Making Causal Inferences for Unmeasured Variables from Correlations Among Indicators

A technique is suggested for making causal inferences from correlations among indicators of variables that themselves have not been measured. The procedure involves developing causal models using both the indicators and the underlying variables, but then ending up with predictions that use only the measured variables. By selecting several indicators of each underlying variable, some indicators being causes and others effects of the unmeasured variable, we can infer a spurious relationship between two unmeasured variables even if the variable producing the spurious relationship remains unidentified. It is argued that we should develop causal models even where not all variables are measurable.