The propensity score is the conditional probability that a unit receives a particular treatment or exposure, given a specified set of observed characteristics. For a binary treatment , with denoting treatment and denoting control, and pretreatment covariates , it is defined as . Its central mathematical property is that conditioning on the true propensity score balances the distribution of those covariates between treatment groups.
Introduced by Paul Rosenbaum and Donald Rubin in 1983, the propensity score became a major tool in causal inference from observational data. It reduces a potentially high-dimensional adjustment problem to one involving a scalar, supporting matching, stratification, weighting, and combinations with outcome models. These methods are widely used in epidemiology, econometrics, and program evaluation, where treatment assignment is not controlled by the investigator.
A propensity score is not itself a measure of treatment benefit, nor does balancing observed characteristics establish that all sources of confounding have been removed. Causal interpretation depends on the treatment definition, the covariates observed, the availability of comparable treated and untreated units, and assumptions connecting observed outcomes to the effects being estimated.
Mathematical definition and the balancing property
Let be a binary treatment indicator and let be a vector of observed pretreatment characteristics. The propensity score is the function
It describes the treatment assignment mechanism at the level of measured covariates. For example, in an evaluation of a job-training program, might include age, previous earnings, education, and employment history. A score of means that units with those characteristics have a conditional treatment probability of under the assignment process represented by the data. It does not mean that the program has a probability of helping a particular participant.
A balancing score is any function satisfying
where denotes conditional independence. Within levels of such a score, the distribution of observed covariates is the same in treated and untreated units. The complete covariate vector is trivially a balancing score, but the propensity score achieves this property with a single scalar:
The result follows from the definition of conditional probability. Because is determined by ,
Also, by iterated conditional expectation,
Thus, after conditioning on the score, knowing the full covariate vector supplies no additional information about treatment assignment. The same reasoning applies to the probability of receiving control.
Rosenbaum and Rubin characterized the propensity score as the coarsest balancing score: any balancing score must contain enough information to determine the propensity score, up to events of probability zero. Other balancing scores can retain additional covariate information. This explains why propensity scores support dimensionality reduction without requiring every treated unit to have an untreated counterpart with exactly the same age, earnings, education, and other attributes.
The theorem concerns population distributions and the true assignment probability. It does not guarantee exact covariate equality in a finite sample, even when the true score is known. Nor does it guarantee that an estimated score balances covariates if its model is misspecified. Two people can also have identical scores while differing substantially in their individual characteristics; balance is a distributional statement across groups, not a claim that matched individuals are otherwise identical.
Causal identification and target effects
The connection between balancing and causal effects is most clearly expressed using potential outcomes. Let denote the outcome a unit would have under treatment and the outcome under control. Only one is ordinarily observed. Under consistency,
A well-defined treatment is essential: if different versions of “treatment” have substantively different consequences, the notation must distinguish those versions or specify the intervention being considered. Standard formulations also exclude interference between units, meaning that one unit’s outcome does not depend on another unit’s treatment.
A central identifying assumption is conditional exchangeability, also called no unmeasured confounding:
It states that, among units with the same observed covariates, treatment assignment supplies no further information about their potential outcomes. This is stronger than observed covariate balance. Another assumption is positivity:
throughout the covariate support relevant to the target population. Positivity requires both treatment possibilities to occur with nonzero probability wherever their outcomes must be compared. Consistency, exchangeability, and positivity together permit identification of causal effects from observed data.
If exchangeability holds given , it also holds given the true propensity score:
Consequently, subject to overlap,
This is the key reduction: adjustment for a scalar can suffice for identification even when exchangeability originally requires a large covariate vector. The reduction does not make exchangeability testable. Missing information about severity, motivation, income, or another common cause of treatment and outcome can invalidate the causal interpretation despite excellent observed balance.
Different analyses target different quantities. The average treatment effect in the population is
whereas the average treatment effect among treated units is
An effect among untreated units or among an overlap population is another possible target. These effects need not coincide when treatment effects vary across people. The required overlap also depends on the target: estimating the ATT primarily requires untreated comparators across the covariate distribution of treated units, whereas the ATE requires support for both potential outcomes across the full target population. Matching, trimming, and weighting can therefore alter not only precision but also the population whose effect is estimated.
