How Classifier Accuracy Measures Polarization in Westminster Politics

🔎 The Problem: Polarization Is Hard To Measure In Parliaments

Measuring how polarized legislators and parties are is essential for understanding political change. In Westminster-style parliamentary systems, however, ideological positions estimated from roll-call votes are often uninformative, making valid measurement difficult.

🛠️ How Polarization Is Measured Here

This approach treats classification accuracy from supervised machine learning as a substantive indicator of polarization. Classification accuracy is defined as the number of correct predictions divided by the total number of predictions. The procedure uses party identification as the label and predicts those labels from legislators' speeches plus information on debate subjects.

• Labels: party identification of members of parliament
• Features: speech text and debate-subject information
• Key intuition: high classifier accuracy (few misclassifications) signals strong between-party discrimination and therefore high polarization; low accuracy (many misclassifications) signals weak between-party differences and low polarization

🔬 Tests and Validation: Simulations and Historical Comparison

The measurement strategy is evaluated through simulation experiments and an empirical application covering 78 years of House of Commons speeches. Estimates are compared with existing qualitative and quantitative historical accounts of British politics for the same periods to assess substantive validity.

📈 Key Findings

• Contemporary British politics is approximately as polarized as it was in the mid-1960s, during the heart of the postwar consensus.
• The classification-accuracy measure is computationally fast and aligns with historical evidence from multiple sources.
• More broadly, technical performance of supervised learning algorithms can convey directly useful information about substantive social-scientific phenomena.

❗ Why It Matters

This method offers a practical, theory-guided tool for measuring polarization in settings where traditional roll-call based measures fail. It also highlights how classifier performance can be repurposed as a meaningful substantive quantity of interest in political science.