Decision Tree Analysis

Thinking Tools5.2 Selecting the Right Decision ToolsAlso known as Expected Value Decision Trees

Developed by Howard Raiffa and Robert Schlaifer (Applied Statistical Decision Theory, 1961) and Ronald A. Howard (Decision Analysis: Applied Decision Theory, 1966), on the expected-utility theory of John von Neumann and Oskar Morgenstern (Theory of Games and Economic Behavior, 1944).

Draws the choices and chance events of a decision as branches, puts a probability and value on each, and rolls them back to compare options by expected value.

Pilot a delivery subscription first, or launch it in every store now?

Squares are choices, shaded boxes are chances. Set the odds with the steppers under the tree: each chance box averages its branches, each choice takes its best, and the first choice marks the better plan. Raise both chances past 80% and the pilot stops paying for itself. Illustrative.

Select a box to read it.Subscription plan. Payoffs are net value over three years, in millions, after every cost including the pilot's.Pilot in two stores $1.5M, Launch in all stores now $0.5M, Do nothing $0.0M. Best: Pilot in two storesLaunch in all stores nowDo nothingTwelve-week pilot. Two stores, twelve weeks, about $0.5M all in. The pilot's cost and delay are already inside each payoff on this branch.50% × $2.5M + 50% × $0.5M = $1.5MAfter a good pilot.Scale to all 40 stores $2.5M, Stop anyway −$0.5M. Best: Scale to all 40 storesStop anywayScale to all 40 stores.Payoff: $2.5MStop anyway.Payoff: −$0.5MAfter a weak pilot.Pivot to office catering $0.5M, Stop −$0.5M. Best: Pivot to office cateringStopPivot to office catering.Payoff: $0.5MStop.Payoff: −$0.5MDemand at launch. No pilot: vans, staff, and marketing for forty stores at once.50% × $3.0M + 50% × −$2.0M = $0.5MStrong.Payoff: $3.0MWeak.Payoff: −$2.0MDo nothing.Payoff: $0.0M

Twelve-week pilot

Pilot works50%Pilot fails50%

Demand at launch

Strong50%Weak50%

An example

A chain of forty bakeries was weighing a bread and pastry subscription with home delivery. The marketing lead wanted to launch in every store at once; the operations lead wanted a pilot first. After two rounds of argument, the finance manager drew it as a tree.

Three branches left the first square: pilot in two stores, launch everywhere now, or do nothing. The room judged launching now a coin flip, worth $3.0 million over three years if demand was strong and a loss of $2.0 million if it was weak. The pilot cost about $0.5 million and a quarter's delay; after a good one, scaling was worth a net $2.5 million, and after a weak one, pivoting the vans to office catering a net $0.5 million.

Rolled back, the pilot was worth $1.5 million and the immediate launch $0.5 million. The finance manager then raised the odds of strong demand on both branches: launching now won only above about 83 percent. Nobody was, and the pilot started in March.

Why it works

A decision tree separates what a team chooses from what happens to it, and puts a number on each. Arguments about a launch usually blend the two, so one person's confidence in demand gets mistaken for a plan. Drawn out, the choices are squares, the uncertainties are circles, and every end has a value in one unit.

Rolling back is where the insight comes from. A circle is worth the probability-weighted average of its branches, and a square is worth its best branch, so a later choice, such as stopping after a weak pilot, is valued today. That is why pilots are often worth more than they look: they buy the right to change course. Moving each estimate until the answer flips then shows which number the decision actually depends on, and so where more data is worth paying for.

The thinking behind it

Howard Raiffa Howard Raiffa, a Harvard decision theorist, helped found decision analysis. With Robert Schlaifer he wrote Applied Statistical Decision Theory (1961), which laid Bayesian foundations for decision trees: assign probabilities to uncertain outcomes, update them with evidence, and choose the act with the highest expected value. The book was published by Harvard Business School. Raiffa later carried the same reasoning into negotiation in The Art and Science of Negotiation (1982).

Ronald A. Howard Ronald A. Howard, a Stanford professor, coined the term decision analysis in his 1966 paper Decision Analysis: Applied Decision Theory and helped found the field. It models a decision through three parts: the alternatives open to the decision maker, the probabilities of what may happen, and the value of each outcome. Those three parts are exactly what a decision tree draws. He also helped develop the influence diagram, a related way of laying out a decision.

John von Neumann John von Neumann and Oskar Morgenstern set out expected-utility theory for rational choice under risk in Theory of Games and Economic Behavior (1944), the book that also founded game theory. Its rule, weigh each outcome by its probability and choose the option with the highest weighted value, is what the tree applies at each circle and square. A circle's value is that weighted average, and a square's value is its best branch, so rolling the tree back from its ends is the rule applied again and again.

How to run it

  1. Draw the choices first. Each option open now is a branch from a square, and doing nothing is one of them.
  2. Add what can happen after each. Chance events are circles; later choices, such as scale, pivot, or stop, are squares after them.
  3. Put a value on every end. Use one unit throughout, such as net value over three years.
  4. Estimate the odds at each circle. Use the record of similar launches where it exists; the branches at each circle add to 100 percent.
  5. Roll back from the right. At each circle take the weighted average; at each square take the best branch.
  6. Find the flip points. Move each estimate until the answer changes, and note how far it had to move.
  7. Spend effort on the estimate that flips it. If a plausible change reverses the choice, get better data on that number before committing.

Where it goes wrong

The numbers look more certain than they are. Odds guessed to the nearest percent get treated as facts. Spot it when nobody can say where a probability came from, and report the flip point next to the answer.

An average hides a loss the business cannot take. A branch that would sink the company gets averaged with good ones and disappears. Spot it when any single end is larger than the business can absorb, and rule that branch out before comparing the rest.

The tree grows past estimating. Thirty ends with guessed odds produce an answer nobody believes. Keep only the branches that could change the choice, and fold the rest into one estimate.

When to use something else

  • Real Options Analysis. When most of the value lies in the right to wait, expand, or abandon over time, and that right needs pricing.
  • Monte Carlo Simulation. When many uncertain inputs interact and the question is the whole range of outcomes, which a single expected value hides.
  • One-Way and Two-Way Doors. When the main question is whether the decision can be undone, and how much care it deserves.

Where it sits in SOLVE eX

Go to the source

  • Applied Statistical Decision Theory, 1961
  • Decision Analysis: Applied Decision Theory, 1966
  • Theory of Games and Economic Behavior, 1944

Back to the catalog