Propensity Score

A propensity score is the probability that a given unit — a person, a customer, a clinic — received a treatment rather than the control, estimated from its observed characteristics. Paul Rosenbaum and Donald Rubin defined it in a landmark 1983 paper as the conditional probability of treatment given a set of covariates, and showed it has a remarkable property: it is a balancing score. Match or group units on this single number and, on average, all the observed covariates that went into it become balanced between the treated and untreated groups too. That is what makes it useful. In an observational study, where nobody randomized who got the treatment, the treated and untreated groups usually differ in ways that muddy any comparison. The propensity score compresses those many differences into one dimension you can adjust for.

You build a propensity score by modeling treatment assignment — often a logistic regression or a machine-learning classifier — using the covariates you believe drove who received the treatment. Then you put it to work in one of a few classic ways Rosenbaum and Rubin described: matching each treated unit to an untreated unit with a similar score, grouping units into strata by score (subclassification), or weighting units by the inverse of their score so the sample mimics a randomized one. Each approach chases the same goal — comparing treated and untreated units that were, as far as the observed data shows, equally likely to be treated. When that works, the remaining difference in outcomes is a cleaner estimate of the treatment's effect, because the systematic differences between the groups have been leveled.

A propensity score and a causal forest often appear in the same causal-inference toolkit, so it helps to be precise about the split. A propensity score is about making groups comparable. It answers how likely a unit was to be treated, and uses that to strip out confounding, typically to recover one average treatment effect. It says nothing, by itself, about whether the effect differs from person to person. A causal forest is about that variation — how the effect changes across units — and it frequently uses a propensity model inside it to handle who got treated before it estimates the differing effects. So the propensity score is the balancing step; the causal forest is the heterogeneity step. You can use the first alone; the second usually leans on something like the first.

The propensity score also differs from a randomized experiment, and honesty about that gap is essential. Randomization balances everything — measured and unmeasured — because chance, not choice, assigns the treatment. A propensity score can only balance what you measured. If some important confounder is missing from the model, the score cannot correct for it, and the comparison stays biased in a way no amount of matching will fix. This is the method's defining limitation and the source of most misuse: a propensity score makes observational data look experiment-like on the observed covariates, but it does not conjure the guarantee that real randomization provides. It narrows the gap between observation and experiment. It does not close it.

Use a propensity score when you cannot randomize but you have rich data on the factors that plausibly drove treatment. Think hard about which covariates belong in the model — the ones that influence both getting treated and the outcome — because leaving out a real confounder undermines everything downstream. After matching or weighting, check that the covariates actually balanced; a propensity model that fails to balance the groups has not done its job, and reporting balance is standard practice. Prefer it when a true experiment is impossible or unethical, and treat its estimate as conditional on the assumption that you measured the confounders that matter. Pair it with sensitivity analysis that asks how strong an unmeasured confounder would need to be to overturn the result.

The classic failures all trace back to unmeasured confounding and overconfidence. Teams treat a propensity-matched comparison as if it were a randomized trial and forget it only balanced the observed variables. They stuff the model with variables that predict treatment but not outcome, or include colliders that inject bias rather than remove it. They skip the balance check, so they never learn the matching failed. And they present a single clean number without the caveat that a hidden confounder could move it. The discipline is to use a propensity score as an honest, assumption-laden way to make observational groups comparable on what you measured — verified by balance diagnostics, hedged with sensitivity analysis, and never mistaken for the ironclad comparability that randomization alone provides.

Propensity score — defined by Paul Rosenbaum and Donald Rubin in 1983 as the conditional probability of treatment given covariates, a balancing score for causal inference in observational data.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where propensity score is a core concern: