A Structural Approach to Legislative Roll Call Vote Prediction

I present a structural approach to the task of legislative roll call vote prediction. Given a voting history, consisting of co-sponsorship information and voting records for past legislation, along with a new piece of legislation with sponsor and co-sponsor information, the task is to predict whether the new legislation will pass or be rejected. I propose an approach that constructs a network representing influence relations based on co-sponsorship information in the voting history. This network is then used to predict the votes of the individual legislators, given sponsorship information for the new legislation. 1. The Task: Legislative Vote Prediction When a new piece of legislation is introduced in the U.S. Senate, it is necessarily sponsored by a particular senator. The legislation may also be co-sponsored by several other senators. After discussion, revision, and other processing, the legislation is voted on and is either passed or rejected based on the number of Yea votes. One key part of political analysis in such legislative bodies is determining whether or not a piece of legislation will pass, given the group of legislators who endorse it. This allows politicians and lobbyists to better allocate their efforts to back legislation that has a greater chance of success, or seek more support when the prospects for their proposed legislation are bleak. Unfortunately, this task of vote prediction is currently based on vague intuitions and rules of thumb of experienced political analysts. Even then, the analyst must spend long hours pouring over the details of voting records and positions every time a new legislator is elected, in order to come up with accurate predictions of that legislator’s voting behavior. It is therefore desirable to automate this process of vote prediction, to take advantage of the speed and flexibility of computer systems, while maintaining nearly equivalent reliability and accuracy to that of the human political analyst. 1.1 Task Formulation In the task of interest here, there is a legislative body, consisting of a set of legislators Xi. The set is assumed to be unchanging, which is the case for example in a typical session of Congress. We are given a vote history for this legislative body, consisting of several pieces of legislation, each accompanied by the name of the legislator who sponsored the legislation, the names of any co-sponsors, and the voting behaviors of all legislators for that legislation. Each legislator may be recorded as voting Yea, Nay, or Not Voting. We are then given a new piece of legislation, with information about who the sponsor and co-sponsors are, but no information about the voting behavior. Based on information gathered in the vote history, and the list of sponsor and co-sponsors for the new piece of legislation, we are asked to make a prediction on whether or not the new legislation will pass. 1.2 Outline of the Approach Thus, this task may be formulated as one of classification, where the instances are drawn from a distribution over the space of possible co-sponsorship combinations, and labeled as Pass or Reject based on some unknown target function, employing hidden variables. However, the task can also be broken down further into the subtasks of predicting the voting behavior of each individual senator and then applying a majority function to determine the overall vote. This approach has the advantage that it allows one to introduce additional structure in the problem based on knowledge of the task, which restricts the space of possible classifiers to be much smaller, without losing the ability to fit the data reasonably well, thus reducing the amount of vote history necessary to learn an accurate classifier (Vapnik, 1998). This type of restriction becomes necessary here, since the vote history would never realistically be larger than 1,000 votes. Specifically, in the approach taken here, I define a notion of influence, a relation between pairs of legislators. I assume that the sponsor and co-sponsors vote Yea, and that they then try to use their influence to convince others to vote Yea. Those who are convinced will vote Yea, and also try to use their influence to convince others to vote Yea. Thus, the vote is propagated through the network of influence relations. The details of the classifier lie in the function that determines whether or not a senator is convinced to vote Yea, based on the votes of those who influence him or her, and also the algorithm used to propagate the vote. As a preliminary approach, I have chosen to use a sigmoid function for the first detail, and a sampling algorithm for the second. The network structure and the parameters for the sigmoid function are learned based on the vote history. 2. Existing Models of Roll Call Voting There has been much work in the political science literature on analyzing roll call voting behavior in legislative bodies. These theories are centered around a Spatial Voting Model (Poole; Ladha, 1994; Clinton, et al, 2004). This framework maps each legislator to an ideal point in a low-dimensional space based on voting history, and likewise maps each piece of legislation to two position points in the space based on how the legislators voted: one point for Yea and one for Nay. Each legislator has some utility function, defined for each point in the space, and chooses to vote according to the position point that maximizes that utility. Generally, an assumption of sincere voting is made, under which the legislators’ utility functions remain static from one vote to the next. This assumption significantly simplifies the analysis, but fails to capture the negotiation and other interactions among legislators. Attempts to relax this assumption, while maintaining the spatial model include (Clinton, et al, 2004; Clinton & Meirowitz, 2003), which alter the utility function to account for these interaction factors. Unfortunately, the spatial voting model is not immediately applicable to the task of vote prediction, since the position points are calculated using knowledge of the voting behavior. Although the ideal points and position points can all be estimated from the historical voting data, the position points of any new pieces of legislation, for which the votes are not yet known, cannot be calculated using existing methods. 3. A Structural Approach As mentioned above, the sincere voting model fails to capture the interactions, negotiations, and influences that are so crucial to a legislator’s voting decision. A more realistic approach should make these interactions a central part of the model. With this motivation in mind, I propose a social network model based on the influence structure of a legislative body. 3.1 Influence Networks In the social networks literature, there has been much work on constructing influence networks for the propagation of ideas and innovations. For example, (Coleman, et al, 1966) provides a classic study. The key to modeling these networks is the temporal order of adoption of an idea, along with the opportunity for an idea to be propagated from one individual to another (i.e., a pre-existing relationship), and some notion of the strength of the influence. The influence relation is then modeled as a directed tie between two actors in the network. 3.2 Co-Sponsorship as Indicative of Influence Discovering the structure of the influence network in a legislative body, given only the vote history, poses some interesting challenges. However, the nature of the task allows one to make the simplifying assumption that all legislators are in contact with all other legislators in the same legislative body. This assumption is justified given the relatively small size of most legislative bodies. This still leaves open the question of how to determine which legislators are related by an influence relationship. The approach I propose is to utilize the sponsorship and co-sponsorship data to create directed edges, which seem to capture some aspect of the influence one legislator has on another. Specifically, legislator A is connected to legislator B with a directed edge A→B if and only if there is some legislation on which A is the sponsor and B is a co-sponsor. The value of the edge is determined by an increasing function of the number of such pieces of legislation. This model of influence, though admittedly not perfect, seems to capture a notion of influence, since legislator A was presumably first to take interest in the issue, and through some method of communication with the others, was able to convince B to be a co-sponsor. It seems natural to conclude that in the future, even when A is unable to convince B to go so far as becoming a cosponsor, he or she would still have a good chance of convincing B to vote a specific way if the A→B link has a high value. 4. Inference: Predicting the Votes Once one has the structure of the influence network, the remaining step is to predict the voting behavior of each legislator, given only information about the sponsor and co-sponsor identities. We can comfortably assume that the sponsor and co-sponsors will vote Yea. Thus, the task can be viewed as one of classifying the remaining legislators as Yea or Nay, given some small number of Yea votes at known positions in the network. This problem can be formulated in a variety of ways, such as information diffusion, transductive inference, or belief propagation.