Predictive Modeling – An Ounce of Prediction

Dr. Jack Mahoney was frustrated. In the fall of 2000, health-care claims at Pitney Bowes, where Mahoney is the global health-care management director, had suddenly spiked. As a self-insured employer, the company foots the health-care bill for most of its 27,000 U.S. employees directly, rather than pay premiums to an insurance company. For the first time in 13 years, Mahoney, who had a reputation for keeping costs down, was going over budget. And he didn’t know why.

The data analysis tools he used explained where his money had gone the year before, but offered no explanation for why costs had risen, and no clear indication what they would be in the future-let alone what he could do about it. As a practicing physician who sees patients in one of Pitney Bowes’ onsite clinics and a passionate advocate for employee health, Mahoney wanted to be sure that when he cut costs, he maximized patient benefits.

What Mahoney needed was a crystal ball. He found the next best thing in the form of predictive modeling: technology that employs either rules-based algorithms or artificial intelligence to predict (in this case) future health-care expenses.

Backed by a health-care savvy CEO, Mahoney and his HR counterpart, David Hom, set aside their pet theories about health-care cost drivers, resisted the temptation of scope creep and convinced skeptical colleagues to trust the models’ predictions. Then they implemented programs based on those predictions that improved care and lowered costs. In an era when health-care costs are swelling 10 percent to 15 percent per year, predictive modeling has helped Pitney Bowes slash the median cost of care for diabetic and asthmatic employees by as much as 15 percent since 2001. And it forecast a dramatic cost increase in one region early enough to give Mahoney and Hom a chance to counteract it.

Now, Pitney Bowes, famous for inventing the postage meter back in 1920, is pioneering the use of predictive modeling in a way that may garner competitive advantage. Hom, who is vice president for employment brand total rewards, wants to pinpoint the ideal combination of benefits that motivate employees to perform better at work-ultimately increasing profits for the $4.6 billion integrated mail services and document management company. Predictive modeling will have the biggest influence on defining the benefits package and the engaged workforce, he claims.

Cynthia Burghard, a Gartner research director specializing in health care, says Pitney Bowes is innovative in its use of predictive modeling because it has developed programs to address the issues its models uncovered. Even among health insurers who use predictive modeling to forecast costs, she says, “There’s a lot of-so now what do I do?”

Symptom: Feverish Costs

The company’s foray into predictive modeling was a journey born out of frustration. Mahoney, a trained epidemiologist, and Hom (both unabashed numbers geeks) had kept Pitney Bowes’ health-care expenses at two-thirds of the industry benchmark. Yet they were constrained by the limitations of retrospective data analysis. At annual review meetings, they got blank stares when they pressed health-care vendors to project costs for the coming year.

Their frustration reached the breaking point in 2000, when the company’s health-care costs spiked. “We knew our population was aging, we were seeing more chronic disease in the population, and the company was growing,” recalls Mahoney. Pitney Bowes was expanding its Management Services division, hiring older workers and more workers from inner cities who didn’t have regular access to health care in the past. Mahoney believed that the cost problem would get worse if he didn’t act quickly.

About the same time when Mahoney saw costs jump, Hom was renegotiating a contract with one of Pitney Bowes’ HMO vendors. After seven years of negotiating favorable rates based on actuarial data, which showed that Pitney Bowes employees tended to be younger (and thus healthier) than average, Hom was stymied. The HMO negotiator had data showing that even though Pitney Bowes employees were younger, they were sicker. And he was using that data to justify a rate increase.

“I said to Jack, ’Son of a gun, every time I said something this guy had an answer. He must be doing something we’re not,’” Hom recalls. That something turned out to be predictive modeling.

The HMO was utilizing algorithms developed by vendor DxCG that apply pharmacy data to predict how sick a given population is likely to get in the coming year. If health plans could use predictive modeling to set their rates, Hom and Mahoney figured that Pitney Bowes could use it internally to forecast future cost spikes in time to take action to minimize them.

