A chapter-by-chapter guide to thinking causally about business metrics — starting from why correlation misleads, through randomization, and into the tools for when a clean experiment isn't available.
Why a correlation on a dashboard can point the wrong way, how to think in parallel universes to define a true effect, and where the number you actually see comes from once bias is added back in.
How a coin flip severs the link between treatment and a person's pre-existing type, a 3-arm coupon test worked end to end, the balance checks a real experiment needs, and what to do when randomizing isn't an option.
Why small-sample rankings are misleading, the one formula (standard error) that tells you how much any estimate could plausibly wobble, and how it turns into confidence intervals, p-values, and a real cost-based decision rule.
The three building blocks of every causal graph — chains, forks, and colliders — and how they let you tell confounding bias (something you should have controlled for) apart from selection bias (something you shouldn't have).
How one line of regression reproduces every hand-computed number from Chapter 3, why adding a control variable is the same thing as 'comparing like-for-like,' and the omitted variable bias formula that tells you which way an uncontrolled estimate is likely to be wrong.
How to weight a regression when all you have is a pre-aggregated table, how to turn categorical variables like device or channel into a regression a computer can run, and why a single regression coefficient can quietly average over very different effects.
Why throwing every available variable into a regression isn't the safe move: a variable that predicts the outcome well tightens your estimate, one that only predicts the treatment can wreck your precision, and one created after the treatment can silently reintroduce bias — including the two-part conversion-rate-times-order-value trap that quietly biases dashboards everywhere.