Randomized experiments are frequently considered as the gold standard for inferring causal effects of a treatment, program or policy. However, randomizing subjects into a treatment and control condition is not always possible (e.g., due to ethical concerns). If a randomized experiment is not feasible we frequently rely on quasi-experimental designs for assessing causal effects of a treatment. This course focuses on the design and analysis of the strongest quasi-experimental methods: regression discontinuity designs, interrupted time series designs, non-equivalent control group designs, and instrumental variable approaches. As with randomized experiments, the goal of these quasiexperimental designs is to estimate the impact of a treatment or policy on quantitative outcomes of interest. Though the focus is on causal description (¿is the treatment effective or not?¿) rather than causal explanation (¿why is the treatment effective¿) analytic techniques for causal explanation, like structural equation modeling, will be briefly discussed.

The course starts with an introduction to the philosophy of causation and then outlines the Rubin Causal Model (RCM) in the context of randomized experiments. RCM is the currently predominant quantitative causal model in statistics and the social sciences. Then we focus on four of the strongest quasi-experimental designs: regression discontinuity designs, interrupted time series designs, non-equivalent control group designs (with an emphasis on propensity score methods), and instrumental variable approaches. For each design, we discuss (i) the basic design idea for identifying the treatment effect, (ii) strategies and design elements for improving the basic design, and (iii) statistical approaches for estimating the effect. For each design we will analyze real data and explore and discuss different analytic strategies.

All analyses of real data will be done in R, a free language and environment for statistical computing and graphics (http://www.r-project.org/). The reason for using R instead of a different software package (e.g., STATA, SAS, or SPSS) is twofold. First, the transparency and flexibility of R, including its powerful graphics, enables us to better understand and individually modify analytic procedures (that would not be possible with ¿canned¿ software packages or procedures). Second, not all techniques required for the course are available in other software packages. 

This is a course of Vienna PhD School of Informatics which is also open for Dr.techn students. This is the visiting course of Media Informatics and Computer Vision.

Lecturer: Prof. Peter Steiner (University of Wisconsin¿Madison)

Registration takes place in TISS