Supervisor: Dr. Russell Zaretzki (Heath Faculty Fellow, Business Analytics & Statistics; Haslam College of Business, University of Tennessee, Knoxville, USA)

Additional Resources: Paper (Document) · Git Repository

Marketing Mix Modeling (MMM) is widely used to allocate media budgets, but it relies on observational data; leaving it vulnerable to confounding from organic demand, ad-targeting mechanisms, and macroeconomic trends. Measuring the true return on ad spend is a notoriously hard problem (Lewis & Rao, 2015): randomized lift tests isolate the causal effect cleanly but are too expensive and slow to run continuously (Gordon et al., 2019), while MMM evaluates all channels at once from historical data but faces well-documented identification challenges of its own (Chan & Perry, 2017). This project asks a direct question: can an observational MMM be trusted to recover causal ROI, and if not, what does it actually take to make it trustworthy?

I don't propose new estimation methodology here. The Bayesian MMM builds on classic market-response modeling (Hanssens et al., 2001) and Bayesian media-mix formulations with adstock and saturation effects (Jin et al., 2017); the backdoor-adjustment framework follows standard causal-DAG theory (Neuberg, 2003) (Hernán & Robins, 2020); and the calibration-via-prior technique follows Zhang et al.'s approach to combining MMM with experimental lift tests (Zhang et al., 2024). The contribution is empirical: using synthetic data with known ground truth, I show concretely when these established methods succeed, when they fail, and what it takes to correct them, then validate the same failure pattern on a real, public dataset.

Navigating real-world deployment requires a unified approach: combining observational MMM with randomized lift testing to get causally robust ROI estimates. Neither approach alone is enough; observational data is scalable but confounded, and experiments are unbiased but too expensive to run for every channel.