## wpiea2026043-source-pdf ## Source details **Canonical URL:** [wpiea2026043-source-pdf](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026043-source-pdf.pdf) ## Other formats - [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026043-source-pdf.pdf.md) - [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026043-source-pdf.pdf.json) --- ### Construction of the AI-narrative database - Data source and coverage - Primary narrative source: Economist Intelligence Unit (EIU) Country Reports. - Empirical focus: subset of 64 advanced and developing economies. - Sample period (unbalanced): 1952:Q1–2023:Q4. - Archive constructed quarterly with one report per country–quarter; when multiple reports in a quarter, retain the chronologically last comprehensive report. - Definition of exogenous actions (Romer and Romer (2010) analogue) - Exogenous action criteria: motive unrelated to contemporaneous macroeconomic conditions (e.g., medium-term consolidation targets, long-run ideological/structural objectives, compliance with law/treaty). - Actions responding to recession, unemployment, overheating, inflation, interest-rate moves, exchange-rate pressure, or financing stress treated as endogenous. - Acknowledging current conditions does not imply endogeneity if stated motive is explicitly non-cyclical. - NLP coding pipeline (fixed GPT–4.1 prompt) - Prompt returns per report: 1. EXOG_NET_SPENDING ∈ {EXP_, CON_, NEUTRAL, UNCLEAR} (optionally append small/large) 2. MOTIVATION: brief phrase 3. COMPONENT (if specified) 4. INTENSITY_ORDINAL ∈ {-10, .. . , 10} (optional) 5. CONFIDENCE (optional) - Aggregation to country–quarter proxy z_{i,t} ∈ {−1, 0, +1}: - +1 if EXOG_NET_SPENDING = EXP_ - −1 if EXOG_NET_SPENDING = CON_ - 0 if EXOG_NET_SPENDING ∈ {NEUTRAL, UNCLEAR} - z_{i,t} interpreted as qualitative proxy for discretionary spending actions (may combine announcement and implementation elements). - Coverage and composition (from 16,029 country-quarter observations) - Nonzero discretionary spending episodes: 8,636 (53.9% of the sample). - Expansions: 4,374; Contractions: 4,262. - Expansions account for 50.6% of all non-zero observations. - Nonzero quarters (z_{i,t} = ±1): 8,636. - Cross-country heterogeneity (nonzero quarters by IMF income group) - Advanced: Nonzero quarters 3,129; Expansion share (%) 45.8 - Emerging: Nonzero quarters 3,587; Expansion share (%) 49.5 - LIC: Nonzero quarters 1,920; Expansion share (%) 60.7 - Motivations by action type (random subsample text-dictionary) - Expansionary actions: nearly 60 percent associated with increases in public investment (infrastructure and capital projects). - Contractionary actions: over two-thirds reflect explicit consolidation measures (spending restraint and deficit reduction). ### Documentation: narrative examples and stylized evidence - Selected classified episodes (representative fragments) - Bangladesh 1980Q2 +1: “five-year plan envisages Tk 300–500 bn in public programs” - Honduras 1978Q1 +1: “US$450 mn over five years in electric-power projects” - Côte d’ Ivoire 1970Q2 +1: “special capital budget; Canadian-supported electrification” - Ghana 2010Q3 +1: “Parliament approves large public-housing program (∼US$10 bn)” - Indonesia 1985Q1 −1: “real development spending cut by about 7%” - Ecuador 1982Q4 −1: “removal of petrol and imported-wheat subsidies” - Example confidence statements in coding outputs - Ghana, 2010Q3: Confidence: 95 percent. - Ecuador, 1982Q4: Confidence: 90 percent. - United Kingdom, 2005Q4: AI assigns exogenous effect as unclear. ### Quarterly macroeconomic and fiscal series - Variables assembled for VAR analysis: - Real GDP (output) and real government expenditure from EIU, Haver, CEIC, and central banks. - Variables expressed in log levels for VARs. - Narrative fiscal impulse series: - Includes only exogenous shocks; endogenous quarters zeroed out. - Final panel dataset = (i) narrative fiscal impulse series (exogenous shocks) + (ii) macroeconomic series for output and other variables. - All subsequent results use intersection of narrative series and available quarterly macro data. ### Validation overview (four complementary assessments) - Validation objectives: 1. Reproduce Romer and Romer (2019) report-level classifications. 2. Replicability of fixed prompt across runs. 3. Empirical mapping: whether proxy predicts measured government spending. 4. External alignment with Adler et al. (2024). --- ### Benchmarking narrative interpretation: comparison to Romer and Romer (2019) - Benchmark design - Used same 200 EIU reports as RR19 for 19 countries. - Model asked the same four RR19 questions; evaluated outputs corresponding to stance and motivation. - Classification taxonomy - Stance labels: Expansionary, Contractionary, Neutral, Unclear. - Motivation taxonomy (RR19 categories 1–9); motives (1)–(3) pertain to expansion, (4)–(8) to contraction. - Validation approach - Panel (a): match on net stance. - Panel (b): match on motivation category (first mentioned). - Agreement statistics (report level) - Expansions: - AI matches the stance in 14 out of 16 cases (87.5%). - AI matches the motivation in 32 out of 34 cases (94.1%). - Including four boundary cases reduces expansion stance match to 14/20 (70%). - Contractions: - AI matches the direction in 18 out of 19 cases (95%). - AI matches the motivations in 44 out of 45 cases (97.8%). - Diagnostic interpretation - Many disagreements arise from reports describing concurrent expansionary and contractionary measures; aggregation rule conservative. - Identified model errors include missed political cue in Korea and mis-sign inference in a United States entry; Finland near-miss in mapping Market Access motive. ### Replicability (51 reruns on 20 reports) - Findings: - Direction of fiscal stance: perfectly stable — for all 20 reports modal share = 1 (identical sign in every run). - Motivation: high replicability with occasional variation. - Intensity and confidence: lower replicability; non-trivial dispersion across runs. - Interpretation: - Coarse classifications (direction) robust; finer quantitative fields more sensitive. ### Correlation with government spending data (local projections, from 3.3) - Local-projection specification: - lnG_{i,t+h} = β^+_h 1{z_{i,t}=+1} + β^-_h 1{z_{i,t}=-1} + Σ_{k=1}^4 [ φ_{h,k} lnG_{i,t-k} + ψ_{h,k} lnY_{i,t-k} + γ_{h,k} z_{i,t-k} ] + α_i + λ_t + ε_{i,t+h}, for h = 0,...,8. - Impact and cumulative estimates (exact numeric statements preserved): - At impact (h = 0), contractions reduce logG by roughly −0.005 (log points; 0.5% level). - Expansions (+1) rise one quarter later with β^+ ≈ 0.002. - Cumulatively, contractions reach about −0.06 by h≈ 8. - Cumulatively, expansions reach about 0.04 by h≈ 8. --- ### External validation against Adler et al. (2024) - Adler et al. (2024) dataset - Covers 31 advanced and emerging economies; reports announced size of discretionary fiscal consolidation packages (% of GDP). - Coverage begins 1978 for advanced economies and 1989 for Latin America and Caribbean, extends through 2020. - Expenditure component denoted spend_c,y: positive values indicate expenditure cuts or restraint (fiscal tightening); negative values indicate net loosening. - Constructed annual proxies from quarterly z_{i,t}: - consol_quarters_c,y = Σ_{q=1}^4 1{indicator_c,y,q = −1} - expand_quarters_c,y = Σ_{q=1}^4 1{indicator_c,y,q = +1} - any_consol_c,y = 1{consol_quarters_c,y ≥ 1} - any_expand_c,y = 1{expand_quarters_c,y ≥ 1} - net_stance_c,y = (consol_quarters_c,y − expand_quarters_c,y)/4 ∈ [−1, +1] - Test 1 — Extensive-margin alignment (hit and expansion rates) - AE_consol_c,y(τ) = 1{spend_c,y ≥ τ} for τ ∈ {0.5, 1.0, 2.0}. - Full merged sample results (Panel A): - τ = 0.5: - N_big = 138 - Hit rate = 0.964 - Expansion rate = 0.246 - N_loose = 1 - Exp. hit rate = 1.000 - Consolidation rate (any_consol = 1) = 0.000 - τ = 1.0: - N_big = 65 - Hit rate = 0.969 - Expansion rate = 0.231 - τ = 2.0: - N_big = 16 - Hit rate = 1.000 - Expansion rate = 0.250 - Panel B (by country group) highlights: - LAC τ = 0.5: Hit rate = 1.000; Expansion rate = 0.296; N_big = 27 - Advanced τ = 0.5: Hit rate = 0.955; Expansion rate = 0.234; N_big = 111 - Additional observation: 23–30 percent of “large consolidation” years also contain at least one expansion quarter in the quarterly series. - Test 2 — Probability of consolidation year (linear probability model) - 1{spend_c,y ≥ 0.5} = α_c + β_1 any_consol_c,y + β_2 any_expand_c,y + ε_c,y - Results (Table 6, column 1): - Coefficient on any_consol_c,y = 0.126 with clustered standard error 0.019 (statistically significant). - Coefficient on any_expand_c,y = −0.073 with clustered standard error 0.020 (statistically significant). - Test 3 — Intensity alignment (continuous regression) - spend_c,y = γ_c + δ net_stance_c,y + u_c,y - Results (Table 6, column 2): - δ = 0.263 with clustered standard error 0.044 (statistically significant). - Interpretation: moving from net_stance = −1 to net_stance = +1 associated with increase ≈ 2 × 0.263 ≈ 0.5 percentage points of GDP in spend_c,y. - Predictability comparisons - Quarterly predictability regressions: only robust predictor of z_{i,t} is its own lag. - Full sample: lagged coefficient = 0.266 ∗∗∗; R^2 = 0.07; constant = −0.0775 ∗∗∗. - Guajardo sample: lagged = 0.308 ∗∗∗; R^2 = 0.10; constant = −0.103 ∗∗. - Adler sample: lagged = 0.298 ∗∗∗; R^2 = 0.09; constant = −0.0915 ∗∗. - Annual aggregation increases predictability: baseline annual aggregation R^2 = 0.125 with GDP Growth (lagged) = 6.030 ∗∗∗ and Expenditure shock (lagged) = 0.324 ∗∗∗. - Overall validation conclusions - Extensive margin: hit rates above 0.95 when Adler reports sizable consolidation. - Intensive margin: δ = 0.263 (s.e. 0.044) links net_stance to spend_c,y. - Predictability: quarterly shocks less predictable from macro lags (R^2 ≤ 0.1) than annual shocks. --- ### Econometric identification strategy (section 4.2) - Identification approach - Use narrative series z_t as observable proxy; estimate country-specific structural VARs with proxy-first (Cholesky) ordering that places proxy first. - First orthogonalized innovation ε_{1t} interpreted as fiscal shock; P_{11} > 0 normalization ties sign to proxy. - Population equivalence: proxy-first Cholesky implementation delivers same relative IRF estimand as IV local projections using z_t as instrument under invertibility. - VAR system and estimation details - Variables included: z_t (proxy), g_t ≡ log G_t, y_t ≡ log Y_t. - Vector x_t = ( z_t, g_t, y_t )′. - Lag order p = 8; effective sample size T_eff = T − p. - Reduced-form VAR: x_t = c + Σ_{ℓ=1}^p B_ℓ x_{t−ℓ} + u_t; Σ_u ≻ 0. - Proxy-first Cholesky identification and admissibility - Cholesky: Σ_u = P P′; orthogonalized innovations u_t = P ε_t. - Identified fiscal shock: ε_{1t}. - Admissibility conditions: accept–reject filter applied to posterior draws to avoid draws where proxy innovation yields negligible spending response or implausible multipliers. - Bayesian reduced-form VAR with Minnesota priors used; draws (B, Σ_u) retained if admissibility inequalities hold. - Objects of interest and multipliers - Orthogonalized IRFs dIRF_y(h), dIRF_g(h) in log points. - Cumulative spending multiplier at horizon h: - cM_h = (1 / ȳ_s_G) · (Σ_{j=0}^h dIRF_y(j)) / (Σ_{j=0}^h dIRF_g(j)), where ȳ_s_G ≡ (Ḡ / Ȳ) average spending-to-GDP ratio. - Summaries reported using posterior medians and posterior uncertainty bands from accepted draws. --- ### Panel proxy-SVAR framework with a qualitative instrument (section 5.2) - Panel VAR with fixed effects - x_{i,t} = Σ_{ℓ=1}^p A_{ℓ} x_{i,t−ℓ} + μ_i + τ_t + u_{i,t}, with x_{i,t} = (g_{i,t}, y_{i,t})′. - Estimation