Narrative Policy Framework

• NPF is a formal framework that decomposes policy narratives into setting, characters, plot, and moral components for empirical and computational analysis.
• It employs Bayesian networks and equilibrium models to quantify causal inferences and measure the impact of narrative framing in policy debates.
• The framework utilizes computational topic models and LLMs to detect narrative shifts, enabling robust analysis of public opinion dynamics and policy polarization.

The Narrative Policy Framework (NPF) is a formal approach for the study and operationalization of policy narratives—structured “stories” that encode how actors frame and influence public policy, debate, and opinion. Combining theoretical formalism with practical textual analysis, the NPF rigorously systematizes narrative structure along four axes—setting, characters, plot, and moral—and increasingly leverages computational models to detect and analyze narrative shifts across large-scale corpora.

1. Structural Elements and Origins of the Narrative Policy Framework

Originating in political science, the NPF provides a schema for decomposing policy-relevant discourses into empirically tractable units. Under the NPF, a narrative is defined as a bounded configuration of four components (per Shanahan et al. 2018):

• Setting: The context, domain, or background where the narrative unfolds (e.g., “during an economic crisis,” “amid a national election”).
• Characters: Actors or forces, both individual and collective, that populate the narrative (such as politicians, institutions, social groups, or abstract entities like “the market”).
• Plot: The sequence of causally or temporally linked events connecting characters, explicitly establishing relationships and dynamics.
• Moral: The underlying “lesson,” takeaway, or policy prescription, which often aligns with an explicit or implicit call to action (e.g., “This proves we must deregulate,” or “Government intervention is required”).

The NPF was developed to formalize how public discourse guides, constrains, and polarizes policy debates by structuring information into culturally resonant story-forms (Lange et al., 25 Jun 2025).

2. Formalization: Bayesian Networks as Narratives

Eliaz & Spiegler (2018) introduce a precise mathematical model for narratives within the Bayesian network (BN) formalism (Eliaz et al., 2018). In this construction:

• State space is defined over a finite set of binary random variables $X = X_1 \times \ldots \times X_n$, where $X_1$ encodes the action/policy variable ($a$), $X_n$ is the outcome ($y$), and $X_2 \ldots X_{n-1}$ denote contextual or intermediate variables.
• Narrative structure is represented by a directed acyclic graph (DAG) $R$ over a selected subset $N \subseteq \{1, ..., n\}$ including $1$ and $n$, with edges $i \to j$ denoting causal claims.
• The subjective narrative joint over $X_N$ is:

$p_R(x_N) = \prod_{i \in N} p(x_i \mid x_{R(i)})$

with $R(i)$ the parents of node $i$. Each factor references the corresponding conditional from the empirical data-generating distribution $p$.

• Beliefs induced by the narrative: The causal DAG $R$ induces specific beliefs about the effect of policy $a$ on $y$:

$$p_R(y \mid a) = \sum_{x_2, ..., x_{n-1}} \prod_{i \in N \setminus \{1\}} p(x_i \mid x_{R(i)})$$

These beliefs may deviate from the empirical $p(y \mid a)$, encoding “distorted” causal inferences consistent with the narrative.

This framework turns each narrative into an object—$(p,R)$—that can be embedded in broader models of policy choice and opinion dynamics (Eliaz et al., 2018).

3. Equilibrium Analysis: Narrative Competition and Polarization

Adopting an anticipatory utility framework, the NPF can model how agents select not just policies, but narrative-policy pairs, favoring narratives that maximize expected positive outcomes:

• Policy variable $d \in D = [\epsilon, 1-\epsilon]$ is the mixed probability of adopting $a=1$.
• Utility computation (gross anticipatory):

$$V(s, d \mid \alpha) = d\,p_R(y=1 \mid a=1) + (1-d)\,p_R(y=1 \mid a=0)$$

• Net utility subtracts intrinsic policy costs $C(d-d^*)$ (centered at a preferred $d^*$):

$$U(s, d \mid \alpha) = V(s, d \mid \alpha) - C(d - d^*)$$

• Equilibrium is a distribution $\sigma$ over narrative-policy pairs $(s,d)$ and a realized frequency $\alpha$, such that
  1. All supported pairs maximize $U(s',d'|\alpha)$.
  2. $\alpha$ matches the expected policy under $\sigma$.

