Using an accessible approach perfect for social and behavioral science students (requiring minimal use of matrix and vector algebra), Holmes examines how propensity scores can be used to both reduce bias with different kinds of quasi-experimental designs and fix or improve broken experiments. This unique book covers the causal assumptions of propensity score estimates and their many uses, linking these uses with analysis appropriate for different designs. Thorough coverage of bias assessment, propensity score estimation, and estimate improvement is provided, along with graphical and statistical methods for this process. Applications are included for analysis of variance and covariance, maximum likelihood and logistic regression, two-stage least squares, generalized linear regression, and general estimation equations. The examples use public data sets that have policy and programmatic relevance across a variety of social and behavioral science disciplines.
Causal Inference Using Control Variables
This chapter reviews how control variables are used to reduce problems of causal inference. It examines trade-offs between different strategies in using control variables. Issues unresolved by control variables are identified and discussed. The various ways in which propensity scores are used to control confounding variables are addressed. Their use with alternative designs will be discussed in Chapter 4.
While it is preferable to control confounding factors with appropriate design characteristics, reasons have already been given why this is not always legal, moral, or feasible. Things that confound inference about causal effects that cannot be controlled by research design need to be controlled statistically. Five approaches are used to statistically control confounding factors: (1) matching, (2) stratifying, (3) weighting, (4) ...