How Bayesian Models Reconstruct Country-Year Public Opinion on Democracy

📌 The Challenge: Macro-Level Opinion Is Fragmented

Many survey responses exist at the microlevel, but reliable country–year panels needed for macro political behavior are scarce. The goal is to measure smooth country–year public opinion series even when surveys are sparse, irregular, or use different items across time and place.

📊 A Modeling Framework That Links Disparate Surveys

A Bayesian dynamic latent trait modeling framework is developed to produce annual, country-level estimates of public opinion by smoothing information across time, space, and survey items. From this framework, six specific models are derived and applied to opinion data about support for democracy.

🧭 How Model Performance Is Assessed

• Models are validated with multiple tests of reliability and validity: internal, external, construct, and convergent validity.
• A held-out test dataset is used to evaluate predictive accuracy.

🔍 Key Findings

• The best-performing model predicts individual response proportions with a mean deviation of 6 percentage points from true proportions in the held-out test dataset.
• Smoothed country–year estimates of support for democracy show both construct and convergent validity: their spatiotemporal patterns and correlations with other covariates align with prior research.

💡 Why This Matters

• Provides a principled way to recover national-level opinion trends from fragmented survey evidence.
• Enables researchers to test macro-level theories of political behavior that require consistent country–year panels, using a validated Bayesian smoothing approach.