A canonical example with $n=3$ clarifies that equilibrium may support two dominant, polarized narratives—one “hopeful” (optimistic about action $a$) and one “pessimistic,” each sustaining policies on opposite sides of the agent’s ideal point (Eliaz et al., 2018).

4. Coding Narratives and Mapping to NPF Axes

The formal model aligns directly with the operational NPF codebook:

• Setting $\leftrightarrow$ Variable space $\{X_i\}$ and their instantiations.
• Characters $\leftrightarrow$ Variable nodes representing actors or forces.
• Plot $\leftrightarrow$ The DAG $R$ (plot-skeleton: selection and sequencing of causality).
• Moral $\leftrightarrow$ Policy beliefs/inferences mapping through the Bayesian network.

Methodological steps for empirical NPF research include the identification of core variables, the coding of distinct narratives as DAGs, the estimation of empirical causal structure from data, and the quantification of how these structures shape policy beliefs and utility (Eliaz et al., 2018). In practice, this enables the analyst to classify and compare narratives, attribute shifts in public opinion to underlying narrative “forms,” and empirically test their influence.

5. Dynamic Detection and Computational Analysis of Narrative Shifts

Recent work demonstrates the application of computational topic models and LLMs for NPF-guided narrative shift detection (Lange et al., 25 Jun 2025):

• Dynamic Topic Modeling: RollingLDA tracks the evolution of topic distributions $\phi_{k,t}$ and $\theta_{d,t}$ across time, integrating a memory parameter to ensure temporal coherence and capture abrupt topical change points.
• Change-Point Detection: Cosine similarity between topical word vectors across time windows, combined with bootstrap-based significance tests, identifies statistically significant topic changes.
• Document Selection: High-impact words (identified via leave-one-out impact) guide the extraction of representative documents exemplifying each shift.
• LLM-Based Interpretation: Selected documents are processed by a LLM (e.g., Llama 3.1 8B) prompted according to the canonical NPF definition—the model must extract the setting, characters, plot, and moral before confirming a narrative (vs. content-only) shift.
• Distinguishing Narrative from Content Change: The classification requires satisfaction of all four NPF structural criteria. Empirical results show that, while LLMs can efficiently extract narrative changes when present, they tend to overidentify narrative shifts due to explanations that fit the NPF schema even when shifts are merely topical or factual (Lange et al., 25 Jun 2025).

Table: Elements and Operationalization in NPF-based Analysis

NPF Axis	Mathematical/Formal Mapping	Empirical/Computational Instantiation
Setting	Variable selection $\{X_i\}$	Domains/topics in text; contextual features
Characters	Nodes in DAG $R$	Explicit actors/forces in discourse
Plot	Edges and topology of DAG $R$	Causal/temporal language; topic co-variation
Moral	Inferred $p_R(y|a)$ via BN	Policy recommendation or judgmental frame

6. Empirical Results, Limitations, and Methodological Guidance

Empirical application to corpora such as The Wall Street Journal (2009–2023) shows that the NPF, operationalized through RollingLDA and LLM-guided narrative extraction, can systematically distinguish between content and narrative shifts when:

• Change-point detection parameters (memory, look-back, significance thresholds) are judiciously chosen.
• LLM prompts strictly enforce the requirement to extract all four NPF narrative components.
• Human annotation confirms that narrative shifts correspond to simultaneous change in setting, characters, plot, and moral axes.

Observed LLM classification achieves $57.4\%$ accuracy and $F_1 = 0.7010$ on narrative shift detection, with errors primarily due to overgeneration—attributable to LLMs' tendency to impose narrative structure on factual event changes (Lange et al., 25 Jun 2025).

The NPF’s formalization as Bayesian networks allows for quantifiable filtering of “perfect” vs. “imperfect” narratives, facilitating the empirical isolation of story-forms that genuinely drive public opinion dynamics and policy polarization (Eliaz et al., 2018). Theoretical and empirical methodologies converge to offer a framework for linking narrative construction, public sentiment, and policy outcomes in a computationally tractable manner.