AIB PIP Claim Screening Experiment Final Report Understanding and

Document Sample
AIB PIP Claim Screening Experiment Final Report Understanding and Powered By Docstoc
					                     AIB PIP Claim Screening Experiment Final Report

               Understanding and Improving the Claim Investigation Process

1. Introduction

1.1 Background

The high cost of auto injury coverages (BI and PIP) continues to be a formidable problem.
Recent research by the AIB in Massachusetts and the Insurance Research Council nationally
have documented the trends and underlying reasons. The cost of BI and PIP are rising, driven
by an influx of strain and sprain claims, many of which are somewhat questionable.

In Massachusetts, these trends have been countered to some degree by enhanced
prosecution efforts spearheaded by the Insurance Fraud Bureau. However, the main line of
defense against suspicious claims remains the front-line claim staff. Most companies have
been attempting to strengthen their cost-containment and fraud resistance capabilities.

To gain some perspective on recent trends in Massachusetts, we have prepared Table 1. This
exhibit is based on Detailed Claim Database (DCD) data1 for PIP claims with injury report dates
during the same four-month period in each of 1994, 1995 and 1996. This period was selected
for reasons that will become clear when we describe the claim screening study design.

For each year, the data were analyzed as of March 31 of the second subsequent year. Only
claims that were closed and reported to the DCD by March 31 were included. Thus, the claims
are restricted to all but late-closing claims and are fairly representative of PIP claims in general.
The comparison of such claims across the three years is informative.

Note first that the number of claims has been increasing, as has the average cost per claim.
The percentage of claims that involve a strain/sprain as the primary injury has remained steady
at about 72 percent. In terms of treatment, the utilization of chiropractors as primary providers
has stayed constant at approximately 30 percent. However, there is increasing use of physical
therapists replacing MDs as primary providers.2 Moreover, the percentage of total medical
charges attributable to PTs has grown from 15 percent in 1994 to 19 percent in 1996. Legal
representation also seems to be rising, from 47 percent in 1994 to 50 percent in 1996, but this
latter increase may be a result of more complete reporting rather than a genuine increase in
legal representation.

  The DCD is a mandated auto injury claim reporting system administered by AIB. The DCD contains claims closed on or
after January 1, 1994. As of June 26, 1998, the DCD had 699,466 claims reported.
  The combined MD and PT primary providers has remained constant at 38 percent of claims.

                                        Claim Screen Experiment and
                              DCD PIP Claims with Injury Report Dates 5/15 - 9/15
                                    Closed Claims Reported at 3/1/(year+2)
                              Year                1994          1995         1996                     CSE
                            Claims                   26,824       29,361       29,590                   2003
                    Average Claim                    $2,222       $2,359       $2,414                 $1,939
                        ET only (%)                    29.9         29.0          31.2                  31.1
                  Strain/Sprain (%)                    72.4         73.5          71.5                  64.3
                       SS Average                    $2,353       $2,489       $2,576                 $2,144

              Primary Medical                  MD             28.2           27.1           25.3         33.9
              Prov. Distribution (%)
                                               CH             29.2           30.7           30.1         24.0
                                               PT              9.9           11.1           12.5         10.1
                                               OT              1.3            0.9            0.8          0.9
                                               NO             31.5           30.2           31.2         31.1
                     Medical Billing
                    Distribution (%)           MD            42.2            39.7           38.1        37.6
                                               CH            40.9            42.1           41.3        40.8
                                               PT            15.3            16.7           19.3        20.0
                                               OT             1.6             1.5            1.3         1.6
                             Lgl (%)                         46.7            49.2           50.0        39.4
                      Lgl $ Average                        $3,321          $3,446         $3,486      $3,252

              Special Handling (%)                            26.0           29.0           34.9         33.9
                         $Sav (%)                             10.3           10.8           14.6         17.8

                                             IME            16.5             18.2           22.2         24.0
                                        $ Sav (%)             4.6             4.8            6.1          7.4
                                             Med            10.0             11.1           14.1         13.8
                                        $ Sav (%)             5.2             5.5            8.6         11.9
                                               SI             2.1             2.5            2.1          2.5
                                        $ Sav (%)             0.7             0.8            0.5          0.3
                                                         Table 1

In terms of industry response, there is some positive news. The utilization of independent
medical examinations (IME) and medical audits to control losses has been increasing
dramatically. IMEs have gone from 16.5 percent in 1994 to 22.2 percent in 1996, while medical
audits rose from 10 percent to 14 percent. Moreover, because these figures are based on all
but late-closing (and therefore complex) claims the true increase may be even greater. On the
other hand, the application of special investigation techniques does not appear to be

These results provide some important context for this discussion of the AIB claim screening
experiment (CSE). Is it possible to identify more claims for which special investigation would be
appropriate? Are IMEs applied in a cost-effective way? The CSE represents the first attempt by

 Full DCD data for claims closing in 1997 versus 1996 show continued increases of 17.3 percent for IME, 26.2 percent for
Medical Audits, but a decrease of 8.3 percent for special investigations.

the AIB to move from retrospective review to testing of potential improvements in a real-time

1.2 Overview of the Claim Screening Experiment

Company procedures for dealing with suspicious claims vary widely. However, all systems rely
heavily on front-line claim representatives or adjusters to identify potentially fraudulent claims.
These claims personnel are given fraud awareness training, based primarily on the traditional
“red flags” disseminated by the National Insurance Crime Bureau and refined by individual

Recent AIB research has shown that the traditional indicators are in fact closely related to claim
suspicion and do often trigger investigative activity. However, we have also learned that many
apparently suspicious claims are slipping through with little or no investigation. We have
speculated that the unstructured reliance on adjuster awareness cannot keep pace with the
contemporary fraud problem. This claim screening experiment was intended to test the
hypothesis that an automated method for early fraud detection could be more cost-effective.

The experiment builds on statistical models previously developed. Using data on PIP claims
from 1993 accidents, we have derived models that can identify claim suspicion. However, these
models were based on a sample of closed claims. The CSE is our initial attempt to test the
feasibility and value of applying the models prospectively at relatively early points in the life of
the claim.

In the CSE we have attempted to simulate the effect of having computer software that could
capture data on key fraud indicators and generate a suspicion rating. Ideally, such a system
would allow convenient entry of information about the claim and routine fraud indicators (red
flags based on facts that would generally be known to the adjuster), and would continually
update a suspicion score as data accumulated. When the score reached some critical level, the
claim would be identified for additional attention (possibly SIU referral).

To implement such a sophisticated system would have entailed a major software development
effort. Therefore, after consultation with company claim managers, we elected to rely on paper
forms rather than computer screens. The resulting data were entered by a data-entry operator
into a simple program on a special-purpose PC provided to each participating company. The
data were downloaded nightly to the AIB for processing. Reports on the resulting suspicion
were then fed back and printed weekly, for use by adjusters and supervisory personnel. This
approach was clearly less elegant and efficient than a fully integrated software implementation
within each company. However, it was hoped that the additional information provided to claim
handlers would add significant value.

To estimate the impact of claim screening we divided all newly arising claims into two groups: a
Test Group and Control Group. We collected data on fraud indicators for both groups, but the
information about suspicion scores was fed back to claim handlers only for the test claims. This

information could in theory help to guide decisions on whether test claims should be further
investigated. However, it was decided to leave decisions regarding investigative activity to the
discretion of the adjusters and their supervisors.

1.3 Conceptual Framework: PIP Claim Flow

The CSE design was based on a conceptual model of the PIP claim adjustment process
introduced in our previous studies. Figure 1 illustrates this claim flow model.

The first stage of this process is a judgment by the front-line adjuster regarding the severity of
injury, the company exposure to claim payment, and the possibility of fraud or buildup. Claims
that appear to be modest and/or legitimate are settled routinely. Claims that raise questions, or
have high loss exposure, are subjected to some form of supervisory review and/or additional
investigation. Based on the results of this second stage of analysis, the claim is either returned
to the mainstream or investigated more thoroughly, possibly with referral to a special
investigation unit (SIU). The results of this investigation will then affect the nature of the
interaction between the insurer and the claimant.

As mentioned above, almost all of the decision-making with respect to investigation is currently
based on adjuster judgment and discretion. Neither of the critical junctures in our conceptual
model are controlled, or even guided, by any formal measurement of suspicion levels. The CSE
was designed to test whether calculating a suspicion score and informing the adjuster about
this score could result in reduced losses for suspicious claims.

                       PIP CLAIM FLOW

                    Claim Arrival

  First Stage      Routine Handling           Routine Settlement

Second Stage      Review/Investigation


                Negotiated Settlement

   Full Payment                             Claim Denial
                       Comp Payment

                                Figure 1


                      Claim Arrival

Routine             Base Information           Express Claims
 Stage                                         Full Settlement

                  Additional Information
                                             Routine Settlement

   Stage              Routine Stage
                     Suspicion Score

                Additional Investigation

                  Investigative Stage                   SIU
                    Suspicion Score

                 Negotiated Settlement             Claim Denial

                                  Figure 2

Figure 2 illustrates the claim flow with suspicion scores added to the processing. Decisions are
then to be made knowing the suspicion scores.

In the next section we discuss the statistical model used in the CSE to generate suspicion
scores. This model was derived from closed claims on which the relevant information was fairly
complete. The CSE is our first opportunity to study claims dynamically, as they arise and
information accumulates over time. The resulting data shed light on the pace at which various
items of data become available. Understanding the dynamics of the claim process may lead to
more sophisticated strategies for improvement.

1.4 Statistical Models for Identifying Claim Suspicion

Based on our previous research, we originally selected a model using the technique of linear
regression analysis with interaction terms. In effect, this model assigned weights to various
fraud indicators and to certain combinations of indicators. The dependent variable in these
models was the level of suspicion on a ten-point scale from zero (none) to 10 (very high).

The interaction terms refer to the pairs of indicators that are weighted. For example, the
combination of “claimant in old, low-value vehicle” and “three or more claimants” added .8 to
the total suspicion score. Either of these factors by itself would not increase the score, only the

This “linear interactive” model appeared to provide the best performance. However, early on in
the CSE it became evident that this model had a flaw. Because of the need to have data on
both factors before assigning any weight to interaction terms, the score was building up too
slowly. We learned of several claims that were clearly suspicious but had zero scores.

To remedy this problem we considered both simple and complex solutions. A mathematically
sophisticated approach would have been to project the probable values for missing information
based on what was already known. Such a methodology would have required more research,
however. We decided on a more straightforward approach. Rather than relying on the linear
interactive model alone, we also computed a score based on one of the alternative simple
linear (no interaction terms) models developed previously. We then used the maximum of the
two scores in the CSE.4 This quick-fix seemed to produce sensible results.

1.5 Focus of this Report

We have termed this the final report, because all but the latest-closing claims are available.
Only 80 of the claims (2.3%) entered into the CSE currently remain open. While some
interesting claims are included in this open subset, the vast majority (96.6%) of the non-dud
claims are closed. Thus, a fair assessment of the payoff from an automated claim screening
should be able to be accomplished using only the closed claims.
 While the maximum score of the two models was calculated and used throughout the CSE, the final suspicion score of
nearly all applicable claims derived from the linear (non-interactive) model with twenty variables.

We know a lot about the investigation that has occurred for most claims. Consequently, it is
possible to reach some meaningful conclusions about the actual and potential impact of claim
screening on investigation. Our analyses of the available data began from this perspective.
However, as we delved into the data, we increasingly realized that the CSE data was extremely
valuable in another way. For the first time (to our knowledge), the “natural history” of claims
evolving over time could be studied in detail.

Our analyses suggest that claim screening technology is unlikely to work without integration
within the company’s computer and workflow systems. Understanding when various kinds of
information arrive in the process can be critical to designing effective improvements. Learning
when and why investigations are currently triggered provides insight into the limitations that
exist. Finding out which types of claims produce favorable outcomes when investigated can
help refine our models for suspicion identification. These are some of the issues we have
addressed in this report.

