Project topics and materials
Get Complete Project Material File(s) Now! »
The French Health Insurance System
This research is conducted in the context of the French health insurance system, which has many specificities. Firstly, the insurance coverage is divided between two main types of entities: the (mandatory) public social security and the (optional) mutuals. Depending on the type of service, the shares of the price covered by each entity vary significantly. Secondly, most reimbursements are made in the context of a third-party payment system, called “Tiers-payant”.8 Service Providers9 (SPs) submit claims to an insurer in the name of their clients, the policyholders, and directly receive the reimbursement. Another reason SPs play a central role in the fraud process is that they act as certifiers, and a policyholder can almost never defraud without colluding with an SP. An SP can even defraud on his own, by submitting fictive claims. Therefore, SPs rather than policyholders are the focus of auditing in this setting. Finally, since a large share of health services are provided by the public sector, many health service providers have fewer incentives to defraud as their salary is fixed and not indexed on their economic activity. For instance, it is very rare for hospital doctors to defraud, and, when they do, their motive is usually to allow the patient to fall within the insurance reimbursement conditions.
This is not the case for opticians who act as privately owned businesses, and we focus on them hereafter. The case of opticians is of particular interest for several reasons. Firstly, social security only covers an insignificant share of the price of optical products, while the mutual reimburses most of the costs. This situation fosters fraud through two channels: it is less risky to defraud when only one entity, the mutual, has incentives to audit, and, because of the bad reputation private insurers have for nitpicking, defrauders may be less prone to suffer from remorse due to the moral cost of cheating.10 Secondly, opticians provide a tangible product (glasses, lenses) in addition to a service. This makes proving fraud easier, as there are receipts and invoices involved Every Claim You Make I’ll be Watching You! The Deterrence Effects of Insurance Claim Audits in the process since the optician himself must purchase the materials from manufacturers. The insurer can then ask for these supporting documents to check the validity of a claim. Finally, French opticians are interesting in that they may also work as audiologists. This allows me to control for a potential generality of the deterrence effects. I can then test if an audit targeting an SP for his audiology activity has some spillovers on his optical claim submission patterns.
Claiming, Detecting, Auditing and Deterring
This study takes place after the deployment of new detection systems developed by IBM France and the mutual PRO BTP. These systems help inform the mutual’s decision to accept or refuse the submission of the claim ex-ante, and to audit the claim ex-post.
Figure 1.1 – Timing of claim submission and auditing
Claiming The process starts with an optician submitting a coverage request electronically to the mutual in the name of one of its policyholders. The policyholder may be visiting the optician for legitimate reasons, or they may be colluding to disguise an illegitimate coverage request as a valid one. A common case involves colluders who request a reimbursement for regular glasses, when the reimbursed is actually used to cover the price of sunglasses which are not covered by the insurance contract.11 Fraud schemes can be very creative and leverage clauses of the contract for other purposes than their original one. In some extreme cases, there may be no policyholder involved, when opticians submit a totally fictive coverage request to the mutual and cash in all the reimbursement.
The insurer receives the request and conducts a quick analysis to determine whether it falls within the terms of the contract and to assess its legitimacy. This is just a preliminary superficial analysis and does not correspond to the actual thorough detection process. Then it either accepts or rejects the request. If it is accepted, it becomes a confirmed claim, the optician proceeds to deliver the service to the policyholder and the reimbursement is directly made to the optician. Detecting At the end of the corresponding month, all accepted claims are thoroughly analyzed in the deployed detection system. The detection process involves several statistical methods such as rule-based red-flagging, anomaly detection and supervised learning. Its output is a flagging of claims depending on the likelihood of being fraudulent, and the results are transferred to an SIU within the mutual.
Auditing The SIU selects claims that it deems suspicious enough, based on the decision systems’ outputs. The selection for auditing criteria are not fixed, and the criteria evolve in time and are based on discussions within the auditing teams. Therefore, while flagging is not random, it is hard to control for it explicitly without having access to an actual systematic selection process.
Deterring Since the introduction of the new detection systems, the changes in the auditing process are twofold: first, because of the rationalization and automation of the process, it became possible to conduct more audits at lower costs. Second, thanks to the more sophisticated analytical process, the audits became more accurate and the probability of finding fraud conditional on an audit increased. According to the theoretical literature on criminal behavior and fraud (Becker (1968), Allingham and Sandmo (1972), Dionne, Giuliano and Picard (2008)), the equilibrium levels of fraud should decrease as a response to the increase in the probability of being investigated. However, this is under perfect information, and auditees should instead be expected to learn progressively about the shift in auditing patterns, through experiencing more frequent audits.
Data
My dataset includes a total of 260,592 observations: it encompasses all the claims submitted by 16,287 opticians to the mutual during the 16 months between August 2016 and November 2017. Table 1.1 defines the main variables and Table 1.2 presents some descriptive statistics.
