Propensity vs uplift modeling

A propensity model predicts how likely an action is. An uplift model predicts how much a treatment changes that action for a person. They sound similar and are often used interchangeably, but they answer different questions, and confusing them quietly wastes a large share of personalization budgets.

What a propensity model does

A propensity model scores the probability of an outcome: how likely this person is to buy, to click, to churn, or to open an email. It is supervised and widely deployed, because the training data is easy to come by: take the history, label who did the thing, and learn to predict it. The output is a ranking of people by likelihood, which feels immediately useful for targeting.

What an uplift model does

An uplift model estimates the causal effect of a treatment on an individual. Rather than asking how likely someone is to act, it asks how much more likely they are to act if you treat them, by comparing the predicted outcome with the treatment against the predicted outcome without it. The output is not a likelihood but a difference: the change the treatment is expected to cause for this person.

The distinguishing test

Distinguishing test: what will happen, or what the treatment changes? Propensity predicts what will happen. Uplift predicts the change caused by the treatment. If a model’s score would be the same whether or not you intend to intervene, it is a propensity model. If the score is defined as the gap between intervening and not intervening, it is an uplift model.

Why it matters in practice

The two can recommend opposite actions on the same people. Targeting the highest propensity often spends the budget on “sure things,” the people who would have converted anyway, so a discount sent to them is margin given away for a sale you already had. Uplift instead targets the “persuadables,” the people whose behavior the treatment actually moves. It also warns you away from “sleeping dogs,” the people a message puts off, whom a propensity model would never flag because they look like good prospects. A campaign optimized for propensity can look efficient on a dashboard while changing very little, because it is measuring attainment rather than influence.

Uplift needs randomized data

The catch is that uplift estimates a causal quantity, and you cannot read a causal effect off observational history. It requires randomized training data, where the treatment was assigned at random so the treated and untreated groups are comparable. That requirement ties uplift closely to experimentation, and it is why uplift is the modeling technique that fits causal measurement most directly. Propensity has no such requirement, which is part of why it is more common and also why it is more often misused: it is easy to build and easy to mistake for a measure of influence it was never designed to be.