State Fiscal Response to Class Conflict: An Agent-Based Model

Author: Asiya Kadivar Institution: Carnegie Mellon University Status: Work in progress (private) — data pipeline validated, response function calibrated (v1), ABM implemented and running (Worker/Firm/State agents, three regimes), first full experiment suite complete. Two known bugs from that first run have been fixed; results below are post-fix. Timeline: July–October 2026 Target Venue: Journal of Artificial Societies and Social Simulation (JASSS) or heterodox economics outlets

Note to anyone encountering this repo mid-development: the results below are a first post-fix pass, not a final finding. Several open questions are flagged explicitly in "Open Questions & Next Steps" — please read that section before citing any number here. This is an active research project, not a finished result.

---

Overview

This project models the US state's fiscal policy response to labor market conditions and conflict intensity using agent-based modeling (ABM). The core innovation: instead of treating fiscal policy (taxation, redistribution) as exogenous or optimizing in a rational-expectations framework, we model the state as a reactive agent that observes economic indicators (inequality, unemployment, protest) and adjusts redistributive spending accordingly.

Central research question: Is state fiscal response to class conflict structural (persistent policy reorientation) or cyclical (temporary pacification)? And what state regime (representative, captured, dictatorship) best explains observed US policy trajectories?

This goal hasn't changed. What's changed since the original draft: the data pipeline (several real bugs fixed), the fitting methodology (more honest about uncertainty and model flexibility), a working ABM with three selectable regimes, a full experiment suite (historical validation, regime comparison, reform-vs-revolution test, sensitivity sweep, permutation null check), and two data-integrity bugs caught and fixed in that suite's first output.

Theoretical Motivation

• Epstein (2002): Civil violence models show how state repression + grievance interact at the micro level
• Acemoglu & Robinson (2000): Redistribution as political equilibrium—elites tolerate it to avoid uprising
• Piketty et al. (2014, 2015): Tax policy responds to political pressure and distributional conflict
• Alesina & Rodrik (1994): Tradeoff between redistribution and growth not mechanically determined—policy choice matters

Gap: No ABM has endogenized state fiscal response to endogenous worker grievance + protest dynamics. Most models treat policy as exogenous or assume rational welfare maximization.

Ideological Stakes

Two interpretations of results:

1. Reform optimism: If redistribution is sticky (ratchets up, persists), suggests genuine policy gains possible via democratic struggle
2. Revolutionary critique: If redistribution is cyclical (state reverses gains when conflict subsides), suggests structural oppression requires systemic change, not reform

We present both without endorsing—the empirics speak. Caveat as of this draft: protest is currently rare-to-near-zero across all three regimes post-fix, which limits how cleanly the reform-vs-revolution test can distinguish these two stories right now — see "Current Results" and "Protest Signal" below.

---

Model Architecture

Agents

Workers (N ≈ 1000–5000)

• Heterogeneous skill, wage, network position
• Grievance = f(wage gap relative to reference point, inequality trend)
• Protest threshold: if (Grievance − Risk) > Threshold, join active protest
• Risk = f(state repression capacity, network visibility)
• Capital-income channel: savings above a consumption floor can be invested (heterogeneous per-agent propensity), earning a stochastic return

Firms (N ≈ 50–200)

• Heterogeneous productivity
• Wage-setting: mechanical response to labor market tightness, not strategic suppression
• Hiring/firing follows profit dynamics (LIFO)
• Owner income + owner investment income feed into the population income Gini alongside worker income

State (N = 1, endogenous decision-maker)

• Observes: Gini(t), unemployment(t), protest intensity(t), wage growth(t)
• Decides: tax rate τ(t), redistribution R(t)
• Response function: calibrated empirically (see Methodology) rather than assumed
• Regimes (implemented): representative, captured (biased toward elite/firm preferences over popular grievance), dictatorship (repression-weighted, largely ignores conflict signal). All three are wired into the ABM and used in regime-comparison experiments below.

Market Dynamics

• Labor clearing: Wages adjust until supply ≈ demand (with frictions)
• Feedback loops:
  • High unemployment → low wages → high relative grievance → high protest
  • High protest → state increases redistribution → worker consumption → demand → hiring
  • Inequality widens → grievance increases independent of absolute wages

---

Data Pipeline (updated)

Source data: FRED (automated pulls), Census/CPS, Tax Foundation, CNTSDA/BLS strike data, SSA/USDA/HHS program spending. Full source list unchanged from original design — see sources.txt.

Fixes applied since first draft:

• gdp_billions was pulling A191RL1Q225SBEA (% change from preceding period) instead of a dollar level. Corrected to FRED series GDP (Billions of Dollars, SAAR, quarterly).
• Annual source files (Gini, EITC, federal tax rate, Medicaid) were being misparsed: bare-year integer columns (1970, 1971, ...) were passed straight into pd.to_datetime(), which silently interprets raw integers as nanosecond epoch offsets rather than years. Fixed at the date-parsing step; parameter override removed so is_fiscal_year is respected per call site.
• protest_intensity_score was structurally stuck near 0.5 for the whole series. Rebuilt to normalize days_idle against total available person-days per quarter, log-transform both components, and min-max rescale — now spans a real 0–1 range.

