Statistics Glossary → Propensity score subclassification (stratification)
Propensity score subclassification (stratification)concept
A method that divides observations into a few strata by their propensity scores, estimates the treatment effect within each stratum and then combines them.
The propensity score is an individual's probability of receiving treatment given their observed characteristics. In subclassification the score distribution is split into S strata by quantiles (usually 5); within each stratum the treated and control groups have similar propensity scores and therefore similar observed characteristics. The effect is computed separately in each stratum and combined with weights based on stratum sizes. If the strata are formed from the quantiles of the whole sample, the average treatment effect (ATE) is estimated; if they are formed from the quantiles of the treated group only, the average treatment effect on the treated (ATT).
According to Cochran's classic finding, five strata remove about 90% of the bias due to an observed confounder, but some imbalance remains within each stratum. Covariate balance should therefore be checked within strata, and if needed the number of strata increased or a regression adjustment made within strata. The extreme strata may contain very few treated or control observations (a common-support problem). Like every propensity score method, it balances only the measured confounders.
An everyday example
A city wants to evaluate a voluntary stop-smoking course; those who join are younger and better educated. The probability of joining is estimated from age, education and income, and participants are split into five propensity strata. Computing the difference in quit rates in each stratum (e.g. 8, 10, 12, 9 and 11 points) and averaging them leads to the conclusion that the course raises the quit rate by about 10 points.
In MerQur
When the stratification method is chosen in propensity score analysis, the 'Number of subclasses' option sets how many strata are formed from propensity score quantiles; for the ATT, the quantiles of the treated group are used.