An experimental result is not a ready-made prior
An experiment estimates an effect for a particular intervention, population and period. A model prior may describe a channel over a broader history or a different commercial outcome. The translation is a modelling judgement that should be visible, rather than a copy-and-paste of the experiment's point estimate.
- Describe the experiment
- Match the estimand
- Represent transfer uncertainty
- Compare prior sensitivity
Start with the calibration guide. Meridian's own calibration guidance emphasises that moving experimental evidence into a prior involves uncertainty beyond the experiment's standard error.
A reproducible workflow makes data checks, modelling assumptions and decision sensitivity visible. The mmm tools, code and platforms learning path connects this idea to the assumptions, practical checks and next steps.
To work through the connected decisions, read MMM calibration: keep an experiment evidence register and Incremental ROAS: report the interval, not just the ratio.
Write an estimand matching sheet
Record the tested channel and campaign, geography, eligible population, treatment contrast, outcome, spend definition, test dates and post-treatment window. Then write the corresponding definitions used by the MMM. Highlight every mismatch before choosing a distribution.
A test of a 20% spend increase measures a local change around an operating point. A full channel holdout measures a different contrast. Neither should be assumed to identify the entire response curve. Similarly, incremental conversions require an explicit value-per-conversion assumption before being expressed as revenue ROI.
Work through an illustrative conversion
Suppose a test estimates 500 incremental orders for £10,000 of incremental spend, with an illustrative interval from 200 to 800 orders. If the relevant net value is £40 per order, the point revenue return is 2, and the translated interval is 0.8 to 3.2 before considering uncertainty in order value or transfer to other periods.
The arithmetic is straightforward, but the judgement is not. If the test ran during a promotion or in unusually strong regions, the model-wide prior may need to be less certain. Document the reason and show the result under alternative plausible prior settings.
Use prior checks and sensitivity before presenting ROI
Inspect what the prior implies for contributions before fitting and compare those implications with the scale of the business. After fitting, examine whether the posterior changes meaningfully from the prior, whether diagnostics are acceptable and whether different reasonable priors change the allocation.
Keep the experimental evidence register separate from the chosen configuration. That makes it possible to update the translation when the outcome definition, currency or model window changes. It also helps identify reused evidence: a result should not be treated as two independent experiments merely because it appears in two reports.
The Meridian walkthrough covers the implementation path. The deliverable here is a traceable chain from experiment to assumption to decision, with no claim that calibration has removed all confounding.
Sources and further practice
This article’s worked example is illustrative, not a measured client result. Use the linked methodology to check the assumptions before applying it.