A Better Mousetrap

A conventional response model technique, such as logistic regression, predicts probabilities by looking at purchase pattern in relation to explanatory elements like historic transaction, demographics and geographics. Customers therefore can be scored in terms of likelihood projection, which allows direct marketers to focus on top-ranked niches and thus raise marketing ROI. This is what every firm dreams of: Mail less and mail smart.

But does this type of model truly address your business issue—measuring a specific campaign and boosting its impact in the next drop? In a real-world environment, customer behaviors are most likely driven by multiple campaigns through different channels, such as direct mail, telemarketing, e-marketing, as well as self-promoted customer service centers and/or referral programs. In consequence, the model built under such a marketing mix can be largely skewed when assessing a single effort.

An effective solution is the use of a differentiation model that is capable of identifying a pure contribution exclusively made by a specified campaign while getting rid of unexpected noises from any other influences.

First, let’s look at how performance is measured for most campaigns.

A Typical Campaign Measurement

Table 1, shown below, illustrates how this approach works toward your marketing objective. In this example, a direct mail campaign is implemented with 50,000 customers treated and another 50,000 pulled as the holdout group. Based on test results displayed in Table 1, the statistical difference between the two groups then is calculated to claim whether the direct mail effort generates a real incremental lift against the holdout group. Assuming the mailing cost absorbs 20 percent of the equivalent response rate, the results for the mailed group in this example, with this cost adjustment, gives an apples-to-apples comparison with the holdout response rate. In this case, no significant lift has been found between the cost-adjusted response rates of 2 percent and 1.86 percent (the z-value, or confidence level, works out to a 95 percent rate). Based on these results, should this marketing initiative be abandoned? Let’s take a closer look.

A “Different” Model

What modeling does is turn no difference into real difference, where an entire customer base is segmented in terms of significant attributes. The direct mail segment in Table 2 (shown on page 110) exhibits a typical segmentation procedure starting with the first entry (age). Similarly, more variables can be selected into a multivariate model like ANN (artificial neural network), CHAID (chi-square automatic interaction detection) or regression analysis.

Many times, a modeling procedure only sticks to the treated segment while ignoring the benchmark, or holdout, segment. Unfortunately, the model developed in lack of the holdout information often describes an overall likelihood of responding to multiple campaigns rather than a single one. A differentiation model solves this puzzle.

According to the direct mail segment in this example (see Table 2), older customers seem more likely to respond to your promotion. By further investigating the holdout segment, you realize those senior sectors virtually did better without receiving this mailing. A comparison between the treated and holdout groups alters the previous conclusion based on the direct mail segment results alone and indicates a revised outcome and a new approach: targeting those groups with consumers between the ages of 36 and 65.

As a basic rule, a differentiation model divides an entire population into two categories—one with significant gain and another with significant loss. An overall response can be enhanced because you keep the positive impact and drop the negative impact. The downward side sometimes could be hard for a company to accept: How can a costly direct marketing effort hurt sales?

One answer: It has been proven that wrong messages are worse than no message to some customers. That is to say, once customers are interrupted or misled by any unwelcome marketing touches, they either change their mind or simply switch to your competitors. To this point, eliminating older and younger sectors has nothing to do with evaluating your customers for response likelihood, since they could be very profitable groups but just don’t like the way in which you approached them.

Getting a Better Result

This is the essence of a differentiation model that explicitly develops a picture of a single campaign’s performance out of a marketing complexity. It is meaningful whenever the holdout response rate runs significantly above zero, which implies that responses are influenced by other media you may be aware of … or not at all. By addressing each campaign issue in a differentiation manner, direct marketers can pinpoint each marketing objective and eventually leverage marketing efficiency as a strategic whole.yy

Amoy X. Yang is a senior database-marketing analyst with Fifth Third Bank in Cincinnati, Ohio. He can be reached via e-mail at amoy_y@yahoo.com.