by least squares after removing fixed effects; remark on state dependence via interactions. - Structural representation and proxy identification - u_{i,t} = B ε_{i,t}; spending shock ε^g_{i,t} defined as first structural innovation with impact vector b ≡ B e_1. - Effective proxy v_{i,t} constructed from qualitative z_{i,t} (and possible lags) with weights w_{i,t}. - Assumption PS (relevance and exogeneity): - (i) E[w_{i,t} ev_{i,t} ε^g_{i,t}] ≠ 0; - (ii) E[w_{i,t} ev_{i,t} ε^j_{i,t}] = 0 for j ≠ g. - Identification (impact vector up to scale): - b ∝ E[w_{i,t} u_{i,t} ev_{i,t}]; normalize b_g = 1. - Two-variable covariance ratio representation (2×2 case): - b̂ = (1, dCov_{S,w}(bu_y_{i,t}, ev_{i,t}) / dCov_{S,w}(bu_g_{i,t}, ev_{i,t}))′. - Instrument processing for qualitative and sparse proxy - q_{i,t} is K×1 vector of instrument components; first-stage projection: - bu^g_{i,t} = π′ eq_{i,t} + e_{i,t}; fitted component v_{i,t} ≡ b π′ eq_{i,t}. - Zeros and missing proxy information conventions: - (i) zeros-as-zero: z_{i,t} = 0 treated literally; - (ii) zeros-as-missing: set w_{i,t} = 0 when z_{i,t} = 0 (but retain quarters in VAR). - Impulse responses and multiplier construction - dIRF(h) = Ψ̂_h b̂, h = 0,...,H. - Group-specific trimmed-mean spending share s̄ ≡ TrimMean_{S,w}(G/Y) used to map to dollar multipliers. - For ∆ log benchmark: cM(H) = (1 / s̄) [Σ_{h=0}^H dIRF_y(h)] / [Σ_{h=0}^H dIRF_g(h)]. - Bootstrap inference - Country-cluster bootstrap: resample countries with replacement, re-estimate VAR, rebuild proxy and impact vector, recompute IRFs and multipliers. - Reported intervals: 90% bootstrap percentile bands. - Practical filter: exclude draws with near-zero denominators or implausible multipliers; bounds stated with results. --- ### Benchmark pooled-panel heterogeneity (Table 12 summary) - Benchmark implementation notes - Transformation: quarterly log changes g_{i,t} = ∆ log G_{i,t}; y_{i,t} = ∆ log Y_{i,t}. - p = 8 lags; zeros-as-zero convention for z_{i,t} in benchmark. - Reported statistics: cumulative multipliers at H = 4 and H = 8 (cM_G(4), cM_G(8)); 90% bootstrap percentile intervals. - Selected group results (exact entries preserved) - Panel A. Trade openness - Closed (26 countries): M_4 = 1.425 * * , M_8 = 1.364 * * ; 90% intervals: (0.437, 1.993) and (0.418, 1.935) - Open (34 countries): M_4 = 0.558 * * , M_8 = 0.649 * * ; 90% intervals: (0.075, 1.094) and (0.088, 1.165) - Panel B. Public debt - Low/No high debt (56 countries): M_4 = 0.738 * * , M_8 = 0.893 * * ; 90% intervals: (0.118, 1.481) and (0.146, 1.669) - High debt (26 countries): M_4 = 0.821 * * , M_8 = 0.851 * * ; 90% intervals: (0.185, 1.856) and (0.188, 2.008) - Panel C. Exchange-rate regime - Fixed (51 countries): M_4 = 1.104 * * * , M_8 = 1.119 * * * ; 90% intervals: (0.372, 1.624) and (0.379, 1.653) - Float (8 countries): M_4 = 0.715, M_8 = 0.703 ; 90% intervals: (-0.053, 1.881) and (-0.019, 1.745) - Panel D. Informality - Low informality (28 countries): M_4 = 1.173 * * , M_8 = 1.195 * * * ; 90% intervals: (0.403, 1.504) and (0.434, 1.494) - High informality (28 countries): M_4 = 0.788 * * , M_8 = 0.798 * * * ; 90% intervals: (0.092, 1.669) and (0.097, 1.687) - Panel E. Labor flexibility - Low (25 countries): M_4 = 1.439 * * * , M_8 = 1.347 * * * ; 90% intervals: (0.555, 1.796) and (0.523, 1.613) - High (24 countries): M_4 = 1.560 * * , M_8 = 1.662 * * ; 90% intervals: (0.410, 1.993) and (0.427, 2.140) - Five benchmark findings (concise) - Trade openness: multipliers smaller in open economies (Closed M_4 = 1.425 vs Open M_4 = 0.558). - Public debt: limited heterogeneity (M_4 = 0.738 vs 0.821; M_8 = 0.893 vs 0.851). - Exchange-rate regime: larger multipliers under fixed regimes (Fixed M_4 = 1.104 vs Float M_4 = 0.715). - Informality: multipliers smaller in more informal economies (Low 1.173 vs High 0.788 at H = 4). - Labor-market flexibility: larger multipliers with more flexible labor markets (High M_8 = 1.662 vs Low M_8 = 1.347). --- ### State dependence: business cycle (sections 6.1–6.2) - Smooth-transition panel proxy-SVAR under growth conditions - Growth index: s_{i,t} ≡ (1/4) Σ_{r=1}^4 ∆ log Y_{i,t−r}, standardized within country to ês_{i,t}. - Logistic low-growth weight: F_{i,t} ≡ 1 / [1 + exp{γ(ês_{i,t} − c)}], with benchmark c = 0 and γ = 1.5. - Representative evaluation points: - F_{HighGrow} = 0.3167 (25th percentile) - F_{LowGrow} = 0.6399 (75th percentile) - Specification: - x_{i,t} = α_i + Σ_{ℓ=1}^p A_{ℓ} x_{i,t−ℓ} + Σ_{ℓ=1}^p B_{ℓ} [F_{i,t} x_{i,t−ℓ}] + u_{i,t}, p = 8. - Zeros-as-missing convention used for identification. - Multipliers across growth conditions (key numeric findings) - ∆ log specification: - High-growth H = 4: −0.01 - Low-growth H = 4: 0.77 - High-growth H = 8: −0.01 - Low-growth H = 8: 0.77 - Log-level specification: - H = 4: High growth = 0.25; Low growth = 0.74 - H = 8: High growth = 0.58; Low growth = 1.39 - Statistical precision: - ∆ log H = 4 bootstrap p = 0.11; H = 8 p = 0.14. - Log-level H = 4 p = 0.39; H = 8 p = 0.24. - Interpretation: - Point estimates suggest stronger multipliers in weak-growth (low-growth) environments; differences suggestive but imprecisely estimated. --- ### Policy uncertainty and fiscal multipliers (section 7) - Uncertainty measures (constructed from EIU text) - WUI: World Uncertainty Index (count of “uncertain” variants scaled by report length). - FUI: fiscal-policy-focused uncertainty index (same method restricted to fiscal passages). - High-uncertainty indicator Unc^m_{i,t}: 1 if four-quarter moving average eU^m_{i,t} ≥ 75th-percentile threshold (benchmark: within-country for WUI; pooled for FUI). - Approximately one quarter of observations classified as high uncertainty for each index. - State-contingent panel proxy-SVAR estimation - Benchmark endogenous vector: x_{i,t} = [∆ log g_{i,t}, ∆ log y_{i,t}]′, p = 4, country fixed effects. - Proxy identification within state κ uses b_{κ,m} ∝ E[u_{κ,m,i,t} z_{i,t}] normalized b_{κ,m,g} = 1. - Multipliers across uncertainty states (Table 14 exact entries) - WUI: - H = 4: Low = 1.04; High = 0.51; High−Low = −0.52 - H = 8: Low = 1.09; High = 0.42; High−Low = −0.67 - FUI: - H = 4: Low = 0.95; High = 0.21; High−Low = −0.74 - H = 8: Low = 1.13; High = 0.18; High−Low = −0.94 - Interpretation: - Multipliers are lower in high-uncertainty quarters for both WUI and FUI; bootstrap intervals for differences are wide and include zero. --- ### Political support and fiscal multipliers (section 8) - Political support measures and coding - Supp_{i,t} ∈ {0,1} from EIU reports via fixed LLM prompt; verbatim evidence stored. - Three political primitives coded {Yes, No, Unclear} with verbatim quotes: - maj_cohesion (working legislative majority/cohesion) - election_proximity (election scheduled in current or next quarter) - unrest_gridlock (material unrest or legislative paralysis) - Missing political text treated as its own state (no imputation); mapping to econometric Supp_{i,t}: 1 if favorable, 0 if unfavorable, missing if null. - Empirical design (pooled LP-IV with interactions) - Cumulative changes defined: - ∆^h y_{i,t} ≡ ln Y_{i,t+h} − ln Y_{i,t−1} - ∆^h g_{i,t} ≡ ln G_{i,t+h} − ln G_{i,t−1} - LP-IV equation includes interactions of ∆g_{i,t} with S_1_{i,t} = 1{Supp_{i,t} = 1} and S_m_{i,t} = 1{Supp_{i,t} missing}; instruments: {z_{i,t}, z_{i,t} × S_1_{i,t}, z_{i,t} × S_m_{i,t}}. - Controls: four lags of ∆y and ∆g plus inflation, unemployment, policy rate, below-trend indicator; standard errors clustered by country. - Suppress multiplier when |ĉRF^g_h(s)| < 0.005. - Main results: political support amplifies implementation and multipliers (Table 15) - h = 4: - Support = 0: suppressed (|ĉRF^g_h(0)| < 0.005) - Support = 1: 1.0590 - Support missing: 1.1078 - h = 8: - Support = 0: 0.1880 - Support = 1: 0.9627 - Support missing: −0.6868 - Interpretation: - Favorable political environment (Supp = 1): scaled dollar multiplier ≈ 1.0–1.1 at horizons of one to three years. - Unfavorable support (Supp = 0): multipliers close to zero or indeterminate at short horizons. - Implementation channel (reduced-form evidence, Table 16) - h = 4: - RF^y_4(0) = 0.00048; RF^y_4(1) = 0.00524; ∆RF^y_4(1−0) = 0.00476 (t = 1.71) - RF^g_4(0) = −0.00330; RF^g_4(1) = 0.01490; ∆RF^g_4(1−0) = 0.01819 (t = 2.94) - h = 8: - RF^y_8(0) = 0.00061; RF^y_8(1) = 0.00621; ∆RF^y_8(1−0) = 0.00560 (t = 1.45) - RF^g_8(0) = 0.00979; RF^g_8(1) = 0.01943; ∆RF^g_8(1−0) = 0.00964 (t = 1.65) - Note: ∆RF_g(1 − 0) = 0.01819 at h = 4 has t = 2.94, indicating a statistically significant difference in cumulative spending reduced form between support and no-support quarters. - Mechanism decomposition: which political dimension matters? (section 8.4) - Reduced-form spending pass-through decomposition (Table 17, exact numeric entries) - h = 4: - z × S1 in RF g: 0.02044 (2.17) - z × ElectYes in RF g: −0.01956 (−2.20) - Net (z × S1) + (z × ElectYes): 0.00088 - h = 8: - z × S1 in RF g: 0.01955 (2.26) - z × ElectYes in RF g: −0.01834 (−2.75) - Net: 0.00121 - Findings: - Favorable political support associated with stronger cumulative spending response at medium horizons (coefficient ≈ 0.02 at h = 4 and h = 8; t ≈ 2.2). - Election proximity associated with substantially weaker spending pass-through (coefficient ≈ −0.018 to −0.020; statistically significant). - Magnitudes comparable: election proximity can nearly offset support advantage. - Robustness: - Sign composition of narrative shocks similar across support states: share of z = −1 is 0.551 when Support = 0 and 0.546 when Support = 1. --- ### Admissibility, robustness, and appendices highlights - Posterior accept–reject restrictions (Draw B / A.3) - Screening conditions: - Average short-run spending response 1/(H_g + 1) Σ_{h=0}^{H_g} θ_{g,h} ≥ g_. - Baseline H_g = 8, g_ = 0 (average response constrained above 0 for first 8 quarters). - Cumulative multipliers constrained to lie within [−2, 5]; short-run H_S = 8; long-run H_L = 20. - Sensitivity to multiplier bounds (A.4) - Wider bounds [−2.4, 6] and tighter bounds [−1.6, 4] tested. - Acceptance rates: Baseline 47%; Wider 50%; Tighter 43%. - Pooled IVW IRFs nearly indistinguishable across bounds. - Bayesian vs frequentist comparisons (B) - Single-country (United States) IRFs: Bayesian and frequentist IRF distributions nearly identical. - Frequentist Minnesota counterpart similar to Bayesian. - Panel VAR and split definitions (C) - Benchmark pooled-panel approach and bootstrap inference details summarized (estimation, proxy mapping, multiplier construction, split cutoffs). - Political support data pipeline (D) - Two-stage fixed-prompt pipeline: - Stage 1: political_environment_favorable_indicator ∈{1,0,null}, verbatim evidence phrases, public_support ∈{HIGH, LOW, MIXED, UNCLEAR}. - Stage 2: code primitives maj_cohesion, election_proximity, unrest_gridlock ∈{YES, NO, UNCLEAR} with exact-quote rule. - Mapping to econometric Supp_{i,t}: 1 if favorable, 0 if unfavorable, missing if null. - Political variables never used to classify fiscal exogeneity; used only to index environment and explore mechanisms. _Italic: Source — wpiea2026043-source-pdf (selected sections and tables summarized from the supplied chapter)._ ### Section 8 studies how political support conditions fiscal transmission. Section 9 concludes. ### Construction of the AI-narrative database ### Data Source - Primary narrative source: Economist Intelligence Unit (EIU) Country Reports. - Coverage features: - Standardized country reports with coverage for many countries extending back to the 1950s. - Reports typically released quarterly; some countries monthly in later years. - Constructed a quarterly archive with