August 2000 Prescription for Online Pharmacy Success Modeling and analysis helps merckmedco.com achieve best-in-class information, resources and services by Edward S. Binkowski and Moshe B. Rosenwein Merckmedco.com, the e-commerce Web site of Merck-Medco Managed Care, L.L.C., services nearly 20 percent of Merck-Medco's customer transactions (including prescription refills) and has been growing rapidly (see Figure 1). Launched in 1998, Merck-Medco's Internet site processes 80,000 online prescriptions per week, yielding some $8.5 million in revenue [1]. Operations research and statistics models and methods have been used to support direct-mail marketing campaigns, promoting merckmedco.com to Merck-Medco's member base. In addition, a forecasting and ad-hoc reporting tool was developed to support merckmedco.com's business strategy function. Figure 1: Internet-based Member Transactions Merck-Medco Medco was founded in 1983 as a single mail-service pharmacy in Elmwood Park, N.J. Its mission was to improve patient health by improving prescription drug care while reducing prescription and overall health care costs. By 1989, the company was providing its clients with fully integrated retail and mail pharmacy services. Merck & Co., Inc. purchased the company in 1993, and it is known today as Merck-Medco. Merck-Medco is the leading U.S. pharmacy-benefit manager (PBM). PBMs work with health plan providers to improve the quality of prescription drug care while working to maintain overall health care costs. Merck-Medco provides pharmaceutical care for more than 52 million Americans on behalf of more than 1,100 health plan clients throughout the United States. Clients include approximately 125 of the Fortune 500 corporations, more than 100 local, state and federal employee and retiree groups, one-third of the nation's Blue Cross/Blue Shield plans, and various union, insurance carrier and managed care plans. Merck-Medco's business may be divided into two components: support of a network of retail pharmacies, and management and operation of a mail service network. Each week, Merck-Medco dispenses some 1.2 million prescriptions through 12 mail-service pharmacies and manages another 5.9 million drug claims through a network of 55,000 retail pharmacies, totaling 370 million prescriptions per year. A member (patient), using the PAID Prescription retail network, presents a plastic card to a retail pharmacist, makes a small co-payment and receives a prescription. Merck-Medco provides clinical expertise, through its information management systems, to help protect patients against potentially harmful drug-drug interactions, as well as potential drug-age, drug-gender and drug-allergy complications. Pharmacists are reimbursed by Merck-Medco (which, in turn, is reimbursed by the patient's health care plan), based on the type of drug dispensed. At the 12 Merck-Medco mail-service pharmacies, registered pharmacists, employed by Merck-Medco, manage every step of the dispensing process. According to a recent article in The Wall Street Journal, the mail-service prescription market is growing by 22 percent annually and accounted for more than 10 percent of the $125.9 billion U.S. prescription market in 1999 [2]. Mail service is one of Merck-Medco's key instruments for providing low-cost prescription drug benefits to its clients. A mail-service prescription (Rx) may be ordered through various channels, depending on whether a prescription is a "new" or "refill" prescription. A new prescription request currently requires a prescriber's original signature or verbal authorization from a directly authorized agent. Thus, a new, mail-service prescription currently may only be ordered via the U.S. mail, fax or phone. In addition to these channels, a refill prescription can be ordered via an interactive voice response (IVR) system or the Internet. merckmedco.com E-commerce supports various strategic initiatives at Merck-Medco, all with the intent and focus on delivering best-in-class information, resources and services to its plan members through online services. The overarching goal is to make merckmedco.com a one-stop shopping, service and convenience center for all of the company's online customers. Some of the initiatives include: an efficient channel for processing customers' mail service prescriptions (i.e., sales). an efficient channel for processing customer service contacts (e.g., order status inquiries, billing inquiries) that traditionally have been handled by teams of customer service representatives in call centers. Along these lines, this channel also helps deliver superior service to consumers, offering 24/7 access to customer service assistance from the comfort and confidentiality of an individual's home or office. an efficient and effective medium for communication with members at their convenience. For example, various Web pages on the site are devoted to health information. an efficient capture of e-mail addresses of members who register on the site. E-mail addresses are a more efficient channel for marketing promotions (i.e., encouraging members to take advantage of their mail-service benefit program for medications they take on an ongoing basis) than direct mail and outbound telemarketing. (A member may choose to receive or not to receive promotions and communications online.) Objectives of Merck-Medco's e-commerce organization include increasing prescription services delivered through the Internet and increasing the Internet's share of customer service transactions. In order to achieve its goals, the organization has launched large-scale, direct-mail promotions to a subset of members, encouraging Internet usage. We, as part of Merck-Medco's Information Research group, supported e-commerce direct mail promotions by developing scoring models to predict which members, from a candidate list, are more likely to respond to a particular promotion. In addition to our support of marketing efforts, we also developed a forecasting and ad-hoc reporting tool to