The two pitched the idea to CEO Michael Critelli, as part of a plan to control future costs and eliminate nasty surprises like the 2000 cost surge. And they promised a 5-to-1 ROI ratio after the first year. Having once served as head of personnel, Critelli is well-versed in the difficulty of managing health-care costs and often speaks eloquently on the pertinence of maintaining employees’ good health to the success of the company. Critelli agreed to sponsor two predictive-modeling projects in part because Hom and Mahoney had a good track record of keeping costs down. But he approved the investments with the caveat that Mahoney and Hom had to deliver predictable cost increases starting in 2001.

Predictive modeling “seemed like a gamble because it was a total unknown,” says Mahoney. “But we knew that the standard actuarial models weren’t working for us.”

To minimize the risk, Mahoney and Hom started with two pilot projects. To protect employees’ privacy, they outsourced the projects (see “A Premium on Health Privacy,” opposite page). They hired DxCG to project future health-care costs by geography and to validate the 2001 budget. And they contracted with Medical Scientists to help identify the factors most likely to lead to an employee turning into a high-cost claimant.

Diagnosis: A Sickly Staff

Hom and Mahoney began meeting with DxCG in the fall of 2000. The following summer, Pitney Bowes turned over its pharmacy and medical claims data to DxCG, which ran the data through its sophisticated cost algorithms. These algorithms can forecast health-care costs for the coming year with much more precision than traditional actuarial models because they take more factors into account. DxCG provided insight into why cost increases varied by geography, enabling Pitney Bowes to fine-tune health plans by region.

The model flagged a looming problem in the New York City area where Pitney Bowes handles mailroom operations or maintenance for 21,000 customer sites. Although actuarial and underwriting models (such as those that factor in age and sex) predicted costs would rise there by just 2 percent, the DxCG model warned of an increase of more than 40 percent. A higher-than-average proportion of the employees in that area had diabetes, asthma or cardiovascular disease, and their medication usage patterns were a harbinger of high costs to come, according to DxCG’s model. If those illnesses aren’t managed carefully, says Mahoney, patients will get sicker and the cost to treat them will escalate.

Mahoney and Hom quickly launched education initiatives to encourage employees to better manage their health. They offered free or low-cost immunizations and health screenings, made sure that employees knew the location of the nearest walk-in medical facility and urged them to rely on it instead of the ER for routine medical care. A group of new employees in New York had previously been on Medicaid. “There are no copays and people use ERs when they could use a doctor’s office,” Mahoney explains.

The DxCG model proved prescient: First- quarter 2003 health-care costs in the New York area were 30 percent to 40 percent over 2002 costs. But thanks to the education initiative, health screenings went up, ER usage was down and overall annual costs came in at just 1 percent above 2002 costs.

When Pitney Bowes’ contract with its New York area health-care vendor came up for bid for 2004, Mahoney and Hom looked for a more proactive provider. “We needed a health plan to reach out to people, to help engage them before they become high-cost claimants,” says Mahoney. Today, employees who sign up for the new plan get a call from a nurse whose goal is to identify the family’s health problems and offer resources to get them under control.

Treatment: Incentives To Stay Healthy

While DxCG was focusing on projecting costs by region, Medical Scientists was developing a model of a high-cost claimant-someone who costs the company $10,000 or more per year in medical claims, workers’ compensation, disability and absenteeism. The software uses several types of artificial intelligence, including neural net technology, to identify patterns among claimants that might otherwise be impossible to spot.

“Everyone has a theory of what drives future costs: smokers, people who don’t eat right, lack of immunizations,” says Mahoney. Those theories tend to color people’s interpretations of the data, he explains. He sees artificial intelligence as a fresh set of eyes, unencumbered by preconceived notions.

Pitney Bowes gave Medical Scientists data on medical, disability and pharmacy claims; encounter data from its seven onsite clinics; and employee data-such as age, gender, ZIP code and salary level. The engine then searched for the combination of variables most likely to cause a person to rack up high expenses.

The analysis produced two findings. First, employees who spent more than $780 on health care were most likely to become high-cost claimants the next year, as were those who spent nothing because they weren’t getting checkups. Second, employees with asthma, diabetes, depression or hypertension who weren’t taking their medicine regularly were also at risk for becoming big spenders.