2. Experimental Design

2.1 Claim Sample

Four large insurers volunteered to participate in the CSE. Each of these companies agreed to
register all PIP claims that arose beginning approximately May 15, 1996 and to continue until a
prespecified goal was achieved. All told, the span of the sample period stretched from mid-May
through mid-September. For purposes of the CSE, a PIP claim was initiated when a notice of
injury was received by the company. Thus, in some instances the accident resulting in the PIP
claim could have occurred prior to May 15, 1996.

The goals for each of the companies and the actual samples obtained are shown in Table 2.
These samples were intended to be roughly proportional to the relative market shares of the
four companies.

                                Claim Screen Experiment Sample
                                                  Goal     Actual
                                Company A           800       531
                                Company B         1000       1520
                                Company C           700       807
                                Company D           500       577
                                   Total          3000       3435
                                            Table 2

Thus our sample size goals were exceeded in three out of the four companies, and the total
sample was about 15% larger than expected.

Note that the sample of PIP claims may not be a random sample of Massachusetts PIP claims
for two reasons. First, the four companies selected may differ from other companies in terms of
the types of claims received and/or their handling of these claims. Second, within our four
companies strict proportionality was not achieved. In particular, Company A was somewhat

under-represented and Company B over-represented. This was the result of the practical
choice of using only one branch office per company. However, for the kinds of data we are
collecting, past experience suggests that differences among insurers are unlikely to be large.
Therefore, treating the 3,435 claims as effectively a random sample should not be misleading.

For the claims for which we have DCD data, however, this issue is compounded by the different
claim-handling practices of the participating companies. Table 1 provides a comparison
between the CSE claims and the entire population of PIP claims, based on DCD data. Once
again, this comparison includes those claims that were closed and entered in the DCD by
March 26, 1998. The results indicate that our CSE sample is broadly similar to claims overall,
but skewed toward less suspicious claims. For example, the average cost is 20 percent lower
($1,939 versus $2,414), the percentage of strains/sprains somewhat lower (64% versus 72%)
and the involvement of attorneys lower (39% versus 50%).

We believe that these differences reflect primarily the fact that Company B has an explicit
policy of rapidly closing simple claims. Most of these simple claims have minimal medical
expenses and no disability. Thus, the sample captured by the DCD has a disproportionate
number of these simple claims. For this reason, quantitative analyses based on CSE claims
must at this point be interpreted with great caution as indicators of population characteristics.5

2.2 Assignment to Test and Control

The unit of analysis for the study was an individual PIP claimant. In principle, each PIP
claimant could be assigned randomly to test or control status. However, we recognized that it
would be impractical to treat individuals differently who had been injured in the same accident.
Therefore, we decided to base the randomization on accident rather than individual.

Specifically, as notification of injuries associated with accidents arrived, we grouped the
accidents into pairs (first and second arriving, third and fourth, etc.). We then assigned the first
or second of these accidents in each pair to the test group and the other accident to the control
group. All claimants involved in a particular accident were assigned to the test or control status
determined by the accident-level randomization.

The primary purpose of the experiment was to test whether information on the value of a
suspicion score would have an impact on claim handling. Specifically, would the quality of
adjuster decisions be improved because of this feedback? To answer this question, we
arranged to transmit back to the adjusters and supervisors the score and associated fraud
indicators for any test claim with moderate to high suspicion. For corresponding control claims
this information was not provided to claim handlers.

2.3 Data Collection Forms

For each claim, data was recorded by the adjusters on the following forms:
    Alternatively, reweighting with population weights may prove to be more accurate but that is beyond the scope of this report.

1. First Report
2. Routine Report
3. Investigation Report
4. Outcome Report

Copies of these forms are included in the Appendix.

Form 1 was completed once for every claim entered in the study and served as an
administration/enrollment form. After downloading the information to the AIB the claim was
officially registered as part of the CSE sample. Form 1 captured basic data about the claimant,
accident and policy.

Form 2 was meant to capture information about the presence or absence of 23 routine
indicators. These indicators are the principal “red flags” that are used informally to identify
suspicious claims. Included among these are the fraud indicators necessary to calculate the
suspicion scores.

Form 3 was used only for claims subjected to some level of investigation. The first part of this
form was a checklist of possible investigative techniques. The second part was a checklist of
“investigative indicators.” These investigative indicators are red flags that can occur as the
result of investigative activity. “Claimant refused to appear or no show for an IME” is an

Form 4 is to be completed when the claim is closed. Form 4 provides information on payments
and on the potential for a BI claim to be made.

2.4 Claim Processing

As explained above, a PIP claim was established for CSE purposes when the injury was first
reported. Claims resulting from injuries between May 15, 1996 and the close of the study
window were eligible, regardless of when the accident may have occurred. The first PIP claim
(individual claimant) for a given accident was assigned to either the test or control group as
described above. All subsequent claims for the same accident were then automatically placed
in the same group.

When a new claim file was originally set up, a copy of Form 1 (First Report) was filled out and
entered into a PC with software provided by the AIB. Each company had one such computer in
the branch office designated to participate in the CSE. The AIB software also performed the
assignment of claims to the test and control groups.

Based on the Form 1 data entered, the computer printed out a set of paper forms with pre-filled
header information (name of claimant, policy number, claim number, etc.). One copy of Form 4
(Outcome Report) and six each of Form 2 (Routine Report) and Form 3 (Investigative Report)
were printed. All of these forms were placed in the claim folder. When any form was completed

by the adjuster, it was given to the data entry clerk. After data entry, the paper forms were

We had intended that Form 2 would be filled out first very soon after the claim (injury) was
reported. Updates would then be submitted at approximately 30-day intervals for the first 120
days, and thereafter when changes occurred. However, we found that compliance with this
schedule was poor. Many claims had no reports until 60 to 90 days from the injury report, and
most claims had no more than two routine reports in all. Although it is possible that claim
information arrived later than expected, we believe the off-line data entry system provided a
disincentive for complete and timely reporting on all claims.

We had expected about 10 percent of the PIP claims to involve only a minimum amount of
treatment and no disability. We designated such claims as “express claims” and allowed the
companies to simplify the handling of these. For such claims, a Form 1 was expected to be
followed by an Outcome Form 4, with no intermediate reports. For the most part, only one of
the four participating companies made extensive use of the express option.

At the end of each business day, the data collected were downloaded to the AIB through an
automatic link. Overnight, the AIB processed the data and updated the CSE database
maintained on the AIB’s mainframe. At the end of each week each company could generate a
set of supervisor/adjuster reports:

•     New Data Report (test group claims only)
•     45-Day Inactivity Report
•     No-Indicator Report

The first of these reports listed all of a supervisor’s open test claims for which new data (i.e.,
fraud indicators or investigation) have been recorded within the previous 7 days. This report
included the AIB model suspicion score along with a list of fraud indicators and any
investigative techniques. The second report listed all claims with no change for at least 45
days. The third report listed claims with no indicators at all. The latter two reports were
intended to monitor the continual flow of data.

These reports could be used by the supervisor and shared with adjusters. Our hope was that
the information would stimulate increased investigation of suspicious claims in the test group.
At the same time, though, adjusters were told not to ignore (or over-scrutinize) other (control)
claims for which no such additional information was provided.

As indicated above, new claims were enrolled until the sample size goals were reached. For
most companies, the actual window of enrollment extended from approximately May 15, 1996 to
September 15, 1996. Each claim included in the CSE was intended to be followed until the
claim was closed. Companies were given the option of deleting claims without activity or filing
outcome reports that showed no activity.6 A total of 446 claims were deleted and 640 claims
had no activity. We have combined both types in defining the Dud category in the next section.
    A few claims with erroneous reporting or with new adjusters unfamiliar with CSE were also deleted.

3. Initial Tracking of Claims

3.1 The PIP Claim Tracks

The design of CSE allowed for the possibility of three types of claims: Duds, Express and the
remainder of claims. The Dud claims are characterized by a report of injury but no follow-up or
pursuit of a claim. No payments are made on Dud claims. Express claims, as we perceived
them, were simple claims which, by their nature, should be suspicion-free and have low loss
exposure to the company.

Both Dud and Express claims should be treated separately from the remaining claims we call
“Target” claims. These are the claims we want to target for information processing and review.

The Dud count included all claims that either were entered and deleted from the database (a
count of those claims was retained) or were closed without payment (CWP) and without any
report of a positive routine indicator. Additionally, companies were requested to review each
CWP that had a positive routine indicator to determine whether it was truly a Dud. Any claim
designated a Dud by the participating companies, and there were many, was also labeled Dud
for CSE.

The Express count was established in a different and somewhat arbitrary manner. The
designation “simple claim” is a fuzzy characterization that can mean different things to different
people. The design of CSE allowed companies to self-select these type of claims. The actual
selections, however, turned out to be quite heterogeneous. We view the Express category to
be a set of claims with pre-determined criteria, claims that can safely bypass the information
processing for high value and suspicious claims.

A subjective review of the data suggested two criteria for Express: Report Date lag from
Accident Date and Expected Value of the Claim. Two values of each criterion were examined:
7 days and 14 days for the report lag; $500 and $1000 for the (expected) claim cost. An
additional requirement that loss adjustment expenses be no more than $25 had the practical
effect of keeping heavily investigated CWP claims in the target group.

The combination of 14 days and $1000 produces the claim track volumes shown in Figure 3
and Table 3. Lowering the dollar criterion to $500 adds about 320 more target claims, making
the express proportion only about 16 percent of the claims. Clearly, other (simple) criteria are
possible to determine the PIP claim type separations.

The triage approach to tracking claims in terms of Duds, Express and Target could apply to
any line of insurance. However, characteristics of three types, and the relative claim volume of
each, will differ. Figure 3 depicts the triage scheme for PIP claims with claim volume estimates
derived from the CSE data.

                               PIP CLAIM TRACKS

                                           Claim Arrival

                                No                                Yes


                                                     No                         Yes

               Duds                                                                  Express
            (No Activity)                                                           (Simple)

                   32%                                   43%                               25%

                                          Claims Handling

            * Express less than $1,000; Reported less than 15 days after accident

                                             Figure 3

Once Express claims are identified, Target claims become those claims with any substantive
activity. The Duds will be the residual type that can be closed according to any workable
schedule (e.g., 180-270 days from Report Date). Provisions should also be made for late
arriving information to change the typing of the claim, much in the same way that previously
closed claims are sometimes reopened.

                                                      PIP Claim Tracks
                                                   Frequency by Company
                       Duds*     Duds %         Express         Express/             Target***      Target         Total
                                 of Total      Inferred**      Inferred %                          % of Total
                                                                 of Total
                  All    1086         31.6%            863             25.1%                   1486        43.3%     3435
           Company A      167         31.5%            109             20.5%                    255        48.0%      531
           Company B      440         28.9%            526             34.6%                    554        36.4%     1520
           Company C      318         39.4%            131             16.2%                    358        44.4%      807
           Company D      161         27.9%             97             16.8%                    319        55.3%      577
         *        Includes 446 Claims deleted from CSE database as duds.
         **       Express Claims inferred using definition of Injury Report Date <15 days from Accident Date;
                   0< (expected) claim payment <1000 allocated expense < $25.
         ***      Includes all 80 claims open as of March 26, 1998.
                                                             Table 3

3.2 Closing Rates of PIP Claim Tracks

Express claims would be expected to close quickly; whereas, dud claims in CSE will close as
dictated by (differing) company policies. Table 4 displays the proportions of closed claims as of
90, 270 and 360 days. About 52 percent of the CSE Express claims were closed by 90 days;
more than 91 percent of Dud and Express claims were closed by 270 day mark. On the other
hand, 93 percent of the target claims remain open and, therefore, are available at 90 days for
the investigative decision to be made (see Sections 5 and 6 below).

