Aaron Strauss spoke today on his work with Kosuke Imai on estimating the optimal order of priority and the optimal approach for contacting voters in a political campaign. They use inferences from field experiments on voter turnout and persuasion and then transfer these findings into a decision-analytic framework. I can’t find a link to the latest version of their article, but a not-too-old version is here.

The talk was fascinating, and a bunch of points came up in the discussion and afterward that I wanted to set down here.

1. Aaron has worked for some Democratic campaigns. We had some discussion about whether he’s worried that candidates on the other side will use these methods now that they are being published. He said he’s not so worried in the short term because people on the other side would still have to learn his methods, and meanwhile he’ll be on to something better.

Long term, though, is more accurate voter targeting a plus, a minus, or simply a zero-sum game? Is it an arms race where both sides have to keep doing this just to keep up?

We discussed a few sides to this point, including the argument that targeting is good because it focuses the campaigns on the actual concerns of the voters (rather than on the concerns of Mark Penn-style gurus), and the argument on the other side that targeting is bad because it empowers existing parties and institutions.

2. Imai and Strauss use a tree-based regression model. In some ways this is good–the discreteness of the model allows the computations to be done more easily and it also automatically allows nonlinearity and nonmonotonicity, which can get you more interesting decisions as compared to linear models, where the optimum is always on the boundary. I recommended trying Bart, which Jennifer and Rob have extended to the problem of estimating treatment interactions in causal models.

3. Imai and Strauss discuss the problem of choosing different treatments for different units. Erwann Rogard, Hao Lu, and I discuss some aspects of this issue in our article, “Evaluation of multilevel decision trees.”

4. I really like their focus on treatment interactions, which reminds me of the “ATE paradox,” which I discussed here but I think is worthy of its own journal article or blog entry at some point.

5. Beyond allowing the treatment effect to vary–that is, including treatment interactions in the model–I suggested to Aaron that they consider treating the “treatment” and “control” conditions asymetrically. Standard statistical theory considers treatment and control to simply be conventional names for two different treatments, but in practice there are differences–for one thing, the correlation between “before” and “after” measurements tends to be higher in the control group than the treatment group. This makes sense, given that the control is typically to do nothing, and treatments are typically active interventions–but this distinction is almost never in a statistical model. (See this article from 2004 for some of my thoughts in this area.)

6. They should also take a look at Rajeev Dehejia’s article, “Program evaluation as a decision problem,” which discusses a similar issue of assigning treatments to some units and not others.

7. Imai and Strauss’s analysis of experimental results from the 2006 campaign suggested that text-messaging is effective for improving the turnout of potential voters aged 20-25, but ineffective or even counterproductive for voters between 18 and 19. I’m highly skeptical of this finding–it seems all too possible for this sort of pattern to come out artifiactually from the fitting of the tree model. That said, their result is consistent with the findings of Sendhil Mullainathan and Ebonya Washington that people who haven’t voted in a presidential election are less politically engaged.

That’s all, I think.