Methodological references

Use these references to assess assumptions and validation choices. They do not specify AMMM4’s API; the source-linked guides describe the current implementation.

• Gelman and colleagues, Bayesian Workflow covers iterative model construction, checking and comparison.
• Jin and colleagues, Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects describes a Bayesian approach to carryover and nonlinear media response.
• Dew, Padilla and Shchetkina, identification of nonlinear and time-varying effects examines response mechanisms that observational MMM data can confound.
• Stan R-hat and ESS reference defines rank-normalised convergence and effective-sample-size diagnostics.
• Stan posterior analysis reference explains Monte Carlo error and effective sample size.

A successful implementation comparison validates only the registered profile. Historical AMMM4 parity claims need retrievable reports and frozen profiles before they can be used as evidence. The FE and CRE guides state which historical claims were not verified by this review; no result automatically qualifies new datasets or interventions.

References for specific modelling choices

• Riutort-Mayol and colleagues, Practical Hilbert space approximate Bayesian Gaussian processes for probabilistic programming supports discussion of basis size, domain boundary and approximation diagnostics in the smooth-baseline example.
• Mundlak, On the Pooling of Time Series and Cross Section Data is the foundational panel-data reference for the unit-summary adjustment idea. AMMM4’s CRE is a specific Bayesian transformed-summary implementation; the citation does not make it a classical specification test.

For calibration, start with the study design and uncertainty report for the actual experiment, then document why its estimand transports to the supported measurement likelihood. No generic literature citation can establish that a particular study matches the model population, intervention or horizon. For FE/CRE qualification, retain the implementation-specific profile and results; neither the panel literature nor deterministic tests substitute for a posterior comparison or recovery/coverage study.