## Heterogeneous Spending, Heterogeneous Multipliers — Working Paper No. WP/2023/052 (wpiea2023052-print-pdf)

## Source details

**Canonical URL:** [Heterogeneous Spending, Heterogeneous Multipliers — Working Paper No. WP/2023/052 (wpiea2023052-print-pdf)](https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023052-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2023/english/wpiea2023052-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2023/english/wpiea2023052-print-pdf.pdf.json)

---

### 2.1 Military Spending — Data, classification, geographic allocation, sample
- Data sources and construction:
  - Data harmonized from NARA (1966−2006) and USASpending.gov (2007−2019).
  - Based on DD-350 and DD-1057 military procurement forms covering approximately 96% of contracts awarded by the DoD.
  - Key fields: contract identification number, dates of action and completion, transaction value, place of performance, Product Service Codes (PSCs) or Federal Supply Codes (FSCs).
  - Year of government spending defined as year of the signature date; modifications reported as negative entries; obligations and de-obligations within 0.5% considered null.
  - Contract dollar values reported in nominal terms and converted to real values using the US BLS Consumer Price Index.

- Classification of goods versus services:
  - PSCs (first digit a letter) classified as “Spending in Services.”
  - FSCs (first digit a number) classified as “Spending in Goods.”
  - Existing 4-digit codes consistent from 1979.

- Geographic allocation and mapping:
  - Allocation based on place where contract tasks are performed (place of performance).
  - USASpending.gov provides city and zip code of performing firm used to identify county; NARA reports county.
  - County-level data aggregated to Core-Based Statistical Areas (MSAs) using NBER spatial crosswalks.
  - If performance location unknown but known to be in the US, the location of the contract recipient is assigned as performance location (following Demyanyk et al. (2019)), with adjustments when postal code missing.

- Outcome variables:
  - Main dependent variable: number of employed persons from BLS QCEW.
  - Second outcome: salary and wages (proxy for labor income) from BEA at MSA level; reported nominal and converted to real.
  - Personal income from BEA used to normalize government spending; converted to real.

- Sample construction and filters:
  - Analysis starts in 1979; sample restricted to 2019 (robustness including 2020 yields similar results).
  - Drop contracts missing locality of performance, year signed, product code, or industry of contractor (industry sometimes imputed).
  - After aggregating by MSA, additional filters:
    - Exclude MSAs with incomplete histories in any outcome variable.
    - Exclude observations with growth rates between consecutive periods >100% or <−50%.
    - Remove MSAs with average population <50,000.
    - Include only MSAs with non-negative ratios of aggregate military spending to personal income <1.5.
  - Final sample: procurement contracts allocated to 334 MSAs from 1979 to 2019.
    - Final sample includes 94% of total contracts and aggregate spending.
    - For goods: sample accounts for about 95% of total contracts and spending on goods.
    - For services: sample includes 85% of contracts worth 92% of total spending on services.

- Descriptive statistics and patterns (1979−2019):
  - Aggregate spending evolution: three sharp rises in the 1980s (Reagan buildup), early 2000s (Afghani and Iraqi wars), and recent years (escalations and competition with Russia and China).
  - Reallocation: shift from goods to services between 1980s and 1990s; until early 1980s goods >70% of procurement value; from 1990s split between goods and services with similar shares.
  - Contract/category summary statistics (Table 1):
    - Share of Spending (%): All 54, Goods 46
    - Share of Contracts (%): All 91, Goods 9
    - Average Value (Thousand USD): All 143, Goods 99, Services 293
    - 90%-to-50%-percentile Ratio: All 5, Goods 24, Services 6
    - Share of Spending for Top 1% (%): All 72, Goods 86, Services 50
  - Key distributional findings:
    - DoD awarded 91% of procurement contracts to purchase goods and 9% to services.
    - Average services contract value ≈ three times average goods contract (Goods 99 Thousand USD; Services 293 Thousand USD).
    - Goods contracts have much fatter right-tail: top decile of goods contracts >40× median; for services top decile 11× median.
    - Top 1% of goods contracts accounts for 86% of goods spending; top 1% of services contracts ≈50% of services spending.
    - Spending concentration across recipients and places of performance is high.

- Geographic heterogeneity:
  - Top 30 MSAs account for >30% of spending on goods and >45% of spending on services.
  - Only 35% of MSAs are in the same quartile for both goods and services spending.
  - Midwest: many MSAs top two quartiles for goods spending but few rank high in services; Midwest has highest concentration of production occupations with average employment share in production jobs almost 50% higher than US average.
  - Distribution of spending at contract and MSA level skewed to the right.

### 4.3 Identification assumptions and threats
- Identification assumption:
  - Conditional on MSA and time-fixed effects, the shift-share instrument is not systematically associated with unobserved political or economic characteristics explaining outcomes.
  - Three main threats identified.

- Threat 1 — Measurement error from outsourcing:
  - Contracts assigned to place of performance; use of sub-contractors outside that place creates geographic misallocation.
  - Measurement error induces attenuation bias in local fiscal multiplier estimates.
  - If correlated with spending category (goods more tradeable than services), interpretation could be biased; authors state they rule out importance of this measurement error in Section 6.2.

- Threat 2 — Geographical spillovers (SUTVA violation):
  - Defense spending in an MSA may affect neighboring localities.
  - Positive spillovers → downward bias; negative spillovers (factor price adjustments) → upward bias.
  - Authors provide evidence in Section 6.2 that estimates are biased downward and can be considered lower bounds.
  - Footnote: limited factor mobility implies negative spillovers are small (Chodorow-Reich (2019)).

- Threat 3 — Shift-share instrument drawbacks:
  - Shift-share IV relies on either shares or shifts uncorrelated with unobserved characteristics (Goldsmith-Pinkham et al., 2020; Borusyak et al., 2022; Adao et al., 2019).
  - Authors' defense:
    - Shifts (national changes in military spending) plausibly exogenous because they respond to international geopolitical events.
    - Measuring outcomes and instruments in changes purges instrument from correlation between shares and unobserved local trends.
    - Locality-fixed effects absorb MSA-specific secular trends.
    - Thus variation is deviations from long-term trend plausibly exogenous.

