Richard asks if the patterns in our red/blue paper can be explained by different income distributions in different states. In particular, the income quantiles we use (which come from the National Election Study, and which we use so we can have comparable estimates across states and across years) are based on national income quantiles. Thus, the people in the highest income category in Mississippi are richer (compared to average Mississippians) than the people in the highest income category in Connecticut (compared to average Connecticutians). Similarly, the people in the lowest income category in Mississippi are richer (compared to average Mississippians) than the people in the highest income category in Connecticut (compared to average Connecticutians).

We were certainly aware of this in writing the paper (in fact, the sizes of the circles in our graph make clear that the relative proportions of each of the categories varies from state to state). The question is: can this explain away our results? As Richard writes,

So here’s an alternative hypothesis for the results they’re getting: in each income quantile the Mississippians feel richer than the Connecticutters in the same quantile and therefore they vote Republican. This can be seen as a slight twist on the economic determinism argument, with subjective, relative income replacing objective, absolute income as the determinant.

So would correcting for this make the effect they found go away? Probably not… or at least it’s not clear that it would. But it would be interesting to see how the model held up on state-specific quantiles and/or COL adjusted ones. Would the top 2% of Connecticutters be more Republican than the top 8%, thus pushing the slope positive and diminishing the overall effect?

My response: this could explain some of the difference in intercepts of the regression lines, but I don’t see it explaining the differences in the slopes, which is what we’re most interested in.

However, it might be a worthwhile calculation to set up a model (e.g., a logistic regression of preference on individual income relative to state-average income) and see how much things change by switching to national-average income as a baseline. Due to discreteness in the data, it would be tough to do this by simply running a new regression, but with a little modeling it shouldn’t be hard to estimate the size of this scaling effect.