KDIR causal knowledge

Automatic identification of quasi-experimental designs for discovering causal knowledge

by Jensen et. al

Introduction

• blackbox approaches (statistical models, artificial intelligence, ...)
• relational dependency network; visualizes the impact of factors (in a way similar to a spreading activation network)
• difference: statistical associated <-> causal effect
• people are interested in causal data
• compare:
  • elections <-> campaign strategy
  • social science <-> social politcy, ..

• problem:
  • elimination of common causes (3rd influence factor; Scheinkorrelation)
  • approaches toward eliminating the effect of common cause
    • control => hold 3rd variables
    • randomization -> compare: sample design (Fischer 1925)
    • Modeling: remove effects of potential common causes
    • quasi-experimental designs
  • emulates control by finding structures in the data¨-> set of conditions to tread the condition as if the data would come from an experiment
  • examples designs:
    • twin design - twins => eliminate the influence of genetics
    • two group pre-test post-test design
    • instrumental variable design
• challenge: algorithms to (automatically) use the idea of quasi-experimental designs
  • requirements: temporal / spatial structure; relational structure
  • number of possible causal models is (3^(N^2))

Method

• use google-scholar queries to compare a topic's popularity

Literature:

• Jensen et al. (2008) Automatic identification of quasi-experimental designs for discovering causal knowledge (KDD 2008)
• work on causal interpretation of Bayes networks