CYRIL TUOHY, managing editor of Risk & Insurance®
In September, Contributing Editor Michelle Kerr interviewed Andrew Pelcin, director of decision support at Aramark, the winner in the for-profit category of the 2009 Theodore Roosevelt Workers' Compensation and Disability Management Award.
Pelcin opened up about the predictive modeling technology that Aramark uses and how it fits in the Teddy Award winner's overall workers' comp and disability program.
Q: How does predictive modeling fit in with other technology that
Aramark risk management is using?
When I was brought on board, I was given only one objective: to rank all Aramark locations by risk. We wanted our management teams, at all levels, to know which locations were more at-risk than others. We also wanted to use a model that could provide suggestions on how to mitigate that risk. The predictive modeling was an outgrowth of our original strategy.
Q: So, that's all a product of predictive modeling?
Yes. Our team launched a new model based on a Bayesian logical model. The technology is provided by third-party vendor named Bayes. We had to do the following to prepare for a change in model:
1. Identify what information we needed to evaluate safety risks.
2. Determine who at Aramark should make the decisions regarding these incidents. For example, who can make the determination about what does and does not pose a significant risk?
3. Decide what the next steps should be once something is determined to be high risk.
We were then able to create a severe case model so that all Aramark employees in the field can be informed and prepared to prevent safety risks.
Q: What different elements go into the model?
We used the Bayesian model because it is able to collect all related data necessary--even if it does not have an effect on the outcome. The model is able to run around 1,000 variables, including demographic information about injured workers, the type of job they were doing, what country they work in and the line of business for which they work.
The model is then able to identify other Aramark employees who "look" just like that individual to see if they have experienced any similar incidents. As a result, Aramark's model is now carrying more than 10 years of claims history.
When a claim is filed, the model can show us whether Aramark has had a similar claim like this in the past, the demographics of that person and the kind of environment in which he worked. It then predicts the severity (ultimate cost) of the current employee injury.
This model has been particularly successful for a company like Aramark. Despite the nature of our business--where employees are working in a variety of different capacities and settings including foodservice, facilities management or uniform services--individual employees are ultimately performing similar tasks across the entire country. This results in large pools of comparative data. The larger the comparative pools, the more accurate the prediction of the final outcome is going to be.
Q: What's the ROI for predictive modeling at
The ability of Aramark's modeling tool to predict the ultimate cost of a claim is accurate more than 95 percent of the time. This allows Aramark to allocate resources most efficiently. For example, we are able scale back the resources dedicated to handling minor claims, but we are also able to identify claims with a high predictability of severity and get them to a nurse case manager in a more timely fashion.
For claims that are somewhere in between the two, this system alerts our internal claims team to perform initial triage to determine whether we should engage a nurse case manager to handle the medical aspects of the claim.
Q: So where you go with this now?
We have so many opportunities to expand the model from evaluating employee risk to evaluating property and liability risk. You can predict, for example, which Aramark locations will sustain property damage in a hurricane. From there, you can ensure that managers at these locations are notified of the impending threat, prepared from a property loss control perspective, and equipped with the appropriate insurance forms and contact information.
November 1, 2009
Copyright 2009© LRP Publications