predict future Web growth (for business planning purposes) and enable reporting and distribution of e-commerce information across the organization. Data Data collection and development of datamarts, needed for modeling and analysis, are critical to our work. Merck-Medco has various databases that capture information about its members' historical (dynamic) behavior, as well as some static data concerning the members themselves. A mail-service database tracks claims placed by (and on behalf of) individual patients. One or more patients are associated with each membership (where a member may be viewed as a head of household). For each claim, there exists: Service date Member and patient identification Drug therapy category (e.g., cardiovascular) Cost Days' supply (quantity of medication dispensed, expressed in days) Order-entry channel (e.g., phone, U.S. mail, Internet, IVR, FAX) New/Refill indicator A customer service history database tracks customer service activity by members (not patients). Typical customer service channels include phone, IVR and Internet. Each record in a customer service history database captures: Member identification Service date Customer service channel Reason code (e.g., sales inquiry, order status inquiry, billing inquiry) In addition to mail-service claims and customer service history data, Merck-Medco tracks previous direct-mail promotions to its members. Merck-Medco also captures certain static information on members, such as address (zip code) data. In addition, each member is associated with a client. Each client has a unique formulary — list of prescription drugs that are available to members — and cost structure. Some clients do not permit communication between Merck-Medco and its membership, thus removing certain groups of members from eligibility for direct-mail promotions. Scoring Models for Promotions: Design and Evaluation For analysis of direct-mail promotions and e-commerce usage, it was useful to build an e-commerce user datamart, grabbing different data from different databases across Merck-Medco. For each member that used the Web, we tracked the number of Internet sales and service transactions by month. In creating such a datamart, we can rapidly calculate and report new e-commerce users by month (i.e., growth of the e-commerce population), shoppers (i.e., users who executed sales transactions) versus all users, and retention metrics (e.g., percentage of all users who used e-commerce within the past three months, past six months, etc.). The e-commerce user datamart may be linked to other data sources in the company. In the first nine months of 1999, there were approximately a half dozen major direct-mail promotions to increase e-commerce usage. All of the promotions focused on those members with an upcoming refill. Beyond that criterion, each promotion used different filters for the ultimate selection of targets, including such factors as level and patterns of past Web usage and specific client attributes. Each promotion produced a significant lift in e-commerce usage; indeed, in all but one promotion, success was extraordinary by typical direct-mail standards. The different criteria used over the applications, however, suggested that the methods employed for targeting members for promotions were sub-optimal. Accordingly, Merck-Medco attempted to meld the experiences of these promotions in developing a system of scoring models with three complementary goals in mind: Efficiency — maximizing the economic effectiveness of promotions with maximum response for minimum cost. Information — predicting alternative promotion outcomes to improve promotion design. Understanding — identifying actual, not assumed, patterns of customer behavior. In the development of a generic scoring model, the threshold question is the determination of what a given behavior is worth. Some scoring models purport to provide a rank-ordering only (e.g., neural-net methods); others try to explicitly model the probability of response (e.g., logistic regression). The ideal scoring algorithm would be one that used value to the corporation as the figure of merit. The models used here employed the volume of Web transactions within a short time period following a promotion as an estimate of the value of customer behavior. Two assumptions underlie this decision. The first is that, since all targeted members had no Web activity in the preceding quarter, these new transactions were caused by the promotion; this will result in a slight overstatement of the success of the promotion, which is corrected by scoring the behavior of a control group. The second is that every Web transaction is a lower-cost substitute for the same transaction by another modality (e.g., a phone customer contact). A more refined model would impute different values to different types of transactions and different drug therapy categories. However, even this gross simplification allows for a much more powerful, interpretable and ultimately rewarding model, just as a credit scoring model is dramatically improved by evaluating amount of delinquency rather than mere incidence of delinquency. Driver Selection: Methodology and Branching Analyses Potential predictors from the e-commerce user datamart were developed in two broad classes: Relatively variable data relating to individual members (date, modality and type of customer service transaction; cost and type of drug therapy). Relatively static data relating to individual members (zip code of member; promotion history and client identification). The collective performance of these possible drivers was investigated in a series of robust, non-linear analyses of families of factors, including such techniques as: classification and regression trees (CART), robust log-linear contingency table analysis plus "smear and sweep," projection pursuit regression and robust "leaps and bounds" regression. Ultimately, the output of these techniques was translated into weighted least squares regression on redefinitions of selected