Historical development and methodological significance
Covariate adjustment predates the propensity score. Researchers used stratification, regression adjustment, standardization, and matched comparisons long before 1983. The difficulty was that matching simultaneously on many characteristics produces increasingly sparse comparison groups. With ten binary covariates, for example, exact cross-classification already creates as many as covariate patterns. This is an instance of the curse of dimensionality: sample requirements rise rapidly as the number of characteristics increases.
Paul Rosenbaum and Donald Rubin’s 1983 paper, “The Central Role of the Propensity Score in Observational Studies for Causal Effects,” established the balancing-score framework and showed how strong ignorability given covariates transfers to the propensity score. Its importance was theoretical as well as practical. It explained why treatment groups could be made comparable on many observed variables through adjustment for a one-dimensional quantity, without requiring a correctly specified model for the outcome at the design stage.
Subsequent work connected the theory to subclassification and matched sampling. Rosenbaum and Rubin’s 1984 work on subclassification and their 1985 work on constructing matched control groups helped establish procedures that became standard in applied research. Dividing observations into score subclasses offered an accessible alternative to complex multivariable outcome modeling. Matching made the construction of a comparison group explicit and separated that task from the subsequent estimation of outcome differences.
An influential application appeared in Rajeev Dehejia and Sadek Wahba’s 1999 study of nonexperimental evaluations of a job-training program. Using data associated with the National Supported Work demonstration and nonexperimental comparison samples, they investigated whether propensity-score methods could reproduce an experimental benchmark. The study became part of a broader debate about observational program evaluation. Subsequent comparisons emphasized that results depend on the comparison population, covariate information, overlap, and implementation; no matching algorithm can universally recover an experimental result from unsuitable observational data.
Later developments broadened the field beyond nearest-neighbor matching. Keisuke Hirano, Guido Imbens, and Geert Ridder’s 2003 work examined efficient estimation using estimated propensity scores. Alberto Abadie and Guido Imbens developed large-sample theory for matching estimators, including results explaining why ordinary bootstrap procedures can fail for fixed-neighbor matching. Kosuke Imai and Marc Ratkovic introduced the covariate balancing propensity score in 2014, explicitly integrating assignment modeling with balance conditions. Research on overlap weighting, flexible nuisance estimation, and doubly robust procedures further shifted attention toward target populations, stability, and direct control of imbalance rather than treatment prediction alone.
Estimating the treatment assignment mechanism
In observational studies, the true propensity score is usually unknown. A common estimator uses logistic regression:
where can include original covariates, nonlinear transformations, and interactions. The coefficients are often estimated by maximum likelihood estimation. For example, age may enter through a spline rather than a straight-line term, while an interaction between prior earnings and employment status can represent different enrollment patterns across employment groups.
Covariate selection is a causal modeling problem, not simply a search for variables that predict treatment. Pretreatment common causes of treatment and outcome are central to confounding control. Variables strongly predictive of the outcome can also improve precision. By contrast, including a variable affected by treatment may block part of the effect of interest or introduce bias. Conditioning on a common effect of other variables can likewise open a noncausal association. Directed acyclic graphs help distinguish these roles; a timestamp preceding treatment does not by itself establish that adjustment is appropriate.
A variable that predicts treatment strongly but has no direct association with the outcome except through treatment is not automatically beneficial. Such a variable can drive estimated scores toward zero or one, inflating weighting variability. In settings with residual unmeasured confounding, adjustment for strong treatment predictors can also amplify bias. Thus, maximizing treatment prediction is not equivalent to minimizing error in a causal effect estimate. A model with excellent discrimination may chiefly reveal that treated and untreated populations have little comparability.
Flexible estimators include boosted trees, random forests, neural networks, and ensembles of multiple algorithms. Regularization can stabilize estimation when the number of candidate predictors is large. Such methods can capture nonlinear assignment mechanisms that simple logistic specifications miss. However, predictive cross-validation criteria do not directly certify covariate balance, overlap, or satisfactory causal performance. Flexible fitting can also produce extreme probabilities or overfit small samples.
Some procedures estimate assignment probabilities while directly targeting balance. Covariate balancing propensity score methods incorporate moment restrictions such as weighted equality of selected covariate means. Other balancing-weight methods construct weights without first treating probability prediction as the primary objective. These approaches are related but not identical: not every set of balancing weights corresponds to a conventional fitted propensity-score model. Regardless of estimation method, the fitted scores are intermediate quantities whose adequacy must be assessed through the comparison groups and weights they generate.