one report per country–quarter; when multiple reports in a quarter, retain the chronologically last comprehensive report. - EIU workflow: country analysts draft, specialists review, centralized editing—promotes comparability across countries and time. - Advantages relative to other sources: (i) higher-frequency coverage for many economies; (ii) long historical time series across advanced, emerging, and low-income economies. - Empirical focus: subset of 64 advanced and developing economies for which consistent quarterly macroeconomic and fiscal data were also collected. ### Information to look for (definition of exogenous actions) - Objective: identify “exogenous” discretionary government spending actions following Romer and Romer (2010). - Exogenous action criteria (motive unrelated to contemporaneous macroeconomic conditions): - Typical non-cyclical motives include medium-term consolidation targets, long-run ideological or structural objectives, or compliance with a law, treaty, or supranational rule. - Treat actions undertaken to counter recession/slowing growth, reduce unemployment, cool overheating, or respond to contemporaneous inflation, interest-rate moves, exchange-rate pressure, or financing stress as endogenous. - Acknowledging current conditions does not imply endogeneity if the stated motive is explicitly non-cyclical. ### Dispassionate and consistent approach to the source (GPT–4.1 prompt and coding) - Fixed GPT–4.1 prompt used off-the-shelf (no task-specific training or fine-tuning); prompt specification and coding rubric invariant across countries and time. - Prompt records whether measures are implemented in the current quarter or announced for future quarters; baseline econometric series uses only discrete sign classification (expansion, contraction, or no action). - Prompt outputs (returned per report): 1. EXOG_NET_SPENDING ∈ {EXP_, CON_, NEUTRAL, UNCLEAR} (optionally append small/large). 2. MOTIVATION: brief phrase citing stated motive(s). 3. COMPONENT (if specified): spending component(s). 4. INTENSITY_ORDINAL ∈ {-10, .. . , 10} (optional). 5. CONFIDENCE (optional): self-assessment of classification confidence. - Coding criteria: action coded as exogenous only when: - (i) stated motive is unrelated to near-term macroeconomic conditions; and - (ii) narrative does not cite contemporaneous growth, inflation, unemployment, interest-rate movements, exchange-rate pressure, or financing stress as the rationale. - Construction of the shock: - Report-level outputs aggregated into country–quarter proxy z_{i,t} ∈ {−1, 0, +1}: - +1 if EXOG_NET_SPENDING = EXP_ - −1 if EXOG_NET_SPENDING = CON_ - 0 if EXOG_NET_SPENDING ∈ {NEUTRAL, UNCLEAR} - z_{i,t} interpreted as a qualitative proxy for discretionary spending actions recorded in real time (may combine announcement and implementation elements). - In econometric analysis, z_{i,t} used as an instrument for innovations to realized government spending; impulse responses capture average effects of the composite policy-action shock component correlated with the narrative proxy. - If direction cannot be determined, model returns UNCLEAR → mapped to z_{i,t} = 0. - When stated motive is not explicitly non-cyclical or linked to contemporaneous macro variables, the prompt classifies the action as non-exogenous (does not contribute to net effect). ### Documentation: narrative evidence and stylized facts - Examples of classified episodes: - Ghana, 2010Q3 — Action: Parliament passed US$10bn housing deal (STX Korea). Stance: Expansion. Motivation: expand housing supply and infrastructure capacity. Timing: implemented. Confidence: 95 percent. - Ecuador, 1982Q4 — Action: removal of petrol and imported-wheat subsidies as fiscal consolidation. Stance: Contraction. Motivation: contain budgetary pressures. Timing: implemented. Confidence: 90 percent. - United Kingdom, 2005Q4 — No explicit spending changes identified; AI assigns exogenous effect as unclear. - Selected concise episodes (representative fragments from reports): - Bangladesh 1980Q2 +1: “five-year plan envisages Tk 300–500 bn in public programs” - Honduras 1978Q1 +1: “US$450 mn over five years in electric-power projects” - Côte d’ Ivoire 1970Q2 +1: “special capital budget; Canadian-supported electrification” - Ghana 2010Q3 +1: “Parliament approves large public-housing program (∼US$10 bn)” - Indonesia 1985Q1 −1: “real development spending cut by about 7%” - Ecuador 1982Q4 −1: “removal of petrol and imported-wheat subsidies” - Aggregate counts and shares (from 16,029 country-quarter observations processed): - Nonzero discretionary spending episodes identified: 8,636 (53.9% of the sample). - Of these, expansions: 4,374; contractions: 4,262. - Expansions account for 50.6% of all non-zero observations. - Coverage summary (nonzero quarters only): - Countries in universe: 64 - Countries with nonzero quarters (this listing): 64 - Sample period (unbalanced; country start dates vary): 1952:Q1–2023:Q4 - Nonzero quarters (z_{i,t} = ±1): 8,636 - Expansions / Contractions (%): 50.6 / 49.4 - Cross-country heterogeneity (nonzero quarters by IMF income group): - Advanced: Nonzero quarters 3,129; Expansion share (%) 45.8 - Emerging: Nonzero quarters 3,587; Expansion share (%) 49.5 - LIC: Nonzero quarters 1,920; Expansion share (%) 60.7 - Motivations by action type (random subsample text-dictionary classification): - Expansionary actions: nearly 60 percent associated with increases in public investment (infrastructure and capital projects). - Contractionary actions: over two-thirds reflect explicit consolidation measures (spending restraint and deficit reduction). - Other categories (social spending, privatization, external financing) appear in both expansions and contractions but are secondary. ### Quarterly macroeconomic and fiscal series - For VAR analysis, assembled quarterly time series for each country: - Real GDP (output) and real government expenditure obtained from EIU, Haver, CEIC, and central banks (detailed sources in Table 4). - Variables expressed in log levels for the VARs. - Fiscal shock series construction: - Narrative fiscal impulse series includes only exogenous shocks; narrative fiscal impulse zeroed out in quarters judged endogenous. - Final panel dataset combines: - (i) narrative fiscal impulse series (with exogenous shocks), and - (ii) macroeconomic series for output and other variables. - All subsequent results use the intersection of the narrative series and available quarterly macro data. ### Validation (overview of four complementary assessments) - Validation exercises aim to evaluate (i) model’s ability to interpret narrative sources like expert coders and (ii) empirical validity of the spending-action proxy: 1. Reproduce Romer and Romer (2019) report-level classifications of fiscal stance and stated motivations to test LLM’s ability to read EIU reports and extract fiscal narratives. 2. Replicability: re-run fixed prompt on same inputs and quantify stability of model outputs across runs. 3. Empirical mapping: test whether the production proxy for exogenous discretionary spending actions predicts subsequent movements in measured government spending. 4. External alignment: compare quarterly sign proxy to action-based annual consolidation measures reported by Adler et al. (2024). *Source: wpiea2026043-source-pdf - Section 8 studies how political support conditions fiscal transmission. Section 9 concludes.* ### 3.1 Benchmarking narrative interpretation: comparison to Romer ### 3.1 Benchmarking narrative interpretation: comparison to Romer and Romer (2019) ### Benchmark design and rationale - RR19 (Romer and Romer (2019)) study fiscal policy responses in episodes of financial distress across 31 OECD countries over 1980–2017 using narrative evidence from EIU Country Reports. - For each episode RR19 examine a fixed sequence of reports (typically nine; one episode includes eleven) and, for each report, answer four questions: - 1. What is the current and/or prospective stance of fiscal policy? - 2. What is the motivation given for the fiscal developments? - 3. Does the EIU mention the debt-to-GDP ratio as a concern or as an underlying motivation? - 4. Is there anything else of note relevant to fiscal policy actions? - The AI exercise uses the same set of EIU reports used by RR19 for the same 19 countries (200 reports in total) and a fixed prompt that mirrors RR19’s four questions to evaluate accuracy for outputs corresponding to questions (1) and (2). ### Classification taxonomy and coding rules - Fiscal stance labels assigned by the model: Expansionary, Contractionary, Neutral, or Unclear. - Motivation taxonomy (RR19’s categories) adopted; multiple motive categories may be recorded: 1. Financial rescue (bank bailouts or financial-crisis-driven stimulus); 2. Countercyclical stimulus (stimulus during a downturn); 3. Politics (stimulus taken to win an election); 4. Market access (austerity due to rising borrowing costs); 5. Conditionality (program requirements associated with official support); 6. Policymaker ideas (domestic policy preferences); 7. EU fiscal rules (adherence to supranational rules); 8. Countercyclical austerity (actions to prevent overheating); 9. Other. - RR19 classify motives (1)–(3) as pertaining to expansionary actions and motives (4)–(8) as pertaining to contractionary actions; motivation categories are not mutually exclusive. - Validation approach: - Panel (a): assess whether model’s net stance matches RR19’s net classification for each report. - Panel (b): assess whether model detects the same motivation category RR19 associates with that report—scored independently of Panel (a). - Validation performed at the report level: for each RR19-listed report, the AI classification is compared to RR19’s human-coded judgment for stance and for the specific motivation first mentioned. ### Validation results: agreement with RR19 classifications - Overall agreement statistics reported: - For expansions: - AI matches the stance in 14 out of 16 cases (87.5%). - AI matches the motivation in 32 out of 34 cases (94.1%). - Note: Four RR19 expansion episodes (Finland 1993Q1, Sweden 1993Q2, Greece 2009M7, and Ireland 2009M7) are excluded from the baseline stance comparison because RR19 classify these as net expansions primarily due to large financial-rescue measures, while the accompanying EIU text describes concurrent conventional fiscal contraction or ambiguous net effects. Including these four boundary cases reduces the expansion stance match rate from 14/16 (87.5%) to 14/20 (70%). - For contractions: - AI matches the direction in 18 out of 19 cases (95%). - AI matches the motivations in 44 out of 45 cases (97.8%). - Diagnostic interpretation: - Many apparent disagreements arise in reports describing both expansionary and contractionary measures simultaneously. - The model generally extracts individual actions and their signs accurately, but the conservative aggregation rule (which does not infer magnitudes when unobserved) can yield a different net stance than RR19’s ex post judgment. - Boundary cases include: - Turkey and Iceland: concurrent increases and cuts lead RR19 to code net expansion while the AI rule assigns contraction or Unclear in absence of textual evidence on relative magnitudes. - Norway: hinges on whether financial-sector support is treated as fiscal spending. - Identified genuine model errors: - Korea: model missed an explicit political-motive cue (“with an eye to elections...”), leading to a misclassified stance. - United States (Table 4 entry): model correctly listed measures but inferred overall stance from remarks about rising deficits and mis-signed the net