Predicting Workers’ Compensation Insurance Fraud Using SAS Enterprise Miner 5.1 and SAS Text Miner

Insurance fraud costs the property and casualty insurance industry over 25 billion dollars (USD) annually. This paper addresses workers' compensation claim fraud. A data mining approach is adopted, and issues of data preparation are discussed. The focus is on building predictive models to score an open claim for a propensity to be fraudulent. A key component to modeling is the use of textual data to enhance predictive accuracy. Binary response modeling is emphasized, but a two-stage modeling approach is briefly addressed. SAS Enterprise Miner 5.1 with SAS Text Miner provides the modeling environment for the analysis. INTRODUCTION Insurance fraud costs the property and casualty insurance industry over 25 billion dollars (USD) annually. Workers’ compensation insurance alone accounts for a sizable portion of this total cost. Workers’ compensation fraud can be divided into three types: • Claimant fraud occurs when an individual falsely claims to have experienced a work-related injury or exaggerates a legitimate injury and files for insurance benefits. • Provider fraud occurs when a provider bills an insurance company for services that were not provided. • Internal fraud occurs when someone inside the insurance organization creates a fictitious claim to route benefits to an accomplice. Sometimes two or more individuals conspire to commit fraud, which leads to fraud cases that have elements of two or three of the possible types of fraud. For example, a claimant and a provider might conspire to commit fraud. A standard operational practice is to have claims adjustors or claims supervisors identify potential fraud cases and route them to a Special Investigative Unit (SIU). The “traditional” SIU employs professionals who have skills in fraud investigation such as surveillance, evidence collection, and evidence analysis. These SIU investigators often have a law enforcement background. One shortcoming of the traditional SIU organization is the absence of computer data processing and predictive modeling professionals. Consequently, details of fraud investigations are rarely stored in electronic form in a modern database. This paper outlines a strategy for detecting claimant fraud in workers’ compensation insurance. It is assumed that a traditional referral process and SIU organization exist. Suggestions are made to help an organization transition to a modern data mining approach to fraud detection. While the primary impact will be on the referral process, there are implications for improved data processing to facilitate future predictive modeling. INSURANCE FRAUD CONTRASTED WITH OTHER FORMS OF FRAUD Modern success stories for data-mining fraud detection include applications in credit-card fraud and telecommunications fraud, which are much easier to model than insurance fraud. Credit-card and telecommunications fraud is easily identified within a relatively short period of time because a customer contests a bill or reports a stolen credit card. Essentially, the absence of a payment is a key component of fraud in credit and telecommunications applications. With insurance fraud, if the fraudulent behavior is not discovered, the insurer never knows that the fraud has occurred. Consequently, an uninvestigated claim cannot be labeled with respect to fraud. Only claims that are investigated by the SIU can be labeled FRAUD=YES or FRAUD=NO. This implies that, out of hundreds of thousands of claims that have detailed claim information, only about 1% can be used to train a predictive model. Contrast this to credit-card fraud where all cases can be included in modeling. Part of a proposed solution to the insurance fraud problem is to process the uninvestigated insurance claims so that some of these cases can also be used in modeling. SUGI 30 Data Mining and Predictive Modeling