                                              PIP Claim Type Closing Rates
                                  Time to Close:                  Duds       Express      Target
                                  From Report Date:
                                              < 90 Days           34.3%         51.8%       7.5%
                                            91-270 Days           58.6%         37.2%      52.4%
                                           271-360 Days            2.2%          3.1%      13.5%
                                              >360 Days            4.9%          7.9%      26.6%
                                                             Table 4

3.3 Time Flow for PIP Claim Tracks

It may be useful to model the expected information flow for the three claim tracks. Table 5
provides a sample of the expected information arrival time together with the open/closed
percentage from the CSE data. Underlying data for table 5, and subsequent tables, are
available in the Appendix. Notice that the target claim type allows for “No Activity” for up to 90
days. While it is not uncommon for medical bills to be “packaged” and arrive from an attorney

after 90 days, legal representation itself is usually known prior to 90 days. All high exposure
and suspected build-up fraud claims should be included in the target claim type.

                                         Time Flow for PIP Claim Tracks
                                                  CLAIM TYPE
        Time                DUD                         EXPRESS                          TARGET
       From                (32%)                           (25%)                          (43%)
       0-14    Report Information (1)          Report Information (2)       Report Information (3)
       Days*   Open 94.8% Closed 5.2%          Open 99.3% Closed 0.7%       Open 99.7% Closed 0.3%
       15-30   No Activity                     App. Info (2)/ ER Bills      App. Info (3)/No Activity
               Open 89.1% Closed 10.9%         Open 89.1% Closed 10.9%      Open 97.9% Closed 2.1%
       31-60   No Activity                     Medical Diagnosis/Bills      Med. Diag./Legal Rep/No Activity
               Open 79.9% Closed 20.1%         Open 63.9% Closed 36.1%      Open 94.0% Closed 6.0%
       61-90   No Activity                     Medical Bills/Wage Loss      Med. Treat./Legal Rep/No Activity
               Open 63.7% Closed 36.3%         Open 47.3% Closed 52.7%      Open 91.3% Closed 8.7%
       91-     No Activity                     Medical Bills/Wage Loss      Med. Bills/Legal/Investigation
       180     Open 16.2% Closed 83.8%         Open 22.7% Closed 77.3%      Open 64.5% Closed 35.5 %
       181-    No Activity                     Medical Bills/Wage Loss      Medical/Wage/Legal/Investigatio
       270     Open 5.7% Closed 94.3%          Open 10.8% Closed 89.2%      n
                                                                            Open 37.9% Closed 62.1%
       271-    No Activity                     Medical Bills/Wage Loss      Medical/Wage/Legal/Investigatio
       360     Open 4.4% Closed 95.6%          Open 7.8% Closed 92.2%       n
                                                                            Open 25.9% Closed 74.1%
       Over    No Activity                     Medical Bills/Wage Loss      Medical/Wage/Legal/Investigatio
       360     Open 0.0% Closed 100.0%         Open 0.0% Closed 100.0%      n
                                                                            Open 0.0% Closed 100%

*      Alternatively 0-7 days
       (1) Accident, vague injury/treatment .
       (2) Minor acc, rapid treatment, probability of build-up small, expected closing < 90 to 180 days, expected
           payment < $1,000. Application and bills confirm expectations.
       (3) All other claim reports and applications.

                                                    Table 5

3.4.   Arising Claims by Accident

The CSE tracked the number of arising claims in each reported accident as part of the
randomization design. About two-thirds (63%) of the arising claims are from single claim
accidents and approximately 14 percent are from accidents with three or more potential
claimants. The patterns of claimants per accident for the three claim tracks are shown in
Tables 6A and 6B.

                                 Percent of Claim Types by Claims per Accident
                             Number         1         2     3 or more Row Sum
                                 Dud         34%       29%       26%      32%
                              Express        27%       25%       18%      25%
                               Target        39%       46%       56%      43%
                                            100%      100%      100%     100%
                                                   Table 6A

                                              Percent of Claims per Claim Type
                                  Number          1         2     3 or more Row Sum
                                      Dud          68%       20%       12%     100%
                                   Express         68%       22%       10%     100%
                                    Target         57%       24%       19%     100%
                                                   63%       23%       14%     100%
                                                         Table 6B

Single claim accidents produce the highest percentage of duds (34%). Substantially more
target claims arise from multi-claimant accidents than from single claim accidents. Perhaps PIP
claimants per accident may serve as a useful indicator in some future refinement of the PIP
claim tracks or model suspicion score.

In our prior studies7 of auto injury claims we observed the key role of the discretionary claim.
Once involved in an accident, individuals with minor or no visible injuries at the scene of the
accident exercise some discretion in filing PIP and BI claims. Those responsible for the
accident, the at-fault drivers, have less incentive to file PIP claims since they are barred from
the tort system.8Conversely, those passengers, pedestrians, cyclists and drivers not at fault
have the free-lottery incentive of non-economic damages from the tort system to file a PIP claim
followed, of course, by a BI claim. In the past we have observed far fewer PIP claims filed by
at-fault drivers, and their passengers, than would be expected in a pure no-fault system.9

Table 7 displays the breakdown of single and multiple claim accidents10 by claimant type. The
data show that operators filing PIP claims are two or three times less likely to be at fault than
not at fault. Passengers in the insured vehicle exhibit claim filing behavior similarly patterned
on the fault of the operator. Target claims for drivers exhibit a somewhat more pronounced
pattern with 73 percent of claims arising when operators are not at fault versus 66 percent of
the express claims.

  Derrig, Richard A. and Herbert I. Weisberg, (1994) "Behavioral Factors and Lotteries Under No-Fault with a Monetary
Threshold: A Study of Massachusetts Automobile Claims", The Journal of Risk and Insurance, 61:2 , 245-275.
  Under MGL.c231 s85, the comparative negligence statute, individuals can recover damages arising from an auto injury only
from parties whose percentage of fault is greater than their own.
  For example, Massachusetts collision claims arising from 1997 accidents show approximately 72% of the operators to be at
   Note that the claims per accident discussed here reflects to view of the PIP carriers only, not the combined number of PIP
claims made to all carriers.

                      Percent of Claimant and Fault Types by Claim per Accident
                                      Target and Express Claims
                Claims per Accident                 1         2       3 or more    Sum
                Operator of Insured Vehicle     81%       51%        32%         67%
                    At Fault                         24%        13%         7%       19%
                    Not at Fault                     57%        38%        25%       48%
                Passenger in Insured Vehicle 13%          47%        66%         29%
                    At Fault                          5%        11%        16%        8%
                    Not at Fault                      8%        36%        49%       21%
                Pedestrian/Bicyclist            5%        1%         0%          3%
                    At Fault                          4%         1%         0%        2%
                    Not at Fault                      1%         0%         0%        1%
                Other                           1%        1%         2%          1%
                    At Fault                          0%         0%         1%        0%
                    Not at Fault                      1%         1%         1%        1%
                All                             100%      100%       100%        100%
                    At Fault                         32%        25%        25%       30%
                    Not at Fault                     68%        75%        75%       70%
                           Note: At Fault is 50% or more fault for the operator.

                                                Table 7

4. Flow of Information

4.1 Routine Indicator Information

One valuable aspect of CSE is the ability to monitor the arrival of information over time. For
this purpose we always measure time in days from when the injury claim was reported to the
company. The timing of the routine indicators is of particular interest, especially if we hope to
base the decision to investigate, at least in part, on the presence of the indicators summarized
in the suspicion score.

CSE recorded the presence of 24 routine indicators, 23 of the usual fraud indicator variety and
the additional indicator of a “new” policy; i.e., not a renewal. The particular indicators were
chosen based upon the outcome of past studies and suspicion score model requirements.

Tables 8A and 8B display the list of 24 Routine Indicators and the proportions of those
indicators in CSE present at Closure, or on March 26, 1998 for the 80 open claims, and in the
most recent study of 1993 PIP claims for perspective.

                                             CLAIM SCREENING EXPERIMENT*
                                              Routine Level Indicators (3/26/98)
        Company      Claims     C1      C2     C3    C4      C5     C6      C7     C8    C9   C10   C11   C12
           A           364      1%     23%    7% 23% 15% 5% 12%                    8%   9%    71%   15%    7%
           B          1080      6%     13%    4%     6%      6%     1%     6%      4%   3%    39%    6%    5%
           C           489      2%     40%    12% 31% 25% 10% 15%                  1%   10%   69%    2%    1%
           D           416      1%     31%    16% 24% 16% 6% 32%                   6%   17%   67%    6%   12%
          ALL*        2349      4%     23%    8% 17% 13% 5% 14%                    4%   8%    55%    6%    6%
                                                STUDY OF 1993 PIP CLAIMS
                                                  Routine Level Indicators
        Company      Claims     C1      C2     C3    C4      C5     C6      C7     C8   C9    C10   C11   C12
           A          226       1%     43%    15% 5% 10% 22% 14%                   3%   7%    54%   51%   15%
           B          294       1%     43%    11% 3% 11% 21% 7%                    6%   9%    57%   65%   14%
           C          193       1%     23%    16% 5%         9% 21% 13%            2%   5%    52%   50%   12%
           D          107       1%     39%    22% 4%         9% 16% 16%            3%   8%    62%   64%   10%
          ALL         820       1%     38%    14% 4% 10% 21% 12%                   4%   7%    56%   58%   13%

* Includes target and express claims

                                              Routine Indicators
                           #1   No plausible explanation for accident
                           #2   Claimant in an old, low-value vehicle
                           #3   Claimant had a history of previous claims
                           #4   Claimant retained a “high-volume” attorney
                           #5   Was one of three or more claimants in vehicle
                           #6   Claimant appeared to be “claims-wise”
                           #7   Insured had a history of previous claims
                           #8   Insured was difficult to contact/uncooperative
                           #9   Accident occurred soon after policy effective date
                          #10   Injury consisted of strain or sprain only
                          #11   No objective evidence of injury
                          #12   Non-emergency treatment was delayed
                                                         Table 8A

                                                CLAIM SCREENING EXPERIMENT
                                                 Routine Level Indicators (3/26/98)
        Company       Claims     C13      C14     C15 C16        C17 C18 C19           C20    C21     C22     C23    New
           A            364      11%      7%      2%     18% 10%         7%     24%    6%     7%      1%      3%     25%
           B           1080      2%       8%      5%     11%     7%      3%     12%    2%     4%      0%      2%     44%
           C            489      1%       5%      2%     23% 18%         5%     34%    5%     3%      1%      0%     39%
           D            416      2%       10%     6%     24% 18%         5%     32%    3%     3%      0%      1%     20%
         ALL*          2349      3%       8%      4%     17% 11%         5%     22%    3%     4%      1%      2%     36%

                                                  STUDY OF 1993 PIP CLAIMS
                                                     Routine Level Indicators
        Company       Claims     C13      C14     C15 C16       C17 C18 C19            C20    C21     C22    C23     New
           A           226       1%       N/A     N/A 21% 11% 11%             4%       N/A    N/A     N/A    6%      28%
           B           294       3%       N/A     N/A 21% 12%           4%    2%       N/A    N/A     N/A    4%      39%
           C           193       2%       N/A     N/A 27% 13%           7%    5%       N/A    N/A     N/A    9%      19%
           D           107       5%       N/A     N/A 22% 15% 10%             1%       N/A    N/A     N/A    13%     22%
          ALL          820       2%       N/A     N/A 23% 12%           7%    3%       N/A    N/A     N/A    7%      29%

•    Includes target and express claims

                                                             Routine Indicators
                       #13        Unusual injury for this auto accident
                       #14        First notice of injury came from an attorney
                       #15        No injury indicated, but treatment began when lawyer got involved
                       #16        Large number of visits to chiropractor
                       #17        Large number of visits to physical therapist
                       #18        Unnecessary or excessive diagnostics
                       #19        Use of “high-volume” medical provider
                       #20        Medical bills vague, inconsistent, poor quality
                       #21        Medical diagnoses/treatment documentation appears “canned”
                       #22        Lost wages statement looked unofficial
                       #23        Long disability for a minor injury
                       New:       Claim was against a new policy
                                                            Table 8B

Evidently, there is a substantive difference between the perceptions of the expert coders of our
1993 Claim Study, generally claims managers with twenty plus years of experience, and the
current adjusters,11 many of whom had less than five years experience. Expert coders had the
luxury of viewing all available information from the (closed) file and forming opinions about the
claim. Claim adjusters, operating in a day-to-day environment with bits and pieces of
information flowing into the company, do not have that twenty-twenty hindsight viewpoint.

