results and
improve the process
• Educating practitioners about the value of
predictive modeling
• Opening up the “black box” approach of data
mining that has generally been applied
Predictive Modeling in Insurance
• Massive amounts of data available
– Accuracy varies
– Much of it is ignored in rating or claims handling
• Innovators
– Use of credit scoring in rating
– Predictive modeling applications
• Underwriting
• Claims handling
• Fraud investigation
• Studies treated as proprietary and not shared or
published
Project Details
• Jointly funded by the National Center for
Supercomputing Applications (NCSA) and ORMIR
• Data set:
– Detail Claim Database created by the Automobile
Insurers Bureau of Massachusetts
• Predictive modeling tool:
– Data-to-Knowledge (D2K) program of NCSA
• Results:
– Papers
– Presentations
Steps in Predictive Modeling
1.
2.
3.
4.
Decide question to be investigated
Access data
Understand your data
Preliminary data mining analysis
•
•
Decision trees
Generalized linear regression
5. Evaluate results and investigate problems
6. Additional data mining analysis
•
•
•
Trees and regression
Neural networks
Other techniques
7. Apply results to insurance operations
8. Evaluate impact of change
Detail Claim Database (DCD)
•
•
Created by the Automobile Insurers Bureau (AIB) of
Massachusetts;
Primary objectives:
–
–
–
–
•
Supporting company claim negotiation and claim denial
activities
Assisting the Board of Registration
Responding to the Division of Insurance and to the
Legislature
Assisting the Insurance Fraud Bureau of Massachusetts in
detecting possible fraud rings
Accessible for all member companies of the AIB and
selected researchers
DCD Observations and Variables
• 491,591 Claim Observations (1/1/94 and subsequent)
• 95 Variables from 5 Categories:
– Policy Information
– Claim Information
• Coverage
• Accident date
• Report date
• Total amount paid
• Type of injury
• Type of treatment
– Outpatient Medical Provider Information (up to 2 providers)
• Provider type (MD, Chiropractor, Physical Therapist, Hospital, Other)
• Amount billed and PIP/MED amount paid
– Attorney Information
– Claim Handling Information
• Type of investigation, if any
Types of Investigations
• Independent Medical Examination (IME)
– 66,876 Requests (16.72%)
– Average Savings $348.71
– Favorable Outcomes (60%)
• Medical Audit (MA)
– 44,099 Requests (11.02%)
– Average Savings $367.08
– Favorable Outcomes (67%)
• Special Investigation (SI)
– 16,668 Requests (4.17%)
– Average Savings $1805.39
– Favorable Outcomes (46%)
Problem – Average Savings values are based on a
formula and may not reflect actual savings
Steps to Avoid Problems with
Recorded Savings Value
1. Use Favorable Outcome as dependent variable
2. Generate value for expected payment
•
•
•
•
Stepwise linear regression (33 steps)
Based on claims not investigated
Apply to IME Requested claims
Compare expected payment to actual payment
Result of IME
Mean (Expected – Actual)
No change recommended
-562
Favorable result
18
Regression Results
MP1_TYPE
CH 690.6566
CO 668.4522
MD
0
MI
0
MO
0
N1 -971.328
N2 -185.653
NO
0
PO 682.4842
PT 644.808
Inj_Type
MJ 3020.227
MM
0
SE
603.998
SS
0
ACCMONTH
Q1
0
Q2
0
Q3
0
Q4 49.58459
Summary of
MP2_TYPE
CH
CO
MD
MI
MO
N1
N2
NO
PO
PT
Coefficients
442.1823
414.3684
0
456.9774
0
-392.816
0
-596.941
458.7032
392.9716
PRIMTYPE
CH
CO
MD
MI
MO
N1
N2
NO
PO
PT
-627.62
-536.863
-287.647
0
0
0
0
-646.206
-375.556
-497.701
Health_I
N 492.2223
U -34.3341
Y
0
0
INJ_GRP
1
2
3
4
-259.887
0
0
0
Em_Treat
B
N
Y
-175.699
-670.978
0
Pol_Type
P
C
-173.8
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
Dependent
Coefficient
1894.004
-173.80
-175.70
-670.98
492.22
-34.33
3020.23
604.00
690.66
668.45
-971.33
-185.65
682.48
644.81
442.18
414.37
456.98
-392.82
-596.94
458.70
392.97
-259.89
1323.34
-627.62
-536.86
-287.65
-646.21
-375.56
-497.70
0.10
49.58
92.12
-21.90
10.05
Variable = Tot_Paid
Variable
Intercept
Pol_Type=P +
Em_Treat=B +
Em_Treat=N +
Health_I=N +
Health_I=U +
Inj_Type=MJ +
Inj_Type=SE +
MP1_TYPE=CH +
MP1_TYPE=CO +
MP1_TYPE=N1 +
MP1_TYPE=N2 +
MP1_TYPE=PO +
MP1_TYPE=PT +
MP2_TYPE=CH +
MP2_TYPE=CO +
MP2_TYPE=MI +
MP2_TYPE=N1 +
MP2_TYPE=NO +
MP2_TYPE=PO +
MP2_TYPE=PT +
INJ_GRP=01 +
ATT +
PRIMTYPE=CH +
PRIMTYPE=CO +
PRIMTYPE=MD +
PRIMTYPE=NO +
PRIMTYPE=PO +
PRIMTYPE=PT +
PRIMBILL +
ACCMONTH=Q4 +
TREATLAG +
REP_LAGT +
CLMT_AGE +
Primary Medical Provider Types and
Attorney Representation Frequency
Prim ary MP_Type v.s. Attorney
100%
90%
80%
70%
CH
CO
60%
MD
MI
50%
MO
PO
40%
PT
Tot al
30%
20%
10%
0%
CH
CO
MD
MI
MO
PO
PT
Tot al
Highest Attorney Representation by
Individual Medical Provider
ATT=1 Total Claims ATT Freq PRIM_TYPE
63
63
100.00%
CO
57
57
100.00%
PO
212
214
99.07%
PO
83
84
98.81%
MD
78
79
98.73%
MO
61
62
98.39%
CH
110
112
98.21%
CO
55
56
98.21%
CH
155
158
98.10%
CH
102
104
98.08%
MO
152
155
98.06%
CH
Injuries’ Seasonality Trend
Decision Tree Example
Injury Type = SS
Y
N
Second Medical Provider
Primary Medical
Provider = CH
N
PIP Coverage
Y
Emergency Medical
Treatment
Decision Tree Approach for IMEs
• Nodes and Favorable Outcomes
–
–
–
–
–
–
–
Strain and sprain only (63%)
Only 1 medical provider (67%)
No emergency room treatment (70%)
PIP claim (72%)
Bill less than $2421 (73%)
Attorney representation (75%)
Accident month November (81%)
Ongoing Research
• New dependent variable for expected savings
– Refine model of expected payment
– Determine estimated savings from investigations
– Generate decision tree based on estimated savings
• Combining variables
– Medical provider type
– Injury type
– Accident quarter (rather than month)
• Examine medical provider/attorney connections
• Suggestions?