More and more consulting groups and insurance carriers are working on a predictive modeling scenario to identify the severity and length of particular claims. An October 2002 study by PWC (PriceWaterhouseCoopers) found 83 variables that could be considered potential predictors. Their attempt was to build a model using existing carrier data that would classify claims as "likely" or "not likely" to become serious. Additionally, data was analyzed using a CART( Decison-tree) model and a logistic regression model. Their outcomes showed 31% of claims as potentially serious. However, of that 31%, 2/3rds were false positives. The models also failed to classify 4% of the known serious claims.
So, we still have not found answers to the prediction issues. Their study points out that additional information such as quantifying the existence of emotional trauma or Evidence Based Medicine from the treating physician could vastly improve the outcomes.
More often than not, injured employees stay out of work longer due to the work environment, the immediate responses of both employer and carrier as well as their internal belief systems. While each of those variables can be globally controlled, quantifying and capturing these data elements will vastly improve the success of a predictive model. To further complicate the data collection, how can it be collected in a streaming mode...to allow for immediate triage?