TV, Paid_Search show the highest ROI with relatively narrow uncertainty bands, suggesting strong evidence of effectiveness. Paid_Social, Display show positive returns but with wider uncertainty— additional experimentation could sharpen these estimates.
All estimates include 80% credible intervals reflecting genuine uncertainty from limited data. Point estimates alone can be misleading—decisions should account for the full range of plausible values.
| Channel | ROI (Mean) | 80% CI | Confidence |
|---|---|---|---|
| Paid_Search | 2.12 | [1.85, 2.41] | Strong evidence |
| Paid_Social | 1.78 | [1.32, 2.28] | Strong evidence |
| TV | 1.45 | [1.10, 1.82] | Strong evidence |
| Display | 1.15 | [0.65, 1.68] | Uncertain |
| Radio | 0.82 | [0.45, 1.22] | Uncertain |
Revenue decomposition breaks down the predicted outcome into component contributions: baseline, trend, seasonality, media channels, and control variables. Each component's contribution sums to the total predicted revenue.
| Component | Total Contribution | % of Total |
|---|---|---|
| Baseline | $680.0M | 84.6% |
| TV | $45.2M | 5.6% |
| Paid_Search | $32.1M | 4.0% |
| Paid_Social | $28.5M | 3.5% |
| Display | $12.8M | 1.6% |
| Radio | $5.4M | 0.7% |
The channel effects below are causal only if every common cause of media spend and the KPI is measured and adjusted for (no unobserved confounding), and one unit's spend does not affect another's outcome (SUTVA). In marketing the dominant hidden confounder is unobserved demand — budgets rise when demand is expected to rise — which no adjustment set can remove. Good fit, tight intervals and passing posterior-predictive checks are all compatible with confounding bias. Effects should be anchored with randomized geo-lift / incrementality experiments where the stakes are high.
This analysis uses a Bayesian Marketing Mix Model with the following components:
Inference via MCMC (4 chains, 2000 samples each, 1000 warmup).
This report follows principles of honest uncertainty quantification: