You Don’t Come into My Journal, Drop a Causal Inference Challenge, and Leave


Martin Gilens and Benjamin Page have a major new piece on the nature of American democracy in the latest issue of Perspectives on Politics. Perspectives comes straight to my mailbox so I always browse it, but this article caught my eye because (1) it’s important and (2) its finding that economic elites and interest groups explain policy action accords with my own subjective beliefs about “how American democracy really works.” From the abstract:

Multivariate analysis indicates that economic elites and organized groups representing business interests have substantial independent impacts on U.S. government policy, while average citizens and mass-based interest groups have little or no independent influence. The results provide substantial support for theories of Economic-Elite Domination and for theories of Biased Pluralism, but not for theories of Majoritarian Electoral Democracy or Majoritarian Pluralism.

Yet what caught my eye and sponsored this post is a quote on pages 572-3.

As noted, our evidence does not indicate that in U.S. policy making the average citizen always loses out. Since the preferences of ordinary citizens tend to be positively correlated with the preferences of economic elites, ordinary citizens often win the policies they want, even if they are more or less coincidental beneficiaries rather than causes of the victory. There is not necessarily any contradiction at all between our findings and past bivariate findings of a roughly two-thirds correspondence between actual policy and the wishes of the general public, or of a close correspondence between the liberal/conservative “mood” of the public and changes in policymaking. Our main point concerns causal inference: if interpreted in terms of actual casual impact, the prior findings appear to be largely or wholly spurious.

What motivates this comment is their finding that mass public opinion predicts policy change in a bivariate regression-type analysis, but when controlling for the preferences of the richest people and of narrow interest groups, that relationship disappears.

I believe that the relationship they report is accurate, and moreover, that their description of the underlying structure politics that that relationship suggests is actually correct (more or less). But I do not think that this kind of statistical analysis shows it, or that the causal inference language that I bolded above is appropriate.

Why? Because these correlations do not correspond to causal questions of the type “what is the effect of an change in mass public opinion on the likelihood that a bill is passed?” Think about it: just what does “actual causal impact” mean? It cannot mean conditional correlation, which is what we are seeing. It must mean something counterfactual. The authors are presumably alluding to the possibility that there is a complex, perhaps unobservable, relationship between mass public opinion and elite opinion/interest group behavior. Perhaps “in the wild” there is little independent variation between mass opinion and the other two, so that it’s unrealistic to think that we could conceptually separate the two. Throughout the text they suggest this is true. But we cannot back out from what they have shown here any conclusion about the causal impact of mass public opinion.

As a further note: even if we didn’t care about causal inference, we should not test competing hypotheses—be they nested or non-nested—through big multiple regression models. We have a range of better procedures for doing that.