UC Santa Barbara

Ecologists are increasingly turning to observational datasets to address hypotheses at large spatial and temporal scales. It is well understood (although often observed in the breach) that applying conventional biometric techniques to such data does not provide causally robust conclusions of the sort obtained from well-designed experiments. Potential pitfalls include omitted variables (which may have causal impacts on both variables being measured and create the correlation between them) and reverse (or even bi-directional) causality. While students are taught that "correlation does not imply causation," they are not taught that, regardless of questions of causality, processes such as omitted variables and bidirectional causality can also severely bias the parameter estimates from regressions on observational data; measurement error in the predictor variables can do the same thing, even in a controlled experiment. One solution to this problem is the method of Instrumental Variables (IV), used primarily by social scientists (especially economists). In this chapter I describe the method and show how it can be used to analyze natural experiments, untangle life-history tradeoffs in unmanipulated populations, and analyze time series data subject to measurement error. The method is technically straightforward; the challenge is in identifying a variable that will serve as a suitable instrument. I describe the qualities of good instruments and give some examples.