### 6.1 Labor intensity — Concept, measurement, and empirical findings
- Labor-intensity measurement:
  - Defined as contribution of employee compensation to value added using BEA annual data on value-added and employee compensation by industry.
  - Industry-level labor intensity assigned to each contract via contractor’s industry.
  - Industries classified as high- or low-labor intensive using median average labor share across 66 macro-industries (data from 1997 onward); classification held fixed pre-1997.
  - Imputation used for pre-1997 and missing contractor industry information.

- Distributional facts:
  - Largest share of government spending directed to labor-intensive industries.
  - Share of goods spending allocated to non-labor-intensive industries is 50% higher than from services (30% vs. 20%).
  - Aggregate labor intensity AggLI_{i,t} = sum_{j∈J} ω_{i,j,t} LI_j with ω_{i,j,t} = G_{i,j,t} / G_{i,t}; services spending aggregate labor intensity ≈10% higher than goods (except 1990s when gap narrowed).
  - Time-varying industry-specific labor intensity (1997−2019) yields consistent results.

- Empirical specification and instruments:
  - Equation (7) separates spending by category i∈{s,g} and labor-intensity p∈{L,N}; instrumented by shift-share instruments Z_{i,p,l,t+h} = s_{i,p,l} (G_{i,p,t+k} − G_{i,p,t−1}) / Y_{l,t−1}.

- Main empirical findings on employment and labor income:
  - Spending in labor-intensive industries generates positive, highly significant effects:
    - Employment multipliers for labor-intensive services spending range between 0.36% and 1.5%.
    - Employment multipliers for labor-intensive goods spending range between 0.07% and 0.54%.
  - Spending in non-labor-intensive industries has negative effects:
    - Negative and significant for goods spending in non-labor-intensive industries.
    - Negative or non-significant for services spending in non-labor-intensive industries (less precise due to small share).
  - Labor income effects mirror employment effects: positive for labor-intensive spending; negative or non-significant for non-labor-intensive spending.
  - Net aggregate goods spending effect: labor- and non-labor-intensive industry effects largely cancel, producing a small aggregate goods multiplier.

- Robustness: tradeability and geographic spillovers (Section 6.2)
  - Concern: goods more tradeable → production outside MSA → measurement error and spillovers.
  - Two empirical strategies:

  1) State-level aggregation (balanced panel: 1,800 observations, 50 states excluding DC, period 1979−2019; Table 2):
    - Employment — Services Spending coefficients (t to t+4): 0.845, 2.736***, 2.989**, 3.430***, 2.846***.
    - Employment — Goods Spending coefficients: 0.240, 0.624, 0.650, 1.150, 1.295*.
    - GDP — Services Spending coefficients: 0.921, 4.882***, 6.525**, 7.897***, 7.154***.
    - GDP — Goods Spending coefficients: -0.113, -0.160, -0.164, -0.110, -0.188.
    - Interpretation: no effect of goods spending on employment or GDP at state level; services spending yields sizable positive effects. State-level estimates larger than MSA-level, consistent with within-state subcontracting attenuating MSA estimates.

  2) Neighboring-locations (50-mile radius) analysis (balanced panel: 6,948 observations, 195 MSAs, period 1979−2019; Table 3):
    - Employment — Services Spending coefficients (t to t+4): 0.109, 0.262***, 0.386**, 0.573*, 0.525*.
    - Employment — Goods Spending coefficients: 0.070, -0.008, -0.217, -0.366, -0.277.
    - Labor Income — Services Spending coefficients: 0.226*, 0.492***, 0.710***, 1.006**, 0.951**.
    - Labor Income — Goods Spending coefficients: -0.070, -0.120, -0.476, -0.709, -0.575.
    - Interpretation: no positive spillovers of goods spending to neighboring localities; services spending exhibits positive outflow effects on employment and labor income. Therefore tradeability and spillovers do not explain differential multipliers; MSA estimates are lower bounds.

- Summary conclusion:
  - Labor intensity of recipient industries is a key driver of heterogeneous local fiscal multipliers: spending in labor-intensive industries (especially services) amplifies local employment and labor income responses.

### 6.3 Productivity Gains — Firm dynamics, innovation, and net productivity implications
- Mechanisms:
  - Government spending may reduce uncertainty about future profits, ease credit constraints increasing turnover, and generate faster growth of incumbents; these channels can affect productivity and thus fiscal multipliers.

- Effects on firm entry and exit (Table 4; balanced panel: 12,024 observations; 334 MSAs; 1979−2019):
  - Establishment entry (services spending):
    - t: -0.180 (0.399)
    - t+1: 0.121 (0.281)
    - t+2: 0.394 (0.256)
    - t+3: 0.637** (0.293)
    - t+4: 0.602** (0.238)
  - Establishment entry (goods spending):
    - t: -0.070 (0.117)
    - t+1: 0.224* (0.135)
    - t+2: 0.155* (0.084)
    - t+3: 0.168 (0.108)
    - t+4: 0.183 (0.125)
  - Exit rate (services spending):
    - t: -0.574 (0.560)
    - t+1: -1.183** (0.542)
    - t+2: -1.102*** (0.317)
    - t+3: -1.104*** (0.295)
    - t+4: -0.775*** (0.256)
  - Exit rate (goods spending): mostly marginal or non-significant effects.
  - Interpretation:
    - Services spending increases entry rates a few periods after shock and decreases exit rates; effects smaller for goods.
    - Increase in entry driven by small-size establishments (results available upon request).
    - Decrease in exit rates may keep low-productivity firms active, affecting reallocation and aggregate productivity.

- Effects on innovation activities (Table 5; balanced panel: 6,716 observations; 292 MSAs; 1979−2006):
  - Number of patents:
    - Services Spending t+4: -5.304** (2.690); earlier horizons mostly negative or non-significant.
    - Goods Spending t: -1.762** (0.851); later horizons small positive or non-significant.
  - Number of citations (quality):
    - Services Spending t+2: -9.563* (5.369); t+4: -14.701** (6.163).
    - Goods Spending: coefficients generally smaller and non-significant; t+4: 2.103 (3.112).
  - Interpretation:
    - Both types of spending associated with declines in patents and forward citations; declines larger after services spending shocks, suggesting worsened innovation among incumbents.