effect. - Finland near-miss: model failed to map language about rising long-term rates and financing pressure to a Market Access motive; this affects motive label but not stance sign. - Overall conclusion from benchmark: - Most apparent disagreements reflect differences in aggregation/judgment in mixed-action passages rather than inability to detect underlying fiscal actions. ### Replicability (results from re-running prompts 51 times on 20 reports) - Experimental setup: - Re-ran the prompt 51 times on a subset of 20 reports using the same prompt used in section 3.1. - Extracted fields: original Romer and Romer questions plus intensity and confidence. - Intensity: integer between -10 to +10 measuring size and direction of fiscal stance. - Confidence: AI’s self-assessed confidence in its assessment of all fields. - Replicability summarized by modal share: fraction of runs in which the most frequently assigned value appears (equals 1 if identical value in every run). - Distributions presented via empirical cumulative distribution functions (ECDFs). - Replicability findings: - Direction of the fiscal stance: perfectly stable — for all 20 reports the modal share equals one (every run assigns the same sign). - Motivation: slightly lower but still high replicability; dominant motivation identified consistently with occasional variation reflecting multiple plausible motives emphasized in narrative. - Intensity and confidence: lower replicability with several cases showing non-trivial dispersion; many reports still display high modal shares but alternative values appear intermittently. - Interpretation: - Direction is a coarse classification closely tied to qualitative narrative; hence highly robust across runs. - Motivation is multi-category and multi-faceted; small differences in textual emphasis across runs can alter which motive is selected as primary. - Intensity and confidence require finer quantitative judgments and are more sensitive to phrasing and ambiguity. ### Correlation with government spending data (local projection evidence, from 3.3) - Empirical specification: - Estimated local-projection equation: lnG_{i,t+h} = β^+_h 1{z_{i,t}=+1} + β^-_h 1{z_{i,t}=-1} + sum_{k=1}^4 [ φ_{h,k} lnG_{i,t-k} + ψ_{h,k} lnY_{i,t-k} + γ_{h,k} z_{i,t-k} ] + α_i + λ_t + ε_{i,t+h}, for h = 0,...,8. - Impact estimates and cumulative responses (exact numeric statements preserved): - At impact (h = 0), contractions (negative shocks) reduce logG by roughly −0.005 (log points; 0.5% level). - Expansions (+1) rise one quarter later with β^+ ≈ 0.002. - Cumulatively, contractions reach about −0.06 by h≈ 8. - Cumulatively, expansions reach about 0.04 by h≈ 8. - Notes on figures: - Figure 2a: local-projection impulse responses of G to expansionary (z_{i,t}=+1) and contractionary (z_{i,t}=-1) narrative shocks; includes country and time fixed effects and listed lag controls; shaded areas denote 68% confidence intervals with standard errors clustered at the country level; horizons are quarters. - Figure 2b: associated cumulative responses. *Source: Authors’ calculations based on Romer and Romer (2019) and EIU Country Reports, as summarized in the source PDF.* ### 3.4 External Validation against Adler et al. 2024 ### 3.4 External Validation against Adler et al. 2024 ### Purpose and data sources - Objective: evaluate external validity of the quarterly narrative fiscal stance series against the annual action-based dataset in Adler et al. (2024). - Adler et al. (2024) dataset: - Covers a set of 31 advanced and emerging economies. - Reports announced size of discretionary fiscal consolidation packages, decomposed into tax measures and expenditure measures, expressed as a percent of GDP. - Coverage begins in 1978 for advanced economies and in 1989 for Latin American and Caribbean economies, and extends through 2020. - Definition: includes only measures primarily motivated by medium-term deficit reduction and debt sustainability; measures taken in response to short-run cyclical weakness are excluded. - Expenditure component in Adler et al. (2024) denoted spend_c,y: size (in percent of GDP) of spending-based consolidation in country c and year y. Positive values indicate expenditure cuts or restraint (fiscal tightening); negative values indicate net loosening. ### Quarterly narrative series and constructed annual proxies - Quarterly narrative coding: - Each quarter coded indicator_c,y,q ∈ {−1, 0, +1}: −1 = discretionary consolidation in government spending; +1 = discretionary expansion in government spending; 0 = no decisive discretionary spending action. - Nonzero actions classified as discretionary consolidation (−1) or discretionary expansion (+1). - Annual proxies constructed (for each country c and year y): - consol_quarters_c,y = Σ_{q=1}^4 1{indicator_c,y,q = −1} - expand_quarters_c,y = Σ_{q=1}^4 1{indicator_c,y,q = +1} - any_consol_c,y = 1{consol_quarters_c,y ≥ 1} - any_expand_c,y = 1{expand_quarters_c,y ≥ 1} - net_stance_c,y = (consol_quarters_c,y − expand_quarters_c,y)/4, which lies in [−1, +1] and equals +1 if all four quarters are consolidation (−1) and none expansion; equals −1 if all four quarters are expansion (+1) and none consolidation. ### Test 1 — Extensive-margin alignment (hit rates and expansion rates) - Definition: AE_consol_c,y(τ) = 1{spend_c,y ≥ τ} for thresholds τ ∈ {0.5, 1.0, 2.0}. - Metrics computed for each τ: - Hit rate: share of (c,y) with AE_consol_c,y(τ) = 1 such that any_consol_c,y = 1. - Expansion rate: share of the same (c,y) such that any_expand_c,y = 1. - Full merged sample results (Panel A, summarized): - τ = 0.5: - N_big = 138 - Hit rate = 0.964 - Expansion rate = 0.246 - N_loose = 1 - Exp. hit rate = 1.000 - Consolidation rate (any_consol = 1) = 0.000 - τ = 1.0: - N_big = 65 - Hit rate = 0.969 - Expansion rate = 0.231 - τ = 2.0: - N_big = 16 - Hit rate = 1.000 - Expansion rate = 0.250 - Panel B: by country group (Advanced / AEs and LAC) - LAC (Latin America / Caribbean): - τ = 0.5: Hit rate = 1.000; Expansion rate = 0.296; N_big = 27 - τ = 1.0: Hit rate = 1.000; Expansion rate = 0.273; N_big = 11 - τ = 2.0: Hit rate = 1.000; Expansion rate = 0.250; N_big = 4 - Advanced economies: - τ = 0.5: Hit rate = 0.955; Expansion rate = 0.234; N_big = 111 - τ = 1.0: Hit rate = 0.962; Expansion rate = 0.222; N_big = 54 - τ = 2.0: Hit rate = 1.000; Expansion rate = 0.250; N_big = 12 - Additional observations: - Between 23 and 30 percent of “large consolidation” years also contain at least one expansion quarter in the quarterly series, reflecting within-year mixtures. - Years with spend_c,y ≤ −τ (net loosening) are rare in the intersection sample (one country–year for τ = 0.5 and zero for τ ≥ 1.0); in the loose case observed, any_expand_c,y = 1 and any_consol_c,y = 0. ### Test 2 — Probability of being identified as a consolidation year (linear probability model) - Estimated model: - 1{spend_c,y ≥ 0.5} = α_c + β_1 any_consol_c,y + β_2 any_expand_c,y + ε_c,y - Country fixed effects included; standard errors clustered by country. - Results (Table 6, column 1): - Coefficient on any_consol_c,y = 0.126 with clustered standard error 0.019 (statistically significant). - Interpretation: conditional on country fixed effects, observing at least one consolidation quarter is associated with a 12.6 percentage point higher probability that Adler et al. (2024) classify the country–year as featuring a sizable expenditure-based consolidation package (≥ 0.5 percent of GDP). - Coefficient on any_expand_c,y = −0.073 with clustered standard error 0.020 (statistically significant). - Interpretation: observing an expansionary quarter is associated with a 7.3 percentage point lower probability that Adler et al. (2024) codes that year as a sizable consolidation. - Both coefficients significant with standard errors clustered by country. ### Test 3 — Intensity alignment (continuous regression linking net_stance to spend) - Estimated model: - spend_c,y = γ_c + δ net_stance_c,y + u_c,y - Country fixed effects included; standard errors clustered by country. - Results (Table 6, column 2): - δ = 0.263 with clustered standard error 0.044 (statistically significant). - Interpretation: net_stance ranges from −1 to +1. Moving from a purely expansionary year (net_stance = −1) to a purely consolidation year (net_stance = +1) is associated, on average within a given country, with an increase of approximately 2 × 0.263 ≈ 0.5 percentage points of GDP in spend_c,y. ### Predictability of quarterly shocks versus annual shocks - Quarterly predictability regressions (equation (5)) assess whether quarterly narrative expenditure shocks z_i,t are predictable from lagged Debt/GDP, Output Gap, GDP Growth, and lagged z_i,t, across three samples: full sample, Guajardo et al. (2014) sample, Adler et al. (2024) sample. - Table 7 summarized results: - Across all three samples, the only robust predictor of z_i,t is its own lag (Expenditure shock (lagged)): - Full sample: lagged coefficient = 0.266 ∗∗∗ - Guajardo et al. (2014) sample: lagged = 0.308 ∗∗∗ - Adler et al. (2024) sample: lagged = 0.298 ∗∗∗ - Coefficients on lagged Debt to GDP, Output Gap, and GDP growth are statistically indistinguishable from zero in these quarterly regressions. - R^2 values: Full sample R^2 = 0.07; Guajardo sample R^2 = 0.10; Adler sample R^2 = 0.09. - Constants: Full sample constant = −0.0775 ∗∗∗; Guajardo = −0.103 ∗∗; Adler = −0.0915 ∗∗. - Comparison: annual predictability of Adler et al. (2024) shocks (Table 8): - Lagged narrative shock highly significant in both primary balance and expenditure specifications. - Lagged GDP Growth contributes to predicting the shocks. - R^2 approximately 0.243 and 0.305 for the two specifications, substantially higher than quarterly regressions. - Aggregation check (Table 9): - Aggregating quarterly shocks to annual frequency (sum within year, range −4 to +4) increases predictability. - Baseline annual aggregation (column 1): - Debt to GDP (lagged) = 0.00545 ∗ - GDP Growth (lagged) = 6.030 ∗∗∗ - Expenditure shock (lagged) = 0.324 ∗∗∗ - Constant = −0.727 ∗∗∗ - R^2 = 0.125 - Contractionary-only and expansionary-only specifications show similar patterns with GDP Growth and lagged shock significant; R^2 values 0.116 and 0.053 respectively. - Interpretation: - Quarterly expenditure shocks contain less systematic information about lagged macroeconomic conditions than annual narrative shocks from Adler et al. (2024). - The weak predictability at quarterly horizon likely reflects higher-frequency measurement; aggregation to annual frequency reveals more systematic relationships with lagged macro controls. ### Overall conclusions from validation exercises - Extensive margin: - When Adler et al. (2024) report a sizable expenditure-based consolidation package (spend_c,y ≥ τ for τ ∈ {0.5, 1.0, 2.0}), the quarterly narrative series almost always records at least one discretionary consolidation quarter in that same country–year (hit rates above 0.95 and reaching 1.000 for τ = 2.0). - These consolidation years only rarely contain exclusively expansionary actions in the quarterly series. - Intensive margin: - Within countries, the annual size of spending-based consolidation (spend_c,y in percent of GDP) in Adler et al. (2024) is increasing in the share of quarters the quarterly database codes as discretionary spending consolidations (δ = 0.263, s.e. 0.044). - Moving from net expansion (net_stance = −1) to net consolidation (net_stance = +1) corresponds to roughly 0.5 percentage points of GDP higher spend_c,y in Adler et al. (2024). - Predictability: - Quarterly shocks exhibit limited predictability from lagged Debt/GDP, Output Gap, and GDP Growth (R^2 ≤ 0.1); only the shock’s own lag is a robust predictor. - Annual aggregated shocks and Adler et al. (2024) annual shocks show stronger predictability from lagged macro indicators (higher R^2). - Interpretation: - The quarterly narrative series and the annual action-based dataset co-move on both extensive and intensive margins and therefore capture the same underlying policy object—discretionary shifts in the government’s expenditure stance aimed at medium-term fiscal adjustment—despite differences in frequency, sample scope, and coding design. *Source: Authors’ calculations and coding using the merged country–year panel described in the text (validation against Adler et al. (2024)).