Hom and Mahoney convened a group of experts-including the chief medical officers of a pharmaceutical company and a health insurance plan-to review the findings. After validating the data, the group brainstormed ways to reduce expenses. Mahoney and Hom prioritized the ideas and launched sweeping changes to Pitney Bowes’ health plans.

First, they redesigned the plans to place greater emphasis on preventive care. Now, employees can get colonoscopies, mammograms and immunizations for free or for a minimal copay. To get employees with common chronic illnesses to take their meds, Mahoney and Hom took an even more radical course of action: They did away with the three-tiered pricing structure for drugs used to treat asthma, diabetes and hypertension. Instead of making employees kick in up to half the cost for brand-name drugs, Pitney Bowes would provide all asthma, diabetes and hypertension drugs at the generic rate of 10 percent. Overall, employees’ out-of-pocket costs for those medications were cut nearly in half.

Mahoney had to convince skeptics that the likely long-term savings from more faithful medication usage justified the expense. Pitney Bowes would be footing a much heftier portion of employees’ drug bills at a time when consultants were forecasting a 22 percent to 24 percent increase in drug costs. And it would also forfeit the rebates that pharmaceutical companies offer for giving their drugs preferred status. Pitney Bowes was potentially adding close to a million dollars to its pharmacy bill.

After phasing in the price cuts over two years, all that Mahoney and Hom could do was wait to see if the model was right. As they waited, they kept a close eye on Pitney Bowes’ pharmacy costs. “We didn’t see the huge spike in pharmacy costs that would have been the indicator of failure,” says Mahoney.

Prognosis: Cautious Optimism

Their patience was rewarded this year with good news: The median cost of care for employees with asthma decreased 15 percent in 2002, while costs for diabetes patients fell 12 percent. Mahoney reports heaving a huge sigh of relief when he realized that reality bore out the Medical Scientists model’s prediction. He attributes the lower costs to people refilling their prescriptions more faithfully, and taking “control” drugs that prevent problems (and costly hospitalizations) instead of more expensive “rescue” drugs. Mahoney says that the rate for patients obtaining control drugs is up approximately 20 percent. In the case of asthma and diabetes drugs, Mahoney reports that Pitney Bowes’ prescription costs are actually down about 10 percent after 18 months because patients need fewer additional medications to treat emergencies or side effects.”In your wildest dreams, [lowering drug prices] is not first thing you’d do,” says Mahoney. “But this turned out to be appropriate.”

One reason that Pitney Bowes achieved such promising results is that Mahoney and Hom didn’t second-guess the data from Medical Scientists and resisted the temptation to lower prices willy nilly. Statin drugs, for instance, are very effective at lowering cholesterol, but their prices were not reduced because the model found no link between patients failing to take these drugs and high costs.

As it turns out, lowering the cost of hypertension drugs did not increase compliance or lower treatment costs. “What’s true for asthma and diabetes may not be true for hypertension,” says Mahoney. Predictive modeling “is going to be the key to helping us understand all that.”

Gartner’s Burghard says that predictive modeling has been used for years by insurance companies to set rates. But calculating the probability that someone will live X number of years to determine his life insurance premium is easier than figuring out what causes health-care costs to increase and how best to control them. Merely investing in predictive modeling won’t yield a return-just a projection that, say, costs will increase by a certain percentage. “The return comes when you figure out why and what you’re going to do about it-and then only if what you do about it makes a difference,” Burghard explains.

Given Pitney Bowes’ success at reining in health-care costs, Hom is launching a predictive-modeling project outside the realm of health benefits. This time, the company will use the technology to comb through data on the company’s benefit options (such as 401(k) plans and insurance elections), customer satisfaction scores, employee survey results and individuals’ performance ratings. The goal is to identify the benefits that are most closely correlated with profitability and external customer satisfaction. “It’s kind of Star Wars stuff,” says Catherine McCabe, a health-care consultant with Mercer Human Resource Consulting.

Pitney Bowes is well-equipped to venture into Star Wars territory because it undertakes new predictive-modeling projects with a clear understanding that they’ll be long-haul propositions. “Patience is a big virtue,” says Mahoney. “You have to be able to stay with something long enough to see results.”