Tables 8A and 8B show some interesting, and important, differences compared to the 1993 PIP
study indicator frequencies. Indicators C4 (claimant retained a “high-volume” attorney), C6
(claimant appeared to be “claims-wise”), C11 (no objective evidence of injury), and C19 (use of
“high volume” medical provider) all showed sharp differences between CSE and the PIP Study.

    It is a misnomer to refer to the CSE coders as the adjuster. In reality, several people may contribute to the claim handling
information due to ordinary personnel changes such as transfers, terminations, illness, vacations and other facts of daily office

The difference in “claims-wise” observations reflects both the experience level and time
constraint differences. Since the prior studies, and the underlying suspicion score model,
placed significant weight on “claims-wise,” this could lead to biased-low routine suspicion
scores. The reliance on the subjective “claims-wise” may have to be curtailed or eliminated in
any future implementation.

The large increases in “high-volume” medical and legal providers reflects a definitional change
from a fixed list of high-volume providers for the PIP Study coders to a subjective “high-volume”
for the CSE adjusters. Both of the subjective “high volume” provider indicators play a
statistically significant role (see Sections 5 and 6) but related objective variables for buildup by
providers may provide equivalent information. So the type of practice, rather than the volume,
seems to drive the investigation decision at the moment.

4.2 Timing of Routine Indicators

The timing of the arrival of information is crucial to the usefulness of a suspicion scoring model.
The data gathered in the CSE indicate that the period 90-180 days is when the decision to
investigate is chiefly made.12 If we fix that interval of time as the (ideal) control point, then we
need to anticipate which indicators will be present for use in the suspicion scoring model.
Table 9 sorts the routine indicators into early (0-30 days), middle (0-90 days) and late (0-180
days) based upon when roughly 50 percent of those indicators arrive. Theoretically, we prefer
to rely on information generally available at the start of the decisionmaking period (90 days)
and either anticipate the arrival of later information through a probabilistic model13 or
reformulate the models to rely only on that information. These alternatives are beyond the
scope of this report (see Section 8).

  Section 5 further refines this observation to distinguish between ordinary and special investigations.
  For example, the probability that “a large number of chiropractor visits” will appear at the late stage may be assessed by
knowing who the chiropractor is at an earlier time and his/her past patterns of treatment.

                                             Timing of Information Received
      All Target Claims (N = 1486)                   % of those reported by day 360 that were reported by:
                                             7        30       90       180       270       360
      Suspicion Indicator Reported               360       Days    Days      Days     Days      Days       Days
      Any Routine Report Received                 1339     18.4%    50.9%    73.7%     93.7%     99.4% 100.0%
      Early Arriving Routine Variables
      Unusual injury for accident (inj11)           67     1.5%     85.1%    70.1%    92.5%   104.5%   100.0%
      No plausible explanation (acc09)              75     4.0%     70.7%    98.7%   109.3%    97.3%   100.0%
      Insured uncooperative (ins06)                 90     7.8%     68.9%    95.6%   100.0%   107.8%   100.0%
      No objective evidence (inj02)                126     3.2%     65.9%    67.5%    82.5%   100.8%   100.0%
      Insured previous claims (ins01)              269    24.5%     59.1%    71.0%    95.2%   102.2%   100.0%
      Soon after effective date (ins07)            146    26.7%     58.9%    78.1%    88.4%   100.7%   100.0%
      Strain or sprain only (inj01)               1013    10.4%     52.6%    68.8%    88.1%    97.2%   100.0%
      Three or more claimants (clt07)              271    11.8%     52.0%    68.3%    99.6%    99.3%   100.0%
      Old, low-value vehicle (acc10)               443    15.3%     49.7%    68.2%    93.2%    98.4%   100.0%
      First notice from attorney (inj13)           172     2.9%     49.4%    93.6%    88.4%    96.5%   100.0%
      Claimant previous claims (clt02)             168    14.3%     46.4%    64.3%    98.8%   100.6%   100.0%
      Non-emergency tx delayed (inj06)             121     9.1%     46.3%    78.5%    83.5%    98.3%   100.0%
      Middle Arriving Routine Variables
      Wage statement unofficial (lw05)              10    10.0%     50.0%   110.0%    90.0%   100.0%   100.0%
      Treatment after attorney (inj14)              83     1.2%     34.9%   100.0%    95.2%   100.0%   100.0%
      Med Dx/Tx appear canned (trt13)               90     3.3%     36.7%    61.1%    73.3%    92.2%   100.0%
      Appeared claim-wise (clt11)                  100    12.0%     33.0%    61.0%   100.0%   101.0%   100.0%
      High-Volume atty retained (clt05)            385     6.0%     37.7%    55.8%    93.0%    97.1%   100.0%
      Late Arriving Routine Variables
      Unnecessary or excess Dx (trt10)             110     1.8%     28.2%    47.3%    79.1%    97.3%   100.0%
      Vague, Inconsist med bills (trt12)            79     2.5%     27.8%    36.7%    78.5%    93.7%   100.0%
      High-Volume med provider (trt05)             494     3.6%     28.3%    44.7%    78.1%    93.5%   100.0%
      Large no. chiro visits (trt01)               377     2.7%     23.9%    44.3%    76.4%    93.6%   100.0%
      Long disability for minor inj (lw06)          39     7.7%     41.0%    76.9%    74.4%    94.9%   100.0%
      Large no. PT visits (trt03)                  260     3.1%     17.3%    35.8%    69.6%    93.8%   100.0%
                                                         Table 9

The indicator groupings make intuitive sense. The early indicators concern the accident itself,
and the notice from an attorney, when present. The middle indicators represent the flow of
medical information, while the late arrivals tend to be those associated with prolonged or
excessive treatment. While the appearance of “high-volume” medical providers is categorized
as late, we hypothesize the arrival of these indicators to be associated with (delayed) packaged
claims in which the attorney is known first (early and middle) prior to the medical package
arriving sometime after 90 days.

4.3 Suspicion Model Information

Table 10 shows that the functioning of the suspicion score regression models in the CSE is
broadly consistent with that of the 1993 PIP Study for closed claims. Both high (7-10) and
moderate (4-6) score levels are approximately the same overall but quite different by company.
Notice that the range of High/Moderate is from 39 percent for company C to 13 percent for
Company B. The low and zero scores should be viewed on a combined basis since the scoring
model implemented in CSE rounded many single indicator claims up to 1.0.

                                 CLAIM SCREENING EXPERIMENT
                                  Model Scoring Levels (3/26/98)
                   Company         Claims     H        M         L       Zero
                         A             352      8%      15%       58%      19%
                         B            1065      3%      10%       32%      55%
                         C             483      5%      31%       49%      16%
                         D             369      4%      30%       53%      14%
                        ALL           2269      4%      18%       43%      35%
                                   STUDY OF 1993 PIP CLAIMS
                                 Interactive Model Scoring Levels
                                                  Interactive Routine Model
                   Company         Claims      H         M         L      Zero
                         A             226      4%      20%       25%      51%
                         B             294      3%      17%       27%      53%
                         C             193      3%      23%       20%      54%
                         D             107      7%      15%       29%      49%
                        ALL            820      4%      19%       25%      52%
                                   STUDY OF 1993 PIP CLAIMS
                                     Coder Suspicion Levels
                   Company         Claims     H        M         L       Zero
                         A             226      8%      17%       15%      60%
                         B             294      8%      22%       14%      56%
                         C             193      5%      20%       25%      50%
                         D             107      5%      24%       15%      56%
                        ALL            820      7%      20%       17%      56%
                                             Table 10

We next turn to the investigative process and its interaction with the CSE design and suspicion
score levels.

5. Utilization of Investigative Techniques

5.1 Types of Investigation Applied

There are several types of activities employed by insurers to investigate claims. Some of these
techniques are used routinely and not necessarily to identify fraud. This investigation is
intended primarily to clarify facts necessary to determine the payment. Other investigative
techniques are focused more specifically on identifying or confirming suspicion.

In this report, we will refer to the first type of investigation as “ordinary” and the second as
“special” investigation. We have divided specific investigative techniques into two categories as
shown in Table 11.

                                         Types of Investigation
                    Ordinary Investigation                    Special Investigation
           Independent Medical Examination (IME) Referral to Special Investigation Unit (SIU)
           Medical Audit                            Examination under Oath (EUO)
           Site Investigation                       Surveillance/ Activity Check
                                                    Accident Reconstruction
                                               Table 11

An independent medical examination (IME) is the most commonly utilized technique. The main
purpose of an IME is to determine the extent of injury in cases with substantial or prolonged
medical treatment. An independent assessment can sometimes provide a basis for questioning
the appropriateness or necessity of continuing treatment. An IME can also be useful if there
exists suspicion regarding the existence or extent of injury. However, the function of the IME is
much broader than to detect fraud.

Medical audits are increasingly being applied to scrutinize medical bills submitted. Once again,
there need be no question of fraud, only a desire to minimize excessive and exclude erroneous
charges. Accident site investigations are most commonly employed to clarify the issue of fault.
Rarely are site investigations invoked because of claim suspicion.

We treat referral to a special investigation unit (SIU) as a technique, although several specific
activities may be conducted within the SIU. An examination under oath (EUO), also called a
sworn statement, can be performed by the SIU or outside. The subject of an EUO might be the
claimant, insured or a witness.

The purpose of surveillance or an activity check is to determine whether the claimant has
misrepresented the extent of disability. Clearly, these would only be applied when there exists
suspicion. Finally, in rare instances, an accident reconstruction expert can be used if there are
serious questions about whether the accident occurred as alleged and/or could have caused
the alleged injury.

5.2 Frequency of Investigation

Table 12 shows the percentages of claims that were investigated using various techniques.
For perspective, these percentages have been compared against the 1993 PIP Study data. The
CSE data include both Target and Express claims, but not Duds. This 2269 sample is
comparable to the 1993 PIP Study sample of 1400 claims.

                                Comparison of Investigation Techniques
                                         1993 PIP Study Vs. CSE
                                       1993 PIP Study                    Claim Screen Experiment
                             Number Percentage % Investigated Number Percentage % Investigated
          Technique         of Claims of Claims         Claims     of Claims of Claims      Claims
 Ordinary Investigation            274       19.6%          85.6%        638      28.1%          90.1%
 Special Investigation              46        3.3%          14.4%         70       3.1%           9.9%
 IME Requested                     n/a          n/a            n/a       505      22.3%          71.3%
 IME Performed                     228       16.3%          71.3%        449      19.8%          63.4%
 Site Investigation                 46        3.3%          14.4%         17       0.7%           2.4%
 Medical Audit                      41        2.9%          12.8%        305      13.4%          43.1%
 SIU Referral                       33        2.4%          10.3%         58       2.6%           8.2%
 Activity Check                     10        0.7%           3.1%          7       0.3%           1.0%
 Claimant Sworn Statement            8        0.6%           2.5%         13       0.6%           1.8%
 Insured Sworn Statement             7        0.5%           2.2%         11       0.5%           1.6%
 Witness Sworn Statement             2        0.1%           0.6%          2       0.1%           0.3%
 Accident Reconstruction           n/a          n/a            n/a         4       0.2%           0.6%
 Surveillance                      n/a          n/a            n/a         7       0.3%           1.0%
 Total Claims                    1400                                   2269
 Total Investigated Claims         320                                   708
 Note: CSE data are based on Target and AIB Express Claims.
                                                 Table 12

The general profile of investigative techniques mirrors that for the 1993 claims. IMEs are by far
the most frequently employed, followed by medical audits and SIUs. Other techniques are
rarely used. Note that the ratio of IMEs to medical audits is lower than in 1993, reflecting the
observed increase in the use of audits.