* ### 4.2 Econometric Identification Strategy ### 4.2 Econometric Identification Strategy ### Identification approach - Use the narrative series as an observable proxy for unexpected spending actions and estimate country-specific structural VARs. - Identify government-spending shocks via a recursive (“internal instrument”) ordering that places the proxy first (Plagborg-Møller and Wolf, 2021). - With the proxy-first ordering, the first orthogonalized innovation (the first Cholesky shock) is interpreted as the fiscal shock: - It may affect all remaining variables contemporaneously. - Remaining innovations are constrained to have no impact effect on the proxy. - In population, this implementation delivers the same relative impulse-response estimand as IV local projections that use the same narrative proxy and, under standard invertibility conditions, coincides with the proxy-SVAR estimand (Plagborg-Møller and Wolf, 2021; Li et al., 2024; Stock and Watson, 2018). ### VAR system, ordering, and lag length - Variables always included for each country: - z_t: narrative spending series (proxy) - g_t ≡ log G_t: real government spending - y_t ≡ log Y_t: real GDP - Vector of observables: - x_t = ( z_t, g_t, y_t )′ (n-dimensional with n ≥ 3; additional macro variables may be included when available) - Lag order: - p = 8 - Effective sample size: T_eff = T − p - Regressor matrix dimension: - k = 1 + n p - Coefficient matrix B is k × n stacking c and B_1,...,B_p ### Reduced-form VAR representation - VAR(p) written as: - x_t = c + Σ_{ℓ=1}^p B_ℓ x_{t−ℓ} + u_t, E[u_t] = 0, Σ_u ≡ E[u_t u′_t] ≻ 0 - Stacked form: - Y = X B + U - Only reduced-form parameters {B_ℓ}_{ℓ=1}^p and Σ_u are needed for identification and impulse-response analysis. ### Proxy-first (Cholesky) identification and impulse responses - Cholesky factorization: - Σ_u = P P′ with P lower triangular and positive diagonal - Orthogonalized innovations defined by u_t = P ε_t, E[ε_t] = 0, E[ε_t ε′_t] = I_n - Identified fiscal shock: - ε_{1t} — the first orthogonalized innovation (standardized innovation in the narrative-proxy equation) - Because P_{11} > 0, a positive ε_{1t} corresponds to a positive impact innovation in z_t — this fixes shock normalization relative to the proxy series. - Orthogonalized impulse responses: - Θ_h(B, Σ_u) is the n×n matrix of orthogonalized impulse responses at horizon h computed from companion form and impact matrix P. - Response of variable j to the identified fiscal shock at horizon h: θ_{j,h} = e′_j Θ_h(B, Σ_u) e_1 - Population equivalence: - Interpreting ε_{1t} as the government-spending shock yields the same population relative impulse-response estimand as IV local projections that use z_t as an instrument (Plagborg-Møller and Wolf, 2021; Li et al., 2024; Stock and Watson, 2018). ### Admissibility conditions and posterior truncation - Economically motivated admissibility conditions are imposed on the implied spending response and the resulting multipliers to avoid inferential artifacts: - Purpose: to avoid drawing conclusions from posterior draws in which the proxy innovation is effectively uninformative about realized spending (e.g., near-zero spending responses) that produce mechanically extreme multipliers. - Implemented operationally as an accept–reject filter (posterior truncation) applied to impulse responses computed under the proxy-first ordering. - Implementation in estimation: - Draw (B, Σ_u) from a Bayesian reduced-form VAR with Minnesota priors (normal-inverse-Wishart). - Compute corresponding impulse responses and multipliers. - Retain a draw if and only if the admissibility inequalities hold (equivalent to conditioning/truncating the posterior on the set of admissible reduced-form parameters). - Notes and robustness: - Admissibility conditions are chosen ex ante based on the existing multiplier literature and applied symmetrically across countries; they do not modify the underlying proxy-based identification of the shock. - Robustness tests to alternative admissibility tests are conducted. - Bayesian estimation relies on a Gaussian working likelihood, potentially misspecified given the narrative proxy is discrete; Appendix B compares to a frequentist counterpart and finds minor differences. ### Objects of interest and summary measures - Orthogonalized impulse responses (dIRF_y(h), dIRF_g(h)) are in log points at horizon h to the proxy-identified shock ε_{1t}. - Level approximations for small responses: - ΔY_{t+h} ≈ Ȳ · dIRF_y(h) - ΔG_{t+h} ≈ Ḡ · dIRF_g(h) - Ȳ and Ḡ are sample averages (or pre-shock means) over the estimation window. - Cumulative spending multiplier at horizon h: - cM_h = (Σ_{j=0}^h ΔY_{t+j}) / (Σ_{j=0}^h ΔG_{t+j}) ≈ (Ȳ / Ḡ) · (Σ_{j=0}^h dIRF_y(j)) / (Σ_{j=0}^h dIRF_g(j)) = (1 / ȳ_s_G) · (Σ_{j=0}^h dIRF_y(j)) / (Σ_{j=0}^h dIRF_g(j)) - where ȳ_s_G = (Ḡ / Ȳ) is the average spending to GDP ratio for a given country. - Summaries reported: - Impulse responses and multipliers summarized using posterior medians and, when reported, posterior uncertainty bands computed from accepted posterior draws. _Italic: Source — wpiea2026043-source-pdf, section 4.2 (Econometric Identification Strategy) and accompanying text._ ### 5.2 A panel proxy-SVAR framework with a qualitative instrument ### 5.2 A panel proxy-SVAR framework with a qualitative instrument ### Panel VAR with fixed effects - Setup: - i ∈ {1,...,N} indexes countries and t ∈ {1,...,T_i} quarters (panel may be unbalanced). - x_{i,t} ≡ (g_{i,t}, y_{i,t})′ collects transformed real government spending and real GDP. Transformation (e.g., ∆ log or log levels) is application-specific and always stated with results. - For a set of observations S, estimate pooled panel VAR with fixed effects: x_{i,t} = Σ_{ℓ=1}^p A_{ℓ} x_{i,t−ℓ} + μ_i + τ_t + u_{i,t}, x_{i,t} = (g_{i,t}, y_{i,t})′. (Equation (10)) - μ_i are country fixed effects, τ_t are time fixed effects (included when indicated), and u_{i,t} are reduced-form innovations. - Estimation by least squares after removing included fixed effects (within transformation). Estimation sample: all (i,t)∈S for which p lags of x_{i,t} are available. - Remark (state dependence and interactions): - Law of motion for x_{i,t} may depend on predetermined state variables via sample restrictions (state-specific subsamples) or interactions (e.g., s_{i,t} x_{i,t−ℓ}). Variants remain linear in parameters and estimated by least squares after fixed-effect removal. Conditional on specified VAR, proxy identification, impulse-response construction, and inference remain unchanged. ### Structural representation and proxy-SVAR identification - Structural representation: - Let u_{i,t} denote reduced-form innovations from estimated panel VAR. Postulate u_{i,t} = B ε_{i,t}, E[ε_{i,t} ε′_{i,t}] = I. - Define spending shock as first structural innovation ε^g_{i,t} with impact vector b ≡ B e_1 (so u_{i,t} = b ε^g_{i,t} + other shocks). - z_{i,t} denotes the narrative proxy. - Proxy moments and weights: - Allow nonnegative weights w_{i,t} (encode subgroup membership, state weights, or proxy missingness). - Construct effective proxy v_{i,t} as scalar instrument from z_{i,t}. Residualize v_{i,t} (and any instrument components) with respect to same fixed effects included in VAR; denote residual by ev_{i,t}. - Assumption PS (proxy relevance and exogeneity): - For environment indexed by (S,w), effective proxy satisfies: - (i) relevance: E[w_{i,t} ev_{i,t} ε^g_{i,t}] ≠ 0; - (ii) exogeneity: E[w_{i,t} ev_{i,t} ε^j_{i,t}] = 0 for all structural shocks j ≠ g. - Identification (impact vector up to scale): - b ∝ E[w_{i,t} u_{i,t} ev_{i,t}]. (Equation (11)) - In practice compute sample analogue using VAR residuals bu_{i,t} and residualized proxy ev_{i,t} on (i,t)∈S: - bc ≡ b E_S[w_{i,t} bu_{i,t} ev_{i,t}], - b̂ ≡ bc / bc_g, where bc_g denotes first element of bc (weighted covariance with spending residual). - Normalization sets spending impact response to unity, b_g = 1. - Two-variable covariance ratio representation: - With two endogenous variables: b̂ = (1, dCov_{S,w}(bu_y_{i,t}, ev_{i,t}) / dCov_{S,w}(bu_g_{i,t}, ev_{i,t}))′. (Equation (12)) - dCov_{S,w}(·,·) denotes weighted sample covariance over (i,t)∈S. - (Equivalently, (12) is the just-identified IV estimand obtained by instrumenting bu_g_{i,t} with ev_{i,t} in a regression of bu_y_{i,t} on bu_g_{i,t}.) ### Instrument processing for a qualitative and sparse proxy - Raw narrative proxy z_{i,t} is qualitative and sparse. Allow effective proxy v_{i,t} to be constructed scalar from multivariate instrument vector q_{i,t} (function of z_{i,t}, e.g., sign indicators and lags). - Let q_{i,t} be K×1 vector of instrument components. Let eq_{i,t} denote residualized version after removing same fixed effects as VAR. - Strengthen relevance via first-stage projection within S: - bu^g_{i,t} = π′ eq_{i,t} + e_{i,t}. (Equation (13)) - Define fitted component v_{i,t} ≡ b π′ eq_{i,t}; ev_{i,t} ≡ v_{i,t} (already residualized). - Two-step construction: - Algebraically equivalent to using multivariate instrument eq_{i,t} directly in proxy-SVAR moments, convenient in implementation and yields scalar effective proxy. - Zeros and missing proxy information: - Two conventions: - (i) zeros-as-zero: treat z_{i,t} = 0 as literal value contributing to proxy moment; - (ii) zeros-as-missing: treat z_{i,t} = 0 as containing no instrument information by setting w_{i,t} = 0 when z_{i,t} = 0, while retaining those quarters in VAR estimation. - Each application states which convention is used. ### Impulse responses and multiplier construction - Let {Â_{ℓ}}_{ℓ=1}^p denote estimated VAR coefficients (or state-dependent counterparts). Moving-average matrices {Ψ̂_h}_{h=0}^H with Ψ̂_0 = I and recursion Ψ̂_h = Σ_{ℓ=1}^{min{p,h}} Â_{ℓ} Ψ̂_{h−ℓ}. - Impulse responses to identified spending shock: - dIRF(h) = Ψ̂_h b̂, h = 0,1,...,H. (Equation (14)) - Map impulse responses into dollar multipliers using average spending share computed from underlying levels data: - s_{i,t} ≡ G_{i,t}/Y_{i,t}. - Environment-specific average spending share as weighted trimmed mean: s̄ ≡ TrimMean_{S,w}(s_{i,t}). - Trim rate and construction of quarterly G/Y (including any smoothing/interpolation) recorded in replication code. - Multiplier formulas depend on VAR transformation: - Benchmark ∆ log representation: impulse responses are for growth rates and are cumulated over horizons. - Let dIRF_g(h) and dIRF_y(h) denote spending and output responses in (14). - Cumulative multiplier at horizon H: cM(H) = (1 / s̄) [Σ_{h=0}^H dIRF_y(h)] / [Σ_{h=0}^H dIRF_g(h)]. (Equation (15)) - For VAR in log levels, use level-log impulse responses at horizon H (without cumulation); each application reports relevant formula when log levels are used. ### Bootstrap inference - Country-cluster bootstrap preserving within-country serial dependence: - For each replication: resample countries with replacement, reconstruct bootstrap panel by stacking full time series of sampled countries, re-estimate panel VAR (or