ABOUT

USING PROPENSITY SCORES WITH QUASI-EXPERIMENTAL DESIGNS

By William Holmes

A propensity score is the probability that a research subject will fall in an intervention group or a non-intervention group, given his or her specific characteristics. When subjects with a given characteristic are more likely to fall into one group, rather than another, this can bias research results. This book examines how propensity scores may be used to reduce bias with different kinds of quasi-experimental designs and to fix or improve broken experiments.

The book covers the causal assumptions of propensity score estimates and their many different uses. It links their use with analysis appropriate for different designs. It addresses assessment of bias, estimation of propensity scores, and improvement of estimates. It shows graphical as well as statistical methods for this process.

Applications are provided for analysis of variance and covariance, maximum likelihood and logistic regression, two-state least squares, generalized linear regression, and general estimation equations. It shows use with correlated and dependent samples. The examples use public data sets that have policy and programmatic relevance across a variety of disciplines. The data can be used to replicate analysis presented to verify procedures or applied to different issues of a particular discipline.

A distinctive feature of the book is its treatment of using propensity scores with missing data and broken experiments to improve estimates of causal effects. Guidelines for using propensity scores with different designs are presented. Cautions on when estimates using propensity scores do not provide reliable causal estimates are also provided.