Decisions Under Incomplete Information

A course about what to do once the data is read: telling a cause from a coincidence, choosing under incomplete information, and deciding whom to trust when checking for yourself is impossible.

WHO IT IS FOR. People whose decisions have consequences: managers, analysts, product managers, founders, researchers, doctors. And anyone tired of hearing "correlation is not causation" without ever being told what it is, then, and what to do about it.

WHAT YOU WILL LEARN. Separate three levels of question: what is associated with what, what happens if we intervene, and what would have happened otherwise — different questions needing different data. Find confounders and, more importantly, colliders: conditioning on a collider manufactures an association that is not there, and it is behind a large share of the false "patterns" people find in everyday observation. Draw a causal graph and use it to decide what may be controlled for and what must not be. Understand why randomisation, natural experiments, regression discontinuity and difference-in-differences exist — and why "it worked for me" is not "it worked".

Then decisions. Expected utility and its systematic violations. Knight's distinction between risk, uncertainty and ambiguity, and the Ellsberg paradox. Value of information and option value: when waiting pays. Reversibility as a speed criterion. The Kelly criterion and the asymmetry in the cost of errors. Game theory where the outcome depends on others: coordination, the prisoner's dilemma, Schelling points, principal-agent, adverse selection. And Goodhart's and Campbell's laws — why a metric breaks precisely when it becomes a target.

WHY IT MATTERS IN LIFE AND AT WORK. Most expensive mistakes are not arithmetic; they are mistakes about causation and incentives: a correlation mistaken for a lever, a metric optimised instead of an outcome, an irreversible decision made at the speed of a reversible one. This course gives you the language to see that in advance and explain it to other people. It also covers how to decide whom to trust: total scepticism is impossible and harmful, so you need a criterion for delegating belief.

WHAT MAKES IT DIFFERENT. It is checkable. The forecast journal you began in part one closes here: you plot your own calibration curve and compute your Brier score — a number instead of a feeling. Exercises have answers known in advance: find the confounder in a real study, decide which design supports a causal claim and which does not.

HONEST ABOUT LIMITS. We say plainly what failed to replicate in behavioural science, and label heuristics as heuristics rather than theories. Models are useful not because they are true — all models are wrong — but because they are wrong predictably.

THIS IS PART TWO. Part one, "Logic as the Foundation of Thinking: How Reasoning Works", covers argument, meaning, probability and statistics: telling the well-founded from the unfounded. If you have not taken it, start there — this course assumes Bayesian updating and base rates are already familiar, and that your forecast journal is already running.