SMABS 2004 Jena University
SMABS 2004 Home Organization About Jena Sponsors Links Contact SMABS Home

European Association of Methodology

Department of methodology and evaluation research

Jena University


© Webmaster

Contributions: Abstract

Testing causal effects

Carmen Brosche
University of Jena
Germany

We study the effect on a outcome or response variable of a treatment 0 relative to another treatment 1 for a population of N units. Because of assigning either treatment 0 or treatment 1 to a unit it is impossible to determine the amount of individual or average causal effects from available observations directly.

Causal inference is based on missing values. Special assumptions for the assignment mechanism, which describes the assigning of treatments to units, allow to construct tests. Rubin (1990) presents an approach for testing the null hypothesis of zero individual causal effect for all units against the alternative of positive individual causal effect. Under the null hypothesis it is possible to determine a p-value for the observations. The p-value of several hypothesised constant causal effects can be used to form confidence limits.

A modification of the assumptions makes it possible to use Rubin's model for a null hypothesis of zero average causal effect without the condition of constant additive individual causal effects. To calculate p-values (and confidence limits) an assumption is necessary limiting the range of the unknown response values. Another extension of Rubin's theory examines the consideration of n-tuples of units. An example illustrates the approaches.