- Net implications for productivity and fiscal multipliers:
  - Services shocks: increase entry, decrease exit, but hurt incumbent innovation.
  - Goods shocks: smaller positive entry effects, marginal exit effects, smaller negative innovation effects.
  - Trade-offs: increased entry may boost multipliers via young innovative firms; reduced exit and lower incumbent innovation may lower aggregate productivity.
  - Overall assessment: productivity-related firm dynamics likely contribute marginally to differences in fiscal multipliers; channels operate in opposite directions and can cancel.

- Policy implications (conclusions):
  - Purchases of services generate positive and significant multiplier effects; shocks to goods spending have smaller and non-significant impacts.
  - Heterogeneity in employment response associates with labor intensity of products demanded by government.
  - Little support for tradeability and productivity improvements as alternative explanations for multiplier differences.
  - Policy recommendation: governments can obtain higher multipliers by reallocating dollars from goods toward services; fiscal authorities should consider both how much to spend and what to purchase.
  - Broader observation: declining fiscal effectiveness in recent decades may reflect changes in composition of spending or economy-wide decline in labor’s role; future work should quantify their contributions to declining national fiscal effectiveness.

### Appendix — Data validation, product and industry classification, imputation, and patent data
- A.1 Military Spending Validation:
  - Aggregate comparison: authors’ data closely match trends and levels used in previous studies.
  - Geographic comparison regression Spend_ours_l,t = β Spend_comp_l,t + α_l + δ_t + ε_l,t:
    - Nakamura and Steinsson (1966−2006, State, 50): β 1.12; 95% C.I. (1.02 - 1.21); Observations 2,050; Within−R^2 0.97.
    - Dupor and Guerrero (1966−2009, State, 50): β 0.94; 95% C.I. (0.89 - 0.99); Observations 2,200; Within−R^2 0.96.
    - Demyanyk et al. (2000−2012, CBSA, 862): β 0.98; 95% C.I. (0.89 - 1.07); Observations 10,636; Within−R^2 0.79.
  - Conclusion: authors’ data highly comparable with prior datasets; largest discrepancy with Nakamura and Steinsson arises for years 2000−2006 where Nakamura and Steinsson underestimate spending.

- A.2 Product Codes Classification:
  - Dataset includes 3,239 distinct 4-digit product codes:
    - 2,547 PSCs identifying services; 692 FSCs identifying goods.
  - PSCs grouped into 24 sub-groups; FSCs into 85 groups.
  - Examples of PSC macro-groups: A - Research and Development; C - Architect/engineer Services; 7A - IT and Telecommunication - Applications; 7K - IT and Telecommunication - Storage.
  - Examples of FSC macro-groups: 10 - Weapons; 13 - Ammunition and Explosives; 15 - Aerospace Craft and Structural Components; 99 - Miscellaneous.

- A.3 Industry Classification by Labor Intensity:
  - BEA data on value-added components for 66 macro-industries starting 1997; compute average labor share and classify industries above median as high-labor intensive.
  - Examples of low-labor intensive industries: 111-112 Farms; 211 Oil and gas extraction; 22 Utilities; 511 Publishing industries.
  - Examples of high-labor intensive industries: 23 Construction; 321 Wood products; 333 Machinery; 611 Educational services; 621 Ambulatory health care services; 722 Food services and drinking places.
  - Distributional finding: out of 66 industries, half classified low-labor intensive and half high-labor intensive; DoD spending: about 85% directed to firms in high-labor-intensive industries.

- A.4 Imputation strategy:
  - Industry information missing for ~5% of contracts after 1979 (accounting for 17% of spending); imputation reduces non-imputed contracts to <1% (<1% of spending).
  - Sequential imputation: assign BEA industry with largest share of spending for product code; tie-breakers use finer industry with highest contract share; if potential industries share same labor-intensity, classify by labor-intensity; otherwise leave not imputed.
  - Validation: for each product code, assigned industry covers on average >70% of spending; for >75% of product codes, assigned industry covers >50% of spending.
  - Pre-1997 labor-intensity classification held fixed; annual reclassification matches full-sample classification >91% of cases.

- A.5 Patent data (PatentsView):
  - PatentsView contains universe of granted patents from USPTO starting 1976; collection date end of 2019.
  - Analysis restricted to utility patents (≈98% of granted patents); application year from 1976; final data exclude patents with DoD economic interest; restrict to patents applied by end of 2006 to address truncation bias.
  - Geolocation: inventor latitude/longitude mapped to MSAs via 2010 TIGER/Line; multi-inventor patents split equally across MSAs by number of inventors.
  - Innovation measures: number of patents (counts) and patents weighted by adjusted forward citations within 5-year window after grant year (quality-weighted).

### Selected additional empirical results and robustness (Tables B.1–B.8, selected cells preserved)
- Local Fiscal Multiplier: Aggregate Spending (Employment; Labor Income; balanced panel: 12,024 observations, 334 MSAs, 1979−2019):
  - Employment — Spending:
    - tt+1: 0.139*** (0.051), F-stat 8.299
    - tt+2: 0.414*** (0.153), F-stat 9.330
    - tt+3: 0.453*** (0.153), F-stat 10.101
    - tt+4: 0.550** (0.236), F-stat 7.680
  - Labor Income — Spending:
    - tt+1: 0.253*** (0.095)
    - tt+2: 0.834*** (0.286)
    - tt+3: 0.873*** (0.285)
    - tt+4: 1.046** (0.424)

- Local Fiscal Multipliers — Spending by Components (Table B.2; Employment & Labor Income):
  - Employment — Services Spending:
    - tt+1: 0.355*** (0.123)
    - tt+2: 0.968*** (0.167)
    - tt+3: 1.191*** (0.199)
    - tt+4: 1.454*** (0.206)
  - Employment — Goods Spending:
    - tt+1: 0.059 (0.041)
    - tt+2: 0.165 (0.128)
    - tt+3: 0.157 (0.121)
    - tt+4: 0.186 (0.177)
  - Labor Income — Services Spending:
    - tt+1: 0.585*** (0.151)
    - tt+2: 1.924*** (0.331)
    - tt+3: 2.424*** (0.403)
    - tt+4: 2.892*** (0.464)
  - Labor Income — Goods Spending:
    - tt+1: 0.078 (0.064)
    - tt+2: 0.269 (0.189)
    - tt+3: 0.227 (0.173)
    - tt+4: 0.260 (0.249)
  - Tests βk_g = βk_s p-values: [.0333], [.0003], [.0000], [.0000], [.0001] (columns corresponding to horizons).

