Myth: Markov chain attribution's 'removal effect' measures incrementality

Myth: Markov chain attribution's 'removal effect' measures incrementality

The question: Markov attribution models the journey as a chain of states and computes each channel's 'removal effect' — how much conversion probability drops if you delete that channel from the graph. That sounds like a causal what-if. Is it?

What the math actually does: a Markov model is fit to observed transition probabilities between touchpoints. The removal effect simulates deleting a node and recomputing conversion probability within that same observed graph. It is an elegant, internally consistent way to allocate credit across paths — and it respects order and interaction effects that linear or Shapley models flatten.

The nuance: 'removal' here means removing the channel from the data structure, not from the real world. The model assumes the remaining transition probabilities stay fixed when a channel disappears — but in reality, cutting a channel changes user behavior, substitution patterns, and the other channels' performance. The removal effect is a within-model counterfactual built on observational transitions; it inherits every confounder in that data. If high-intent users disproportionately pass through brand search, Markov will assign brand search a large removal effect that an experiment would not confirm.

Bottom line for practitioners: Markov is a strong descriptive and credit-allocation tool, arguably the best of the algorithmic family for honoring sequence. But its removal effect is a simulation inside observed data, not a measured causal lift. Use it to allocate, then calibrate the suspicious channels against a real holdout.