Choose a marketing method by estimand, bias and decision.
Answer:
A method is relevant only for the quantity it estimates. Attribution distributes observed conversions; an experiment estimates a causal effect for a population and period; MMM decomposes aggregate variation under assumptions; a survey measures declarations; a choice model projects trade-offs among options. Apparent precision or popularity cannot compare these different objects. Start with the estimand: effect of what, on whom, versus what, when and for which decision.
Method and decision:
Specify treatment, outcome, unit, population, comparator and horizon. Map selection, targeting, seasonality, anticipation, interference, incomplete measurement and composition change. Compare cost, delay, power, grain and repeatability, then pair a main method with independent calibration evidence. Choose the method whose estimand exactly matches the decision; investigate disagreement instead of averaging incompatible figures.
Worked example:
A search campaign has attributed ROAS 6.2. A pause test in comparable areas estimates incremental ROAS 1.8 ± 0.6. Annual MMM estimates 2.4 ± 0.9 including some lagged effects. Attribution distributes €620k of observed sales for €100k spend; the test estimates €180k sales that would not otherwise occur, interval €120k–€240k; MMM estimates €240k over a longer horizon but partly confounds branded search and video activity. At 35% margin, central incremental contribution from the test is €63k, or €0.63 per euro spent. The attributed figure pilots queries and traffic quality, not a budget increase; the test supports the immediate marginal decision and MMM informs annual balance.
Limits:
No map replaces examination of real design and data. The theoretically best method may be infeasible or underpowered. Effects rarely transport across segments, periods and intensities without assumptions.