Misix Library

Following the Breadcrumbs: An Example of Decision Tree Analysis in Action

July 29, 2015

Imagine you’re the lead underwriter at a major insurance carrier. The number of policyholders filing multiple claims per year is on the rise, and you’ve been tasked with reassessing premiums to ensure they reflect the level of risk the organization is facing. Easier said than done. Identifying risk-prone policyholders after a claim is filed is simple, but this forces the organization into a reactionary position and does nothing to protect it upfront.

What if you could identify the characteristics of policyholders that file multiple claims per year? With this information in hand, you could then identify other policyholders with similar characteristics and set their premiums knowing they are the most likely to file multiple claims per year. Problem solved!

For those unfamiliar with data science this may sound like a fantasy, but leveraging historical data to make predictions about future cases can certainly be accomplished, though it isn’t necessarily easy. To do this, we need a tool that allows us to segment data into a logical structure, with each piece guiding us to the next in a meaningful and statistically significant way. That’s where decision tree analysis (DTA) comes into play.

 

What is decision tree analysis?

In a nutshell, DTA is a method of classifying individual elements into groups based on their characteristics. These groups are created based on a target dependent variable (the decision question itself; using our previous example, “Will this policyholder file multiple claims per year?”) and a set of independent variables. In simpler terms, DTA allows us to pose a question or scenario and analyze the impact a host of variables could have on the outcome in order to determine the most appropriate course of action. DTA is utilized in countless situations – from policy making to risk assessment, and everything in between – but for the purposes of this article we’ll focus on two aspects of DTA that can significantly enhance an organization’s marketing efforts:

1. Segmentation. Identifying homogeneous groups.

2. Prediction. Constructing rules for making predictions about individual cases and applying said rules to real-world applications.

The easiest way to understand DTA is to see it in action, so let’s take a look at a decision tree we created for this discussion.

 

Decision tree analysis in action

We began with information sourced from the National Household Education Survey, which collects data on the educational activities of the U.S. population. These specific data are from a survey of parents of children age 3 to 6 regarding their child’s readiness for school. The researchers examined the factors that relate to whether or not a child can “write (his/her) first name, even if some of the letters aren’t quite right or are backwards.” This will serve as our dependent variable.

It’s important to note that this won’t be a rigorous analysis; we aren’t going to include any of the available demographic or socio-economic data. Rather, we are going to examine how a child’s interactions with other family members relate to their ability to write their name.

To collect their data, researchers asked the following questions:

  1. When you or someone in your family reads to (CHILD), how often do you …Ask (CHILD) to read with you?
  2. When you or someone in your family reads to (CHILD), how often do you …Stop reading and ask (CHILD) to tell you what is in a picture?
  3. When you or someone in your family reads to (CHILD), how often do you …Stop reading and point out letters?
  4. When you or someone in your family reads to (CHILD), how often do you …Talk about the story and what happened when the book is done?
  5. About how many books does (CHILD) have of (his/her) own (including those shared with brothers or sisters)?
  6. Does (CHILD) actually read the words written in the book, or does (he/she) look at the book and pretend to read?
  7. Which television networks or channels, for example ABC, Nickelodeon, Discovery Channel, or PBS, does (CHILD) watch at least once per week?
  8. In the past week, has anyone in your family done the following things with (CHILD)?   Played board games or did puzzles with (CHILD)?

Using the data aggregated by the researchers, we created the decision tree you see below. It may look complicated at first, but have no fear. We’ll walk through it together.

 

The results

While there are several methods for presenting a decision tree, we feel they’re easiest read in a top-down format, similar to a flowchart. In this layout, you’ll see the decision question (dependent variable) listed at the top. In our case, that question was “Can the child write his/her name?” The response percentage was as follows:

  • Yes (64.9%)
  • No (35.1%)

So we know that nearly two thirds of the children in our sample can write their own names. That may answer the question at its highest level, but it doesn’t provide insight into what’s influencing the results. If we segment the children by their parents’ responses to the survey questions, however, we can see the entire picture.

