🧠What the model does
A new model extends the standard vote-choice framework to jointly generate roll-call votes and legislative speech from a common set of underlying preference parameters. Both votes and text are modeled as manifestations of the same latent preferences, allowing speech to inform estimates that traditionally rely only on votes.
🧰 How preferences are estimated
- Estimation uses a sparse Gaussian copula factor model that:
- Infers the number of latent dimensions automatically
- Is robust to outliers in the data
- Accounts for zero inflation common in text counts
- This approach links spatial preference estimation (from votes) with latent factor structure recovery (from text) in a single statistical framework.
📦 What data was analyzed
- Roll-call votes and floor speeches from recent sessions of the U.S. Senate were used to illustrate the estimator's performance.
🔎 Key findings
- Two stable latent dimensions consistently emerge:
- A conventional ideological dimension
- A second dimension that reflects Senators' leadership roles
- The method can leverage shared speech patterns to impute missing vote information, recovering reliable preference estimates for rank-and-file Senators even when only leadership votes are available.
💡 Why it matters
- Jointly modeling votes and text provides a fuller, more robust picture of legislative preferences.
- The estimator is practical for research settings with noisy text data or incomplete voting records, enabling better inference about legislators' positions and role-based behavior.
