Decision Making
A decision is a commitment to one of several possible actions in the face of uncertainty about their consequences. Good decision-making is not the same as getting good outcomes — luck intervenes — but consistently good decisions produce better outcomes over time. This skill covers the structured methods decision scientists use to bring rigor to choices that matter.
Agent affinity: kahneman-ct (System 1 / System 2 allocation), tversky (expected value and biases), paul (integration with elements of reasoning)
Concept IDs: crit-decision-frameworks, crit-calibrated-confidence, crit-intellectual-humility
The Decision Toolbox at a Glance
| # | Method | Purpose | When to use |
|---|---|---|---|
| 1 | Expected value calculation | Weigh probabilities and payoffs | Repeatable decisions with quantifiable outcomes |
| 2 | Decision trees | Map sequential choices and chance nodes | Multi-stage decisions with contingencies |
| 3 | Multi-criteria decision analysis (MCDA) | Weigh multiple incommensurable criteria | Choices involving trade-offs across dimensions |
| 4 | Pros and cons with weights | Simple MCDA for everyday decisions | Personal choices, not enough data for formal analysis |
| 5 | Pre-mortem | Imagine failure to surface risks | Before committing to a major plan |
| 6 | Reversibility check | Decide how much deliberation is needed | Every decision |
| 7 | Two-way door vs. one-way door | Distinguish easily-undoable from locked-in | Speed decisions for reversible, delay for irreversible |
| 8 | Minimax regret | Minimize worst-case regret instead of maximizing expected value | Extreme uncertainty; loss aversion is justified |
| 9 | Satisficing | Pick the first option meeting minimum criteria | When the cost of searching exceeds the benefit of finding the best |
| 10 | Stopping rules | Decide in advance when to stop deliberating | When analysis paralysis is a risk |
The Fundamental Distinction — Good Decision vs. Good Outcome
A decision can be good even if the outcome is bad, and a decision can be bad even if the outcome is good. Confusing these is the root of most decision-making errors.
Good decision, bad outcome. You evaluated the options carefully, weighed the probabilities, chose the highest expected value action. The low-probability bad outcome happened anyway. This is not a decision error; it is luck. Recording it as a decision error would corrupt future decisions.
Bad decision, good outcome. You took a reckless action that had a high probability of failure. It worked out anyway. Do not learn "reckless action is good" from this. Over time, bad decisions produce bad outcomes on average.
Discipline. Evaluate decisions by the process at the time, not by the outcome in retrospect. Annie Duke calls this "resulting" — judging decisions by outcomes — and identifies it as a primary corruption of the decision-making process.
Method 1 — Expected Value
Pattern: For each option, compute the sum over outcomes of (probability of outcome) × (value of outcome). Choose the option with the highest expected value.
Formula. EV(option) = Σ P(outcome_i) × V(outcome_i)
Worked example. Deciding whether to accept a 50% chance of winning $1000 or a guaranteed $400.
- Option A (gamble): 0.5 × $1000 + 0.5 × $0 = $500
- Option B (guaranteed): 1.0 × $400 = $400
EV favors option A. But EV is only the right criterion when the decision repeats many times. For a one-shot decision, loss aversion and risk tolerance matter.
Limitations.
- Requires probability estimates that are often unavailable or unreliable.
- Treats all values as commensurable (all convertible to a single currency).
- Ignores risk aversion, which is a legitimate preference for one-shot high-stakes decisions.
- Ignores variance — two options with the same EV but different variance are not equivalent for most humans.
Method 2 — Decision Trees
Pattern: Draw a tree with choice nodes (where the decision-maker picks) and chance nodes (where the world picks). Compute expected value back from the leaves to the root.
Worked example. Should you enter a new market?
Enter market (cost $1M)
├── Market succeeds (p=0.4) → +$5M
│ └── Net: +$4M
└── Market fails (p=0.6) → $0
└── Net: -$1M
Do not enter
└── Net: $0
EV(Enter) = 0.4 × $4M + 0.6 × (-$1M) = $1.6M - $0.6M = $1M EV(Do not enter) = $0
Expected value favors entering, but risk tolerance, capital at stake, and opportunity cost all modify the final decision.
When trees help. Multi-stage decisions with contingencies. The tree forces explicit statement of all probabilities and payoffs, which exposes hidden assumptions.
Method 3 — Multi-Criteria Decision Analysis
Pattern: When a decision involves multiple criteria that cannot be converted to a single number (money, time, quality, risk, ethics), use a structured weighting.
Steps:
- List the criteria that matter.
- Assign each a weight (either by importance ranking or by paired comparison).
- For each option, score it on each criterion (0-10 scale or similar).
- Compute weighted sum: score(option) = Σ weight_i × rating_i
- Compare.
Worked example. Choosing a job.
| Criterion | Weight | Job A | Job B | Job C |
|---|---|---|---|---|
| Salary | 0.3 | 8 | 6 | 9 |
| Growth | 0.2 | 7 | 9 | 5 |
| Work-life balance | 0.2 | 5 | 8 | 3 |
| Mission alignment | 0.2 | 6 | 9 | 4 |
| Commute | 0.1 | 8 | 6 | 9 |
- Job A: 0.3(8) + 0.2(7) + 0.2(5) + 0.2(6) + 0.1(8) = 6.8
- Job B: 0.3(6) + 0.2(9) + 0.2(8) + 0.2(9) + 0.1(6) = 7.6
- Job C: 0.3(9) + 0.2(5) + 0.2(3) + 0.2(4) + 0.1(9) = 6.0
The process does not replace judgment — the weights and scores reflect subjective values. But the structure prevents one criterion from dominating the decision by loudness or salience.
Method 4 — Pre-Mortem
Pattern: Before committing to a decision, imagine the decision has been made and failed spectacularly. Ask the team: "What were the reasons for failure?" Then treat the reasons as risks to mitigate.
Why it works. Standard risk analysis asks "what could go wrong?" which triggers defensiveness and optimism bias. Pre-mortem asks "why did it go wrong?" which treats the failure as a fact and surfaces reasons that would otherwise be suppressed.
Worked example. Planning a product launch. Pre-mortem question: "It is six months from now and the launch failed. What happened?"
Surfaced reasons:
- Delays in manufacturing caused missed launch window
- Marketing messaging did not resonate with target demographic
- Competitor launched similar product two weeks earlier
- Initial reviews were negative due to Day 1 bug
- Key partnership fell through
Each surfaced reason becomes a risk mitigation. Pre-mortem typically generates 30-50% more risk identifications than standard risk reviews.
Method 5 — Reversibility and the Two-Way Door
Pattern: Distinguish decisions by whether they can be easily reversed. Spend deliberation in proportion to reversibility.
- Two-way door (reversible). You can change your mind later at low cost. Decide quickly. The cost of delay usually exceeds the cost of a wrong choice.
- One-way door (irreversible). Once taken, the decision cannot be undone or can be undone only at great cost. Deliberate thoroughly. The cost of a wrong choice exceeds the cost of delay.
Worked example. Deciding which text editor to use (two-way door) should take a few minutes at most. Deciding whether to move to another country for a job (one-way door for at least months, often years) warrants substantial deliberation.
Common error. Treating reversible decisions as irreversible. This produces analysis paralysis on low-stakes choices and leaves no time for the decisions that actually matter.
Method 6 — Minimax Regret
Pattern: Instead of maximizing expected value, minimize the maximum regret across scenarios. Useful when uncertainty is extreme and loss aversion is justified.
Worked example.