Thread
Attribution and mix models
Dividing an observed outcome across channels, and the counterfactual that none of these methods contains.
Multi-touch attribution splits conversions across the touchpoints that preceded them. A mix model splits an aggregate outcome across spend, seasonality, and everything else in the regression. They work at opposite ends of the data, one on individual journeys and one on weekly totals, and they share a property that decides how far either can be trusted: both divide something that already happened. Neither observes the world where the spend was absent. That makes them useful for allocation and weak for causal claims, and it is why the interesting question about each is what has to be true for its division to be right.
Attribution
-
Attribution Is Not Incrementality
Attribution assigns credit for conversions you observed. Incrementality estimates the ones that would not have happened otherwise. The gap between them is where budgets go wrong.
-
What Multi-Touch Attribution Actually Computes
Seven credit rules run on the same journeys, from last touch to Markov chains and Shapley values, against the effects that generated them. What changes between rules, what does not, and why the sophisticated ones fail in the same place.
Allocation
-
Adstock, Saturation, and Honest Response Curves
Lagged effects and diminishing returns are where most mix models smuggle in their conclusions. How to tell a fitted curve from an identified one, and what variation would identify it.
-
Calibrating an MMM to Experiments
Which experiment to run, how to carry its estimate into the model, and how hard to lean on it. Anchoring a channel's level drags its slope down with it, and the slope is what the media plan runs on.
-
Marginal Return, Not Average Return
The channel with the best cost per acquisition is often the worst place to put the next dollar. Averages describe money already spent. Budgets are decisions about money that has not been spent yet.
-
What a Marketing Mix Model Can and Cannot Do
A mix model is a good allocation tool and a weak causal one. Where its identification would have to come from, why a Bayesian version does not create any, and what makes one worth acting on.
- Worked example
An Experiment-Calibrated Marketing Mix Model
A mix model that fits three years of weekly data almost perfectly and gets the channel contributions badly wrong, and what one geo holdout does about it.
Systems
- System
Media Taxonomy and Spend Spine
One vocabulary for channels and tactics, one spend series reconciled to what the business paid, joined to the outcome the budget is meant to move and what it is worth, at a grain that supports the designs you want to run.
- System
Response Curve Estimation
Adstock, saturation, and priors, plus the harder question underneath them: what in the spend history is doing the identifying, and where does the curve stop being estimated and start being assumed.
- System
Calibration
Folding measured effects into the model as priors, and deciding what happens when the experiment and the model disagree. The step that closes the loop between testing and planning.
- System
Counterfactual Traffic
What to do when the change had to ship everywhere and there is no other half: modeling what the pages would have done, and knowing when a core update has broken the comparison.
- System
Paid and Organic Together
Some of an organic gain on a term you also buy is your own paid traffic arriving free, and reporting the two channels separately can count the same click twice. How much is measured, not assumed, and who owns the number.
- System
Sourced, Influenced, and What Replaces Them
Two reporting conventions that divide pipeline without a counterfactual, why both are attribution wearing a B2B accent, and the evidence ledger that should sit above them.