Practical issues in creating a decision tree

Before talking about how to create a decision tree, it is important to know a few basic principles of how they work.  Here is a sample decision tree:

Note that:

• Time flows from left to right
• Three major types of nodes: terminal (red triangle), chance (green circle), and decision (blue square). Of course, different computer programs may represent terminal, chance and decision nodes differently.
• Branches emanating from a decision node must represent all the options.
• Options must be distinct and not overlap.
• Branches emanating from a chance node must represent all the possible outcomes, must be non-overlapping, and must sum to 1.0.  In the above tree, the “#” sign indicates “1 – (sum of the other branches)”.

 The first step is to structure the problem and create the decision tree.  Ask yourself (and your consultants):

• What are the alternatives?
• What are the relevant outcomes?
• What influences the outcomes?
• Which perspective are you going to take?

 Then, once you have identified and bounded the problem:

• Obtain probabilities
• Estimate utilities
• Do the analysis (including sensitivity analyses)
• Choose a preferred action

Below, we will consider each of these steps in a little more detail.

Creating the decision tree: alternatives and outcomes

The most important thing about creating a model is to make sure that you have considered all reasonable strategies.  When we first created our model for strep, we didn’t consider the strategy of “No treatment”.  However, it turns out that this is a reasonable strategy in some situations.  To cover your bases, talk to experienced clinicians, researchers with an expertise in the field, and even patients.  A physician in fulltime practice will often have useful insights that completely elude the academic physician, and vice versa.  Talk to both.  

Selecting a Perspective

The utility can be assigned from the perspective of the patient or from the perspective of society.  This is particularly important when we consider cost in the next module (Cost-Effectiveness Analysis) but must also be considered in assigning utility.  For example, when treating strep throat, do we consider the benefit to society of returning a patient to work sooner, or only the benefit to the patient in reduced pain and other symptoms?

Assigning probabilities

Probabilities are generally drawn from the medical literature. For example, a recent meta-analysis found that a reasonable probability of strep is 10% in an adult population, and 25% in a pediatric population.  You may also have these data from your own practice setting.  Or, you may have to make a SWAG:  Stupid Wild Ass Guess.  This is also known as getting the consensus of an expert panel.

Assigning utilities

There are a number of approaches to assigning utilities. Generally one can consider each or all of these factors:

• quality of life
• quantity of life
• cost

In the next module (Cost-Effectiveness Analysis), we will talk in more detail about how to evaluate the utility of a given health state.

Determining a range for sensitivity analysis

That same literature search or expert panel can be a source of a reasonable range for your sensitivity analyses.  In the case of strep, studies have found probabilities of strep as low as  4% and as high as 40%.  This would be a reasonable range for sensitivity analyses.  Or, you could consider patient subgroups with more or fewer symptoms, and use a range of 2% to 75% based on McIsaac’s clinical decision rule.