variant), rebuild effective proxy (including first-stage (13) when used), recompute impact vector (12), impulse responses and multipliers (14)–(15). - Reported intervals: 90% bootstrap percentile bands. - Two-sided bootstrap p-values for H_0: M(H) = 0 computed from empirical distribution of bootstrap replicates. - Practical filter for ratio instability: - Because (15) is a ratio, small number of bootstrap draws can yield unstable values when cumulative spending response in denominator is close to zero. - Apply mild acceptance filters excluding draws with near-zero denominators or implausibly extreme multipliers. Specific bounds stated with results and are not consequential for reported patterns. ### Implementation in this section: pooled benchmarks within groups - For a given characteristic, let G denote corresponding subsample of country–quarter observations. Within each group G, set S = G and estimate two-way fixed-effects panel VAR (10) on stacked country–quarter sample. - For Table 12: - Benchmark transformation: quarterly log changes, g_{i,t} = ∆ log G_{i,t} and y_{i,t} = ∆ log Y_{i,t}. - Identification uses narrative proxy z_{i,t} ∈ {−1, 0, +1} via effective proxy construction in Section 5.2.3: - Form q_{i,t} from positive/negative narrative indicators and a small number of their lags. - Estimate first stage (13) within G after removing same fixed effects, and use fitted component v_{i,t} in (12). - Benchmark adopts zeros-as-zero convention for z_{i,t}. - Reported statistics: - Cumulative multipliers at one- and two-year horizons: cM_G(4) and cM_G(8), computed from (15) and scaled by group-specific trimmed-mean spending share s̄_G = TrimMean_G(G/Y). - Uncertainty bands: 90% country-cluster bootstrap percentile intervals constructed as in Section 5.2.5. ### Benchmark results (Table 12: Benchmark Panel Heterogeneity by Structural Characteristics) - Panel A. Trade openness - Closed (26 countries): M_4 = 1.425 * * , M_8 = 1.364 * * ; 90% intervals: (0.437, 1.993) and (0.418, 1.935) - Open (34 countries): M_4 = 0.558 * * , M_8 = 0.649 * * ; 90% intervals: (0.075, 1.094) and (0.088, 1.165) - Panel B. Public debt - Low/No high debt (56 countries): M_4 = 0.738 * * , M_8 = 0.893 * * ; 90% intervals: (0.118, 1.481) and (0.146, 1.669) - High debt (26 countries): M_4 = 0.821 * * , M_8 = 0.851 * * ; 90% intervals: (0.185, 1.856) and (0.188, 2.008) - Panel C. Exchange-rate regime split - Fixed (51 countries): M_4 = 1.104 * * * , M_8 = 1.119 * * * ; 90% intervals: (0.372, 1.624) and (0.379, 1.653) - Float (8 countries): M_4 = 0.715, M_8 = 0.703 ; 90% intervals: (-0.053, 1.881) and (-0.019, 1.745) - Panel D. Informality split - Low informality (28 countries): M_4 = 1.173 * * , M_8 = 1.195 * * * ; 90% intervals: (0.403, 1.504) and (0.434, 1.494) - High informality (28 countries): M_4 = 0.788 * * , M_8 = 0.798 * * * ; 90% intervals: (0.092, 1.669) and (0.097, 1.687) - Panel E. Labor flexibility - Low (25 countries): M_4 = 1.439 * * * , M_8 = 1.347 * * * ; 90% intervals: (0.555, 1.796) and (0.523, 1.613) - High (24 countries): M_4 = 1.560 * * , M_8 = 1.662 * * ; 90% intervals: (0.410, 1.993) and (0.427, 2.140) - Notes on group assignments and construction: - Trade openness: country classified as open if median trade-to-GDP ratio (over estimation sample) exceeds 60 (percent). - Public debt: high-debt state defined as country–quarters belonging to episode with debt-to-GDP > 70 (percent). - Exchange-rate regime: country classified as floating if share of quarters coded as floating (float_dummy = 1) exceeds 0.25. - Informality: country classified as high informality if its median informality measure is above overall median; median cutoff reported as 25.25 in text. - Labor-market flexibility: country classified as high flexibility if median labor-reform index exceeds overall median; median cutoff reported as 0.73 in text. - Entries report cumulative multipliers as ratio of cumulative response of real GDP to cumulative response of real government spending, scaled by group-specific average spending share (Equation (9) in source). Parentheses report 90% country-cluster bootstrap confidence intervals. - *, **, *** denote rejection of H_0: M = 0 at the 10%, 5%, and 1% levels, respectively. ### Five benchmark findings (summary of Table 12 results) - Trade openness: - Multipliers are smaller in more open economies, reproducing IMV regularity. - One-year multiplier: 1.425 (closed) versus 0.558 (open). Two-year: 1.364 (closed) versus 0.649 (open). - Interpretation: consistent with import-leakage attenuating domestic output response in open economies. - Public debt: - Limited debt-state heterogeneity in pooled benchmark. - One- and two-year multipliers similar across debt states: M_4 = 0.738 (low/no-high-debt) versus 0.821 (high-debt); M_8 = 0.893 versus 0.851. - In pooled design, multipliers vary more sharply with openness and exchange-rate regime than with debt state. - Exchange-rate regime: - Multipliers larger under fixed exchange-rate arrangements (canonical IMV benchmark). - Fixed regimes: M_4 = 1.104 and M_8 = 1.119 (both statistically different from zero). - Floating regimes: M_4 = 0.715 and M_8 = 0.703, but inference weaker due to small floating-regime subsample. - Key regularity: ranking fixed > float is strongly supported. - Informality: - Multipliers smaller in more informal economies. - One-year multiplier declines from 1.173 (low-informality) to 0.788 (high-informality); similar gap at two years (1.195 versus 0.798). - Informality emerges as quantitatively important for fiscal transmission in pooled benchmark. - Labor-market flexibility: - Multipliers larger in countries with more flexible labor markets. - Gap modest at one year (1.560 versus 1.439) and more pronounced at two years (1.662 versus 1.347). ### Lessons so far - Pooled panel benchmarks reproduce canonical IMV regularities: - Smaller multipliers in more open economies. - Larger multipliers under fixed exchange-rate arrangements. - Additional concordant results with literature linking multipliers to informality and labor market reforms. - These results provide an external-validity check for the narrative database used. - Next focus: time-varying state dependence — whether effectiveness of exogenous spending actions varies with higher-frequency macroeconomic states: - (i) slack and monetary-policy constraints around the effective lower bound, - (ii) alternative measures of policy uncertainty, - (iii) political support and cohesion. ### 6 State of the business cycle ### Motivation - Literature documents spending multipliers tend to be larger during recessions and when increased spending is followed by monetary policy accommodation (e.g., Ramey and Zubairy (2018) and cited literature). - This section asks whether causal effects of narrative discretionary spending actions vary systematically with growth conditions. - Use continuous, predetermined growth index and estimate smooth-transition panel proxy-SVAR; useful because narrative proxy is noisy and sample splitting can sharply reduce precision. ### 6.1 Smooth-transition panel proxy-SVAR under growth conditions - Framework applies panel proxy-SVAR from Section 5.2 with two modifications relative to pooled benchmark: - Allow panel VAR dynamics to vary smoothly with predetermined growth index. - Implement proxy identification using regime weights (high- versus low-growth) in proxy moment and construction of environment-specific spending shares. - Growth index and smooth regime weights: - Let Y_{i,t} denote real GDP and define quarterly output growth ∆ log Y_{i,t}. - Summarize local growth conditions by lagged four-quarter moving average: s_{i,t} ≡ (1/4) Σ_{r=1}^4 ∆ log Y_{i,t−r}. (Equation (16)) - s_{i,t} is predetermined with respect to quarter-t shocks. - Standardize s_{i,t} within country: ês_{i,t} ≡ [s_{i,t} − E(s_{i,·})] / sd(s_{i,·}). - Map ês_{i,t} into smooth low-growth weight F_{i,t} ∈ (0,1) using logistic transition: F_{i,t} ≡ 1 / [1 + exp{γ(ês_{i,t} − c)}], γ > 0. (Equation (17)) - F_{i,t} larger in weak-growth environments. - Benchmark: c = 0 and γ = 1.5, so F_{i,t} ≈ 0.5 at average standardized growth and transitions smoothly across growth conditions. - Reporting evaluates impulse responses and multipliers at two representative values of F_{i,t}: the 25th and 75th percentiles of sample distribution. In the data these correspond to: - F_{HighGrow} = 0.3167 and F_{LowGrow} = 0.6399 (denoted “HighGrowth_p25” and “LowGrowth_p75”). - Smooth-transition panel VAR: - z_{i,t} ∈ {−1, 0, +1} denotes narrative spending-action proxy. - Estimate pooled smooth-transition panel VAR with country fixed effects. Endogenous vector: x_{i,t} ≡ [x^g_{i,t}, x^y_{i,t}]′, where either (x^g_{i,t}, x^y_{i,t}) = (∆ log G_{i,t}, ∆ log Y_{i,t}) or (log G_{i,t}, log Y_{i,t}), with G_{i,t} real government spending and Y_{i,t} real GDP. - With p = 8 lags, specification: x_{i,t} = α_i + Σ_{ℓ=1}^p A_{ℓ} x_{i,t−ℓ} + Σ_{ℓ=1}^p B_{ℓ} [F_{i,t} x_{i,t−ℓ}] + u_{i,t}. (Equation (18)) - α_i are country fixed effects. u_{i,t} are reduced-form residuals. - When F_{i,t} = 0 (high growth) lag coefficients equal A_{ℓ}; when F_{i,t} = 1 (low growth) they equal A_{ℓ} + B_{ℓ}. - Estimation by least squares after removing α_i. - Proxy identification with regime weights: - Identification follows Section 5.2.2 using narrative proxy as external instrument and allowing proxy moment to vary with growth environment via weights. - Define regime weights: - w_{Low,i,t} ≡ F_{i,t}, - w_{High,i,t} ≡ 1 − F_{i,t}. - Zeros-as-missing convention: quarters with z_{i,t} = 0 retained in VAR estimation but contribute no instrument information; operationally replace regime weights by w_{κ,i,t} · 1{z_{i,t} ≠ 0}. - Let ez_{i,t} denote proxy residualized with respect to included country fixed effects. - For each regime κ ∈ {High, Low}, impact vector identified (up to scale) by weighted proxy moment: b_{κ} ∝ E[w_{κ,i,t} u_{i,t} ez_{i,t}], normalized by b_{κ,g} = 1. (Equation (19)) - Thus spending impact response normalized to one; output impact response scaled relative to spending. - Relation to existing literature: - Smooth-transition specification parallels Auerbach and Gorodnichenko (2012) modeling fiscal multipliers varying smoothly with business cycle via logistic transition on lagged output growth. - Key difference: embed smooth-transition within panel proxy-SVAR framework and allow external-instrument moment to vary with regime weights. *Source: IMF staff working paper section 5.2–6.1 (A panel proxy-SVAR framework with a qualitative instrument; Smooth-transition panel proxy-SVAR under growth conditions).