SecondLevel

The second level represents the characteristic that most statistically significantly relates to a person’s response to the decision question. In this case, we found that the frequency of a family member asking the child to read along with them most significantly relates to whether or not the child can write their name. This characteristic was separated into four nodes based on how often the family member asked the child to read with them:

  • 2 – Sometimes
  • 3 – Never
  • 1 – Usually
  • -1 – Inapplicable

Each of these nodes has a different distribution of response with respect to the decision question.

  • Of the respondents who said they “usually” ask their child to read along with them, 75.6% reported their child can write his/her first name (Node 3).
  • Of the respondents who said they “never” ask their child to read along with them, only 54.9% said their child can write his/her first name (Node 2).

This doesn’t mean the frequency of asking children to read along with you causes children to be able to write their first name. Rather, the two are related: the rate of successful name-writing is statistically significantly different for those who “usually” ask their child to read along and those who “never” ask their child to read along.

We then repeat the process of growing out the branches of the tree for each node until there are no longer any characteristics left that are significantly related to the decision question.

By examining the bottom-most row of nodes we can see which of our children are most likely to be able to write their name.

Node23

Node 23 shows us that all 69 respondents who met the following criteria had a child who could write their name:

  • “Usually” ask their child to read along with them.
  • Child actually reads the words written in the book, or both reads the words and looks at the book and pretends to read.
  • Watches PBS television network at least once per week.
Node17

What about those children least likely to be able to write their own name? Node 17 shows that 68% of respondents who met the following criteria had a child who could not write their name:

  • “Never” ask their child to read along with them.
  • “Sometimes” or “usually” stop while reading to child and point out letters.
  • “Never” talk about the story and what happened when the book is done.

 

What are the benefits of decision tree analysis?

As we see when reviewing the results of our example, decision tree analysis can simplify the decision making process in several ways:

 

  1. It prioritizes characteristics based on their impact on the decision criterion.
    • In our case, we found the frequency of a family member asking the child to read along with them has a greater impact than any other characteristic.
  2. It omits characteristics that do not significantly impact the decision criterion.
    • This is especially useful in cases where two characteristics are highly correlated.
      • For example, suppose we find single people experience more traffic accidents than married people; however, single people are more likely to be younger, and age is a more significant indicator of whether a person has been in an accident, so marital status should be omitted from the model.
  3. It quickly classifies responses across many observations and many characteristics and pinpoints the combinations that maximize the decision criterion.
  4. For continuous variables, it finds the thresholds at which groupings become statistically significantly different.
    • In this example, three distinct groupings branch off from Node 1 based on the number of books the child has on his/her own. While other techniques are able to identify marginal effects, they lack the ability to pinpoint exact thresholds.

 

Decision tree analysis and marketing

Analyzing what influences the ability of a child to write his or her name is surely a worthwhile objective, but we’re marketers, and we tend to view the world through that unique lens.

For example, imagine the level of segmentation seen in our example applied to your database of contacts. You could identify the prospects most likely to purchase your product, purchase a second product in the future, purchase an extended warranty, and so on. By segmenting based on statistical significance, a decision tree can help you understand the variables with the greatest influence over your outcomes.

Similarly, the predictive attributes of DTA offer marketers a powerful tool when setting strategies. Gaining even a rough understanding of anticipated campaign performance can help marketers prioritize audiences, set realistic budgets and much more. The target audience is a small group with a lower likelihood of purchasing? Probably wise to invest less in the campaign. The target audience is a small group with a higher likelihood of purchasing? Then it’s time to break out the checkbook and swing for the fences.

 

Before we close, it’s important to note that even the most well-conducted decision tree analysis does not represent any sort of guarantee. While aggregating and processing historical data can be tremendously useful, it simply can’t account for unknown future variables. But for anyone who understands the technique’s capabilities, decision tree analysis can take much of the guesswork out of the decision making process.

 

How do you use decision trees? Let us know in the comments below.

 

For updates on our blog and other information, follow us on Twitter.@misixanalytics

Data drives your business, data should drive your marketing. Find out how Misix helps businesses redefine their marketing in a data-driven world by visiting https://misix.com/process.