5.1 Research Extensions
Message wearout. Thus far, we have not considered the possibility that a customer may be exposed to a message more than once. Advertising theory would
predict a wearout effect associated with an increasing number of repetitions of a message.
One way a marketer could handle ad wearout would be to simply expose a customer to a different message after a certain number of exposures to the original message. An alternative method would be to incorporate ad wearout directly in the regression models by allowing for a curvilinear effect of an additional covariate representing the number of times the customer has been exposed to the message. Of course, both approaches would require that the marketer track the number of exposures to each message for each
customer.
Deciding when to end a pre-test. Although our research is motivated by the opportunities afforded by new media, our approach has a more traditional application that can be exploited by marketers. The typical approach for conducting a pretest is to choose an arbitrary experimental period and then use the pretest results to select a message for the remainder of the marketing campaign. This is similar to the approach taken in
medical trials, in which researchers choose an experimental phase and a terminal phase in which the treatment phase with the higher mean in the experimental phase is used
exclusively during the terminal phase. This sequential medical trials problem has been extensively studied in the statistics literature. Lai, Levin, Robbins and Siegmund (1980) show how to choose the length of the experimental period (i.e., determine a stopping rule) such as to maximize the expected reward for the entire trial (total number of patients treated). A key insight in their approach is that the length of the experimental period should depend on the total number of patients treated.
In certain marketing situations it may not be possible to continually learn and update the parameters of the customer response models. Nevertheless, the marketer can adopt a two-stage pretest approach in which the length of the experimental period is chosen according to Lai et al’s stopping rule. We could adopt their approach, and also extend the theory to handle covariates.
5.3 Concluding Remarks
It is worth emphasizing that our methodology applies to just about any marketing condition, not just advertising. For example, the choice could involve the right content or information to provide a particular customer. Further, our approach is applicable to other media environments besides the Internet. For example, in the typical database marketing example, a cataloger decides to send a particular catalog to a customer based on a model using data from the database. The only difference is that the decisions are not made in real-time, but in waves. The Digital Impact example is a hybrid: the medium is the Internet, but the decisions are made in batches (waves) rather than in real-time.
This paper contains both academic and managerial contributions. On the
academic side, we provide a theoretical framework for investigating a problem that is of high relevance given the recent emergence of significantly different media environments.
As for Internet marketing, theory lags far behind practice. Furthermore, we have
proposed a procedure to solve a realistic bandit problem that incorporates covariates. In doing so, we are filling an important gap in the statistical decision theory and information theory literatures.
The ideas presented in this paper should be of interest to managers for at least two reasons. First and foremost, we offer a procedure that improves expected response rate in a wide variety of marketing applications. Therefore, managers who adopt our approach stand to gain in economic terms. Second, managers can use our ideas to quantify the return on investment of a direct response marketing campaign in interactive media. The implications of these findings for managers will be actionable strategies involving dynamic changes in decisions involving message allocation, advertising creative and execution, promotion, information content choices, and other marketing mix decisions.
References
Ansari, Asim and Carl Mela (2000) "E-Customization," unpublished working paper.
Bather, J. A. (1980), “Randomized Allocation of Treatments in Sequential Medical Trials," Advances in Applied Probability, 12, 174-182.
Bather, J. A. (1981), “Randomized Allocation of Treatments in Sequential Experiments," Journal of the Royal Statistical Society, B:43, 265-292.
Berry, Donald A. and Berry Fristedt (1985), Bandit Problems: Sequential Allocation of Experiments, Chapman and Hall, London.
Blattberg, Robert C. and John Deighton (1991), “Interactive Marketing: Exploiting the Age of Addressability,” Sloan Management Review, Fall, 5-14.
Blattberg, Robert C. and John Deighton (1996), “Manage Marketing by Customer Equity,” Harvard Business Review, July-August, 136-144.
Clayton, Murray K. (1989), “Covariate Models for Bernoulli Bandits,” Sequential Analysis, 8, 405-426.
“Determinants of Click-Through Rates: Some Preliminary Results,” Infoseek Network Advertising Monograph Series, Infoseek Corporation, 1996.
Gittins, J. C. (1979), “Bandit Processes and Dynamic Allocation Indices,” Journal of the Royal Statistical Society, B: 41, 148-177.
Gudmundsson, O., Hunt, M., Lewis, D., Marshall, T. and M. Nabhan (1996),
“Commercialization of the World Wide Web: The Role of Cookies,” Working Paper, Owen Graduate School of Business, Vanderbilt University.
Hoffman, Donna L. and Thomas P. Novak (1996) “Marketing in Hypermedia Computer-Mediated Environments: Conceptual Foundations,” Journal of Marketing, 60 (July), 50-66.
“How to Market and Sell in a Cyberworld,” Direct Marketing, October 1996, pp. 26-27.
“The Internet’s Expanding Role in Building Customer Loyalty,” Direct Marketing, November 1996, pp. 50-54.
Judson, Bruce (1996), NetMarketing: Your Guide to Profit & Success on the Net, Wolff New Media LLC, New York.
Jupiter Communications (1997), “Ad Boom Foreseen Overseas,” Press Release, February 11, 1997.
Lai, T. L. (1987), “Adaptive Treatment Allocation and the Multi-Armed Bandit Problem,” Annals of Statistics, 15 (3), 1091-1114.
Lai, T. L., Levin, B., Robbins, H. and Siegmund, D. (1980), “Sequential medical trials,”
Proc. Natl. Acad. Sci. USA 77: 3135-3138.
Lai (1992), “Certainty Equivalence with Uncertainty Adjustments in Stochastic Adaptive Control,” in Stochastic Theory and Adaptive Control, T. Duncan and B. Pasik-Duncan, eds., Springer-Verlag, New York, 270-284.
Lodish, Leonard (1985), The Advertising & Promotion Challenge: Vaguely Right or Precisely Wrong, Oxford University Press, New York.
“Online Technology Ushers in One-to-One Marketing,” Direct Marketing, November 1996, pages 38-40.
Sarkar, Jyotirmoy (1991), “One-Armed Bandit Problems with Covariates,” The Annals of Statistics, 19 (4), 1978-2002.
“Web Ads – A Lot of Growing to Do, ” The San Jose Mercury News, Business section, June 7, 1997.
Woodfroofe, Michael B. (1979), “A One-Armed Bandit Problem with a Concomitant Variable,” Journal of the American Statistical Association, 74, 799-806.