- Cumulative and outlays specifications (selected):
  - Cumulative Employment — Services Spending tt+1: 0.355*** (0.123); Goods tt+1: 0.059 (0.041).
  - Outlays Employment — Services tt+1: 0.283*** (0.089); Goods tt+1: -0.108 (0.124).
  - Excluding top 1% contracts by value (Table B.5) — Employment — Services tt+1: 1.348*** (0.300); Labor Income — Services tt+1: 2.059*** (0.750).

- Labor Intensity heterogeneity (Table B.8; balanced panel: 12,024 observations, 334 MSAs, 1979−2019):
  - Employment — Services Spending — High Labor Intensity:
    - tt+1: 0.365** (0.141)
    - tt+2: 0.978*** (0.172)
    - tt+3: 1.183*** (0.257)
    - tt+4: 1.554*** (0.302)
  - Employment — Goods Spending — High Labor Intensity:
    - tt+1: 0.073** (0.032)
    - tt+2: 0.283* (0.150)
    - tt+3: 0.302** (0.148)
    - tt+4: 0.537* (0.309)
  - Employment — Goods Spending — Low Labor Intensity:
    - tt+1: -0.062 (0.074)
    - tt+2: -0.190* (0.113)
    - tt+3: -0.175* (0.090)
    - tt+4: -0.331** (0.148)
  - Labor Income — Services Spending — High Labor Intensity tt+4: 2.968*** (0.572).
  - Labor Income — Goods Spending — High Labor Intensity tt+4: 0.795** (0.391).
  - Interpretation: labor-intensity drives heterogeneity; high-labor-intensive spending (especially in services) yields large positive employment and labor income multipliers; low-labor-intensive spending can yield negative local effects.

- Additional robustness:
  - Pre-determined shift-share instrument specifications (Table B.6) and concentration ratio regressions (Table B.7) reported; evidence indicates monopsony concerns limited (weak negative relationship between employment concentration and measured labor intensity).

*Italic source attribution: wpiea2023052-print-pdf — Heterogeneous Spending, Heterogeneous Multipliers, Working Paper No. WP/2023/052*

### 2.1    Military Spending

### 2.1    Military Spending

### Data sources and construction
- Data harmonized from two sources: the National Archives and Records Administration (NARA) for 1966-2006, and USASpending.gov for 2007-2019.
- Data based on DD-350 and DD-1057 military procurement forms that account for approximately 96% of contracts awarded by the DoD.
- Key contract fields available: contract identification number, dates of action and completion, transaction value, location where the contract is performed, Product Service Codes (PSCs) or Federal Supply Codes (FSCs).
- Year of government spending defined as the year of the signature date (award or modification).
- Modifications of existing contracts reported as negative entries; obligations and de-obligations within 0.5% of each other are considered null and void.
- Contract dollar values reported in nominal terms and converted to real values using the US Bureau of Labor Statistics’ Consumer Price Index.

### Classification of goods versus services
- Product Service Codes (PSCs) and Federal Supply Codes (FSCs) are 4-digit alphanumeric classifications indicating deliverable type; existing codes are consistent from 1979.
- PSCs: first digit is a letter → classified as “Spending in Services.”
- FSCs: first digit is a number → classified as “Spending in Goods.”
- Appendix A.2 reports major product codes in the two categories (not reproduced here).

### Geographic allocation and mapping
- Geographic allocation based on the location where contract tasks are performed (place of performance).
- USASpending.gov provides city and zip code of the performing firm; these are used to identify county.
- NARA reports county where the firm performs the contract.
- County-level data aggregated into Core-Based Statistical Areas (MSAs) using spatial crosswalks from the National Bureau of Economic Research.
- If a contract is known to be performed in the US but performance location unknown, the location of the contract recipient is assigned as performance location (following Demyanyk et al. (2019)), with adjustments for city information when postal code is missing.

### Outcome variables (economic outcomes)
- Main dependent variable: number of employed persons from the Bureau of Labor Statistics Quarterly Census of Employment and Wages (QCEW).
- Second outcome: salary and wages (proxy for labor income) from the Bureau of Economic Analysis at MSA level; reported in nominal terms and converted to real terms.
- Personal income from the Bureau of Economic Analysis used to normalize government spending; converted to real terms.
- Personal income measures households’ income from salary and wages, Social Security and other government programs, dividends and interest, business ownership, and other sources.

### Sample construction and filters
- Start of analysis: 1979 (product codes before 1979 are not consistent with those after that date).
- Sample restricted to 2019 to avoid the effect of the COVID-19 pandemic (robustness check including 2020 yields similar results).
- From universe of DoD procurement contracts, drop contracts missing: locality of performance, year signed, product code, or industry of contractor (industry sometimes imputed using product codes; contracts with missing industry after imputation excluded).
- After aggregating spending by MSA, additional filters:
  - Exclude MSAs with incomplete histories in any outcome variable.
  - Exclude observations with growth rates between two consecutive periods greater than 100% or smaller than −50%.
  - Remove MSAs with average population smaller than 50,000 inhabitants.
  - Include only MSAs with non-negative ratios of aggregate military spending to personal income smaller than 1.5.
- Final sample: procurement contracts allocated to 334 MSAs from 1979 to 2019.
- Final sample includes 94% of the total number of contracts and the aggregate spending.
  - For spending on goods: sample accounts for about 95% of total contracts and spending on goods.
  - For spending on services: sample includes 85% of the contracts worth 92% of total spending on services.

