Calculating a Realistic Return Rate Using Customer Defection and Historical Data in the Continuing Education Industry

By Jacob Ensign, Business Analyst, JMH Consulting, Inc.

The article “Computing True Return on Investment Using Customer Lifetime Value in the Continuing Education Industry: A First Glance”, discusses the relationship between return on investment (ROI) in marketing and the customer lifetime value (CLV). The article provides a best method for computing CLV. In particular, this method lets your business model drives the methodology for computing CLV. However, attempting to compute CLV in a rigorous way leads us to another concept: customer defection.

Companies already use customer defection as a way to measure how healthy their business is because it is a clear measure of perceived on-going value that the company offers its customers. It also serves as a predictor of diminishing revenue and profits because existing customers tend to generate more revenue and higher profits than new customers do.

For companies in most industries, calculating CLV and recognizing when a customer defects is a straightforward exercise. If I am the sugar supplier for pie baking companies across the States, I know that customer defection is the same as a customer not renewing a contract. If I am the owner of a restaurant, I know that customer defection means families stop eating my food (and probably going to another restaurant down the street).

Unfortunately, for us in the continuing education industry, it is not so straightforward. Customer engagements in continuing education and lifelong learning are not as regular as customer engagements in other industries. For all businesses, the rate of customer defection relates closely to the rate of return (ROR). After all, if a customer does not defect, then the business retained her/him.

Why we need to know customer defection to compute customer lifetime value

As we discuss in the “First Glance” article, CLV is a necessary tool for computing true ROI of your marketing efforts. Furthermore, you can easily determine your CLV from the ROR of your customers. In other words, computing ROR is almost the entire battle.

ROR is easy to compute, right? Count every registrant in your database that shows up more than once, and then divide that number by the total number of registrations. This gives you the proportion of registrants that returned—the rate of customer return. Like this:

Organizations in our industry define ROR this way, but as seen in the “A First Glance” article, it is not this easy. This formula ignores the student who is taking her first class right now, because it does not take into account the next several months in which the student may or may not register for additional classes. Thus, behavior of students several years ago inflates the ROR of this equation, and it weighs current behavior less heavily.

This becomes especially critical when you think about the lifelong learning organization that continually works to improve ROR. If you come up with a great idea to improve ROR, this equation will probably not reflect those improvements for years to come. It will feel like climbing out of a dark well. Thankfully, there is a better way.

If we understand when a customer defects, i.e. is no longer counted as an active customer, then we can calculate ROR more precisely and measure the affect that our ROR-improving efforts have.

How to define customer defection

The real issue in determining customer defection is how long you allow a customer to wait before returning, i.e. a time limit for labeling a customer a “defector”. Consider some extreme cases:

• Student 1 took a Spanish class last fall; he enjoyed it so much that at the end of the class, he went straight to your website and enrolled in an intermediate level Spanish class that very same quarter.
• Student 2 took a class on Performance Management; she will return to take another management class when her schedule permits—in about five years.

It is easy to tell that student 1 is a non-defector and that student 2 is probably a defector. As an organization, we would hurt our bottom-line by spending our resources continually marketing to student 1 as an active student. More specifically, when should we retire student 2 to the defector list? That is, when should we start treating the first student like a new student again? Your unique business operations and processes must answer the question for you.

A data-free approach

Consider a hypothetical situation in which your organization sends a catalog of all upcoming classes quarterly. How do you determine which past students go onto the mailing list? If you pull all registrations from two years back, then two years is a good cut-off. Your process of only including students from within two years maintains that if students have not come back within two years, despite your quarterly mailing, they have defected. That is, your process of pulling a mailing list pre-determined the two-year cut-off for defection. For you, this is a good approach because it fits with your business model; however, this approach can generate some additional questions.

• Is this the optimal amount of time? Is there a better cut-off that will improve revenues and profitability?
• Is there another operation or process that should weigh in to this decision? Ultimately, it depends on the complexity of your organization’s processes, your course offerings, and your methods of marketing.
• How did you make that decision to begin with, and how can the methodology for that decision impact the bottom line?
• In our web-connected world, communication occurs much, much faster than we can control with brochure mailing. Thus, what about search-engine optimization (SEO) projects, pay-per-click (PPC) campaign overhauls, and email marketing efforts?

Clearly, there are some shortcomings in this approach to computing ROR, but it gives us a great start. This approach can give you a general idea of how “healthy” your continuing education programs are. Nevertheless, for computing a rigorous ROI to make those all-important marketing decisions, this approach is lacking.

A slightly more data-intense approach to determining customer defection

There is a straightforward method for using historic data to compute a simple ROR. The idea is to examine the behavior of all registrants who did come back to take another class. In this case, a behavior is the number of days a registrant waits before registering for another class. For each registration that had another registration following, we will compute the time it took the registrant to return and then build a distribution chart of all customer behaviors. We will use this data, then, to determine what behaviors reflect defection and non-defection behavior.

Hypothetically, if the distribution were normal, it would look like the chart in figure 1. In a normal distribution, we can compute the standard deviation (denoted byσ). In this case, standard deviations are “slices” of the total distribution, which are measured in time, e.g. 60 days. Three standard deviations to the right of the middle of the distribution (+3σ) represent the limits of natural variance; that is, anything further to the right of +3σ is outside of what we accept as normal behavior. If the standard deviation were 30 days, then +3σ would be 180 days.

