Inverse probability weighting (IPW) is a statistical technique that makes observational data behave more like data from a randomized experiment. It works by assigning each person in a study a weight based on how likely they were to receive the treatment they actually got. People who were unlikely to end up in their group get larger weights, and people who were very likely get smaller ones. The result is a rebalanced dataset where background characteristics no longer predict which group someone belongs to, letting you estimate the true effect of a treatment or exposure.
The Core Problem IPW Solves
In a randomized trial, a coin flip decides who gets the treatment and who doesn’t. That randomness ensures the two groups look similar on average: same mix of ages, health conditions, habits, and everything else. Any difference in outcomes can be attributed to the treatment itself.
Observational data doesn’t have that luxury. Doctors choose treatments based on patient characteristics. Sicker patients may get a more aggressive drug, older patients may avoid surgery, and so on. These factors that influence both who gets treated and how they fare are called confounders. If you simply compare outcomes between treated and untreated groups without accounting for confounders, you’ll get a biased answer. IPW is one of the most widely used tools for removing that bias.
How the Weights Are Calculated
IPW works in two steps. First, you estimate each person’s propensity score: the probability that they would receive the treatment they actually received, given their observed characteristics like age, sex, disease severity, and other relevant factors. This is typically done with logistic regression, though machine learning methods can also be used.
Second, you turn those probabilities into weights by taking the inverse. For someone in the treated group, the weight is 1 divided by their propensity score. For someone in the untreated group, the weight is 1 divided by (1 minus their propensity score). A treated person with only a 20% predicted chance of being treated gets a weight of 5, meaning they “count” as five people in the analysis. A treated person with a 90% chance of being treated gets a weight of about 1.1, barely inflating their contribution at all.
The logic is intuitive once you see it. People who received the treatment despite being unlikely candidates for it carry more information about what the treatment does across a broader population. Upweighting them, and downweighting the people who were almost certain to receive it, balances out the confounders that drove treatment decisions in the first place. The weighted dataset is sometimes called a “pseudopopulation” in which confounders are equally distributed across exposed and unexposed groups.
Why Stabilized Weights Are Preferred
The basic (unstabilized) weights described above have a practical problem: they inflate the effective sample size. The sum of all the weights ends up larger than the actual number of people in the study, which tricks standard statistical software into thinking you have more data than you do. That leads to artificially narrow confidence intervals and an inflated rate of false-positive findings.
Stabilized weights fix this by multiplying the basic weight by the overall probability of treatment (ignoring individual characteristics). For treated individuals, the stabilized weight is the overall treatment probability divided by their individual propensity score. For untreated individuals, it’s one minus the overall treatment probability divided by one minus their propensity score. The result is a set of weights that center around 1 and preserve the original sample size in the pseudopopulation. Stabilized weights also reduce the impact of extreme values, produce appropriate variance estimates directly from standard regression models, and require no additional correction steps.
The Positivity Assumption
IPW rests on a critical requirement called positivity: every person in the study must have had some nonzero chance of receiving either treatment. If a subgroup of patients would never receive a certain drug (say, because of a strict contraindication), their propensity score is effectively zero, and their weight shoots toward infinity.
When propensity scores land very close to 0 or 1, even a few individuals can receive enormous weights that dominate the entire analysis, inflating variance and distorting the estimated treatment effect. This is one of the most common reasons IPW analyses go wrong. Researchers handle it through trimming (removing individuals with extreme scores, such as those above 0.9 or below 0.1) or truncation (capping weights at a fixed value). Alternative weighting schemes like overlap weights assign near-zero weight to individuals who strongly violate positivity, sidestepping the problem more gracefully.
Structural violations are more serious. These occur when the treated and untreated groups differ so fundamentally in their characteristics that no amount of reweighting can make them comparable. In those cases, the treatment effect for the full population simply isn’t identifiable from the data.
Checking Whether IPW Worked
After applying weights, you need to verify that the pseudopopulation is actually balanced. The standard diagnostic is the standardized mean difference (SMD) for each covariate, calculated before and after weighting. An SMD below 0.1 is a commonly used threshold indicating adequate balance. If some covariates remain imbalanced after weighting, the propensity score model likely needs revision, perhaps by adding interaction terms, nonlinear terms, or additional confounders.
Inspecting the distribution of weights themselves is also important. A few very large weights signal that the model is relying heavily on a small number of individuals, which makes the results fragile. Comparing stabilized and unstabilized weight distributions, checking maximum weight values, and looking for outliers are all routine parts of a responsible IPW analysis.
IPW vs. Propensity Score Matching
Propensity score matching is probably the better-known alternative. It pairs treated individuals with untreated individuals who have similar propensity scores, then discards unmatched people. IPW takes a different approach: nobody gets discarded. Every person in the dataset stays in the analysis, just with a different weight. This is a significant advantage when the sample is small or when losing participants would reduce statistical power.
Matching does have strengths of its own. When the propensity score model is correctly specified and treatment selection is strong, full matching (a variant that uses all subjects by forming matched sets of varying sizes) tends to produce estimates with lower bias than IPW. That’s because IPW is more vulnerable to extreme weights in those settings. In practice, many analysts run both methods as a sensitivity check, expecting consistent results if the analysis is sound.
IPW for Missing Data
Beyond treatment effects, IPW is widely used to handle missing data. The idea is the same: if certain types of participants are more likely to drop out of a study or have incomplete records, the remaining complete cases are not representative. You can model the probability of being observed (rather than the probability of being treated), then weight the complete cases by the inverse of that probability. People who are similar to those who dropped out but happened to stay get upweighted, correcting the bias.
This approach works when data is “missing at random,” meaning the chance of being missing depends on observed variables but not on the missing values themselves. It’s most straightforward when the missing data pattern is monotone (people who drop out stay out). Handling more complex, nonmonotone patterns with IPW is an active area of methodological work.
Augmented IPW and Double Robustness
Standard IPW has a notable weakness: if the propensity score model is wrong, the weights are wrong, and the treatment effect estimate is biased. Augmented inverse probability weighting (AIPW) addresses this by combining IPW with a separate model that directly predicts the outcome. The result is called “doubly robust” because the estimate remains valid as long as at least one of the two models is correctly specified.
If the propensity score model is accurate, AIPW behaves like standard IPW. If the propensity score model is inaccurate but the outcome model is correct, the estimator falls back on the outcome model and still produces consistent results. You don’t need to know in advance which model is right. This safety net has made AIPW increasingly popular, particularly in health economics and comparative effectiveness research where model misspecification is a constant concern.