Panel Paper: Using Predictive Analytics for Early Identification of Short-Term Disability Claimants Who Exhaust Their Benefits

Saturday, November 10, 2018
8206 - Lobby Level (Marriott Wardman Park)

*Names in bold indicate Presenter

Kara Contreary1, Yonatan Ben-Shalom1 and Brian Gifford2, (1)Mathematica Policy Research, (2)Integrated benefits Institute


Short-term disability insurance (STDI) pays partial wage replacements to employees temporarily unable to work due to “off-the-job” medical conditions. Most STDI policies replace wages for a fixed period, such as six months. Because wages are replaced only partially, STDI claimants have an incentive to return to work. Those who are unable to return before benefits expire may be at higher risk of job loss and receipt of long-term disability insurance (LTDI) or Social Security Disability Insurance (SSDI) benefits. An STDI claim can be an early identification point of workers with medical conditions who could, with adequate support, remain in the workforce. However, little is known about the factors influencing STDI duration or the transition to LTDI or SSDI benefits. Furthermore, careful timing and targeting of interventions is critical to efficiency; some workers may return to work without intervention, while others may not benefit from it. In this paper, we: (1) compare the performance of alternative models using information in claims data to predict exhaustion of STDI benefits; and (2) assess if waiting for some claims to resolve without intervention can improve the efficiency of targeting individuals for early intervention aimed at helping them remain in the workforce.

Data. We use Integrated Benefits Institute (IBI) Health and Productivity Benchmarking Data from 2011 through 2015, including 820,751 closed STDI claims from 8,587 small, medium, and large businesses associated with 9 disability insurance carriers and third-party leave administrators. The data include claim outcomes and claimant, employer, and insurance plan design characteristics. The primary outcome of interest is exhaustion of the STDI benefit.

Methods. We fit several predictive models to the data, including logistic regression, regularized logistic regression (using an elastic net), and random forests. We randomly divide our sample into training and testing sets, and select a model based on the area under the receiver operating characteristic curve. Individuals are flagged as having a high probability of exhausting their benefits using a predicted probability threshold, which was chosen to balance the tradeoff between sensitivity and specificity. We report the predictive performance of our models when applied to the test set. We perform the analysis for claims with benefit duration of 26 weeks, first using the full sample, then sequentially eliminating claims that resolved within 2, 4, and 6 weeks. Comparing across durations illustrates the potential efficiency gains of waiting to allow some claims to resolve on their own.

Results. In the logistic regression, the factors most strongly associated with exhaustion of STDI benefits are age, diagnosis, and employer industry. Waiting to allow some claims to resolve without intervention improves the efficiency of targeting efforts. Modeling based on observable factors helps further narrow the target population, with the machine learning techniques we use expected to outperform logistic regression in predictive performance.

Conclusions. Our approach could represent significant savings through efficient targeting of interventions to those STDI claimants who are most likely to benefit from them. We simulate the cost and benefit of several existing early intervention proposals under our predictive modeling framework.