Propensity score matching
Propensity Score Matching (PSM, German roughly paired assignment based on inclination scores ) is a form of matching for estimating causal effects in non- experimental observational studies.
PSM was introduced in 1983 by Paul Rosenbaum and Donald Rubin.
application
It is used in the social sciences to estimate the effect of an intervention (e.g. a policy measure). The basic problem of causal analysis is that it is not possible to simultaneously measure how an individual behaves with and without the intervention. Participants and non-participants already differ before the intervention. Sampling bias occurs if participants are not randomly assigned to the intervention ( randomization ).
PSM tries to reduce this bias and to imitate randomization by forming pairs of people who are as identical as possible and for them the effect of the measure is compared.
execution
First, the propensity scores are estimated. The basic problem that is solved by this one-dimensional measure is the so-called curse of dimensionality . Originally, people are sought who are similar or identical in all the variables considered. This is partially possible for categories such as gender. For metric variables such as age and income and every other variable, the problem arises of finding a person with the same gender, identical age (to the exact day) and income (except for the euro).
To make the treatment and control groups comparable, a single value between zero and one is calculated based on covariates . This is usually done using a logistic regression .
There are a number of methods for matching, for example nearest neighbor , caliper and radius, stratification and kernel-based methods. The matching quality is checked afterwards.
literature
- Caliendo, M., & Kopeinig, S. (2008). Some practical guidance for the implementation of propensity score matching. Journal of economic surveys, 22 (1), 31-72.
Individual evidence
- ^ Rosenbaum, PR, & Rubin, DB (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70 (1), 41-55.