Annual vs. Quarterly (design decision, not a limitation to hide)

The state's fiscal decisions (redistribution, Gini, tax rate) are sourced from genuinely annual data; quarterly figures in the raw CSV are forward-filled duplicates for calendar alignment, not four independent observations. Current approach: collapse to one observation per year for fitting the response function, while aggregating quarterly-resolution predictors (protest, GDP growth) within each year first.

---

State Response Function — Methodology (v1)

$$R(t) = f(\text{Protest}(t-\tau), \text{Gini}(t), \text{Growth}(t))$$

We fit a family of candidates — linear, logistic (saturating), exponential, and a GAM — and compare via AIC/BIC rather than assuming which shape is "right." The exponential form wins on AIC (n=63 annual observations, 1963–2025); GAM has higher raw R² but its flexibility is penalized appropriately given sample size. Durbin-Watson on the best model is 0.79, indicating residual autocorrelation typical of a multi-decade macro series — point estimates are usable, but standard/CI estimates from this fit may be too tight.

Protest Signal (open question, not yet resolved)

Across every functional form tried, the protest coefficient's confidence interval crosses zero, including under HAC correction. A permutation null check on the historical d_protest(t-1) → d_redistribution(t) relationship (2000 shuffles) confirms this isn't a modeling artifact: observed correlation = −0.0192, p = 0.87 — indistinguishable from chance. Two live hypotheses, not yet distinguished:

1. The effect is genuinely weak/absent at this level of aggregation.
2. It's a measurement problem — annual-only, strike-only (CNTSDA) severity data for large stretches of the series.

---

Current Results (first post-fix experiment suite)

Two bugs were caught in the first full run of experiments.py and are fixed as of this update:

• Gini out-of-range values (means/CIs above 1.0, e.g. 2.68): caused by unfloored negative investment/owner income feeding the Gini calculation. Fixed by clipping the income vector at 0 before computing Gini (not by altering the underlying wealth/investment mechanics, which are left intact) plus an assertion that Gini stays in [0,1].
• Dictatorship reform-vs-revolution test was invalid: with protest near-zero for that regime, the 90th-percentile "spike" threshold itself computed to 0, so the spike condition matched every tick rather than real outliers, producing a fabricated monotonic "response" curve. Fixed by requiring the threshold to be strictly positive and the tick itself nonzero; dictatorship now correctly reports 0 usable spike events rather than a fake trajectory.

Post-fix findings:

• Regime-comparison Gini is now bounded and plausible (0.471–0.476 across regimes), consistent with real US income-Gini figures (~0.47–0.49).
• Protest share collapsed to near-zero across all three regimes (0.0000–0.0041), down from ~0.20–0.24 pre-fix. This tracks: the old out-of-range Gini was driving macro_reference = avg_wage × (1 − gini) negative, artificially inflating worker grievance. With Gini fixed, protest now reflects the model's actual (currently low) grievance-risk gap.
• Regime differentiation in Gini is small and CIs largely overlap (captured 0.4709 vs. representative 0.4760) — theoretically we'd expect captured regimes to produce visibly worse inequality; this convergence is a genuine open question, not settled.
• Reform-vs-revolution event study is no longer degenerate for dictatorship (0 events, correctly reported as insufficient data) but is now event-count-thin for captured/representative too (463 and 388 pooled spikes off a ~0.1–0.4% protest base) — treat those curves as suggestive, not conclusive, until protest calibration is revisited.
• Permutation null check remains solid and consistent pre- and post-fix (p = 0.87).

---

Open Questions & Next Steps

1. Protest signal (measurement vs. real absence): improve resolution before drawing conclusions — recover quarterly granularity for post-1988 strike data, and evaluate swapping/supplementing CNTSDA with ACLED for broader, more recent event coverage.
2. Near-zero post-fix protest — calibration sanity check (new): confirm whether the current near-zero protest share is a correct reflection of the fixed grievance dynamics, or whether the logistic protest-decision steepness/threshold parameters need revisiting now that the earlier (buggy) grievance inflation is gone. Needed before leaning on the reform-vs-revolution event counts above.
3. Regime differentiation in Gini — sanity check (new): verify that regime actually drives income/wealth dynamics and not only protest suppression (via police_intensity/risk perception). If captured vs. representative Gini stays this close after the protest-calibration check above, that's a substantive finding worth investigating in the regime response-function weighting itself, not just a threshold artifact.
4. Period-by-period breakdown: the current fit assumes one fixed relationship across the full 1960–2025 span. Rolling-window or regime-split fits (e.g. pre/post-1980) rather than assumed stability.
5. Indirect / compositional correlation: test whether protest correlates with shifts in spending composition (e.g. EITC up while Medicaid flat) or with volatility of spending, rather than only its level.
6. Full per-permutation BVAR refit for the null check: current version permutes a single lagged correlation rather than refitting the full shrinkage system each permutation — a real but smaller-scope check; a full refit would be more rigorous.

