Selecting Instrumental Variables: A Graph-Theoretic Approach

Abstract

An algorithm is developed for the empirical selection of valid instrumental variables based on graph-theoretic causal search (or Bayes-net) methodologies for directed cyclical graphs. The method also helps to identify causal structure when the underlying relationships are simultaneous or cyclical, as well as recursive. The method is illustrated using cross-country data set with 87 variables for 72 countries.

Keywords: instrumental variables, Bayes nets, causal search, directed cyclical graphs, cross-country growth

JEL Classification: C01, C30, C36