Over the considered period, there have been 1,088,680 claims submitted to the mutual. Figure 1.2 shows their distribution. The average optician submitted 4.18 claims per month. Standard deviation is 6.03, meaning claim count data is over-dispersed. Decomposing the variation in claims into between and within variation reveals heterogeneity with regards to the individual characteristics of opticians and to the timing of claims. First, the between standard deviation is 5.18. This is due to heterogeneity in the claiming profile: some opticians submit large monthly numbers of claims, reaching a maximum of 101 claims, while others only seldom submit. This is a consequence of many factors: some opticians correspond to major nationwide chain stores while others can be rural small businesses. Some opticians may also interact more often with policyholders of the mutual than others. This heterogeneity in claiming intensity might also be a consequence of fraud: assuming honest claiming is independent of fraud, then defrauders would submit more claims than honest opticians on average. Second, the within standard deviation of 3.08 is a consequence to the seasonality of claiming patterns, as illustrated in Figure 1.3a. Opticians submit fewer claims during months with low economic activity, especially during August, as opposed to June and July. Submissions are also larger during particular periods such as December, which is partly related to policyholders trying to use their insurance before the The volume of suspicious12 claims remains marginal when compared to the volume of non suspicious ones: a monthly average of 3.80 claims per optician are believed to be honest, while a monthly average of 0.37 claims per optician are believed to be fraudulent. On average, the fraudulent claims represent 8.8% of the submitted claims, which is consistent with the 10% estimate commonly used in the industry. Both fraudulent and honest claim counts are overdispersed, in both the between and the within dimensions. However, while non suspicious claims display similar time patterns to the general claims (see Figure 1.3b), suspicious ones have a higher within standard deviation than their between standard deviation. This is because opportunistic fraud timing depends highly on the season. As illustrated in Figure 1.3c, the number of suspicious claims is remarkably larger during June and July. This is because an important share of fraud for vision insurance consists in purchasing sunglasses, which are seldom covered by the policy, while declaring regular glasses to the insurer, in order to reduce the former’s price.
Regarding audits, a monthly average of 0.0097 claim per optician is audited. Auditing is therefore very rare, as it is costly and capacity constrained. It is also endogenous and targeted towards suspicious individuals, which explains the large standard deviation at 0.1256. A between standard variation of 0.04 shows that some opticians receive particularly intense scrutiny, and the maximum number of audits received by a given optician at a given month is as large as 16.
within variation is even larger with a 0.12 corresponding standard deviation, meaning auditing is also heterogeneous in timing. This is illustrated in Figure 1.3d. Understanding the concentration of auditing in some months rather than others is more subtle than the concentration of claims. It is an indirect function of the claiming pattern, as audit target claims in the months following submission, but it is also a function of unobservable random events such as a sudden change in the auditing policy of the mutual.
The audits can be of different types. Warning letters are basic deterrence tools that are sent after some suspicious general claiming patterns are detected, but without specific enough fraud signals at the claim level. If specific signals are observed, the SIU sends letters requesting supporting documents for the corresponding claims. These documents can be purchasing orders for glasses or lenses, or quotes for the offered service. The document requesting letters may also mention that the policyholder will be asked whether he did receive the service/product. The latter is one of the cases where fraud is the less tedious to prove, especially when the optician defrauded on his own, without colluding with the policyholder. A third type, reminder letters, complementing the document request, may be sent in case of no response from the optician. This lack of response may for example be due to opticians not accepting to comply because they still do not perceive the threat as credible. Finally, an optician may receive a letter requesting documents for an audiology claim. It is not clear if such letter has an effect at the mutual level, where it would deter fraud even for the optical claims, or if it shifts fraud from audiology claims to optical ones.
Another quantity of interest is the cumulative number of incurred audits. In the context of the deployment of the new detection systems, the increase in the probability of being audited is only perceived by auditees through actually received audits. Thinking of auditees as Bayesians who update their belief accordingly, the higher the number of audits received in the past, the higher the updated probability of being audited, and the lower the probability of fraud. During the considered 16 months period, an average optician had received an average of 0.26 audits in the previous months.
Table of contents :
Introduction
1 Every Claim You Make I’ll be Watching You! The Deterrence Effects of Insurance Claim Audits
1.1 Introduction
1.2 Research Setting and Empirical Strategy
1.3 Empirical Results
1.4 Discussion: Limits and Extensions
1.5 Conclusion
2 Preliminary Investigations for Better Monitoring: Learning in Repeated Insurance Audits
2.1 Introduction
2.2 Single Period Auditing
2.3 Two-Period Auditing: The Learning Effect
2.4 Variable Claim Value
2.5 Conclusion
2.6 Appendix
3 Should I Stalk or Should I Go? An Auditing Exploration/Exploitation Dilemma
3.1 Introduction
3.2 The Model
3.3 Exploration vs Exploitation: Auditing to Separate the Wheat from the Chaff
3.4 The Extents of Learning
3.5 Conclusion
3.6 Appendix
4 You’ll Give Yourself a Bad Name: Reputation Effects in Repeated Audits
4.1 Introduction
4.2 Model
4.3 Reputation-based Deterrence
4.4 Auditing with Deterrence and Learning Effects and the Restless Bandit Problem
4.5 Conclusion
4.6 Appendix
Conclusion
Bibliographie
Liste des tableaux
Liste des figures