5.3 Investigation versus Suspicion

Our previous research indicated that the level of claim suspicion was related to the likelihood
of investigation. Table 13 displays the relationship between investigation and the routine
suspicion score. This analysis is restricted to Target claims.

                      Routine Suspicion Score and Investigation Comparison
                                      All Closed Target Claims
                     Suspicion                                         Number of
                       Level         Investigation Level at Final       Claims
                                   None        Ordinary        Special
                         None       89%             11%            0%        245
                           Low      54%             41%            5%        660
                      Medium        32%             63%            5%        403
                          High      26%             58%           16%         98
                             All    52%             43%            5%       1406
                                  Test Level Closed Target Claims
                                                                       Number of
                                     Investigation Level at Final       Claims
                                   None        Ordinary        Special
                         None       90%             10%            0%        121
                           Low      52%             42%            5%        309
                      Medium        33%             62%            5%        183
                          High      24%             61%           15%         54
                             All    52%             43%            5%        667
                                 Control Level Closed Target Claims
                                                                       Number of
                                     Investigation Level at Final       Claims
                                   None        Ordinary        Special
                         None       89%             11%            0%        124
                           Low      56%             40%            5%        351
                      Medium        31%             64%            5%        220
                          High      27%             55%           18%         44
                             All    52%             43%            5%        739
                                              Table 13

Overall, 48% percent of the Target claims had some form of investigation. Of these 43 percent
received ordinary investigation and 5 percent special investigation. The likelihood of
investigation was clearly related to the level of suspicion. Only 11 percent of claims with no
suspicion and 46 percent of those with low suspicion were investigated. However, 68 percent of
moderately suspicious claims and 74 percent of highly suspicious claims were investigated.

When restricted to special investigation, a similar trend exists. However, only the highly
suspicious claims have a substantial probability (16%) of special investigation. Both low and
moderate suspicion resulted in a 5 percent rate.

These results suggest that suspicion certainly plays an important role in decisions regarding
investigation. However, it is surprising that so few suspicious claims were investigated. Only
about half of the claims found by the model to be highly suspicious were subjected even to
ordinary investigation (primarily an IME), and only 16 percent to a special investigation.

5.4 Timing of Investigation

In Section 4 we discussed the timing of information that becomes available to the company.
Indicators about the accident and claimant tends to arrive first, followed by indicators related to
the injury and attorney, and finally indicators about lengthy medical treatment. This pattern of
information flow can be compared with the timing of investigation.

Figure 4 shows the relationship between information and investigation for the Target claims.
The higher line on Figure 4 graphs the number of claims for which at least one routine report
(Form 1) was received as of various time points (7 days, 30 days, 60 day, etc.). The lower line
shows the number of claims with at least one investigative report. In theory, a report should
have been completed shortly after the information first became available. In fact, there may
have been a substantial lag in some cases. Thus, the graphs can be regarded as
approximations that are biased low.

                           Timing of Information and Investigation for Target Claims

                100%                                                                                                                   1270



                              Any Routine Report Received





                                                    Any Investigative Report Received

                  0 Days         7 Days         30 Days          60 Days          90 Days         180 Days            270 Days   360 Days

                                                Any Routine Report Received       Any Investigative Report Received

                                                                  Figure 4

Figure 4 suggests that the timing of investigation tends to track quite closely the receipt of
information. As expected, investigation lags somewhat behind receipt of the fraud indicators.
Note especially the relatively small number of investigations prior to 90 days, despite the
availability of at least one routine report for most claims. There is, however, a jump in
investigations after 90 days.

Figure 5 displays the timing of investigation by type of investigation for all Target claims that
had some form of investigation within 360 days. The graph shows the number of claims with at
least one technique by the indicated time point (7 days, 30 days, etc.) The three lines represent
any technique, ordinary technique and special technique.

                                 Timing of Investigation Decision for Target Claims

                 100%                                                                                                       399



                                                                            Any Technique

                                                                                               Ordinary Technique



                  20%                                                                           Special Technique

                    0 Days      7 Days      30 Days          60 Days        90 Days         180 Days       270 Days   360 Days

                                             Any Technique      Ordinary Technique     Special Techinque

                                                                Figure 5

The pattern for all investigations is dominated by the ordinary techniques, which constitute the
bulk of ordinary investigations, and especially by IMEs. It is clear that IMEs occur typically
between 90 and 180 days after the injury report. This is consistent with the flow of information,
since details of medical treatment are often unavailable until at least 90 days. Thus, the IME is
used in a largely reactive manner to curtail further treatment. One possible benefit of early
detection of suspicion might be to allow a more timely intervention in cases where buildup or
fraud can reliably be anticipated.14

For special investigation techniques, the pattern is quite different. The investigation might be
invoked at any point, and a relatively higher proportion of special investigations occur earlier.
Since special investigations typically center on instances of planned fraud, this result is also
consistent with the flow of information. That is, fraud indicators pertaining to the accident and
claimant are likely to arrive early and perhaps trigger suspicion about a staged or caused
accident. Also information not captured by our set of fraud indicators may play a role in these
specific cases.

Table 12 examines the patterns of both suspicion and investigation over time. This table is
restricted to Target claims for which a routine report was filed. At each time point, the status of

  These situations could be indicated by high build-up scores for the medical providers, the attorney, if any, or the premium
town. See generally, sections 6 and 7.

each claim has been determined. Of particular interest is the extent to which suspicious claims
have been investigated. It appears that investigation lags behind the (potential) knowledge of
suspicion. For example, at 30 days 2.4 percent were investigated, but another 4.5 percent were
already found to be moderately suspicious. Arguably, only one-third of the claims that qualified
for an investigation were actually flagged. Up to 90 days, this apparent deficit was maintained.
This finding raises the possibility that many additional fruitful investigations could occur much
earlier on.

                     Patterns of Suspicion and Investigation Report Date to 360 Days
                   Status           7 Days    30 Days 90 Days 180 Days 270 Days 360 Days
         No routine report               922       575       351         81        11   0
         Susp= 0, not investigated        52        83        95         74        71  67
         Susp=L, not investigated        188       447       526       527        510 507
         Susp=M, not investigated         16        53       113       170        188 189
         Susp= H, not investigated         0         3        19         38        33  28
         Ordinary Investigation            6        16        60       258        329 347
         Special Investigation             6        13        26         42        48  52

         Total                        1190        1190      1190      1190    1190    1190
                                     Profile of Status Versus Time
         No routine report           77.5%     48.3%    29.5%         6.8%    0.9%    0.0%
         Susp= 0, not investigated    4.4%      7.0%     8.0%         6.2%    6.0%    5.6%
         Susp=L, not investigated    15.8%     37.6%    44.2%        44.3%   42.9%   42.6%
         Susp=M, not investigated     1.3%      4.5%     9.5%        14.3%   15.8%   15.9%
         Susp= H, not investigated    0.0%      0.3%     1.6%         3.2%    2.8%    2.4%
         Ordinary Investigation       0.5%      1.3%     5.0%        21.7%   27.6%   29.2%
         Special Investigation        0.5%      1.1%     2.2%         3.5%    4.0%    4.4%
         Total                        100%      100%     100%         100%    100%    100%
                                               Table 14

5.5 Treatment versus Control

Increasing the number of appropriate investigations was a goal of the CSE. Therefore, we
expected that the test claims would have more investigations. Table 13 shows the relationship
between suspicion and investigation separately for test and control claims. The results are
remarkably similar. It appears that the feedback of suspicion scores to adjusters had no effect.
Claim screening per se, without other changes in systems and procedures may not be sufficient
to increase investigation.

6. Multivariate Modeling of Investigation

6.1 Approach to Model-Building

We know that investigations are prompted in part by the level of suspicion evidenced. However,
suspicion is not the only factor considered by adjusters. To understand the decision-making
process in more detail, we have developed multivariate statistical models. The main dependent

variable (outcome) of interest is whether or not an investigation occurred. Because this
dependent variable is a simple dichotomy (yes or no), logistic regression is the most
appropriate statistical technique to apply.

In our prior study of PIP claims arising from 1993 accidents15, we modeled the investigation
process as dependent upon:

•    Suspicion of fraud and build-up
•    Possibility of subrogation to another carrier
•    Financial exposure (BI) to the PIP carrier
•    Coordination of medical payment with health insurer
•    The size of the (expected) claim payment

Overall, we found the likelihood of investigation increases with the suspicion level16 and the
financial exposure to a possible BI claim. Neither the claim payment size, the possibility of
subrogation, nor potential coordination with health insurers provided significant contributions to
the overall models. For IMEs, however, both suspicion scores and the size of claim payments
did correlate with IME likelihood.

Using logistic regression, we adopt a “stepwise” approach, in which various combinations of
predictors were tried. We present several final models that we believe best describe the
investigation process. The following variables were considered for inclusion in the models:

•    Routine suspicion score at close
•    All individual routine fraud indicators
•    Availability of coordinated health insurance
•    Injury type
•    Medical claim payment
•    Whether the medical bills involved emergency treatment only
•    Whether or not the claimant could file a BI claim
•    Medical provider Buildup Profile average score
•    Attorney Buildup Profile average score
•    Premium Town Buildup Profile average score
•    Test versus Control group
•    Company operational differences

The models were based on target and express claims combined, and were further restricted to
claims with the necessary data. In particular, some of these variables were only available for
DCD matched claims. Therefore, the quantitative results should be interpreted with the
realization that the claim population underlying specific models may be different.

   Derrig and Weisberg (1995), "A Report on the AIB Study of 1993 Personal Injury Protection Claims - Identification and
Investigation of Suspicious Claims", AIB Filing on Cost Containment and Fraudulent Claims Payment Filing, (Docket R95-12),
   Four different suspicion models based upon fraud indicators were tested, 10-variable, 20-variable and interactive regression
models and a CART model, all with similar results.

The Buildup Profile17 is a measure of apparent buildup that can be calculated for each claim
based on its DCD data. The AIB computes and updates for all medical providers, attorneys and
towns the average value of this measure. This average can be interpreted as a rough index of
the extent of buildup among the claims for that provider or locale. We use the index based
upon 1995 closed claims as a proxy for the information available to the adjuster, or branch
office, on patterns of practice current at the time the claim is opened. This information type
variable was not available at the time of the study of 1993 PIP claims cited above.

6.2 Determinants of Investigation

We first consider the logistic model to explain what factors increase the chances of any
investigation. Table 15 summarizes results for five sets of models. The table shows the
concordance, the intercept parameter, and the odds ratio of each explanatory variable
statistically significant at the .10 level. We interpret the concordance percentages as relative
goodness-of-fit parameters for the models.

The odds ratio shown for the significant variables, such as Strain (=1 if strain or sprain claim, 0
otherwise), provides the relationship of the likelihood of a positive outcome to a negative
outcome in the presence of the parameter. For example, if the Strain odds ratio is 1.71, this
means that investigation is 71 percent more likely for strain and sprain claims than other claims,
all other model parameters equal. For continuous variables such as suspicion score or medical
loss payment, the odds ratio value represents approximately the percentage increase in
positive outcome probability per unit increase in the continuous variable.