Matching and subclassification
Propensity-score matching constructs treated and untreated groups with similar estimated assignment probabilities. In nearest-neighbor matching, a treated unit is paired with one or more untreated units whose scores are closest under a specified distance. Matching may be with replacement, allowing an untreated unit to serve as a comparator several times, or without replacement. Replacement often improves local comparability when suitable controls are scarce, but repeated use of the same controls affects precision and the dependence structure of the resulting estimator.
A caliper places an upper limit on acceptable distance. Matching is often performed on the logit of the estimated propensity score rather than on the raw probability. A frequently studied rule uses a caliper of times the standard deviation of the logit score. This is an empirical convention, not a theorem ensuring unbiasedness. Its performance depends on the covariate distribution, treatment prevalence, sample size, and outcome process. A narrow caliper can reduce poor matches while leaving many treated units unmatched.
The matching algorithm also matters. Greedy matching selects available matches sequentially and can depend on processing order. Optimal matching chooses a collection of matches to minimize an overall distance criterion. Full matching creates matched sets containing at least one treated and one untreated unit and permits variable treated-to-control ratios. Exact restrictions can be combined with propensity-score distances—for example, requiring matches within the same region or calendar period while using scores to address remaining covariates.
Subclassification, or stratification, instead divides the score distribution into intervals and estimates treatment contrasts within each interval. If strata approximate sufficiently narrow score levels, treated and untreated units within them should have similar observed covariate distributions. An overall effect is obtained by averaging stratum-specific effects, with weights chosen for the target estimand. ATE estimation uses target-population stratum frequencies, while ATT estimation uses the distribution of treated units across strata.
Five subclasses became common in early applications, but five is not a universally adequate number. Broad intervals can retain substantial residual confounding, while excessively fine intervals may contain only treated or only untreated units. Both matching and subclassification therefore involve compromises between local comparability and usable sample size. Propensity-score closeness should also be checked against the original covariates: nearly equal estimated probabilities do not ensure that an important prognostic characteristic is adequately balanced.
Matching commonly targets the ATT when treated units are retained and controls are selected to resemble them. If treated units are discarded for lack of matches, however, the result generally concerns the retained treated population. The distinction becomes substantive when excluded units are older, poorer, more severely affected, or otherwise systematically different from those retained.
Weighting and the construction of target populations
Inverse probability weighting uses the propensity score to create a weighted population in which treatment is independent of observed covariates. For the ATE, treated units receive weight , and untreated units receive weight . Under the identification assumptions,
For instance, a treated observation with a treatment probability of receives weight , representing treated potential outcomes among units whose observed treatment is relatively uncommon.
An unnormalized estimator averages the inverse-weighted outcomes over the original sample size. A normalized, or Hájek-type, estimator instead divides each treatment group’s weighted outcome total by its total weight:
Normalization changes finite-sample behavior while preserving the same large-sample target under appropriate conditions. Stabilized ATE weights use for treated units and for controls. In separate normalized group means, these constant numerators cancel.
Weights can be adapted to other populations. ATT weighting assigns weight to treated units and to controls. More generally, a nonnegative tilting function defines a target covariate distribution proportional to times the original population distribution. Corresponding weights are
The resulting effect is
ATE weights use , and ATT weights use .
Overlap weighting uses , giving treated units weight and untreated units weight . These bounded weights emphasize units with substantial probability of receiving either treatment and downweight regions where assignment is nearly deterministic. Under suitable conditions, an unpenalized logistic maximum-likelihood fit with an intercept yields exact finite-sample equality of overlap-weighted means for the covariates included in its linear predictor.
Weighting does not manufacture missing comparisons. Extreme inverse weights indicate that a small number of unusual observations represent large portions of the target population. Trimming observations or capping weights can improve stability, but the consequences differ: trimming can define a restricted target population, while ad hoc weight truncation generally introduces bias relative to the original weighting estimand. An overlap effect is a deliberately different estimand, not merely a more precise estimate of the full-population ATE.
Outcome models and doubly robust estimation
Propensity-score methods can be combined with models for the outcome. After matching, regression adjustment may account for residual covariate differences or improve precision. Within subclasses, outcome models can adjust for remaining imbalance. Weighted regression is another possibility, but its interpretation depends on the weights, model specification, and effect scale. A coefficient from a nonlinear model is not automatically the marginal treatment effect targeted by the weighting design.
