Blog
Decision Trees: Give Predictions a “So What”
Posted by: John Dillard | Posted on: March 1st, 2012 | 4 Comments
In large corporations and in government clients, I often observe a great deal of consternation and uncertainty when it comes to assessing future scenarios, and more importantly, what to DO about those scenarios once they have been identified. Executives really don’t have time to sit down and analyze every “what if” and their teams sometimes don’t know how to go about it, much less how to adjust strategy and/or operations as a result of those futures.
Source: Harvard Business Review
What’s worse, I can’t tell you how many “futures studies” and scenario analyses I’ve seen that — while incredibly interesting — don’t really tell anyone what to do right now about it.
What do I do if our competitors figure out a way to apply carbon nanotube research to a key product line?
What happens to the physician practices my hospital just acquired if medicare payments get cut?
What difference does it make if Brazil nationalizes our supplier of an important raw material?
How will I change if my adversary discovers a way to deny or deceive my best intelligence collection method?
In short, those studies often don’t make the leap from “interesting” to “useful” very often. There is no “so what.”
That is very, very bad news if you either a) spent a lot of money on such a study or b) need to *do* something, like make an investment choice or pick a strategy.
There is hope for the frustrated, overextended, and action-oriented exective. Many tools and techniques are out there that can help, but the one that I think is cleanest and (more importantly) user-friendly — for moderately complex choices — is the decision tree. This post is the first of several I’ll write on the topic, so let’s start with the basics.
Decision Trees are a way to think through the probability of a number of different scenarios, and, in turn, determine the potential financial (or otherwise quantitative) impact of each. They can be applied to almost anything, from the likelihood of terrorist events and their consequences to starting a new business to the possible scenarios that play out if you expand an existing business. I’ve mostly used them in sessions to explore possibilities (and get the perception of the leaders of what’s going to happen), or with R&D clients to think about potential project financial value. It’s a way of applying solid risk analysis to a variety of business problems, whether they’re organizational, national security, or investment-related. The flexibility of the tree is quite remarkable.
In the next post on these, I’ll explore a little more about why you might need one, and several examples applied to both commercial and government problems.
Vernon J. Menard -
January 20, 2012
Great introduction to the topic! I look forward to more posts and to learn how I might use decision tree analysis for strategic planning in my businesses and even at home.
John Dillard -
January 20, 2012
Thanks for the feedback, Vernon. I haven’t tried using them at home yet. . . but perhaps I’ll start planting the seeds for future college selection for my daughters.
Bruce Altmann -
April 5, 2012
Good stuff.
When probabilities are assigned at the “chance nodes”
(and they should be for real decisions);
this % has a rather large impact on actually using the tree.
In my experience even when this % starts with some real math – upon review, it always seems to fall back under the executive influence.
“well maybe the %’s are really more X, Y, Z”
Which is of course based more on decision bias than reality.
The goal of course – is to help the customer avoid this “drift”
May I ask – How do you approach creating the probability % results – to help avoid this political drift?
John Dillard -
April 5, 2012
Bruce – thanks so much for your comments and questions.
You’re absolutely right that the chance nodes can distort the model significantly if they’re chosen poorly.
In most situations where I’ve used the tree approach, that element of subjectivity is unavoidable. It’s just preferable to some other, even more subjective method that the executive / decision-maker would otherwise use. There are situations, however, where you really can assign probabilities that are purely quantitative — but in most cases you’re talking about more high volume decisions where there is a probability that you can assign with a certain level of statistical confidence, depending on your tolerance for error. To get there, you’d need a data set on each node from which you could actually make such statistical judgments, and that’s a pretty tough thing to do with a large tree with many nodes.
So we’re back to the biases issue, which you correctly identified as the big challenge for a more subjective, high-level model. To counter that, the first thing to do in my vierw is to make sure that the team assigning the probabilities is aware of the possible biases. When I was an analyst at CIA and first used this tool, analytical bias training was one of the first things we learned. We wrote a series of posts on that — check it out at http://bigskyassociates.com/2009/09/best-of-the-blog-decision-making-traps/ .
The second tactic is to use more than one method of assigning the probability, if you have the time. For example, you might try to use both a Delphi Method for canvassing expert opinion, and also conduct a traditional scenario analysis, and combine the results. No matter what, though, the user needs to recognize that Decision Trees have an element of subjectivity. In our view, they are better than the alternatives for inherently subjective decisions.
Leave a Comment