### Descriptive statistics and patterns (1979−2019)
- Aggregate spending evolution: three sharp rises
  - 1980s: Reagan military buildup
  - Early 2000s: Afghani and Iraqi wars
  - Recent years: escalations in military buildups and competition with Russia and China
- Reallocation across categories:
  - Shift from goods to services between the 1980s and the 1990s; product classification harmonized so changes are not due to reclassification.
  - Until beginning of the 1980s, spending on goods was more than 70% of procurement value.
  - Starting with the 1990s, spending split between goods and services with similar shares.
- Comparison with private consumption:
  - Private consumers increased share of expenditures allocated to services starting in the early 1970s; services fraction higher in private consumption than DoD spending.
  - Government spending is more goods-driven than private consumption; reallocation between categories in government spending takes longer.
- Contract- and category-level summary statistics (Table 1):
  - Share of Spending (%): All 54, Goods 46
  - Share of Contracts (%): All 91, Goods 9
  - Average Value (Thousand USD): All 143, Goods 99, Services 293
  - 90%-to-50%-percentile Ratio: All 5, Goods 24, Services 6
  - Share of Spending for Top 1% (%): All 72, Goods 86, Services 50
- Key distributional findings:
  - DoD awarded 91% of procurement contracts to purchase goods and 9% to acquisition of services.
  - Average value of a services contract is about three times the average goods contract (Average Value: Goods 99 Thousand USD; Services 293 Thousand USD).
  - Goods contracts have a much fatter right-tail: top decile of goods contracts over 40 times greater than median; for services top decile 11 times greater than median.
  - Top 1% of goods contracts accounts for 86% of goods spending; top 1% of services contracts accounts for around 50% of services spending.
  - Results indicate granularity and concentration of spending both across contract recipients and place of performance.

### Geographic heterogeneity
- Final sample computed over 334 MSAs.
- Top 30 MSAs in awarded spending account for more than 30% of spending on goods and more than 45% of spending on services.
- Only 35% of MSAs are in the same quartile for both goods and services spending.
- Most MSAs in the Midwest are in the top two quartiles of spending on goods but few rank high in services spending.
- The Midwest has the highest concentration of production occupations with an average employment share in production jobs almost 50% higher than the US average.
- Distribution of spending at both contract and MSA-level is skewed to the right.

### Empirical strategy
- Identification exploits variation in military procurement spending across time and localities to estimate local fiscal multipliers.
- Benchmark specification (panel regression in changes, horizons k = {0,...,4}):
  - v_{l,t+k} − v_{l,t−1}  over v_{l,t−1}  = β^{k} (G_{l,t+k} − G_{l,t−1}) / Y_{l,t−1} + α^{k}_{l} + δ_{t+k} + ε_{l,t+k}
  - v is outcome of interest in location l at horizon k.
  - G_{l,t+k} − G_{l,t−1} measures change in military spending normalized by lagged personal income Y_{l,t−1}.
  - Locality-fixed effects α^{k}_{l} and time-fixed effects δ_{t+k} included.
  - Standard errors clustered by locality.
- Interpretation:
  - β^{k} quantifies multipliers over k years.
  - Since spending normalized by personal income, for employment multiplier β^{k} is the change in employment growth rate produced by a one-percentage-point increase in government spending (normalized by personal income) at the local level.
- Augmented specification to estimate heterogeneous effects of goods and services (equation 2):
  - v_{l,t+k} − v_{l,t−1}  over v_{l,t−1}  = β^{k}_{g} (G^{g}_{l,t+k} − G^{g}_{l,t−1}) / Y_{l,t−1} + β^{k}_{s} (G^{s}_{l,t+k} − G^{s}_{l,t−1}) / Y_{l,t−1} + α^{k}_{l} + δ_{t+k} + ε_{l,t+k}
  - G^{g} represents spending in goods; G^{s} represents spending in services; G^{g} + G^{s} = G for each period and locality.

### Instrumental variables research design
- Concern: Military spending may be endogenous to local political or economic shocks.
- Instrument constructed following Nakamura and Steinsson (2014) — a shift-share (Bartik) instrument combining nationwide changes in procurement with locality comparative advantage.
- Instrument for total spending (equation 3):
  - Z_{l,t+h} = s_{l} (G_{t+k} − G_{t−1}) / Y_{l,t−1}
  - s_{l} is the average share of spending in locality l, capturing comparative advantage in receiving military contracts.
  - Conditional variation arises from national-level changes in military spending; locality-fixed effects remove level differences.
- Instruments for goods and services (equation 4):
  - Z^{g}_{l,t+h} = s^{g}_{l} (G^{g}_{t+k} − G^{g}_{t−1}) / Y_{l,t−1}
  - Z^{s}_{l,t+h} = s^{s}_{l} (G^{s}_{t+k} − G^{s}_{t−1}) / Y_{l,t−1}
  - s^{g}_{l} and s^{s}_{l} are shares of spending in goods and services captured by location l.
- Clustering and robustness:
  - Standard errors clustered by locality.
  - For narrower geographic units, robustness checks cluster at the state level; results remain unaffected (available upon request).

*Source: wpiea2023052-print-pdf - 2.1    Military Spending*

### 4.3    Identification assumptions and threats

### 4.3    Identification assumptions and threats

### Identification assumption
- The identification assumption: conditional on the MSA and time-fixed effects, the instrument is not systematically associated with any unobserved political or economic characteristics that may explain the outcomes of interest.
- The identification strategy may suffer from three main identification threats.

### Threat 1 — Measurement error from outsourcing of military procurement
- Contracts are assigned to their place of performance (where the product is assembled or processed); use of sub-contractors outside the place of performance would geographically misallocate part of outsourced military spending.
- This measurement error will create an attenuation bias in estimates of the local fiscal multiplier.
- A more serious interpretational threat arises if the measurement error is systematically correlated with a specific spending category:
  - Physical goods are more tradeable than services, so the measurement error could be correlated with spending on goods.
- The authors state they rule out the importance of this measurement error in Section 6.2.

