Triangulation does not mean averaging every available figure.
Triangulation does not mean averaging every available figure. First verify that methods measure the same effect on the same population, horizon and spending variation. Divergence may reveal bias, an estimand difference or real heterogeneity. Quantitative synthesis is legitimate only after this mapping.
Decision and method.
Base decisions on the most identifiable compatible evidence and use divergence to define subsequent tests rather than force one number. For each method create a card with treatment, counterfactual, outcome, population, period and unit; exclude measures answering another question. Document selection, confounding, contamination, measurement error, specification and extrapolation, noting shared weaknesses. Put estimates and uncertainty on a common scale. Combine only compatible estimates by inverse-uncertainty weighting; otherwise retain several figures and make the decision robust to their range.
Worked example.
An experiment estimates iROAS 1.40 with standard error 0.255; a compatible MMM gives 1.80 with 0.306. Attribution reports 3.20, but measures credited conversions, not causal effect. Inverse-variance weights are 15.38 and 10.68. Synthesis = (1.40 × 15.38 + 1.80 × 10.68) ÷ 26.06 = 1.56. Standard error = √(1 ÷ 26.06) = 0.196, approximate interval 1.18–1.95. Attribution remains useful for journeys but not causal average; budget relies on 1.56 and tests robustness down to 1.18.
Checks and limits.
Align treatment, population, outcome and horizon before comparison; document data and bias dependencies; only compatible estimates enter quantitative synthesis; test the decision across the interval. Inverse-variance weighting assumes comparable, weakly dependent estimates. Narrow intervals may still be biased by poor identification, and transport from a local experiment to national budget needs justification.
Resources and sources.
Download the triangulation register. Related: choose the right method, define causal effect, calibrate MMM, choose a protocol. Sources: Gordon et al. (2019); Shadish, Cook and Campbell (2002); Salaka and Prabhu (2011).