DEEPCAUSAL Procedure
Overview: DEEPCAUSAL Procedure
The DEEPCAUSAL procedure estimates the average causal effect and performs policy evaluation and policy comparison by using deep neural networks (DNNs). DNNs overcome several technical difficulties in the big data era, including the following:
the huge amount of discrete or continuous potential covariates
the unknown nonlinear relationships among the covariates, the outcome, and the treatment assignment
A problem in applying DNN to causal inference is interpretability. To solve this problem, the DEEPCAUSAL procedure applies the DNNs via a two-step semiparametric framework (see the section Details: The DEEPCAUSAL Procedure) and gives inferential results for the parameters of interest through the corresponding influence functions.
The DEEPCAUSAL procedure can estimate 11 types of causal effects (or parameters of interest). These include full-population causal effects, such as the average treatment effect (ATE), and subpopulation causal effects, such as the average treatment effect for the treated (ATT) and the average treatment effect for the untreated (ATU). For more information, see the sections Full-Population Average Effect Parameters and Subpopulation Average Effect Parameters. The DEEPCAUSAL procedure can also perform policy evaluation and policy comparison. For more information, see the section Policy Evaluation and Comparison. PROC DEEPCAUSAL supports the SCORE statement, which enables you to save the causal model specifications and estimation results for scoring and policy evaluation and comparison without needing to reestimate the DNNs for these models.
Tools for causal analysis of nonrandomized data are available in the following procedures:
The CAUSALGRAPH procedure enables you to analyze graphical causal models and to construct sound statistical strategies for causal effect estimation. For more information, see Chapter 37, The CAUSALGRAPH Procedure (SAS/STAT User's Guide).
The PSMATCH procedure enables you to perform propensity score analyses and to assess covariate balance. For more information, see Chapter 101, The PSMATCH Procedure (SAS/STAT User's Guide).
The CAUSALTRT procedure enables you to perform estimation of a causal effect. For more information, see Chapter 39, The CAUSALTRT Procedure (SAS/STAT User's Guide).
The CAUSALMED procedure enables you to decompose a (total) causal effect into direct and indirect effects. For more information, see Chapter 38, The CAUSALMED Procedure (SAS/STAT User's Guide).
The MODEL, PANEL, QLIM, and TMODEL procedures in SAS/ETS and the CPANEL and CQLIM procedures in SAS Econometrics support endogenous variables and instrumental variables. For more information, see Chapter 24, MODEL Procedure (SAS/ETS User's Guide), Chapter 25, PANEL Procedure (SAS/ETS User's Guide), Chapter 27, QLIM Procedure (SAS/ETS User's Guide), Chapter 39, TMODEL Procedure (SAS/ETS User's Guide), Chapter 11, CPANEL Procedure, and Chapter 12, CQLIM Procedure.
The VARMAX procedure supports Granger causality tests. For more information, see Chapter 42, VARMAX Procedure (SAS/ETS User's Guide).
PROC DEEPCAUSAL requires SAS Cloud Analytic Services (CAS) in order to run, and it does the following:
enables you to run on a cluster of machines that distribute the data and the computations
exploits all the available cores and concurrent threads