### Threat 2 — Geographical spillovers (violation of SUTVA)
- A violation of the stable unit treatment value assumption would occur if defense spending in a specific MSA affects neighboring localities that did not receive any contracts.
- The direction of bias depends on spillover effects:
  - Positive spillovers (increased demand for final consumption or intermediate goods in neighboring localities via input-output linkages) would cause a downward bias in estimates.
  - Negative spillovers (increases in spending affecting factor prices and allocation of production factors across MSAs) could cause an upward bias in estimates.
- The authors provide evidence in Section 6.2 that their estimates are biased downward and can be considered as lower bounds.
- Footnote/contextual note: limited factor mobility makes negative spillovers small (Chodorow-Reich (2019)).

### Threat 3 — Shift-share instrument drawbacks from recent econometric literature
- The empirical strategy uses a shift-share instrumental variable with several shifts (e.g., as many asG t+k −G t−1 differences can be computed in the data) and one single share for each MSA (the time-invariant measure of comparative advantage).
- Recent literature (Goldsmith-Pinkham et al., 2020; Borusyak et al., 2022; Adao et al., 2019) highlights drawbacks of shift-share instruments.
- Borusyak et al. (2022) show that the shift-share IV requires either the shares or the shifts to be uncorrelated with unobserved characteristics that may affect local outcomes.
- Authors' defense of exogeneity:
  - The shifts are argued to be a reasonable source of quasi-random variation because military buildups respond to international geopolitical events rather than to unobserved local factors (automation, trade competition, national fiscal policy) that could have heterogeneous impacts across MSAs.
  - Exogeneity of the shifts would assure instrument validity, even when MSAs with high comparative advantage experience different economic trends than MSAs with low comparative advantage.
  - As both outcomes and instruments are measured in changes, the research design purges the instrument from potential correlation between shares and unobserved local economic trends.
  - Locality-fixed effects absorb any MSA-specific secular trend.
  - Therefore, the variation provided by the instrument consists of deviations in spending from its long-term trend, which is plausibly exogenous to local economic characteristics.

*Italic source attribution: wpiea2023052-print-pdf - 4.3    Identification assumptions and threats*

### 6.1    Labor intensity

### 6.1    Labor intensity

### Labor-intensity concept and measurement
- Labor intensity defined as the contribution of employee compensation to value added, using BEA annual data on value-added and employee compensation by industry.
- Industry-level labor intensity assigned to each contract based on the contractor’s industry.
- Industries classified between labor- and non-labor-intensive (examples in text): healthcare, education, hospitality, and food service classified as labor-intensive; manufacturing and retail classified as non-labor-intensive.
- Imputation used for industry assignment of contracts awarded before 1997 because the data contain 66 NAICS codes starting from 1997.
- Potential confounder: imperfect competition (monopsony) could lead industries that use labor intensively but pay employees little to be misclassified. To assess this, employment concentration shares used as proxies for imperfect competition.
  - Regression of labor intensity on measures of concentration (after absorbing year and industry fixed-effects) documents a weak negative relationship between employment concentration and labor intensity at the industry level, interpreted as evidence that monopsony power is not an important confounder.

### Distributional facts about spending and labor intensity
- Figure 6 highlights:
  - The largest share of government spending is directed to labor-intensive industries.
  - The share of goods spending allocated to non-labor-intensive industries is 50% higher than the share allocated from spending in services (30% vs. 20%).
  - Aggregate labor intensity measure constructed as AggLI_{i,t} = sum_{j∈J} ω_{i,j,t} LI_j for i∈{s,g}, with ω_{i,j,t} = G_{i,j,t} / G_{i,t}.
  - The aggregate labor intensity is higher for services spending than for goods spending.
  - Apart from the 1990s (in which the gap narrowed), the aggregate labor intensity measure is about 10% higher for services spending than goods spending.
- Time-varying industry-specific labor intensity used for robustness over 1997−2019 yields consistent results with the average-industry measure.

### Empirical effects by labor intensity (specification and instruments)
- Specification (equation (7)) separates spending by category i∈{s,g} and by industry labor-intensity p∈{L,N}; independent variables instrumented by shift-share instruments (equation (8)) where Z_{i,p,l,t+h} = s_{i,p,l} (G_{i,p,t+k} − G_{i,p,t−1}) / Y_{l,t−1}.
- Intuition: If labor intensity drives fiscal multipliers, spending in labor-intensive industries should generate large and positive effects regardless of spending category.

### Main empirical findings on employment and labor income
- Spending directed to labor-intensive industries generates positive and highly significant effects on employment and labor income.
  - Employment multipliers for labor-intensive services spending range between 0.36% and 1.5%.
  - Employment multipliers for labor-intensive goods spending range between 0.07% and 0.54%.
- Spending in non-labor-intensive industries has negative effects for both goods and services spending, although significance differs:
  - Negative and significant for goods spending in non-labor-intensive industries.
  - Negative or non-significant for services spending in non-labor-intensive industries (less precisely estimated because the share of services spending directed to these industries is relatively small).
- Effects on labor income:
  - Spending directed to labor-intensive industries generates increases in labor income.
  - Effects of spending in non-labor-intensive industries are either negative or non-significant.
- Interpretation: When a large share of spending passes through to households via increases in labor income, households raise consumption and firm production, generating sizable employment responses.
- Net effect for aggregate goods spending: comparable magnitude but opposite-signed effects from labor- and non-labor-intensive industries largely cancel out, producing a small aggregate goods multiplier.

### Robustness: tradeability and geographic spillovers (section 6.2)
- Concern: Goods are more tradeable than services; production could occur outside the MSA (or US), generating geographic spillovers and measurement error in local allocation.
- Two empirical strategies applied to assess geographic allocation and spillovers.

1) State-level aggregation comparison
- Rationale: Geographic misallocation of contracts less important at state level; one-for-one relation between prime contracts allocated to a state and shipments from that state in prior literature.
- Table 2 state-level estimates (balanced panel: 1,800 observations, 50 states excluding DC, period 1979−2019):
  - Dependent variable: Employment — Services Spending coefficients across horizons (t to t+4): 0.845, 2.736***, 2.989**, 3.430***, 2.846***.
  - Dependent variable: Employment — Goods Spending coefficients: 0.240, 0.624, 0.650, 1.150, 1.295*.
  - Dependent variable: GDP — Services Spending coefficients: 0.921, 4.882***, 6.525**, 7.897***, 7.154***.
  - Dependent variable: GDP — Goods Spending coefficients: -0.113, -0.160, -0.164, -0.110, -0.188.
- Interpretation: No effect of goods spending on employment or GDP at state level, while services spending yields sizable and significant positive effects. State-level estimates are larger than MSA-level estimates, consistent with within-state subcontracting attenuating MSA estimates (MSA estimates are lower bounds).

