The study of causal relationships is one of the most important endeavors in empirical science. Knowledge of whether a treatment (or an intervention) would have a causal effect on an outcome enables you to make informed decisions or take actions that could lead to anticipated results about the outcome. All the following examples involve an investigation of the putative causal effect of a treatment variable on an outcome variable:
Does smoking cause lung cancer?
Can a new policy boost the small business economy?
Does attending a private high school lead to greater success in college than attending a public high school?
When you use randomized controlled trials (RCTs) to study causal relationships, statistical procedures for estimating causal effects are relatively easy to apply. In a simple RCT, each subject (or research unit) is randomly assigned to either the treatment or control condition. The purpose of randomization is to make the two groups comparable by design. Then the observed differences in the outcomes between the treatment and control groups are attributed solely to the causal effects of the treatment. You can use statistical procedures such as the ANOVA, GENMOD, GLM, LIFEREG, REG, and TTEST procedures in SAS/STAT software, among others, to estimate the causal effects and test statistical significance.
But when conducting RCT is not practical, you would need to draw causal inferences from observational data or imperfectly randomized experiments. With observational data, the statistical associations among the observed variables reflect not only causal influences but also noncausal, or confounding, factors. This is the explanation behind the maxim that association is not causation. To infer causal effects from observational data, you might need to rely on specialized methods that can somehow address the confounding associations and other causal inference issues. In this regard, several procedures available in SAS/STAT can mitigate these issues and help make valid causal inferences:
The CAUSALGRAPH procedure applies the theory of graphical causal models to guide you with valid modeling strategies for estimating causal effects.
The CAUSALMED procedure posits a mediator in the causal process through which a treatment causally affects an outcome, and it estimates the direct, indirect, and various causal mediation effects.
The CAUSALTRT procedure estimates causal treatment effects by propensity score weighting, outcome regression, and doubly robust methods.
The PSMATCH procedure provides a variety of tools for propensity score analysis and produces output data sets to which you can apply propensity score matching, weighting, and stratification methods in order to estimate causal effects.
The main features and usages of these procedures are described in the next few sections. For an introduction to causal inference with data from observational or imperfectly randomized experiments, see Berzuini, Dawid, and Bernardinelli (2012), Hernán and Robins (2020), Imbens and Rubin (2015), and Morgan and Winship (2015) and references therein.
For valid causal inference and effect estimation from data to be possible, some assumptions must be made about the statistical techniques and methods that these procedures use. These assumptions are laid out within the potential (or counterfactual) outcomes framework (Neyman, Dabrowska, and Speed 1990; Pearl 2001; Robins and Greenland 1992; Rubin 1990) or the structural causal model framework (Pearl 2009b). It is important to emphasize that these assumptions are in addition to those of standard statistical modeling, such as independent observations or distributional assumptions. It is essential for the validity of the causal inferences that are obtained through the causal procedures that these assumptions be tenable. For more information about these assumptions and the related theoretical frameworks, see the following sections for the individual procedures:
PROC CAUSALGRAPH: Causal Graphs and Potential Outcomes
PROC CAUSALMED: Causal Mediation Effects: Theory, Definitions, and Effect Decompositions
PROC CAUSALTRT: Causal Effects: Definitions, Assumptions, and Identification
PROC PSMATCH: Observational Studies Contrasted with Randomized Trials and Propensity Score Analysis