As figure 1 shows, in a normal distribution, there are extremely low numbers of registrations beyond the third standard deviation (+3σ). Therefore, if the distribution were normal, we could use the +3σ limitation to create our time limit for customers to be considered defectors. In other words, using the example, any customer who waits longer than 180 days to register for a new class is a defector.

Unfortunately, for any continuing education or lifelong learning organization, the distribution will not be normal. In fact, figure 2 illustrates a skewed distribution using a real set of data. This is data characteristic of such organizations as it skews heavily to the right side of the distribution. In this case, the standard deviation is approximately 132 days. If we used the approach that we use on normal distributions by multiplying the standard deviation three times, then the defector behavior would occur after 395 days, about 13 months. However, since the distribution skews heavily, the standard deviation is not a good measure for variation, so we need a different approach.

Instead, we can construct a chart in a grid (see figure 3). Along the vertical axis, we will again plot registration data, and along the horizontal axis we will plot behavior data (how long the registrant waited to re-register). This time, however, we will use percentages. Thus, the vertical axis contains the percentage of all return registrations, and the horizontal axis contains the percentage of all different behaviors (or the amount of time a registrant waits to return for another class).

Now our goal is to pick a behavior that represents the extreme of non-defector behavior: everything to the left is non-defector, everything to the right, defector. For example, if 395 days is the extreme of non-defector behavior, then any registrant waiting more than 395 days is a defector, any registrant waiting less than 395 days is not.

Like the center of a normal distribution, we are looking for the largest “lump” of people while allowing a narrow range of possible behaviors. Since the cut-off behavior is to represent the extreme of “typical behavior”, it will ideally:

1. Maximize the number of return registrations
2. Minimize the number of behaviors that are considered non-defective (so we avoid allowing extreme behaviors)

In figure 3, a grid that plots the same data as figure 2. We want to pick a point on the line that is as close to the upper-left corner as possible, i.e. as close to 100% of registrations and 0% of behaviors as possible. That way, it will satisfy the two cases above for defining “typical behavior”.

The dotted line on the chart connected the upper-left corner of the grid to the nearest point on the data curve. Thus, the point on the curve where the dotted line crosses represents our desired point: somewhere around 83% of registrations and 16% of behaviors. In this particular set of data, this means 145 days is the cut-off.

We can see how this is a good point by comparing some points on either side of it. Think about the following examples like this: We can choose what cut-off divides defector from non-defector behavior. By moving this margin to the right (larger numbers), we collect more registrants, and by moving it to the left, we give up registrants. In essence, we can “buy” a larger group of registrants, by “spending” more in the behavior margin. Therefore, the goal is to get the registrants, for the least behaviors.

If we consider 90% of registrants to be non-defectors, we open up the cut-off from 145 days (16% of behaviors) to 244 days (26% of behaviors). Doing this, we increased the non-defector behavior margin by 10 percentiles, but only increased the number of registrations by seven percentiles.

Let us examine a point lower than 83%, say 80%. If we consider 80% of all registrants to be non-defectors, we only allow 13% of behaviors. Moving it from 75% to 83% of all registrations corresponds to a six-percentile increase in the non-defector behavior margin (10% to 16%), and captures eight more percentiles of registrants. By contrast, moving from 75% to 83% of registrants is relatively “inexpensive” compared to the move from 83% to 90%. Thus, the cut-off at 16% (144 days) is an efficient one.

Computing customer lifetime value

With this new definition of customer defection, we can now compute a believable ROR. Now, to compute the most recent ROR, we can count all registrations from any time range up to 144 days before today. Count all of the people in that dataset that took another class within 144 days, and then divide by the total number of registrations in the dataset.

However, this model gives us some more valuable insight into the behaviors of the common registrant. Typically, if a customer is going to return to our organization for more lifelong learning or continuing education, s/he will do so within 144 days (for this particular organization). That gives us an idea of what type of window we have to work with for marketing to active (non-defector) registrants. We can safely assume that registrations that do not occur within that time imply that we can treat the customer like a potential, not past, student.

Conclusions

While it is difficult to derive some common business key performance indicators in our industry, they are not out of reach. In fact, it is critical that lifelong learning and continuing education organizations begin to think about their operations as rigorously as for profit businesses, in particular in light of recent economic conditions. As we saw in the “A First Glance” article, a crucial concept in marketing, ROI, depends upon customer lifetime value, which in turn depends upon our ability to determine which types of registration behavior reflect defector behavior and which do not. This article represents a best practice in the computation of ROR using well-defined customer defection.

A quick word of caution, however: The dataset we examine in this article is an example of one organization in our industry. While the overall trends (skewness) may be similar in different locations and between different groups (professional versus enrichment program attendees), the actual data themselves will vary widely. We will explore the differences between offering types in a later article. Ultimately, however, you must examine your own customer behavior patterns to make these determinations for your organization. As discussed in “A First Glance”, you must also be conscientious of differences between attendees of disparate programs.

Finally, this method of determining customer defection has applications in other aspect of our industry. For example, information sessions meant to funnel people into your certificate programs have lifecycle, as do customers. In a subsequent article, we will apply the same method used in this article to determine when an information session stops converting attendees into program registrants.