2) Neighboring-locations (50-mile radius) analysis
- Specification (equation (9)) estimates spillovers to neighboring MSAs (balanced panel: 6,948 observations, 195 MSAs, period 1979−2019); instruments in equation (10).
- Table 3 neighboring-location estimates:
  - Dependent variable: Employment — Services Spending coefficients (t to t+4): 0.109, 0.262***, 0.386**, 0.573*, 0.525*.
  - Dependent variable: Employment — Goods Spending coefficients: 0.070, -0.008, -0.217, -0.366, -0.277.
  - Dependent variable: Labor Income — Services Spending coefficients: 0.226*, 0.492***, 0.710***, 1.006**, 0.951**.
  - Dependent variable: Labor Income — Goods Spending coefficients: -0.070, -0.120, -0.476, -0.709, -0.575.
- Interpretation:
  - No positive spillover effects of goods spending from a locality to neighboring localities.
  - Services spending exhibits positive and significant outflow effects on neighboring locations for employment and labor income.
  - These results imply tradeability and geographic spillovers of goods cannot explain the differential multipliers between goods and services; small goods effects are not due to goods generating multipliers in neighboring locations.
  - The presence of some geographic spillovers for services suggests MSA-level estimates are downward biased and should be considered lower bounds.

### Summary conclusion
- Services and goods spending differ in the labor intensity of the industries that receive contracts.
- Labor intensity matters for quantifying local fiscal multipliers: spending in labor-intensive industries amplifies local employment and labor income responses, producing larger local fiscal multipliers.
- The heterogeneity in labor intensity is an important driver of the differences in local fiscal multipliers between services and goods spending shocks.
- Monopsony concerns appear limited based on the weak negative relationship between employment concentration and measured labor intensity.
- Tradeability and geographic allocation issues do not account for the weaker local multipliers of goods; neighboring-location analysis shows no offsetting spillovers for goods, while services show positive spillovers.

*Source: IMF working paper chapter 6.1, "Labor intensity" from wpiea2023052-print-pdf*

### 6.3    Productivity Gains

### 6.3    Productivity Gains

### Mechanisms linking government spending to productivity
- Government spending may:
  - reduce uncertainty about future profits;
  - ease credit constraints leading to higher turnover rates of firms;
  - generate faster growth of incumbents.
- These channels, typical in growth models with firm heterogeneity, may generate productivity improvements.
- Heterogeneity in fiscal multipliers could be partially explained by differences in productivity gains across types of spending.

### Effects on firm entry and exit (Table 4)
- Data and estimation details:
  - Balanced panel: 12,024 observations; 334 MSAs; period 1979−2019.
  - Instruments computed as in equation (4); standard errors clustered by MSA.
  - Regressions include MSA and year-fixed effects. F-statistics for the first stage reported.
- Entry rate (establishment entry):
  - Services Spending:
    - t: -0.180 (standard error 0.399)
    - t+1: 0.121 (0.281)
    - t+2: 0.394 (0.256)
    - t+3: 0.637** (0.293)
    - t+4: 0.602** (0.238)
  - Goods Spending:
    - t: -0.070 (0.117)
    - t+1: 0.224* (0.135)
    - t+2: 0.155* (0.084)
    - t+3: 0.168 (0.108)
    - t+4: 0.183 (0.125)
  - F-statistics (columns t to t+4): 6.325, 12.472, 14.034, 17.648, 29.439.
  - Interpretation:
    - Establishment entry rate increases a few periods after a shock to services spending.
    - Effects are smaller after shocks to goods spending, consistent with higher entry costs faced by goods producers.
    - Note: increase in entry rates is driven by entry of small-size establishments (results available upon request).
- Exit rate:
  - Services Spending:
    - t: -0.574 (0.560)
    - t+1: -1.183** (0.542)
    - t+2: -1.102*** (0.317)
    - t+3: -1.104*** (0.295)
    - t+4: -0.775*** (0.256)
  - Goods Spending:
    - t: 0.159 (0.240)
    - t+1: -0.232 (0.189)
    - t+2: -0.058 (0.187)
    - t+3: -0.194 (0.194)
    - t+4: -0.224 (0.195)
  - F-statistics (columns t to t+4): 6.325, 12.472, 14.034, 17.648, 29.439.
  - Interpretation:
    - Estimates suggest a decline in exit rates after shocks to services spending.
    - Shocks to goods spending have marginal effects on exit rates.
    - Potential implication: government demand may keep low-productivity firms active that otherwise would have exited, affecting allocation of production factors and aggregate productivity.

### Effects on innovation activities (Table 5)
- Data and estimation details:
  - Balanced panel: 6,716 observations; 292 MSAs; period 1979−2006.
  - Instruments computed as in equation (4); standard errors clustered by MSA.
  - Regressions include MSA and year-fixed effects.
- Number of patents:
  - Services Spending:
    - t: -3.294 (2.458)
    - t+1: -0.984 (1.648)
    - t+2: -3.072 (2.059)
    - t+3: -3.100 (2.230)
    - t+4: -5.304** (2.690)
  - Goods Spending:
    - t: -1.762** (0.851)
    - t+1: -0.438 (0.629)
    - t+2: 0.171 (0.506)
    - t+3: 0.598 (0.731)
    - t+4: 1.066 (0.878)
  - F-statistics (columns t to t+4): 7.628, 16.917, 17.043, 26.988, 28.410.
- Number of citations (quality of patents):
  - Services Spending:
    - t: -9.660 (6.328)
    - t+1: -3.428 (4.291)
    - t+2: -9.563* (5.369)
    - t+3: -9.037* (5.200)
    - t+4: -14.701** (6.163)
  - Goods Spending:
    - t: -1.915 (4.660)
    - t+1: -1.042 (2.438)
    - t+2: -0.850 (2.759)
    - t+3: -0.186 (4.524)
    - t+4: 2.103 (3.112)
  - F-statistics (columns t to t+4): 7.628, 16.917, 17.043, 26.988, 28.410.
- Interpretation:
  - Both types of spending show declines in number of patents and forward citations.
  - Declines are larger after shocks to services spending compared to goods spending.
  - Suggests a worsening of innovation activities among incumbent firms following services spending shocks.