  The zero to six buildup profile score is based on the unweighted sum of six objective factors identified to be closely related
with an expert judgment of build-up. See Derrig and Weisberg, 1996 A Report on the AIB Study of 1993 Personal Injury
Protection and Bodily Injury Liability Claims, Coping with the Influx of Strain and Sprain Claims, on Cost Containment and
Fraudulent Claims Payment, DOI Docket R96-36, July 12, Boston, page 65.

                                                           Decision to Investigate
                                     Concordance, Intercept and Odds Ratio Effect of Significant Variables
I. Target + Express Claims (2269)
          Dep. Variable = Any Investigation
Medical (Count)              Concord     Intrcpt     Suspicion    BI-Same       BI-Other       Strain        Emer only   Med_Loss   Build-
A. All (2269)                85%         -2.2        1.54         1.34          X              1.71          0.29        1.16       X
B. All/PrimMed (1116)        79%         -3.8        1.32         X             X              X             X           1.10       2.41
C. All/Town (2111)           85%         -3.2        1.51         X             X              1.59          0.30        1.19       1.55
D. All/Atty (768)            66%         -0.1        1.32         X             0.62           X             0.31        X          X

II. Target Claims (1406)
           Dep. Variable = Any Investigation
A. All (1406)                71%           -1.1      1.38         X             X              1.56          0.35        X           X
B. All/PrimMed (889)         71%           -2.6      1.26         X             X              X             X           X           2.00
C. All/Town (1293)           75%           -1.9      1.38         X             X              1.36          0.34        X           1.38
D. All/Atty (734)            67%           0.2       1.30         X             0.59           X             X           .92         X

III. Target + Express Claims (2269)
            Dep. Variable = Any Investigation, Add Co. Dummies + Routine Indicators
A. All (2269)                 88%          -1.8     1.26        X               X              1.67          0.28        1.22        X
B. All/Prim.Med (1116)       85%         -2.9       1.26         X              X              X             X           1.13        2.27
C. All/Town (2111)           90%         -1.2       1.25         X              X              1.31          0.22        1.24        X
D. All/Atty (768)            79%         0.9        1.14         X              0.72           X             0.24        X           X

IV. Target Claims (1406)
           Dep. Variable = Any Investigation Add Company + Routine Ind: Dummies
A. All (1406)                80%          -0.8     1.27         X               X              1.60          0.33        X           X
B. All/Prim Med (889)        81%         -1.8       1.26         X              X              X             X           X           2.03
C. All/Town (1293)           83%         -0.2       1.21         X              X              X             0.28        1.07        X
D. All/Att (734)             78%         0.8        1.43         X              0.63           X             0.43        X           X

V. Target Claims (1406)
           Dep. Variable = IME Requested; Add Company + Routine Ind: Dummies
A. All (1406)                75%       -2.0      1.16         X              X                 1.48          0.34        1.0         X
B. All/Prim Med (889)        73%         -2.8       1.17         X              X              X             X           X           1.66
C. All/Town (1293)           75%         -1.0       1.15         X              X              X             0.31        1.01        0.76
D. All/Att (734)             65%         -0.7       1.11         X              X              X             X           X           X

          Note:      Model B adds Primary Medical Provider Build-up Score, Model C adds Town Build-up Score and Model D adds
                     Attorney Build-up Score. Medical Loss is expressed in Thousands of Dollars.
                                                                 Table 15

For each of the final combinations of claim sample (target and express or target only),
dependent variable (any investigation or IME requested), and independent variable set (with or
without company and routine indicator dummies), model parameters are shown without any
build-up profile scores (A) or with the build-up scores for primary medical provider (B), premium
town (C), and attorney (D).

The models in Table 15 apply to all claims, target and express, or to target claims only as
indicated. The first four sets of models use the decision to do any investigation as the
dependent variable while the fifth panel analyzes the IME technique. The parameter values in
Table 15 reflect the principal variables, with the effect of company and routine indicator
dummies shown separately in Table 16.

Strain and sprain claims are about 60 percent more likely to be investigated than other injury
types. Claims with emergency treatment only are substantially less likely to be investigated
than those with outpatient treatment where the potential for build-up exists. The size of the
medical payment (Med-Loss) and the financial exposure (BI same) are significant only when
express claims are included in the sample.

The average build-up score for the primary medical provider (models B) provides a substantial
summary distinction among investigated claims. Model IB shows a 25 percent probability of
investigation of a typical $2,500 claim. A target claim with a suspicion score of 5 (moderate)
and medical build-up score of 3 would be 4 times more likely to be investigated than a target
claim with a suspicion score of 2 (low) and a medical build-up score of 2. Evidently, our two
summary statistics, suspicion score and medical provider build-up score, give good coverage to
the range of probabilities to investigate target claims.

Table 16 displays the additional odds ratio effects of including company and routine fraud
indicator dummies in the logistic models. The company ratios, when any investigation is the
dependent variable, all reflect the large proportion of medical audits performed by company 4.
Model VA shows that company 2 does more investigations and company 1 does fewer
investigations than companies 3 and 4, all other parameters equal.

We interpret the odds ratios of the routine indicators as representative of the relative weights
given to the indicator by the adjuster in relation to the weight in the suspicion model. For
example, adjusters placed much less emphasis on M150, no plausible explanation for the
accident, than did the suspicion model. On the other hand, adjusters gave much greater weight
to M195, strain and sprain only claims, than our model did. These observations can lead to
suspicion model refinements provided they point toward investigations that produce favorable
outcomes. We turn to the question of what determines favorable outcomes of investigations.

                                                                 Decision to Investigate
                                              Log Odds Effect of Company and Routine Indicator Dummies

                                 Susp            Target + Express (2269)                  Target (1406)                 Target (IME) (1406)
Dummy Var.                       Weight   IIIA     IIIB      IIIC     IIID        IVA    IVB      IVC     IVD    VA        VB       VC        VD

Company 1                        X        0.10     0.07      0.06     0.06        0.13   0.07     0.08    0.08   0.58      0.38     0.44      0.48
Company 2                        X        0.33     0.27      0.20     0.27        0.46   0.32     0.26    0.30   1.34      X        X         X
Company 3                        X        0.28     0.27      0.17     0.21        0.33   0.25     0.20    0.25   X         X        X         X
M150 No plaus explanation        1.6      0.58     0.46      0.50     0.39        0.48   0.34     0.41    0.26   0.44      0.53     0.45      0.54
M155 Clmnt old, low value        0.3      X        X         X        X           X      X        X       X      X         X        X         X
M160 Clmt history prev           0.2      X        X         X        X           X      X        X       X      1.45      2.03     1.60      X
M165 Clmnt high volume           0.8      1.92     X         1.70     1.68        1.69   X        1.51    X      X         X        X         X
M170 3 or more claimants         0.5      1.46     X         1.58     X           1.35   X        1.51    X      X         X        1.48      X
M175 Clmnt appear claim          1.3      X        X         X        2.81        X      X        X       X      X         X        X         X
M180 Insd history prev claims    X        1.73     1.74      2.00     X           1.59   1.84     1.88    1.62   1.42      X        1.50      1.58
M185 Ins dif to                  1.8      X        X         X        X           X      X        X       X      X         X        X         X
M190 Acc soon after pol eff      X        X        1.98      X        X           X      1.87     X       X      X         1.86     X         X
M195 St/Sprain only              0.4      2.86     2.08      3.17     1.62        2.08   1.68     2.60    1.58   1.56      1.45     1.76      X
M200 No objective evid inj       X        X        X         X        X           X      X        X       X      X         X        X         X
M205 Non-Emer treat delayed      0.3      0.53     0.35      0.47     0.44        0.45   0.38     0.47    0.45   0.56      X        X         X
M210 Unusual injury for Acc      1.6      2.16     3.15      2.19     2.37        2.13   3.05     2.15    X      2.64      3.39     2.67      2.82
M215 First Notice from           X        3.41     1.83      2.49     X           2.13   X        1.91    X      1.48      X        X         X
M220 Treat began with            X        0.51     X         X        X           X      X        X       X      X         X        X         X
M225 Large # Chiro visits        2.6      X        X         X        X           X      X        X       0.56   X         X        X         X
M230 Large # PT visits           1.8      X        X         X        X           X      X        X       X      X         X        X         X
M235 Unness/Excess               0.9      0.60     X         X        X           0.51   X        X       X      X         X        X         X
M240 High volume med prov        0.8      1.46     X         1.54     1.52        X      X        1.45    X      1.52      X        1.51      1.40
M245 Med bills vague/incon       X        X        X         X        X           X      X        X       X      X         X        X         X
M250 Med diag/treat canned       X        0.52     X         0.37     X           X      X        0.50    X      X         X        X         X
M255 Lost wages look             1.3      .27      0.18      0.25     X           X      0.19     0.24    X      X         X        X         X
M260 Long disability minor inj   0.8      X        X         X        X           0.47   X        X       0.39   X         X        X         X
Mnew New Policy                  X        X        X         X        X           X      X        X       X      X         X        X         X

      Note: Suspicion Model Parameters are from the 20 variable dominant Model.
                                                                      Table 16

      6.3 Determinants of Favorable Outcomes

Investigated claims may or may not produce results advantageous to the carrier. In this section
we explore the characteristics of the claims with a favorable outcome.

The dependent variable was whether or not the DCD report indicated that billing or treatment
was curtailed or damages mitigated. Note that these outcomes were self-reported by the
companies and represent their own assessments of success. It is possible that company
reporting practices may not be completely uniform.

Our analysis concentrates on IMEs. Our data set consists of only target claims for which the
company DCD reporting of IME requested was correct, providing a subsample of 440 target
claims. The results, in the absence of routine indicator dummies, are shown in Table 17 using
two definitions of favorable. One, we call "soft" favorable uses the qualitative responses of
treatment curtailed or damages mitigated. We also report a "hard" favorable, claims where
dollar savings were reported as the difference between medical bills and medical payments.

                                   IME with a Favorable Outcome
                                   Conditioned on IME Requested
                             Parameters for Favorable Outcome Model
                                Variable            Value/Odds Ratio
                                                   Soft         Hard
                         Intercept                     0.85          -0.78
                         Suspicion                     1.15           1.16
                         Medical Payment               0.84           0.86
                         Company 3                        X           0.33
                                              Table 17

The suspicion score plays a role in determining favorable outcomes. Moderately suspicious
claims have about a 50 percent higher likelihood of a favorable, soft or hard outcome, than low
suspicious claims. The medical size of claim variable is more problematical with coefficients
that imply fewer favorable results for increasing claim payments. This may reflect some
circularity (favorable outcome claim payments have lower claim payments) or may indicate that
the middle level claims, usually strains and sprains, are those most fruitful for investigation.
Company 3 appears by virtue of the differing outcomes to have underreported the dollar
savings values in DCD records. All other principal variables in Table 15 were not significant.

To test the circularity hypothesis, we also calculated model values excluding the medical loss
variable, since both med-loss and BI same were significant variables in the decision to
investigate (Table 15, Model IA above). We expect the BI same company or BI other company
variables to then become significant. Table 18 shows the results of those regressions.

                                   IME with a Favorable Outcome
                                   Conditioned on IME Requested
                              Parameters for Favorable Outcome Model
                                  Variable          Value/Odds Ratio
                                                     Soft       Hard
                           Intercept                     0.70      -0.78
                           Suspicion                     1.14       1.16
                           BI Same                       0.50       0.57
                           BI Other                      0.67          X
                           Company 1                        X       0.18
                           Company 2                        X       0.64
                           Company 3                     0.42          X
                                              Table 18

Table 18 shows unanticipated results for the BI variables; i.e., fewer favorable outcomes
occurred on the BI eligible PIP claims than on those not reported as eligible. A review of the
data underlying the 440 claim subsample showed that to be true. It may be the case that
additional savings accumulate to the BI claim and are not reported here, the IME prevented the
claim from exceeding the tort threshold or there may be another explanation of which we are
unaware. A review of all associated BI claims may shed light on this result in a follow-up study.
(See Section 9).

