Statistical Modeling and Inference for Social Science

We have ignored the elephant in the room long enough. Social science theory does not just imply relationships. It often implies causal relationships – that a change in an explanatory factor causes a change in the dependent variable for some specified reason. Yet all the methods and concepts we have covered are simply about assessing relationships, causal or not. We have tools to describe relationships in observed data, and tools to make inferences about whether the relationships observed in data hold in the underlying process that generated the data or whether they can be ascribed to idiosyncratic chance. We have tools to make a best guess about the relationships that hold in the data-generating process (DGP) and quantify our uncertainty about those guesses. But none of this deals in a sustained fashion with whether relationships are causal relationships.

When we spin theories, we often spin them in causal terms. And when we make policy recommendations, we often do so on the belief that the recommended intervention will cause a change in some important outcome. But causation is not just a theoretical concept. “Correlation does not imply causation” has reached the status of a truism. In light of this maxim, it might seem that the best approach to inferring causation from correlation is not to try it. But this causal nihilism is not reasonable. Claims of causation are more credible in some empirical research than others. We need to ask why this is so, and if possible, try to make empirical research with causal aspirations follow a template that makes causal claims credible.