variables. These methods were employed not to create exotic high-order "magic bullet" interactions of complex factors, but rather to achieve two more down-to-earth benefits. First, beyond the normal data laundering activities, considerable attention was devoted to locating multivariate outliers and downweighting their influence. Thus, the models, when they did fail, were allowed to fail spectacularly in a small proportion of the cases — with the advantage of fitting the bulk of the data more accurately. Second, and more importantly, was the identification of behaviorally distinct subclasses of customers that might require distinct models for more accurate prediction. For example, in the first promotion studied, a single driver overwhelmed all others. Keying on this driver would select only 2 percent of members in the overall promotion but would account for 32 percent of new Web transactions with an overall response rate of 40 percent! The celebration was short-lived, however, once that driver was identified as member Web usage two quarters ago (again noting that the screen for all promotions was no Web usage in the preceding quarter). In other words, if a member had ever engaged in e-commerce, any promotion served as a highly effective reminder even when not intended for that goal. But attracting this group is preaching to the choir, an attractive target in building up nominal usage but not for increasing market share. In the end, a hierarchy of four user types was developed, each type representing a different business strategy and each type predominantly on a different kind of information for its score: Current e-commerce users — reward them. Prior e-commerce users — remind them. Other customer-service users (phone and IVRU) — relocate them. Non-customer-service users (mail) — recruit them. For the first three groups, the level and type of customer service provided the most critical data. For the fourth group (without a customer-service history), characteristics of the client and the local neighborhood proved to be the most useful drivers. (These same factors did provide a similar but subordinate role for the first three groups.) Performance and Validation Do these models work? Compared to the alternative in place (using local area demographics alone), the scoring models incorporating customer service history achieved the same number of responses with a promotion of only one-sixth the size. However, predicting the result of a given promotion, using data from that promotion, is, of course, suspect. Confirmation on a take-out sample is much more methodologically reliable but has limited efficiency and ignores the largest sources of variability. That is, a particular promotion will not be run again. Models developed from it will be valuable only if they are useful for other promotions. The models created on a single early promotion were cross-validated with all other promotions with consistent, dramatic success for each member type (excluding, of course, the first group, current e-commerce users). After the passage of time and the vagaries of alternative promotion protocols, the largest non-specific source of variation was the client. The scoring models were also validated by using values based on one selection of clients to predict behavior for an excluded client ("jackknifing"). Future Opportunities (i.e., Current Failures) All of the types of potential drivers played some role in the development of the scoring models, except the drug-therapy category, i.e., type and cost of drug, factors for which initial expectations were quite high. Current research continues on the search for relevant patterns. It may well be that the greatest benefit to be derived from this type of information will be in the refinement of the value estimate for a given transaction. A PC-based forecasting and ad-hoc reporting tool was also developed to support e-commerce business planning. The forecasting system embeds a generic forecasting engine — ForecastPro — developed by Business Forecast Systems, Inc. of Cambridge, Mass. The tool is capable of forecasting e-commerce demand by week, by client and by transaction. The tool supported development of e-commerce forecasts for the year 2000. "Models and analysis developed by Merck-Medco's Information Research organization have supported our e-commerce organization for more than a year," says Amy Foley, director of e-commerce marketing for Merck-Medco. "During this period of extraordinary growth in visitors and shoppers to merckmedco.com, quantitative marketing analysis has provided a value-added lift to our direct-mail promotions in particular, and has provided us with insightful information regarding users and non-users of our site." References E. Corcoran, "The E Gang," Forbes, July 24, 2000. G. Harris, "How Merck Unit Beat Dot-Coms in Web Foray," Wall Street Journal, April 13, 2000. Acknowledgements and Dedication We greatly benefited from interactions with Jill Blank, John Daly, Stu Feldman, Amy Foley, Andy Kuo and Gerald Silverman of Merck-Medco Managed Care, L.L.C. and Khasha Dehnad of AIMS Consulting. The authors wish to dedicate this article to the memory of our good colleague and friend, Dr. Gerald B. Silverman. Edward Binkowski is a statistician working in New York City. He received his Ph. D. in Statistics from Princeton, and has taught at Cornell Medical School, Fordham Law School and Hunter College. Moshe Rosenwein is director for Information Research in Merck-Medco's Information Strategy and Development unit. He received his Ph.D. in Decision Sciences from the University of Pennsylvania. Table of Contents OR/MS Today Home Page OR/MS Today copyright © 2000 by the Institute for Operations Research and the Management Sciences. All rights reserved. Lionheart Publishing, Inc. 506 Roswell Street, Suite 220, Marietta, GA 30060, USA Phone: 770-431-0867 | Fax: 770-432-6969 E-mail: lpi@lionhrtpub.com URL: http://www.lionhrtpub.com Web Site © Copyright 1999, 2000 by Lionheart Publishing, Inc. All rights reserved.