Guide: Lift Test Calibration

This guide explains how experimental lift evidence can inform MMM parameter learning.

What This Guide Covers

• Required lift-test data structure
• Where lift calibration fits in the AMMM workflow
• How to validate effect on posterior estimates

Why Use Lift Tests

MMM is observational. Lift tests add causal evidence at specific spend deltas, which can stabilise saturation/effect parameters when integrated correctly.

Expected Lift Data Columns

Each row should represent one lift experiment and include:

• alignment coordinates (for example channel identifier),
• x (baseline activity),
• delta_x (intervention change),
• delta_y (observed incremental outcome),
• sigma (standard error of lift estimate).

Use original (unscaled) units consistent with model inputs/target.

Integration Pattern

Lift data is passed in Python when building the model, not as a standalone YAML key. In V2 workflows this is typically done in custom API scripts that instantiate the model directly.

If you are using runme.py only, treat lift calibration as an advanced extension path.

How to Verify Impact

Run two comparable fits:

1. with lift calibration;
2. without lift calibration.

Compare these artefacts:

• 20_model_fit/model_trace.png
• 20_model_fit/model_summary.csv
• 20_model_fit/posterior_forest.png
• 60_response_curves/response_curves.png

Then review diagnostics:

• 50_diagnostics/convergence_report.json
• 50_diagnostics/calibration_report.json

Practical Caveats

• Poor-quality lift inputs can degrade model quality.
• Coordinate mismatches silently invalidate calibration intent.
• Diagnostic adequacy remains mandatory even with experimental evidence.
• Lift-informed posteriors still require decision-layer sensitivity checks.