Chapter 3: A Technical Deep Dive Into MMM

Our framework is built on top of Meridian, a powerful, open-source framework for building sophisticated MMMs. It’s based on a Bayesian framework and offers a robust and flexible approach to modeling.

Core Technical Features:

• Bayesian Causal Inference: Meridian uses Bayesian statistical inference for parameter estimation, it incorporates principles from causal modeling.
• Hierarchical Geo-Level Modeling: Meridian is designed to handle large-scale, geo-level data. This allows for more granular insights and can lead to more robust models by leveraging the additional information present in geographical variations. National-level modeling is also supported.
• Incorporation of Priors: The Bayesian model allows you to incorporate existing knowledge about your media performance through the use of ROI priors. This can be derived from past experiments, industry benchmarks, or internal expertise.
• Reach and Frequency Data: You can optionally include reach and frequency data as model inputs to gain deeper insights, particularly for video advertising.
• Model Transparency: The model’s assumptions for valid causal inference are made explicit and can be reproduced on any machine. The use of MCMC NUTS (No-U-Turn Sampling) for efficient exploration of model parameters ensures that the results are robust, with good convergence.
• Model Evaluation: Meridian includes tools for reporting model fit statistics, both in-sample and out-of-sample, allowing for robust model comparison and selection.

Technology Stack: Meridian utilizes TensorFlow Probability and its XLA compiler for efficient MCMC sampling.

Key Methodological Papers and Resources:

• Geo-level Bayesian Hierarchical Media Mix Modeling: This paper provides the foundational methodology for Meridian’s geo-level modeling approach.
• Bayesian Hierarchical Media Mix Model Incorporating Reach and Frequency Data: This paper details the methodology behind incorporating reach and frequency data into the model.
• Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects: For a deeper understanding of the Bayesian methods used for handling adstock and saturation effects.

Getting Started and Implementation:

• GitHub Repository: The complete open-source code is available on GitHub.
• Meridian Documentation: The full documentation is located here.

Getting Started Colab: A Colab notebook is available to quickly get started with Meridian using sample data.