What decision tree analysis is (and which decision tree we mean)
Search 'decision tree' today and most of what you find is a machine-learning algorithm — a classifier that splits data on features. This page is about the older, human tool: the decision-analysis tree, the technique canonized in Howard Raiffa's Decision Analysis (1968) and the Harvard/Stanford decision-analysis tradition, for reasoning through choices under uncertainty.
The grammar is small. Decision nodes (drawn as squares) are points where you choose. Chance nodes (circles) are points where the world chooses — an approval comes through or doesn't, demand is high or low — each branch carrying your estimated probability. Terminal nodes (leaves) carry the payoff of ending up there. Time flows left to right: first commitment first, downstream consequences after.
The analysis is the rollback: starting from the leaves and working backward, each chance node collapses to its probability-weighted average, each decision node to its best available branch — because when you get there, you'll pick the best branch. What arrives at the root is each initial option's expected value, and often something more useful: the discovery that a staged strategy ('pilot first, then decide') dominates both bold commitment and doing nothing. Where it sits in the toolbox: decision-making models; for the static version of the money question see cost-benefit analysis.
When to use it — and when not to
Decision trees earn their keep when:
- ✓The decision is sequential. Choices now open or close choices later — pilots, staged investments, option-to-expand structures. This is the tree's home turf, and it's what flat comparisons can't see.
- ✓Uncertainty is discrete and estimable. Regulatory approval, deal-close, technical milestone — events you can honestly put a probability range on.
- ✓You suspect information has value. Trees price the 'run the study first' branch — the expected value of buying knowledge before committing, one of decision analysis's genuinely deep contributions.
And where they mislead:
- ✗Bet-the-company stakes. Expected value is the right criterion for repeatable decisions; for one-shot survival stakes, a 10% chance of ruin is not offset by a juicy mean. Risk attitude belongs in the analysis, and naive EV ignores it.
- ✗Probabilities nobody can defend. The tree's output inherits the quality of its estimates; 0.6 typed into a branch is a claim, and if no one can say why, the arithmetic is theater. (This is the argument-tree connection below.)
- ✗Combinatorial explosion. Model every uncertainty and the tree becomes unreadable. The craft is choosing the two or three uncertainties that actually drive the decision.
- ✗Deep ambiguity. When you can't even enumerate the outcomes, probabilities are premature — that's scenario planning's territory.
Step by step, with a worked example
Illustrative scenario: an invented biotech-tools startup deciding whether to launch its product now or run a certification study first. The procedure:
- 1Frame the root decision. Three branches from the first square: launch now; run the certification study first (one year, at a cost); shelve the product.
- 2Map the uncertainties per branch. Launch now → chance node: market accepts uncertified product (estimated 40%) or rejects it (60%). Study first → chance node: certification granted (70%) or denied (30%); if granted, a further decision square: launch certified (higher acceptance, estimated 80%) or license the technology out.
- 3Attach payoffs at the leaves. Each end state gets a value: certified-launch success is the big payoff; failed uncertified launch is a loss plus reputation damage (costed explicitly); shelving is zero minus sunk development.
- 4Roll back. Average each chance node by its probabilities; at each decision square keep the best branch. Suppose the numbers make study-first worth more than launch-now despite the year's delay — the certification's boost to acceptance more than pays for the wait.
- 5Stress the estimates. At what acceptance probability does launch-now win? If the answer is 55% and your honest range is 30–50%, the conclusion is robust; if the flip point sits mid-range, the decision hinges on an estimate — and that estimate deserves the scrutiny of step 6.
- 6Interrogate the load-bearing numbers. The 70% certification estimate came from somewhere — comparable approvals? counsel's judgment? Make the basis explicit and challengeable before the arithmetic gets believed.
Decision trees and argument trees — complements, not rivals
The names invite confusion, so here is the clean split. A decision tree computes: given these probabilities and payoffs, which strategy is worth most? An argument tree reasons: are these probabilities and payoffs actually credible? In decision-quality terms, the decision tree structures alternatives (including the staged ones flat analysis misses) and information (estimates made explicit); the argument tree supplies sound reasoning about exactly those inputs:
Each probability → a claim
"Certification succeeds with ~70% likelihood" becomes an argument carrying its basis — comparable approvals, counsel's view — with challenges and support attached, and a rating that reflects the group's scrutiny.
Each payoff → a mini-case
The big-success leaf value is a forecast; its assumptions (pricing, adoption rate) become sub-arguments someone can attack before the number does its work.
The recommended strategy → root claim
"Run the study first" sits at the root with the rollback as headline evidence — and the sensitivity flip-points as named open questions.
Estimates move → the tree updates
When counsel revises the approval odds, the claim updates, the rollback recomputes, and the recommendation's visible support shifts — no stale spreadsheet.
The decision tree does the arithmetic; the argument tree earns the inputs. Run the numbers where they can be challenged — see decision quality for why the reasoning link is the one no framework supplies.
Decision trees vs the alternatives
| If your question is… | Reach for | Why not a decision tree |
|---|---|---|
| One-shot comparison, uncertainty minor | Cost-benefit analysis | Tree machinery adds nothing without branches |
| Flexibility itself is the value — wait, stage, abandon | Real options thinking | Options logic extends the tree with market-priced framing |
| Outcomes can't be enumerated, let alone probabilized | Scenario planning | Probabilities on unknowable futures are false comfort |
| The discounting of far-off payoffs is the crux | NPV & IRR: the limits | Trees inherit every discounting judgment call |