Technical Summary: Function-Valued Causal Influence in Nonlinear Time Series

1. Problem Statement

The paper addresses a fundamental representational gap in nonlinear causal discovery for time series. While modern machine learning models (e.g., Neural Additive Vector Autoregression) successfully learn rich, state-dependent representations of temporal dependence, the standard practice for interpreting these models is to summarize causal relationships using scalar edge scores.

The authors argue that reducing complex, state-dependent causal functions to single scalar magnitudes constitutes a severe information bottleneck. This practice:

• Obscures functional structure: It collapses essential dynamics such as thresholds, saturation, asymmetry, and sign changes.
• Conflates variation with noise: Scalar scores (typically standard deviations of contributions) upper-bound the variance of the true function-valued influence, conflating between-state variation with within-state residual noise.
• Hides regime dependence: Relationships that are strong only in specific regimes (e.g., high vs. low democracy) may receive identical scores to weak, globally active relationships, leading to misleading policy or theoretical interpretations.

2. Methodology

2.1 Theoretical Formalization

The authors formalize function-valued causal influence for additive, contribution-decomposable autoregressive models.

• Model Class: They focus on models where the prediction $\hat{X}^{(j)}_t$ is a sum of source-specific functions: $\hat{X}^{(j)}_t = \beta_j + \sum_i f_{ij}(X^{(i)}_{t-K:t-1})$.
• The Causal Object: Instead of a scalar, the causal influence of variable $i$ on $j$ is defined as the conditional expectation of the contribution function: $g_{ij}(x) = E[\text{contrib}_{ij,t} \mid X^{(i)}_{t-1} = x]$.
• Scalar Limitation: They demonstrate mathematically that a scalar score $S_{ij}$ (defined as the standard deviation of contributions) satisfies $S_{ij}^2 \geq \text{Var}(g_{ij}(X^{(i)}_{t-1}))$. The gap represents the expected within-state residual variance, proving that scalar scores are low-dimensional projections that discard functional information.

2.2 Estimation Framework: Lag-Aggregated ICE

To estimate these functions directly from trained models without retraining, the authors propose an intervention-style framework based on Individual Conditional Expectation (ICE):

• Intervention: For a trained model, the lagged history of a source variable $i$ is replaced with a fixed value $x$ (or a sequence of values), while all other inputs remain unchanged.
• Lag-Aggregation: To obtain a variable-level response rather than a lag-specific one, the authors intervene on all lags of the source variable simultaneously ($X^{(i)}_{t-1} = \dots = X^{(i)}_{t-K} = x$).
• Estimator: The function-valued influence is estimated as the average change in the target prediction over time windows:
  $$\hat{g}_{ij}(x) = E_t \left[ \hat{X}^{(j)}_t(X^{(i)}_{t-1:t-K} \leftarrow x) - \hat{X}^{(j)}_t(X_{t-K:t-1}) \right]$$
• Regime-Conditioning: The framework extends to estimate $g_{ij}(x \mid r)$ by partitioning data into regime bins (e.g., low/medium/high levels of the target variable), revealing state-specific activation patterns.

2.3 Experimental Setup

• Synthetic Systems: A controlled three-variable system ($X \to Y \to Z$) was generated where $X$ influences $Y$ via one of four distinct mechanisms: Linear, Thresholded, Saturating, and Sign-Changing. All mechanisms were calibrated to have comparable overall scalar scores.
• Real-World Application: A panel dataset of 139 countries over 35 years (Varieties of Democracy project) was analyzed using Neural Additive Vector Autoregression (NAVAR). The study focused on the causal influence of institutional factors (e.g., judicial constraints, freedom of expression) on "clean elections."

3. Key Results

3.1 Synthetic Experiments: The Scalar Bottleneck

• Indistinguishable Scores: Despite fundamentally different underlying mechanisms, the four synthetic systems produced scalar causal scores that were tightly clustered (means ranging from 0.626 to 0.650 with overlapping standard deviations).
• Divergent Functions: While scalar scores were identical, the estimated ICE curves revealed qualitatively distinct behaviors:
  • Linear: Monotonic response.
  • Thresholded: Flat near zero, sharp activation after a critical point.
  • Saturating: Diminishing marginal returns at high magnitudes.
  • Sign-Changing: Negative effect near the origin, positive in the tails.
• Recovery Quality: The ICE estimator successfully recovered the true functional forms with high fidelity (Pearson correlations $r \geq 0.968$ across 15 runs), confirming that the information is present in the model but lost in scalar summarization.

3.2 Democratic Development Case Study

• Scalar Graph: The scalar causal graph suggested a dense network of weak, uniform influences between democratic institutions.
• Function-Valued Insights: ICE analysis revealed substantial heterogeneity invisible to scalar scores:
  • Threshold Effects: The influence of "freedom of expression" and "judicial constraints" on election quality was negligible in low-regime contexts but became strongly positive only after crossing a moderate institutional baseline.
  • Asymmetry & Saturation: "Legislative constraints" showed strong gains from low to moderate levels but saturated at high levels.
  • Regime Reversal: In a broader survey of 50 edges, 98% exhibited "regime reversal," where the direction of causal influence flipped between low- and high-democracy contexts.
• Conclusion: Scalar summaries systematically missed these regime-specific and asymmetric structures, which are critical for substantive political science interpretation.

4. Key Contributions

1. Formalization of Function-Valued Influence: The paper defines causal influence in nonlinear time series not as a scalar magnitude but as a state-dependent function $g_{ij}(x)$, formalizing the limitations of scalar aggregation via the law of total variance.
2. Practical Estimation Framework: It introduces a computationally tractable, intervention-style method (Lag-Aggregated ICE) to recover these functions directly from trained additive autoregressive models without architectural changes.
3. Empirical Demonstration of Representational Collapse: Through synthetic and real-world experiments, the authors demonstrate that edges with indistinguishable scalar scores can encode radically different causal mechanisms, and that scalar summaries systematically obscure regime-dependent behaviors (thresholds, saturation, sign changes).
4. Application to Social Science: The paper provides a concrete example in democratic development where function-valued analysis reveals policy-relevant mechanisms (e.g., minimum institutional baselines) that scalar approaches fail to detect.

5. Significance and Claims

The paper claims that the opacity often attributed to nonlinear causal models stems less from the models themselves and more from how their outputs are summarized.

• Reframing Causal Discovery: The authors argue that causal discovery should be viewed as a problem of representation. By retaining the function-valued structure learned by models, researchers can recover interpretable, regime-specific mechanisms.
• Limitations of Scalar Scores: The paper asserts that scalar scores are insufficient as primary summaries for nonlinear causal structures. They are useful for coarse screening but fail to capture how causality operates (e.g., when effects activate, saturate, or reverse).
• Policy and Theory Implications: In theory-driven domains like the social sciences, function-valued analysis is presented as essential for distinguishing between mechanisms that require minimum thresholds for intervention versus those with uniform marginal returns.
• Modesty: The authors explicitly state that their approach does not claim to identify structural or interventional causality in the counterfactual sense. The analysis remains within the predictive, Granger-style framework, characterizing what the model learns about directed predictive relationships. They also note that the method relies on additive, contribution-decomposable models, and extending this to non-additive architectures remains future work.