7. Benefits and Cost of Investigation

Are the usual investigative techniques cost-effective? This section addresses this critical issue.
The primary focus will be on independent medical examinations, which account for the majority
of total investigative expenses. We derive estimates of the savings attributable to IMEs, and
the loss adjustment expense incurred.

7.1 Expenses Associated with Investigation

The CSE adjusters recorded data on the presence of various investigative techniques, and the
total LAE for the claim. In theory, the LAE should equal the sum of the specific expenses
associated with each of the individual techniques plus a small amount of miscellaneous
expense. We can represent the relationship between the total LAE and the techniques present
as a linear function:

LAE = M + T1*C1 + T2*C2+ .....

Here M represents the miscellaneous expense and Ti is an indicator of whether or not
investigative technique i occurred. Ti has the value 1 if the technique was used and zero
otherwise. Ci represents the cost associated with technique i.

To estimate the averages for the Ci’s, we can employ multiple linear regression. This approach
has been taken for the total sample of all claims with non-zero reported LAE, and for each of
the four CSE companies individually. The results are displayed in Table 19.

                               Regression Estimated Investigative Costs
                  Technique       Company Company          Company     Company    Total
                                      1           2            3          4
              IME No Show                80         119                     124      92
              IME Completed             371         349          268        340     325
              EUO                       338         733                             600
              Surveillance              309         275                             233
              Acc. Reconstr.            661                                         370
              Med Audit                             166          173                 73
              Miscellaneous              21          29          104         52      49
                                              Table 19

The only investigative “technique” omitted from this table is referral to an SIU. We did not have
data for the expenses associated with SIU investigation, since the amount of such expenses
was not generally known to the adjuster recording the expense information.

Note that these averages represent per-claim amounts, not per-unit costs. For example, the
average of $121 for an IME “no show” is the average payment to the IME provider, and may
include charges for multiple appointments.

Note also that some of the variability among the companies is attributable to the small numbers
upon which some of these regression coefficients were based. Blank cells represent cases in
which the sample was too small to produce a statistically significant regression coefficient.

As a reality check on these results, we asked the participating companies for information about
their typical costs for various investigative activities. The results are summarized in Table 20.
Three companies replied by the time of this report.

                           Company Information on Unit Investigative Costs
                          Technique      Company        Company       Company
                                              2             3            4
                      IME No Show                 75             80          75
                      IME Completed              270           350          350
                      EUO                        428           400          900
                      Surveillance               275           464          900
                      Acc. Reconstr.             615           -----       1000
                      Med Audit               38/195             58     29/300
             Note: Company averages will depend upon mix of medical audit techniques
                                             Table 20

These company estimates are quite consistent with the empirical estimates we derived using
regression. Note that the per-unit cost for an IME no-show of $75 to $80 is approximately 2/3
of the $121 per-claim charge. Thus, an average of 1.5 no-shows per claim is implied.

7.2 Supplementary File Review

Our preliminary examination of the data revealed that IMEs were often requested when the
apparent suspicion level was low. To understand in more detail the IME process, we collected
supplementary data on a sub- sample of 175 CSE claims. This sample included approximately
equal numbers of the following:

•   Low suspicion and no IME requested
•   Low suspicion and IME requested
•   Moderate/High suspicion and no IME requested
•   Moderate/High suspicion and IME requested

The first of these categories was large and required a random sample to select approximately
45 claims. For the other three categories, we attempted to obtain data on virtually all claims in
the category.

Once the supplementary sample was selected, the claim files were obtained and reviewed by a
group of experienced claim managers. A Supplementary Data Collection form was completed
for each file. A copy of this form can be found in the Appendix.

The primary purpose of this form was to capture detailed information on the IME process. In
particular, how does the decision to use an IME relate to the level of claim suspicion? Are

IMEs being utilized enough on high suspicion claims and too much on low suspicion claims? Is
there potential for significant improvement?

7.3 Supplementary Sample Analysis

To analyze the role of claim suspicion relative to IMEs, we divided the supplementary sample
into four subsets based on the suspicion level according to the statistical model:

•   Essentially no suspicion (score = 0 or 1)
•   Low suspicion (score = 2 or 3)
•   Medium suspicion (score = 4 to 6)
•   High suspicion (score = 7+ )

For each of the four levels, we further divided claims according to four possible outcomes with
respect to the use of IMEs:

•   IME not requested
•   IME requested but not completed
•   IME completed with positive outcome
•   IME completed with negative outcome

A positive outcome for the IME was defined to be any situation in which either the submitted
bills were reduced, future billing or treatment was curtailed, or damages were mitigated for the
associated BI claim.

For each of the four suspicion levels and outcomes we have estimated several characteristics:

•   Average indemnity payment (medical + wages + other)
•   Average loss adjustment expense (LAE)
•   Percent that deserve IME (per supplementary sample coder)
•   Average days from first report to IME request
•   Percent of IME requests deemed to be “timely”

The results are shown in Table 21. In interpreting these results, we must be aware that the
sample is deliberately “biased” to over-represent claims that are suspicious and/or on which
IMEs were requested. Consequently, this data cannot be used to estimate directly various
quantitative relationships, but rather to generate hypotheses that can be tested in the overall
CSE claim population. Moreover, the small sample sizes also require caution in the
interpretation of statistics in the table.

                                           CSE Supplementary Sample
                                              IME Characteristics
          Suspicion = None ( 0 -1 )
                                       Claims    Percent    Indemnity     LAE  % Should Avg.Lag    %IME
                                                                                 IME     to IME    Timely
          IME not Requested                 25       57%         $1,493    $14       8%
          IME Req but not                    0        0%
          IME Req and Completed
              Positive Outcome              10       23%         $4,510   $476      90%       77     67%
              Negative Outcome               9       20%         $2,984   $375      67%      111     50%
      Total                                 44      100%         $2,484   $193      39%       93     59%
      Suspicion = Low ( 2-3)
                                       Claims    Percent    Indemnity     LAE  % Should Avg. Lag %IME
                                                                                 IME     to IME  Timely
          IME not Requested                 18       46%         $2,292    $41      44%
          IME Req but not                    2        5%         $4,486   $280     100%
          IME Req and Completed
              Positive Outcome              10       26%         $2,734   $377      90%       75     77%
              Negative Outcome               9       23%         $3,253   $467      78%       82     44%
      Total                                 39      100%         $2,740   $238      67%       78     61%
      Suspicion = Medium ( 4-6)
                                       Claims    Percent    Indemnity     LAE  % Should Avg. Lag %IME
                                                                                 IME     to IME  Timely
          IME not Requested                 26       47%         $2,844   $145      73%
          IME Req but not                    1        2%             $0    $75     100%
          IME Req and Completed
              Positive Outcome              13       24%         $2,971   $362     100%       63     80%
              Negative Outcome              15       27%         $4,729   $404      93%       72     53%
      Total                                 55      100%         $3,336   $266      85%       68     66%
      Suspicion = High ( 7 -10 )
                                       Claims    Percent    Indemnity     LAE  % Should Avg. Lag %IME
                                                                                 IME     to IME  Timely
                 IME not Requested          16       43%         $3,739    $17      63%
                    IME Req but not          4       11%         $4,132   $238     100%
            IME Req and Completed
                   Positive Outcome         15       41%         $3,467   $354      87%       63      62%
                  Negative Outcome           2        5%        $11,375   $906     100%       61     ---
                               Total        37      100%         $4,084   $226      79%       63      62%
                                                     Table 21

Certain patterns appear to cut across all four suspicion levels. First, claims on which an IME
was requested tend to result in larger indemnity payments than those without IMEs. At first
consideration, this finding seems counter-intuitive. However, the larger payments result from
the selection process. Claims with little potential for large medical expenses in particular are
unlikely to elicit an IME.

This point is clarified by comparing indemnity of the claims with positive IME results against
those with negative results. With the possible exception of the completely non-suspicious
claims, the positive IME outcome results in a lower payment. Since the claims with an IME

requested and completed can be assumed fairly homogeneous, this difference is an unbiased
estimate of the IME’s effect.

In terms of loss adjustment expense, the results are generally consistent with the regression
and company estimates obtained in section 7.1. Recall that we estimated $50 per claim in
miscellaneous and $327 for an IME completed. The variation in Table 22 is not surprising
given the small samples in many subgroups.

For the three categories with some significant suspicion, the supplementary sample coders
found more claims suitable for an IME than did the CSE adjusters. For claims with virtually no
suspicion, the coders recommended somewhat fewer. This particular result is summarized in
Table 22

                             Actual versus Recommended IME Requests
                        Suspicion        Actual %       Coder Suggested
                          Level          IME Req          % IME Req
                       None                43%                39%
                       Low                 54%                67%
                       Medium              53%                85%
                       High                57%                79%
                                             Table 22

Although the exact numbers may be incorrect because we are dealing with a biased sample,
the inference is that more IMEs may be warranted, but primarily on suspicious claims.

Finally, there seems to be a relationship between the success of the IME and the timeliness. In
general, the average lag between the injury report and IME request is somewhat shorter for
claims with a positive result. Similarly, the percent evaluated as being timely by the coder is
higher for positive IMEs.

To understand more precisely what is considered to be timely (according to the supplementary
sample coders), we examined the relationship of timeliness to the lag. The results are shown in
Table 23.
                           Timeliness versus Lag from Injury Report for IMEs
                         Days from Injury Report to    Number            %
                                IME Request             Claims        Timely
                        30 or less                            15          100%
                        31 - 60                               29           86%
                        61 - 90                               15           80%
                        91 - 120                              15           40%
                        Over 120                               8           13%
                        Total                                 82           72%
                                              Table 23

It is clear that the dividing line between IMEs considered timely and those not timely occurs at
around 90 days. If the IME is not requested within 90 days, the timeliness and (presumed)
likelihood of success decline precipitously. Overall, 59 out of 82 claims, or 72%, were timely.

While the fact that most IMEs were assessed to be timely is encouraging, there is a negative
aspect to this finding also. Our supplementary sample coders appear to accept the “fact of life”
that IMEs can rarely be performed within the first 30 days from the injury report, let alone the
accident itself. As a consequence, a pattern of heavy treatment may already be established
prior to the IME. In effect the horse had begun to bolt before the barn door is closed.

The levels of indemnity payment in Table 21 further support this view. Even when the IME is
successful, the payment is ordinarily well in excess of $2,000. Thus, the IME is being used
primarily to cut off future buildup after the tort threshold has been reached. There appear to be
few instances in which the IME is invoked very early as a means to prevent buildup from lifting
an undeserving claim over the tort threshold.

7.4 Cost versus Benefit of Independent Medical Exams

The supplementary sample analyses suggest the hypothesis that IMEs generate savings for
some claims, particularly those with medium to high suspicion. In this section we use the CSE
and DCD data to test these hypothesis more systematically. Table 24 presents the results of
the cost-benefit study. The table includes data on all Target claims, which encompass nearly
all of the IMEs requested.

To estimate the IME savings for the claims with positive IME results, we subtracted the average
indemnity for these claims from the average indemnity for claims with completed IMEs but
negative outcomes. For claims in which an IME was not requested and for those with negative
outcomes, we assumed that the IME savings were zero. In the full CSE sample there were also
a substantial number of claims for which an IME was requested but apparently not completed.
For these, savings were also estimated as the difference between the average indemnity paid
and the average for the claims with completed IMEs but negative results.

