Choice experiments for political microtargeting
07 March 2015

I am reading these days some literature on conjoint experiments in which respondents are asked to choose, rank or score a set of alternatives for which some attributes are randomized. In the typical application the number of factors to randomize is too large for your usual full factorial design and you need to resort to slightly more sophisticated experimental designs that can give you, at least, the main parameters with a limited number of experimental runs. These stated choice studies are particularly useful to guide decision-making in the design of products when you want to choose among a large number of candidate profiles. You ask each respondent to perform a small number of tasks in which you present them with a choice between a number of options with randomization in the attributes. From the set of all responses, you can then estimate a carefully parametrized utility function that may have originated the observed data and retrieve quantities such as the elasticity of the demand for each attribute.

Precisely because of that, it seems like a good tool for political microtargeting, at least within the field of political persuasion, and I am surprised it not used more often. In the classical application of GOTV, you want to find the individuals with the highest elasticity of their voting behavior to a particular form of campaigning. In other words, you want to know for a given individual the change in the probability of supporting your candidate if she was to be contacted by, for instance, a canvaser. The structure of the experiment is simple and it only tests one manipulation (contact) on one outcome (expected vote, proxied by partisanship).

But one can also think about estimating not just the effect of the contact but also of the effect of the message, which tends to be multidimensional. Instead of running separate experiments and thus assuming that the effect of each treatment is independent, one may well run only one survey with which to evaluate the relative response of each sociodemographic group to each component of the message on individual choice.