Polarization is an emergent outcome of political opinion dynamics, yet the causal processes linking individual-level opinion updates to collective consensus remain poorly understood. To address this, we conceptualize political discourse as hierarchical coarse-graining where individual-level opinions are compressed and aggregated into group- and system-level representations. In this framework, upward causation refers to individual updates influencing group-level opinion, while downward causation captures how dominant collective views shape individual opinion updating. Using the Continuous Opinions and Discrete Actions (CODA) model, we simulate Bayesian belief updating, social identity–driven information integration, and migration between identity groups to capture coarse-graining across micro-, meso-, and macroscales. Our results show that connectivity between identity groups shapes overall opinion dynamics, producing three distinct regimes: independent (local convergence), parallel (rapid local and global convergence), and iterative (slow global convergence). In the iterative regime, low connectivity fosters transient diversity, reflecting a more informed consensus. Across all regimes, time-scale separation gives rise to downward causation that ultimately drives agents toward consensus. Critically, any coherent connectivity between identity groups leads to downward causation. These findings highlight the key role of downward causation, identity group structure, and inter-group connectivity in shaping the dynamics underlying polarized political discourse.