The US Capitol building in Washington, DC

Absence Makes the Ideal Points Sharper: An Ordinal IRT Model for Complete-Data Rollcall Vote Modeling

Estimating legislator ideal points from roll-call voting data has become a powerful tool for understanding legislative behavior and polarization within legislatures over time. However, the dominant approaches in the field tend to ignore both absences and abstentions in roll-call votes even though it is very likely that absences and abstentions represent a form of strategic behavior. In emerging democracies like Tunisia, this lacuna is particularly problematic because legislator absence may occur very frequently, and resulting estimates that exclude absences are either very imprecise or significantly biased.

As a remedy for this situation, I put forward a Bayesian IRT model that can handle legislators absences as a separate category of data for determining legislator ideal points. The estimation uses the concept of a hurdle model to deflate the probabilities of legislator’s votes by the probability of absences. The model produces a single set of ideal points, but utilizes different parameters for bill absence points. Compared to DW-NOMINATE, this model tends to produce more moderate estimates of US Congresspeople’s ideal points because it can model roll-call votes where there are very few opposing votes. For parliamentary data, the model provides much more precise estimates, especially in legislatures with very high rates of absence. Additionally, the model can incorporate absentions as a middle category between yes and no votes for legislatures with high rates of abstentions.

An R package providing this model, along with other standard IRT ideal point models, is under development. Current source code is available from my Github page. The estimation is done via MCMC within the Stan modeling language.

Link to paper here.