A particularly important combination is the augmented inverse-probability-weighted estimator. Define outcome regressions
For the ATE, the estimator is
The first difference predicts each unit’s two potential-outcome means. The remaining terms correct those predictions using inverse-weighted observed residuals.
This construction exhibits double robustness: under the causal identification assumptions and suitable regularity conditions, it is consistent if either the propensity-score model is correct or both treatment-specific outcome regressions are correct. If the score is correct, weighted residual corrections repair misspecified outcome predictions in expectation. If the outcome regressions are correct, their residuals have conditional mean zero, so an imperfect propensity model need not invalidate consistency. Double robustness does not mean that arbitrary errors in both models cancel, and it does not address unmeasured confounding.
With sufficiently accurate nuisance estimates, augmented estimators can achieve the semiparametric efficiency bound for the ATE in the standard nonparametric model. Their leading remainder often depends on a product of propensity-model and outcome-model estimation errors. This helps explain why moderately accurate estimators of both components can support strong large-sample performance. It also explains why the consistency property alone is insufficient to guarantee a valid confidence interval: convergence rates and other regularity conditions matter.
Cross-fitting separates nuisance-model training from evaluation. The data are partitioned into folds; propensity and outcome models are trained outside each fold, then used to construct effect-estimation contributions inside it. This limits overfitting-induced dependence and makes flexible machine-learning methods easier to combine with asymptotic inference. Related approaches include targeted maximum likelihood estimation, which updates an initial outcome fit using information from the assignment mechanism. Such methods still depend on meaningful overlap: an outcome model can smooth across sparse regions, but doing so introduces extrapolation rather than direct empirical comparison.
Balance diagnostics and practical overlap
The primary diagnostic question is whether the adjusted treatment groups are comparable on relevant observed characteristics. Treatment-model accuracy is secondary to this purpose. A high area under a receiver-operating-characteristic curve indicates strong discrimination between treated and untreated units, which may reflect severe separation rather than a desirable causal design. Conversely, a modestly predictive model can be adequate if it produces good balance in the covariates needed for confounding control.
A common measure is the standardized mean difference. For a continuous covariate ,
where the means may be weighted or computed in a matched sample, and is a specified reference standard deviation. Possible denominators include the pooled pretreatment standard deviation or a target-group standard deviation. Using a fixed denominator before and after adjustment makes changes easier to interpret. An absolute standardized difference below is a common convention, not proof that remaining confounding is negligible.
Equality of means is only one aspect of balance. Covariate variances, tails, categorical proportions, nonlinear transformations, and interactions may also matter. Distribution plots and empirical cumulative distribution functions can reveal differences concealed by average summaries. A “love plot” displays standardized differences across many covariates before and after adjustment. Formal balance-test p-values are generally less informative: they combine imbalance with sample size, so a large study can detect trivial differences and a small study can fail to detect important ones.
Overlap diagnostics examine the score distributions in both treatment groups and the resulting weights or matched distances. A region populated almost entirely by treated units cannot support a direct treated-versus-control comparison without strong extrapolation. Structural positivity violations occur when treatment is impossible or certain for a covariate-defined group. Practical violations arise when both treatments are theoretically possible but one is represented by very few observations in the available sample. Estimated scores away from zero and one do not rule out a structural violation if the assignment model is wrong.
Weight concentration can be summarized by an effective sample-size diagnostic,
usually computed separately by treatment group. If all weights are equal, this equals the number of observations. If a few observations dominate, it can be far smaller. This expression is a useful description of weight dispersion, not a complete substitute for an estimator-specific variance calculation.
Diagnostics also reveal changes in the population under study. Reporting how many units remain after matching or trimming, and how their characteristics differ from those excluded, distinguishes internal comparability from population representativeness. Excellent balance among retained units can coexist with limited relevance to the original population.
Statistical uncertainty and sensitivity to hidden bias
Uncertainty calculations must reflect the estimator actually used. Weighted observations are not equivalent to independent observations sampled in proportion to their numerical weights. Matching can reuse controls, stratification introduces estimated groupings, and all these methods may use estimated assignment probabilities. Analyses that ignore such features can misstate the standard error, sometimes substantially.
