Data-Driven Modeling of U.S. Ideological Dynamics

TL;DR

The paper tackles how political opinions evolve in the United States by embedding survey-derived measures of dissonance ($d = p - g$) and exposure into a one-dimensional, bounded ideological framework. It constructs reaction and exposure surfaces from high-resolution data and develops a family of drift-based dynamical models in a domain $g \in [-1,1]$, incorporating noise and optional tribal biases. A simple model $g' = a d (1-g^2)$ suggests polarizing tendencies, but augmentation with saturating dissonance, centralizing forces, tribalism, and party cohesion yields equilibria that better reflect observed distributions, highlighting a theory–experiment loop for refining ideological dynamics. The work aims to predict polarization risk and inform interventions to sustain constructive political discourse, while openly acknowledging data limitations and model freedoms that invite further empirical validation.

Abstract

The dynamics of political opinion are a critical component of modern society with large-scale implications for the evolution of intra- and international political discourse and policy. Here we utilize recent high-resolution survey data to quantitatively capture leading-order psychological and information-environmental patterns. We then inform simulations of a theoretical dynamical framework with several different models for how populations' ideology evolves over time, including a model which reproduces current macro-scale ideological distributions given the empirical micro-scale data gathered. This effort represents an attempt to discover true underlying trends of political reasoning in general audiences, and to extrapolate the long-term implications of those trends as they interact with the political exposure landscape. Accurate modeling of this ecosystem has the potential to predict catastrophic outcomes such as hyperpolarization, and to inform effective intervention strategies aimed at preserving and rebuilding constructive political communication.

Figures (12)

• Figure 1: Agreement Curve: Agreement versus subjectively experienced dissonance for 23,736 statement-observer events. The black curve is a moving median (with window width of 7), with 25$^{th}$ and 75$^{th}$ percentiles (dotted). Horizontal striation occurs near multiples of ten (visible from survey software), not labels on the slider; there were seven labels per slider, roughly corresponding to positions $0, \pm16, \pm32,$ and $\pm48$.
• Figure 2: Linear Agreement Model: As Fig. \ref{['fig:reaction_scatter']}, but with the data "folded" so that left and right halves are overlaid, and a least-squares linear fit: $a = 22.46 - 0.78d$. This fit appears to overestimate the "spine" of the data for high distances, due to both edge effects and non-Gaussian-distributed errors. This fit was found by restricting the fitting data to $d<80$, which was the steepest result found for any such data-cutoff.
• Figure 3: Reaction Probability Surface: Heat-map view of a quasi-kernel-density surface generated by "fuzzing" each data point from Fig. \ref{['fig:reaction_scatter']} into a two-dimensional Gaussian distribution with standard deviations $\sigma_x = \sigma_y = 7$ in the $x$ and $y$ directions, to reflect the inherent uncertainty in each quantity, then normalized so that each vertical slice yields a probability distribution of agreement for that level of dissonance.
• Figure 4: Reaction Probability Distributions: Three particular vertical slices of Fig. \ref{['fig:reaction_surface']}, showing the implied probability distributions for each level of dissonance on the $[-50,50]$ scale: 0 (in line with the observer's self-identified ideology), +65 (much more conservative than the observer) and -35 (significantly more liberal, just past the uncertainty boundary of $\approx\pm 30$).
• Figure 5: Difficulty of quantifying tribalism: Moving-median agreement curves for a) in-group percepts ($N = 4054$) and b) out-group percepts ($N= 4025$). Where the window contains fewer than 20 data points, the median and percentile curves are omitted. We see the challenge of isolating the effect of tribalism, due to the strong correlation between ideology and party: in-group statements tended to be low-dissonance and out-group statements the opposite. In addition, sample sizes are significantly smaller due to subdividing: participants must be assigned the marked condition (50%), identify as Strong or Lean towards one of the parties (77%), and view a statement from one of the non-Centrist statement pools ($30/68 = 44\%$ each), resulting in the use of only about 17% of our original dataset to inform each panel.