Propensity score matching
Propensity score matching (PSM) is an observational method for estimating causal effects when randomization is not available. It models each unit's probability of being exposed (the propensity score) from observed characteristics, then compares exposed and unexposed units with similar scores. By balancing the groups on measured confounders, PSM approximates an experiment. This page explains the technique and its key limitation — unmeasured confounding.
How matching works
First, a model estimates each unit's propensity score: the probability it was exposed, given observed characteristics. Then exposed and unexposed units with similar scores are matched, creating comparison groups balanced on those characteristics.
The idea is that, conditional on the propensity score, exposure resembles random assignment among matched units, so the outcome difference approximates the causal effect.
The confounding limit
PSM can only balance the variables you measured. If something unobserved drove both exposure and the outcome — a hidden confounder — matching cannot remove its bias, and the estimate is off. This is the central caveat versus a true randomized experiment, which balances unobserved factors automatically.
In marketing, PSM is a useful fallback where experiments are impossible, but its estimates should be treated as conditional on the assumption that key confounders were captured.
- Match on probability of exposure, not raw traits
- Balances measured confounders between groups
- Cannot fix unmeasured confounding
How it appears in analytics and logs
A difference in outcomes between matched exposed and unexposed groups estimates an effect, but only as well as the observed covariates capture why exposure occurred.
Diagnostic use case
Estimate the effect of an exposure (such as seeing a campaign) from observational data when a randomized experiment was not run, by matching on the likelihood of exposure.
What WebmasterID can help detect
WebmasterID's first-party, observed segment attributes can supply the covariates a propensity model needs while keeping the analysis on aggregated, privacy-minimized data.
Common mistakes
- Assuming PSM removes bias from unobserved confounders.
- Matching on variables affected by the exposure itself.
- Treating an observational PSM estimate as experimental proof.
Privacy and accuracy notes
PSM works on aggregated, modeled comparisons; it should use minimized, consented data and cannot correct for confounders it never observed. This is educational, not legal advice.
Related pages
- Difference-in-differences for measurement
Difference-in-differences (DiD) is a quasi-experimental method that estimates the causal effect of an intervention — like turning a campaign on in some regions — by comparing how a treated group changed against how an untreated control group changed over the same time. By differencing out both pre-existing gaps and shared time trends, DiD isolates the incremental effect. This page explains the method, its key assumption, and where it fits in measurement.
- Deterministic vs probabilistic matching
Identity resolution in attribution uses two approaches. Deterministic matching links touchpoints when they share a known, persistent identifier (a logged-in user ID, a hashed email). Probabilistic matching infers that two touchpoints belong to the same user from circumstantial signals — IP, device, behavior — without a confirmed identifier. The two differ sharply in accuracy and privacy posture.
- Incrementality testing: what attribution cannot tell you
Incrementality testing measures the lift a channel actually causes by withholding it from a control group and comparing outcomes. It answers the question every attribution model dodges: would this conversion have happened anyway? It is causal where attribution is merely correlational, but it requires deliberate experiment design.
- Attribution analytics
First-party covariates for observational comparison.
Sources and verification notes
- NIST/SEMATECH e-Handbook of Statistical MethodsReference for matching and comparison-group statistical methods.
Last reviewed 2026-06-24. Facts are checked against primary/official sources where available; uncertain specifics are marked “Data not yet verified” rather than guessed.