---

Data Sources & Rationale

(Unchanged from original — automated FRED pulls for unemployment, LFPR, wages, job openings, GDP, UI/workers' comp/veterans benefits/Medicaid; manually collected Gini, wage percentiles, SNAP, EITC, federal tax rate, strikes/protests. See sources.txt for full listing.)

---

Mathematical Foundations

Worker Grievance

$$G_i(t) = \frac{\text{reference_wage}(t) - w_i(t)}{\text{reference_wage}(t)}$$

where reference wage = local average wage × (1 − Gini).

Worker Protest Decision

Probabilistic (logistic), not a hard cutoff — see worker.py for the full form:

$$p_i(t) = \sigma\big(k \cdot [(G_i(t) - R_i(t)) - \theta_i]\big)$$

Source: Epstein (2002) civil violence model; adapted for labor context.

State Response Function

$$R(t) = \beta_0 + \beta_1 \cdot \text{Protest}(t - \tau) + \beta_2 \cdot \text{Gini}(t) + \beta_3 \cdot \text{Growth}(t) + \epsilon(t)$$

Fit via multiple functional forms (linear, logistic, exponential, GAM), compared on AIC/BIC with HAC-corrected uncertainty.

---

Known Limitations & Caveats

1. Firm wage-setting: mechanical, not strategic. (Mitigation: sensitivity test with mild monopsony — included in the lightweight sweep above; more exhaustive sweep still open.)
2. State as single agent (partially addressed): three regime types now implemented, but this still doesn't fully model Treasury/Congress/Fed contestation.
3. Protest data resolution: see "Protest Signal" and Open Question #1 above — actively being worked on, not a settled limitation.
4. Causality unclear: does redistribution cause lower protest, or do high-protest periods select into high redistribution? ABM assumes reaction; history is endogenous.
5. Demand-side omitted: no consumption dynamics, credit, or asset bubbles modeled. Scope boundary, not a bug.
6. Sample size: n≈63 annual observations for a 3–4 parameter response-function fit is workable but not large; treat current coefficients as a first pass.
7. Reform-vs-revolution event counts are currently thin given near-zero post-fix protest (see Open Question #2) — don't over-read the current spike-response curves until that's resolved.

---

Ideological Transparency

This project is grounded in heterodox political economy (Marxist-adjacent but not doctrinaire). We test whether capitalism naturally stabilizes via redistribution (reform) or requires systemic change (revolution).

We are not neutral on these questions, but the model is. Results will speak for themselves — which is also why the honest-uncertainty reporting above (wide CIs, unresolved protest signal, thin post-fix event counts) is being kept in this README rather than smoothed over.

---

References (Preliminary)

• Acemoglu, D., & Robinson, J. A. (2000). Political losers as a barrier to economic development. The American Economic Review, 90(2), 126–130.
• Alesina, A., & Rodrik, D. (1994). Distributive politics and economic growth. The Quarterly Journal of Economics, 109(2), 465–490.
• Epstein, J. M. (2002). Modeling civil violence: An agent-based computational approach. Proceedings of the National Academy of Sciences, 99(Suppl 3), 7243–7250.
• Piketty, T., Saez, E., & Stantcheva, S. (2014). Optimal taxation of top labor incomes. Journal of Political Economy, 122(2), 231–271.
• Manning, A. (2011). Imperfect competition and macroeconomics. The Economic Journal, 121(554), 45–65.

---

Repository Structure

state_response_abm/
├── README.md (this file)
├── us_state_response_data.py (data collection — GDP + date-parsing fixes applied)
├── state.py (data aggregation + response function fitting + State agent)
├── firm.py / worker.py (ABM agents, regime-aware)
├── abm_model.py (Mesa ABM implementation — done; Gini clipping fix applied)
├── experiments.py (historical validation, regime comparison, reform-vs-revolution,
│                    sensitivity sweep, permutation null check — done;
│                    spike-detection threshold fix applied)
├── data/
│   ├── us_state_response_data.csv (raw quarterly output)
│   └── manual_collections/ (wage percentiles, Gini, protest data)
└── results/
    ├── plots/ (calibration fits, time series) [TO DO]
    └── experiments/ (regime comparison outputs, first post-fix run complete)

---

Last updated: August 2026 Status: ABM implemented; first full post-fix experiment suite complete. Protest-calibration sanity check and regime-Gini differentiation are the top open items before drawing paper-level conclusions.