* ### 6.2 Multipliers across growth conditions ### 6.2 Multipliers across growth conditions ### Multipliers across growth conditions (specifications and estimation) - Two pooled-panel specifications that are scale-invariant across countries are used: - Benchmark VAR in growth rates (∆ log). - VAR in log levels. - In both specifications z_i,t = 0 is treated as missing in the proxy moment. - Estimation: smooth-transition panel proxy-SVAR with country fixed effects and p = 8 lags. - Growth conditions: - s_i,t is the lagged four-quarter moving average of ∆ log real GDP, standardized within country. - F_i,t is the logistic low-growth weight with γ = 1.5 and c = 0. - “High growth” evaluates the VAR at F_HighGrow = 0.3167 (25th percentile of F). - “Low growth” evaluates the VAR at F_LowGrow = 0.6399 (75th percentile of F). ### Key findings: multipliers by growth condition - Under the ∆ log specification: - High-growth multiplier at H = 4: −0.01. - Low-growth multiplier at H = 4: 0.77. - High-growth multiplier at H = 8: −0.01. - Low-growth multiplier at H = 8: 0.77. - Under the log-level specification: - H = 4: High growth = 0.25; Low growth = 0.74. - H = 8: High growth = 0.58; Low growth = 1.39. - Two patterns: - Point estimates suggest stronger multipliers in weak-growth (low-growth) environments. - Differences are suggestive but imprecisely estimated: - For the benchmark ∆ log specification, H = 4 (bootstrap p = 0.11) and H = 8 (p = 0.14). - For the log-level specification, H = 4 (p = 0.39) and H = 8 (p = 0.24). - Table 13 (cumulative real spending multipliers): - 1-year (H = 4): Real (∆ log VAR) High growth −0.01; Low growth 0.77. - 1-year (H = 4): Real (log-level VAR) High growth 0.25; Low growth 0.74. - 2-year (H = 8): Real (∆ log VAR) High growth −0.01; Low growth 0.77. - 2-year (H = 8): Real (log-level VAR) High growth 0.58; Low growth 1.39. --- ### 7 Policy uncertainty and fiscal multipliers ### Uncertainty measures (construction) - Source: EIU Country Reports; broad WUI and fiscal-policy-focused FUI (text-mined). - WUI: World Uncertainty Index computed by counting frequency of the word “uncertain” (and variants) in EIU reports, scaled by report length and rescaled. - FUI: same text-mining approach restricted to fiscal-policy and public-finance passages. - High/low uncertainty states: - For each index U^m_{i,t}, compute four-quarter moving average eU^m_{i,t} ≡ 1/4 ∑_{s=0}^3 U^m_{i,t−s}. - Binary high-uncertainty indicator Unc^m_{i,t} ∈ {0,1} using 75th-percentile thresholds. - Benchmark WUI: threshold computed within-country: Unc_WUI_{i,t} ≡ 1{ eU_WUI_{i,t} ≥ Q^0.75(eU_WUI_{i,·}) }. - Benchmark FUI: pooled 75th-percentile rule applied to eU_FUI_{i,t}. - In the data, roughly one quarter of observations are classified as high uncertainty for each index (among defined observations). ### State-contingent panel proxy-SVAR (estimation) - For each uncertainty concept m∈{WUI, FUI} and each state κ∈{Low_m, High_m}, estimate pooled panel VAR on S_{κ,m}. - Benchmark endogenous panel vector x_{i,t} ≡ [∆ log g_{i,t}, ∆ log y_{i,t}]′ with p = 4 lags and country fixed effects. - Proxy identification within state: proxy v_{i,t} = z_{i,t} (residualized); impact vector b_{κ,m} ∝ E[u_{κ,m,i,t} z_{i,t}] normalized by b_{κ,m,g} = 1. - Cumulative multipliers constructed in dollars using cumulative-response mapping with state-specific trimmed mean spending share s̄^G_{κ,m}. ### Multipliers across uncertainty states (key results) - Table 14 (cumulative multipliers at H = 4 and H = 8): - WUI: - H = 4: Low = 1.04; High = 0.51; High−Low = −0.52. - H = 8: Low = 1.09; High = 0.42; High−Low = −0.67. - FUI: - H = 4: Low = 0.95; High = 0.21; High−Low = −0.74. - H = 8: Low = 1.13; High = 0.18; High−Low = −0.94. - Two patterns: - For WUI, spending multipliers are lower in high-uncertainty quarters than in low-uncertainty quarters at both horizons (H = 4 and H = 8). - For FUI, the preferred specification also shows substantially smaller cumulative multipliers in high-uncertainty quarters at both horizons. - Precision: - Bootstrap-based uncertainty intervals for the high-minus-low differences are wide and include zero, reflecting limited precision when splitting the sample. --- ### 8 Political support and fiscal multipliers ### Political support measures and coding - Narrative political-environment indicator Supp_{i,t} ∈ {0,1} from EIU Country Reports: - Supp_{i,t} = 1 indicates a broadly favorable political backdrop; Supp_{i,t} = 0 indicates unfavorable. - Coding performed with a fixed, pre-specified large-language-model prompt; model returns verbatim excerpts retained for auditability. - Political primitives coded as {Yes, No, Unclear} with verbatim evidence: - (i) majority/cohesion (working legislative majority or stable coalition), - (ii) election proximity (national election scheduled within the current or next quarter), - (iii) unrest or severe gridlock (material unrest or legislative paralysis). - Missing political text is treated as its own state (no imputation). ### Empirical design: pooled panel local projections with IV - Narrative proxy z_{i,t} instruments within-country change in government spending. - Define ∆g_{i,t} ≡ ∆ ln G_{i,t}; ∆y_{i,t} ≡ ∆ ln Y_{i,t}. - Cumulative changes from t−1 to t+h: - ∆^h y_{i,t} ≡ ln Y_{i,t+h} − ln Y_{i,t−1}. - ∆^h g_{i,t} ≡ ln G_{i,t+h} − ln G_{i,t−1}. - Political-state indicators: - S_1_{i,t} ≡ 1{Supp_{i,t} = 1}. - S_m_{i,t} ≡ 1{Supp_{i,t} missing}. - Omitted baseline state is Supp_{i,t} = 0. - Pooled LP-IV equation (for each horizon h): - ∆^h y_{i,t} = α_i + τ_t + β_h ∆g_{i,t} + θ^1_h (∆g_{i,t} × S_1_{i,t}) + θ^m_h (∆g_{i,t} × S_m_{i,t}) + Γ_h X_{i,t−1} + ε_{i,t+h}. - Instruments: {z_{i,t}, z_{i,t} × S_1_{i,t}, z_{i,t} × S_m_{i,t}}. - X_{i,t−1} includes four lags of ∆y_{i,t} and ∆g_{i,t} and standard macro controls (inflation, unemployment, policy rate, and a below-trend indicator). - Standard errors clustered at the country level. - Cumulative multiplier at horizon h: - b̂^m_h(s) ≡ (1 / s^G) × [ĉRF^y_h(s) / ĉRF^g_h(s)], s ∈ {0,1,m}, where s^G is the sample median of G/Y. - Suppress b̂^m_h(s) when |ĉRF^g_h(s)| < 0.005. ### Results: political support amplifies spending pass-through and multipliers - Table 15 (pooled LP-IV cumulative multipliers at h ∈ {4, 8}): - h = 4: - Support = 0: (suppressed; cumulative spending reduced form |ĉRF^g_h(0)| < 0.005). - Support = 1: 1.0590 (note: table entry shows as "1.0590" paired with "1.1078" formatting; reported text: at horizons of one to three years the scaled dollar multiplier is approximately 1.0–1.1). - Support missing: 1.1078 (table formatting presents entries as "h = 4–1.05901.1078"; interpret as Support=0 suppressed, Support=1 1.0590, Support missing 1.1078). - h = 8: - Support = 0: 0.1880. - Support = 1: 0.9627. - Support missing: −0.6868. - Notes: multipliers are suppressed (“–”) when |ĉRF^g_h(s)| < 0.005; suppression occurs for Support=0 at h = 4. - Interpretation: - In favorable political environments (Supp_{i,t} = 1), multipliers are consistently positive and economically meaningful: at horizons of one to three years the scaled dollar multiplier is approximately 1.0–1.1. - Under unfavorable support (Supp_{i,t} = 0), multipliers are close to zero (and occasionally slightly negative). - Implementation channel (Table 16: reduced-form evidence) - h = 4: - RF^y_4(0) = 0.00048; RF^y_4(1) = 0.00524; ∆RF^y_4(1−0) = 0.00476 (t = 1.71). - RF^g_4(0) = −0.00330; RF^g_4(1) = 0.01490; ∆RF^g_4(1−0) = 0.01819 (t = 2.94). - h = 8: - RF^y_8(0) = 0.00061; RF^y_8(1) = 0.00621; ∆RF^y_8(1−0) = 0.00560 (t = 1.45). - RF^g_8(0) = 0.00979; RF^g_8(1) = 0.01943; ∆RF^g_8(1−0) = 0.00964 (t = 1.65). - Notes: ∆RF_g(1 − 0) = 0.01819 at h = 4 has t = 2.94, indicating a statistically significant difference in the cumulative spending reduced form between support and no-support quarters at medium horizons. - Conclusion: - The main driver of political-state dependence is the realized spending response: high-support quarters translate the narrative shock into a substantially larger cumulative spending response, producing positive output responses. Low-support quarters show small and unstable cumulative spending reduced forms at short horizons, making multipliers not well-defined there. *Source: IMF staff calculations and tables as presented in the supplied chapter section.* ### 8.4 Which political dimension matters? Elections attenuate spend- ### 8.4 Which political dimension matters? Elections attenuate spending pass-through ### Empirical approach - Estimated reduced-form pooled local projections for cumulative spending ∆h g i,t on z i,t interacted with: - observed support (S1 i,t), - primitive indicators (majority/cohesion, election proximity, unrest/gridlock), - controls for missingness and “unclear” classifications, - country and time fixed effects, and baseline macro controls. - Sample restricted to observed support (Support ∈ {0, 1}). - t-statistics in parentheses are clustered by country. ### Key findings on political dimensions and spending pass-through - Favorable political support is associated with a significantly stronger cumulative spending response at medium horizons even after controlling for primitives: - Coefficient on z i,t × S1 i,t in the spending reduced form is about 0.02 at h = 4 and h = 8 (both t ≈ 2.2). - Election proximity is independently associated with substantially weaker spending pass-through: - Coefficient on z i,t × ElectYes i,t in the spending reduced form is around −0.018 to −0.020 at h = 4 and h = 8 (both statistically significant). - The magnitudes of the support advantage and the election proximity penalty are comparable, implying that election proximity can nearly offset the implementation advantage associated with political support. ### Numerical estimates (Table 17: Decomposing political support into primitives: reduced-form spending pass-through) - Horizon h = 4: - z × S1 in RF g: 0.02044 (2.17) - z × ElectYes in RF g: -0.01956 (-2.20) - Net: (z × S1) + (z × ElectYes): 0.00088 - Horizon h = 8: - z × S1 in RF g: 0.01955 (2.26) - z × ElectYes in RF g: -0.01834 (-2.75) - Net: (z × S1) + (z × ElectYes): 0.00121 ### Robustness on shock composition - Verified that the sign composition of narrative shocks is essentially identical across observed support states: - Share of z = −1 shocks is 0.551 when Support = 0 and 0.546 when Support = 1. - Interpretation: support-state results are not driven by simple sign selection across shocks. ### Implication for state-dependent fiscal transmission - State dependence operates primarily through implementation: narrative spending shocks translate into a substantially stronger realized government-spending response when political support is high, with correspondingly larger output responses. - Election timing can attenuate the effectiveness of discretionary spending actions by weakening spending pass-through. *Source: Authors’ calculations (Table 17 and accompanying text).* ### 2. Draw B ### 2. Draw B ### A.3 Impulse responses, multipliers, and accept–reject restrictions - For each draw (B^(s), Σ_u^(s)), compute Cholesky factor P^(s) from Σ_u^(s) = P^(s) P^(s)′ and orthogonalized impulse responses and multipliers using the companion-form representation implied by (6) and the impact matrix P^(s). - Denote θ_j,h^(s) as the response of variable j at horizon h to shock ε_1t. - Apply proxy-based ordering supplemented with weak admissibility conditions to rule out draws where the identified innovation produces negligible spending responses and economically implausible (small-denominator) multipliers. - Formal screening conditions: - Let g be the index of government spending and y the index of real GDP. - Impose a weak lower bound on the average short-run spending response: - 1/(H_g + 1) Σ_{h=0}^{H_g} θ_{g,h}(B, Σ_u) ≥ g_. - Screen multipliers by requiring a positive cumulative spending response at the screening horizons and restricting cumulative multipliers to lie in pre-specified ranges at a short horizon H_S and a long horizon H_L: - M(H;B, Σ_u) = (1/ s̄_G) [Σ_{h=0}^H θ_{y,h}(B, Σ_u)] / [Σ_{h=0}^H θ_{g,h}(B, Σ_u)]. - Baseline specification: - Average response of government spending constrained to be above 0 for the first 8 quarters (H_g = 8, g_ = 0). - Cumulative multipliers constrained to lie within the range [−2, 5]. - Short run horizon H_S = 8; long run horizon H_L = 20. ### A.4 Robustness to alternate multiplier ranges - Sensitivity assessment: widen and tighten baseline interval [−2, 5] by 20%, yielding: - Wider bounds: [−2.4, 6] - Tighter bounds: [−1.6, 4] - Holding the spending-response condition fixed (H_g = 8, g_ = 0). - Results: - Pooled IVW impulse responses (figure 6) are nearly indistinguishable from baseline; substantial overlap in