assumptions

Gabriel 

Yesterday at the Harvard-MIT economic sociology workshop, I saw David Brady from Duke present some work on an approach to poverty that takes both micro (human capital, family structure, etc.) and macro (welfare state, economic growth, etc) seriously. In a nutshell, he finds that in the OECD, left-wing governments imply a large welfare state which implies not only low poverty, but ameliorates the micro predictors like female-headed household and low education. Furthermore, he found that the main effect of unionization is mediated by the state (i.e., collective bargaining is peanuts and the real action is unions getting lefties in office). Brady cast the research as testing liberal economic theory (growth solves poverty by increasing demand for labor), structural theory (liberal theory is basically right but it doesn’t work for all groups), and power relations theory (poverty is solved when left-wing parties take power and solve it) — with the results most supporting power relations, being ok for structural, and not at all for liberal. He had some very well designed models and good data and basically did great stuff with the error terms, intervening variables, and all that, but there was a serious problem, at least rhetorically, and possibly epistemologically as well.

The problem was the dependent variable. Brady measures poverty as a binary variable defined as disposable income less than half of the context-specific median. This means there are two ways for “poverty” to increase: for the bottom to lose disposable income or for the median to gain it. This leads to conclusions that are a bit hard to swallow on the face of it, like that Ireland has recently experienced a sharp increase in poverty, whereas the conventional way to describe that island is that in the last twenty years it has experienced explosive economic growth and now has large positive migration. Despite the counter-intuitiveness of a celtic tiger whose poor were better off when the economy was in the toilet, the definition of <.5contextmedian(disposable) is very defensible based on a large and solid literature across the social sciences (which Brady knows infinitely better than I do) on both bidding up of scarce positional goods as well as less tangible and more subjective aspects of relative deprivation. Basically, it is true that as the saying goes, America is a country where the poor people are fat — but this doesn’t mean that at least in some meaningful sense they aren’t really poor.

Nonetheless, even if you think this definition of poverty has a lot going for it, there is an inescapable sense that we’re witnessing a bait-and-switch where we’re measuring inequality and calling it poverty. It’s a lot less interesting to demonstrate that left-wing parties reduce inequality than to say they reduce poverty. At least two people in the room (myself and the guy who beat me to asking the question) were thinking this and Brady said he gets that question all the time. His answer was basically that absolute measures of poverty are both philosophically unjustifiable and methodologically problematic in comparative research because of things like which basket of goods (notably health care) the state provides directly. I thought this was a convincing argument for making <.5contextmedian(disposable) his preferred measure but a rather weak and defensive reason for making it the only measure. A big part of the problem is that he cast his project as testing liberal economic theory vs. power relations theory, but he uses a definition of the outcome preferred by the latter. From a liberal perspective, this is a cop-out since they never claimed that they would compress the income distribution, only that they would promote Pareto-efficient growth. You can use happiness research to demonstrate that this goal will not maximize subjective utility of the bottom quartile, but you can’t use it to demonstrate that it won’t succeed on its own terms. To see if liberal economics accomplishes what it wants to accomplish, you need a liberal outcome; whether it is right to want to accomplish that is a different question.

More generally, say that you build an analysis on some assumptions that you really like and you have all sorts of good reasons to favor, but that some people are just too stubborn to take them seriously. If these people are just marginal cranks, you can ignore them. But if they are not, then you have a problem. You can either convince them to adopt your assumptions (good luck), or you can, for the sake of argument, adopt their flawed assumptions and see if your results are robust. If they are robust, you win, since even on their terms you’re still right. If they aren’t robust, then you have a problem, but at least you’ve narrowed and clarified it. It’s not necessarily worth doing this if it would be extremely cumbersome (radically different model specification, nontrivial data collection) but if it’s something simple (log something, use Huber-White standard errors, break a continuous variable into a dummy set to look for nonlinear effects) then there’s really no excuse. You don’t have to make the main presentation their way, but there’s nothing so convincing as a footnote saying “I tried it the other way and the results are robust.” Basically, when you’re dealing with some obstinate asshole (especially if that asshole is a peer reviewer) it’s a good idea to be the adult and let them be the stubborn one rather than engage in a shouting match of the deaf. Would it really kill you to say, I still think I’m right, but here’s what it looks like if we assume you’re right? In the specific case of Brady, I think his research is solid, but it would be devastating if he added an appendix where he does all his models the same, but uses an absolute definition of poverty to show that his analysis is robust to assuming the importance of relative deprivation. (Or is it? There’s only one way to find out.)