For weighting and augmented estimators, common approaches include influence-function calculations, joint estimating equations, and appropriate sandwich variance estimators. Treating fitted propensity scores as known is not generally equivalent to accounting for their estimation. In some important settings, estimating the propensity score can improve asymptotic efficiency relative to using the true score in a particular inverse-weighting estimator. This is possible because estimation imposes sample relationships that can remove otherwise extraneous variability; nuisance estimation is not always simply an added source of noise.
Bootstrap resampling is useful for many smooth estimators when the entire fitting procedure is repeated appropriately. It is not universally valid. Abadie and Imbens showed that the ordinary nonparametric bootstrap can fail for standard nearest-neighbor matching estimators with a fixed number of matches. The nonsmooth changes in selected neighbors are central to the problem. Matching-specific variance estimators or other theoretically justified procedures are therefore distinct from simply resampling matched observations.
Residual dependence must also be represented. Cluster-level treatment assignment, repeated observations on the same person, or outcomes correlated within schools, hospitals, or regions can require cluster-aware inference. The number of independent clusters, rather than just the total number of records, can govern precision. Similar care is needed when estimating treatment effects within subgroups: good overall balance does not guarantee balance or adequate overlap inside every subgroup.
Sensitivity analysis addresses a different uncertainty: departures from the assumptions needed for causal identification. Rosenbaum’s matched-study sensitivity framework uses a parameter to bound how much treatment odds could differ between otherwise comparable units because of an unobserved covariate. A value corresponds to equal treatment odds within the relevant matched comparison; larger values allow greater hidden departure from randomized assignment. Bounds on tests or effect statements can then be examined across values of .
Other sensitivity analyses specify the strength of associations between an unmeasured confounder, treatment, and outcome, or introduce bias functions directly. These analyses do not estimate hidden confounding from observed balance. They make explicit how conclusions change under stated violations. Confidence intervals based on sampling uncertainty and sensitivity intervals based on identification uncertainty answer different questions and should not be interpreted interchangeably.
Extensions and boundaries of application
For treatments with more than two categories, assignment probabilities become a collection
Their components sum to one, leaving at most independent dimensions. Pairwise comparisons may require different overlap populations, and a scalar probability for one treatment generally does not balance every possible treatment contrast simultaneously. Multinomial models, generalized balancing weights, and multigroup matching extend the binary framework, but the target intervention and population must be specified for each contrast.
For a continuous exposure , the generalized propensity score is commonly defined using the conditional treatment density,
This is a density, not a probability of receiving an exact continuous dose; it can exceed one and has units determined by the exposure scale. Methods based on it estimate dose–response relationships under continuous-treatment analogues of exchangeability and positivity. The binary balancing theorem does not transfer unchanged into a single universal scalar that solves all continuous-exposure adjustment problems.
Longitudinal settings present another complication. A covariate measured after an earlier treatment may influence later treatment and the outcome while also being affected by that earlier treatment. Ordinary adjustment can then mishandle treatment–confounder feedback. Marginal structural models, developed by James Robins and collaborators, use products of time-specific treatment weights, often together with censoring weights, to estimate effects of treatment histories under sequential exchangeability. The treatment mechanism must be modeled at each relevant time, conditional on the observed history.
Propensity ideas also appear in adjustment for missing outcomes, study participation, and transport between populations. In these settings, an observation or participation probability replaces the treatment probability in part of the weighting construction. Identification requires assumptions appropriate to selection, not merely the treatment-assignment assumptions. A treatment propensity score cannot on its own correct informative loss to follow-up, differential outcome measurement, or selection into a database.
Applications in health research include comparisons of treatments in registries and administrative records; policy applications include educational programs and labor-market interventions. Their credibility depends on measured covariates, temporal alignment, intervention definitions, and the availability of valid comparison groups. Propensity adjustment alone does not resolve reverse causation, immortal-time bias, or exposure misclassification. In particular, assigning follow-up time incorrectly can create apparent treatment advantages even after baseline groups are closely balanced.
There is also methodological debate about propensity-score matching itself. Scalar nearest-neighbor matching is not automatically superior to direct multivariate matching, regression adjustment, or weighting. Pruning observations solely on score distance can sometimes worsen balance on important covariates or discard information without a compensating gain. These concerns distinguish a particular implementation from the underlying balancing theorem: the theorem is exact under its definitions, while an applied procedure must still establish that its estimated scores, retained population, and uncertainty calculations support the intended causal comparison.