posterior bands across all horizons. - Acceptance rates: - Baseline specification: 47% - Wider bounds [−2.4, 6]: 50% - Tighter bounds [−1.6, 4]: 43% - Conclusion: inference is not driven by exact choice of multiplier endpoints; results robust to economically reasonable perturbations of admissible multiplier range. ### B Comparing Bayesian and frequentist estimates - Motivation: examine whether Gaussianity assumption implicit in Bayesian estimates holds in practice; frequentist estimation does not impose parametric distribution on innovations and uses bootstrapping of empirical residuals. - Procedure: - Estimate baseline three-variable Bayesian VAR with priors approximating a flat (diffuse) prior. - Estimate frequentist counterpart via OLS, resample reduced-form residuals to generate artificial data, re-estimate VAR on simulated samples; retain bootstrap draws only if they satisfy same stability and admissibility constraints as Bayesian analysis. - Single-country focus: United States. - Finding: IRF distributions under Bayesian and frequentist procedures are nearly identical (Figure 7a). - Frequentist Minnesota counterpart: - Estimate penalized least squares using Minnesota-style ridge penalty; resample residuals and re-estimate penalized VAR on simulated samples; retain bootstrap draws only if they satisfy same constraints. - Finding: Bayesian and frequentist Minnesota IRFs are very similar (Figure 7b). - Conclusion: Gaussianity assumption underlying Bayesian inference provides a good approximation in this application. ### C Benchmark pooled panel heterogeneity: specification, splits, and inference - Objective: mirror pooled quarterly perspective of Ilzetzki et al. (2013) while remaining consistent with identification that relies on a qualitative quarterly narrative proxy used as an external instrument. - C.1 Panel VAR specification and estimation - For each split, estimate pooled panel VAR on stacked country–quarter observations in group G: - x_{i,t} = Σ_{ℓ=1}^p A_ℓ x_{i,t−ℓ} + μ_i (+ τ_t) + u_{i,t}, x_{i,t} = (g_{i,t}, y_{i,t})′. - Country fixed effects μ_i always included; time fixed effects τ_t included in benchmark for Table 12 and varied in robustness checks. - Estimate by least squares after removing fixed effects (within transformation); with two-way fixed effects, demeaning over country and time. - Estimation sample is unbalanced; use all available country–quarter observations that permit p lags of x_{i,t}. - Benchmark transformations: quarterly log changes (∆ log) for both spending and output: - g_{i,t} = ∆ log G_{i,t}; y_{i,t} = ∆ log Y_{i,t}. - Note: alternative transforms used in robustness include detrended log levels. - Reported multiplier is constructed from estimated impulse responses of transformed variables. - C.2 Proxy-SVAR identification in the pooled panel - Let b u_{i,t} = (b u_{g,i,t}, b u_{y,i,t})′ denote residuals from (28) in group G. - Let z_{i,t} denote quarterly narrative proxy (ordered sign indicator). Identification proceeds group-by-group using z_{i,t} as external instrument for spending shock. - Baseline covariance mapping: impact vector of spending shock proportional to covariance between proxy and reduced-form innovations within group G: - b_G ∝ E_G[u_{i,t} z_{i,t}]. - Sample analogue bc_G = E_G[b u_{i,t} z_{i,t}]; normalize so impact response of spending equals one: - b̄_G = bc_G / bc_{G,g} = (1, dCov_G(bu_{y,i,t}, z_{i,t}) / dCov_G(bu_{g,i,t}, z_{i,t}))′. - Benchmark instrument construction for Table 12: - Positive/negative instrument construction with lags: form indicators for positive and negative narrative innovations and include a small number of lags in first-stage projection for spending residual. - Denote q_{i,t} vector collecting these instrument components. - Estimate within group G after removing fixed effects: - b u_{g,i,t} = π′ q_{i,t} + e_{i,t} - Fitted component b v_{i,t} = b π′ q_{i,t} is effective proxy; replace z_{i,t} by b v_{i,t} in mappings. - Benchmark treats zeros in z_{i,t} as literal zeros (rather than missing). - C.3 Impulse responses and multiplier construction - Given {b A_ℓ}_{ℓ=1}^p and b b_G, compute moving-average coefficients {b Ψ_h}_{h=0}^H and obtain impulse responses: - d IRF_G(h) = b Ψ_h b b_G, h = 0,1,...,H. - Let d IRF_{g,G}(h) and d IRF_{y,G}(h) denote spending and output responses. - Cumulative multipliers at horizon H: - c M_G(H) = (1/ s̄_G) Σ_{h=0}^H d IRF_{y,G}(h) / Σ_{h=0}^H d IRF_{g,G}(h). - Table 12 reports c M_G(4) and c M_G(8), and medium-run summary: - M_{G,4:8} = (1/5) Σ_{H=4}^8 c M_G(H). - Spending share scaling term s̄_G defined as TrimMean_G(G/Y) where trimmed mean drops small fraction of largest and smallest observations. - C.4 Split definitions and cutoffs - Implement heterogeneity by estimating (28) separately on groups defined by structural characteristics; table reports number of countries contributing observations to each group. - Trade openness: - Measured as trade-to-GDP (imports plus exports divided by GDP). - Countries assigned to Closed if openness below cutoff; Open if at or above cutoff. - Public debt: - Measured as debt-to-GDP; debt split defined at country–quarter level. - Classify High debt when debt-to-GDP exceeds cutoff, subject to a minimum-duration rule to avoid transient reclassifications; remaining observations assigned to Low/No high debt group. - Informality: - Measured using country-level informality index coded so higher values = more informality. - Countries split into low- and high-informality groups using cutoff reported in Table 12. - Labor-market liberalization: - Measured using labor reform (liberalization) index coded so higher values = more liberal labor markets. - Countries split into less- and more-liberal groups using cutoff reported in Table 12. - C.5 Bootstrap inference and acceptance - Inference for each group-specific multiplier based on country-cluster bootstrap designed to preserve within-country serial dependence. For each bootstrap replication b: 1. Sample countries with replacement from set of countries in group G. 2. Construct bootstrap panel by stacking full time series for each sampled country (retain original within-country time ordering). 3. Re-estimate (28) and recompute proxy-SVAR mapping (29)–(30) (including first stage (31) when applicable). 4. Recompute impulse responses and multipliers (32)–(33). - For each reported statistic, 90% confidence interval is the bootstrap percentile band. - Bootstrap p-values for testing H_0: M = 0 computed from empirical distribution of bootstrap replicates (two-sided). - Acceptance and plausibility filters: - Exclude bootstrap draws with near-zero denominators or extreme ratios to avoid unstable ratios when cumulative spending response in denominator of (33) is close to zero. - Reported acceptance rate is fraction of bootstrap replications retained after filters. - Table 12 reports results for specifications with high acceptance rates; baseline results not sensitive to reasonable variations in these filters. ### D Building the political support narrative data - Purpose: document fixed prompts, coding criteria, and mapping rules used to construct: - Quarterly indicator political_environment_favorable_indicator ∈{1, 0, null} of whether political environment is favorable for passing and sustaining fiscal measures. - Three primitive components: maj_cohesion, election_proximity, unrest_gridlock. - Pipeline overview: two-stage fixed-prompt non-adaptive pipeline; all model outputs stored with short verbatim evidence excerpts for ex-post auditing. - Stage 1 (support indicator + verbatim evidence) - From each country–quarter EIU Country Report excerpt, a model produces: 1. political_environment_favorable_indicator ∈{1, 0, null} 2. political_environment_favorable_explanation (short narrative 2–4 sentences) 3. political_environment_favorable_evidence: compact verbatim evidence bundle (short quotations copied exactly from EIU excerpt referring to majority/cohesion, election timing, unrest/gridlock) - Also produce public_support ∈{HIGH, LOW, MIXED, UNCLEAR} and public_support_evidence: at most one short phrase (≤ 30 words) copied verbatim; if UNCLEAR return empty string. - Favourable political environment indicator rules: - political_environment_favorable_indicator = 1 if report clearly indicates all three jointly: 1. executive backed by a working majority in legislature (single party or stable coalition); 2. no national election scheduled in current or next quarter; 3. report does not highlight material political unrest, sustained gridlock, or legislative paralysis that threatens government’s capacity to act. - political_environment_favorable_indicator = 0 if report clearly indicates at least one of these conditions fails. - political_environment_favorable_indicator = null if report does not provide sufficient information. - Outputs to return: - public_support and public_support_evidence - political_environment_favorable_indicator and political_environment_favorable_evidence (up to three short phrases ≤ 30 words each) - political_environment_favorable_explanation (2–4 sentences) - Notes: base classifications strictly on provided excerpt; evidence fields must be copied verbatim; prefer MIXED/UNCLEAR/null when information incomplete. - Stage 2 (mechanism primitives) - Input: JSON object where evidence string is concatenation of political_environment_favorable_evidence quotations. - Task: code three primitives using only evidence string: 1. maj_cohesion: Is there a working legislative majority or stable coalition during the quarter? 2. election_proximity: Does a national election occur or is it scheduled within the window [q_start, next_q_end] (current quarter through next quarter)? 3. unrest_gridlock: Is there major unrest or severe political/legislative gridlock that materially hinders policy? - Outputs: JSON with "results": array; each element contains maj_cohesion, election_proximity, unrest_gridlock ∈{YES, NO, UNCLEAR} and corresponding quote fields. - Exact-quote rule: If label is YES or NO, quote must be copied exactly as substring from evidence; if UNCLEAR, quote must be empty string. - Stage 2 does not affect baseline political-environment indicator or fiscal narrative classification; used only for mechanism and diagnostics. - D.4 Mapping to econometric variables - Map political_environment_favorable_indicator_{i,t} ∈{1,0,null} to econometric support indicator Supp_{i,t}: - Supp_{i,t} = 1 if political_environment_favorable_indicator_{i,t} = 1 - Supp_{i,t} = 0 if political_environment_favorable_indicator_{i,t} = 0 - Supp_{i,t} = missing if political_environment_favorable_indicator_{i,t} = null - Baseline panel treats missing support as distinct state; robustness checks may drop missing-support quarters. - Stage 2 outputs primitives maj_cohesion_{i,t}, election_proximity_{i,t}, unrest_gridlock_{i,t} ∈{YES, NO, UNCLEAR} and corresponding quote fields. - In pooled regressions, primitives form indicator variables (e.g., 1{YES}) with additional controls for missing/unclear classifications as described in main text. - D.5 Separation from fiscal narrative coding - Political-context variables never used to classify fiscal actions as exogenous or endogenous. - Narrative exogeneity screen for fiscal measures based solely on stated motivations in fiscal text. - Political variables used only to index environment in which already-identified exogenous spending actions occur and to explore mechanisms through primitives. *Source: AI Meets Fiscal Policy: Mapping Government Spending Actions Across 64 Countries — Working Paper No. WP/2026/043 (2. Draw B).* --- _Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026043-source-pdf.pdf_