To estimate the net savings, after accounting for expenses, we subtracted from the indemnity
savings the estimated LAE. To back out miscellaneous expenses clearly unrelated to IMEs, we
subtracted the average LAE for claims in which no IME was requested. We assumed that the
remaining component of LAE was related to the IME, although some other expenses may also
be included. Note that the amounts of LAE appear generally consistent with the results in
Section 7.1. In particular, the averages for situations in which an IME was performed are only
slightly higher than the $327 average found above.

This estimated LAE for IME was subtracted from all claims with an IME requested.
Consequently, for claims with a negative result, the net “savings” was negative, reflecting the
fact that nothing was actually saved but additional expenses were incurred.

                                   CSE Estimates of Savings for Claims with IME
       Target Claims
       Suspicion = None ( 0 - 1)
                                         Claims     Percent   Indemnity   LAE    IME Sav Net IME Net %
                                                                                          Sav     Sav
           IME not Requested                  501      81%       $2,565      $27      $0       $0   0%
           IME Req but not Completed           13       2%       $3,171     $122    $587    $492   13%
           IME Req and Completed
               Positive Outcome                56       9%       $3,317     $371      $441      $97      3%
               Negative Outcome                52       8%       $3,758     $388        $0   ($361)    -11%
       Total                                  622     100%       $2,745      $90       $52    ($11)     -0%
       Suspicion = Low ( 2-3)
                                         Claims     Percent   Indemnity   LAE    IME Sav Net IME Net %
                                                                                          Sav     Sav
           IME not Requested                  172      61%       $2,970      $73      $0       $0   0%
           IME Req but not Completed           17       6%       $2,490     $163    $909    $819   25%
           IME Req and Completed
               Positive Outcome                56      20%       $3,066     $373      $333      $33      1%
               Negative Outcome                38      13%       $3,399     $410        $0   ($337)    -11%
       Total                                  283     100%       $3,018     $183      $121      $11      0%
       Suspicion = Medium ( 4-6)
                                         Claims     Percent   Indemnity   LAE    IME Sav Net IME Net %
                                                                                          Sav     Sav
           IME not Requested                  195      48%       $3,252      $96       $0      $0   0%
           IME Req but not Completed           15       4%       $2,234     $102   $1,889 $1,883   46%
           IME Req and Completed
               Positive Outcome               123      31%       $3,137     $389      $986     $693    18%
               Negative Outcome                70      17%       $4,123     $381        $0   ($285)    -7%
       Total                                  403     100%       $3,330     $235      $371     $232     7%
       Suspicion = High ( 7 -10 )
                                         Claims     Percent   Indemnity   LAE    IME Sav Net IME Net %
                                                                                          Sav     Sav
           IME not Requested                   38      39%       $3,612      $25       $0      $0   0%
           IME Req but not Completed            9       9%       $3,673     $230   ($376) ($581)  -19%
           IME Req and Completed
               Positive Outcome                31      32%       $3,374     $368     ($77)   ($420)    -14%
               Negative Outcome                20      20%       $3,297     $637        $0   ($612)    -23%
       Total                                   98     100%       $3,478     $277     ($59)   ($311)    -10%
       GRAND TOTAL                           1406                $3,019     $163     $150       $42      1%
1. IME savings are calculated as the difference between the category indemnity and the IME completed negative
    outcome indemnity.
2. IME savings include the benefits and cost of any other investigative techniques which may have been applied.
                                                     Table 23

The main result is that overall the net savings from IMEs are modest. Indeed, IMEs as currently
employed seem to be essentially a breakeven proposition for claims with no suspicion or low
suspicion levels. Only for the medium suspicion level do we find a substantial 7% net savings.

For high suspicion claims, we estimate a negative 10% net benefit, but this finding must be
interpreted with caution. For high suspicion claims, the IME represents only part of the
investigative story. As shown in Table 13 above, 16% of these claims have special
investigation in addition to ordinary investigation, compared with only 5% for low and medium
claims. Moreover, some claims under special investigation may be among the 80 claims that

remain open and are not included in this analysis. Therefore, we must treat the current
estimates as very provisional, pending further analyses and data.

On a per-claim basis the largest savings come from claims in which an IME is requested but not
completed. Presumably, the claimant or provider may “voluntarily” limit treatment in these
situations rather than submitting to the IME. In some instances the claim may even be

In the supplementary sample analysis described in Section 7.2 we noted that the coders
recommended more IMEs, especially for the suspicion claims. Thus, there appears to be
potential for increased savings beyond the 7% currently achieved for these moderately
suspicious claims. On the other hand, it can be argued that IMEs should not be attempted for
claims with less suspicion. Even without accounting for additional overhead costs related to
IMEs, there appears to be no net PIP benefit for such claims, although there may be an
additional positive effect on an associated BI claim.

We also observed that under current practice, positive outcomes consist primarily in curtailing
future payments on fairly large claims. It is tempting to speculate that early identification of
potential claim buildup might lead to dramatically increased savings. To achieve such savings
would require a more proactive system for identifying within the first 30 days those claims
likely to end up with medium suspicion. Such a system based on a predictive statistical model
could in theory further reduce the losses for suspicious claims, while expediting low-suspicion

8. Summary of Findings

8.1 The PIP Claim Process

The CSE data have provided a wealth of information about the PIP claim handling process. On
a basic level, we found that newly arising PIP claims can be divided into three main categories:
Duds, Express and Target. From a process standpoint, we view their type of categorization as
universal across lines of insurance. The Target PIP claims comprise approximately 42 percent
of the total and include nearly all claims with potential for loss reduction.

For the Target claims, we studied the flow of information on various fraud indicators. While
there is great variability across claims, a pronounced overall pattern emerged. Information
tends to arrive in three phases. Early indicators (usually within 30 days from report of injury)
pertain to the accident, claimant and insured. The middle phase indicators (30 to 90 days)
concern the nature of the injury, attorney representation and some initial treatment. Finally, we
have later-arriving (90 to 180 days) information about the details of lengthy medical treatment.

In terms of investigation, we have drawn a distinction between ordinary and special
investigation. Ordinary investigation (primarily IMEs and medical audits) occurred in
approximately 43 percent of Target claims, while special investigation was used for only 5
percent. Ordinary investigation is intended to control medical expenses, whether or not the

claim is suspicious. Preliminary data from the DCD suggest that the use of IMEs and medical
audits has been increasing.

The timing of investigation parallels the arrival of information as described above. Special
investigation can occur at any time, triggered by early or later-arriving indicators. Ordinary
investigation tends to occur later, with a large number starting between 90 and 180 days.
These IMEs and medical audits are implemented primarily to limit further medical expenses.

The investigation decision is rather complex. The overall level of claim suspicion certainly plays
a significant role. However, the carrier’s financial exposure, including potential BI claims also
has an influence. In addition, the adjuster’s awareness of the particular medical providers,
attorney and premium town may play a role. In particular, if the medical providers are
associated with a propensity for buildup claims, an IME or medical audit appears more likely.

We prepared a cost-benefit analysis for claims with IME’s. Our results show that IME’s
performed on claims with little or no suspicion are, at best, a break-even proposition for the PIP
claim. On the other hand, claims with moderate suspicion levels produce net savings after
expenses of about 7 percent. Thus, the potential exists of additional favorable IME’s for
moderately suspicion claims. High suspicion claims are more complex usually involving SIU
referral with results that are not as easily evaluated.

8.2 PIP Claim Screening

With respect to claim screening methods, our results indicate that merely providing feedback to
adjusters on suspicion scores and indicators does not seem to be an effective intervention. We
cannot at this point discern any impact of this information on adjuster behavior.

We remain optimistic, however, about the potential of claim screening. The CSE design
embodied some practical compromises that allowed us to move ahead quickly. Of critical
importance was the decision to “simulate” a process of data collection and feedback that
might eventually become part of a company’s computer and workflow systems. We hoped that
such a self-contained, paper-based system would be good enough to have an impact.

The evidence suggests otherwise. In hindsight it seems clear that “the system is the solution.”
The CSE was an add-on that fell outside the central claim processing operation in several
important respects, including:

•   Adherence to the reporting schedule was suggested but not required
•   No specific actions were suggested or required in response to suspicion
•   Feedback to the adjusters was not immediate
•   Information provided to adjusters was based on supervisory discretion
•   There was no monitoring of compliance with recommended procedures

Conversations with claim managers and supervisors suggested that the front-line claim
handlers did not perceive any major benefit from the additional data received, and were

unhappy with the additional work. One exception mentioned was the benefit to inexperienced
adjusters. Senior claim staff generally felt that they already knew which claims needed
investigation, and the scores simply validated their intuition.

On the other hand, our previous and current research, reinforced by national studies
undertaken by the Insurance Research Council, suggest that there is potential for improvement.
Specifically, a majority of claims with moderate to high suspicion are not currently being
investigated or are being investigated later than necessary. So while individual claim handlers
may be doing more, the magnitude of the problem has grown dramatically. New systems are
needed to enable and encourage adjusters to close the gap that has been created. We remain
convinced that automated early detection within a more structured and user-friendly framework
might increase the number of useful investigations.

8.3 Lessons Learned

The claim screen experiment, taken as a whole, provides lessons for those willing to develop
real-time automated screening systems. While the lessons listed below pertain to
Massachusetts auto injury claims, some lessons are universal to any automated claim
classification process. We have learned:

      •   Automated claim screening should add value.
      •   On-line system is preferable to an off-line system.
      •   Instantaneous feedback to decision makers is necessary; delays disrupt the claim
          processing flow.
      •   Data quality needs to be monitored.
      •   Statistical models “smooth” adjuster input and compensate for observational
          differences among adjusters
      •   Investigation using independent medical examinations and other techniques appears
          cost justified only for moderately suspicious claims.
      •   Medical, Attorney and Town buildup scores can be helpful early identifiers of
          potentially fruitful investigations.

9. Future Directions

9.1 Analysis of Associated BI Claims

The analysis explored in this study of PIP claims is incomplete as respects the total
compensation available to claimants. More than half of PIP claimants go on to file BI claims
against either the same carrier, when the insured operator is at fault, or another carrier,
especially when the driver is less than 50 percent at fault. The savings that accrue to the PIP
claim through curtailed or disallowed treatment as a result of investigation have an even
greater effect upon the BI claim. General, non-economic damages depend directly on the

claimed medical charges18 so that reduced PIP medical payments lead to reduced BI
payments. It seems worthwhile to complete the picture on the CSE cohort of claims by
analyzing the accompanying BI claims. This can be accomplished through a matching process
with 1996 BI DCD claims as they close during the next two years. We look forward to acquiring
that data.

9.2. Model Requirements

We have identified two key modeling efforts in building an effective claim screen. The first
concerns the appropriate early variables for determining express claims, claims that should be
adjusted without investigation. While the time from accident to injury report is an important
variable, others should be identified with, perhaps, build-up scores playing a role.

The second effort should be devoted toward minimizing the use of subjective indicators in the
development of a suspicion score. While a completely objective set of indicators may prove
impractical to find, our reliance on purely subjective indicators should be reduced as much as
possible. We observed an alarming lack of agreement between the adjusters and the coders
on the presence of individual indicators in our supplementary sample (Section 7.2), even when
the indicators were fairly objective (claimant in old low-valued vehicle). Fortunately, the
suspicion categories produced by the fairly robust 20-variable model applied to either set of
responses were quite similar. Evidently, different adjusters may attribute the same underlying
uneasiness about the claim to different sets of routine indicators. The use of “data missing“
techniques, such as neural networks, may show the way to more practical and efficient
suspicion models.

  For example, see Table 7 in Derrig, Richard A. and Herbert I. Weisberg, (1994) "Behavioral Factors and Lotteries Under No-
Fault with a Monetary Threshold: A Study of Massachusetts Automobile Claims", The Journal of Risk and Insurance, 61:2 ,