### Net implications for productivity and fiscal multipliers
- Empirical summary:
  - Shocks to services spending: increase establishment entry rates, decrease exit rates, and hurt innovation activities of incumbent firms.
  - Shocks to goods spending: positively affect entry rates but effects are small; marginal effects on exit rates; smaller negative effects on innovation.
- Mechanism trade-offs:
  - Increased entry rates may generate positive multiplier effects due to young and innovative firms entering the market.
  - Declines in exit rates may hinder reallocation of factors to more productive firms, reducing aggregate productivity.
  - Declines in incumbent innovation may lower economy-wide productivity.
- Overall assessment:
  - These channels operate in opposite directions and their effects can cancel out.
  - Contribution of productivity-related firm dynamics to differences in fiscal multipliers is expected to be relatively marginal.

### Policy implications (from conclusions)
- Purchases of services generate positive and significant multiplier effects; shocks to goods spending have a smaller and non-significant impact.
- Heterogeneity in employment response is associated with labor intensity of the products demanded by the government.
- Little support found for tradeability and productivity improvements as alternative explanations for multiplier differences.
- Policy recommendation:
  - There is room for governments to redesign fiscal spending and obtain higher multipliers by reallocating dollars from goods toward services.
  - Fiscal authorities should design interventions considering not only how much to spend but also what to purchase.
- Broader observation:
  - The declining effectiveness of fiscal policy in recent decades may be due to changes in the composition of spending or an economy-wide decline in the importance of labor in production.
  - Future work should quantify the extent to which these structural changes have contributed to declining fiscal effectiveness at the national level.

*Source: wpiea2023052-print-pdf - 6.3    Productivity Gains*

### References

### References

### Key bibliographic works cited
- Adao, R., Kolesár, M., and Morales, E. (2019). Shift-share designs: Theory and inference. The Quarterly Journal of Economics, 134(4):1949–2010.
- Alonso, C. (2017). Cutting back on labor intensive goods? implications for fiscal stimulus. Mimeo.
- Auerbach, A., Gorodnichenko, Y., and Murphy, D. (2020). Local Fiscal Multipliers and Fiscal Spillovers in the USA. IMF Economic Review, 68(1):195–229.
- Barnichon, R., Debortoli, D., and Matthes, C. (2020). Understanding the Size of the Government Spending Multiplier: It’s in the Sign. Working Papers 1145, Barcelona Graduate School of Economics.
- Barro, R. (1981). Output Effects of Government Purchases. Journal of Political Economy, 89(6):1086–1121.
- Bartik, T. J. (1991). Who Benefits from State and Local Economic Development Policies? Number wbsle in Books from Upjohn Press. W.E. Upjohn Institute for Employment Research.
- Basso, H. S. and Rachedi, O. (2021). The Young, the Old, and the Government: Demographics and Fiscal Multipliers. American Economic Journal: Macroeconomics, 13(4):110–141.
- Blanchard, O. and Perotti, R. (2002). An empirical characterization of the dynamic effects of changes in government spending and taxes on output. The Quarterly Journal of Economics, 117(4):1329–1368.
- Boehm, C. E. (2020). Government consumption and investment: Does the composition of purchases affect the multiplier? Journal of Monetary Economics, 115(C):80–93.
- Born, B., Juessen, F., and Müller, G. J. (2013). Exchange rate regimes and fiscal multipliers. Journal of Economic Dynamics and Control, 37(2):446–465.
- Borusyak, K., Hull, P., and Jaravel, X. (2022). Quasi-experimental shift-share research designs. The Review of Economic Studies, 89(1):181–213.
- Bouakez, H., Rachedi, O., and Santoro, E. (2020). The sectoral origins of the spending multiplier. Mimeo.
- Buchheim, L., Watzinger, M., and Wilhelm, M. (2020). Job creation in tight and slack labor markets. Journal of Monetary Economics, 114(C):126–143.
- Chodorow-Reich, G. (2019). Geographic cross-sectional fiscal spending multipliers: What have we learned? American Economic Journal: Economic Policy, 11(2):1–34.
- Chodorow-Reich, G., Feiveson, L., Liscow, Z., and Woolston, W. G. (2012). Does State Fiscal Relief during Recessions Increase Employment? Evidence from the American Recovery and Reinvestment Act. American Economic Journal: Economic Policy, 4(3):118–145.
- Cole, H. L. and Ohanian, L. E. (2004). New Deal Policies and the Persistence of the Great Depression: A General Equilibrium Analysis. Journal of Political Economy, 112(4):779–816.
- Conley, T. G. and Dupor, B. (2013). The American Recovery and Reinvestment Act: Solely a government jobs program? Journal of Monetary Economics, 60(5):535–549.
- Corbi, R., Papaioannou, E., and Surico, P. (2019). Regional transfer multipliers. The Review of Economic Studies, 86(5):1901–1934.
- Cox, L., Müller, G., Pastén, E., Schoenle, R., and Weber, M. (2020). Big G. NBER Working Papers 27034, National Bureau of Economic Research, Inc.
- Demyanyk, Y., Loutskina, E., and Murphy, D. (2019). Fiscal stimulus and consumer debt. Review of Economics and Statistics, 101(4):728–741.
- Dolls, M., Fuest, C., and Peichl, A. (2012). Automatic stabilizers and economic crisis: US vs. Europe. Journal of Public Economics, 96(3):279–294.
- Dupor, W. and Guerrero, R. (2017). Local and aggregate fiscal policy multipliers. Journal of Monetary Economics, 92(C):16–30.
- Dupor, W., Karabarbounis, M., Kudlyak, M., and Mehkari, M. S. (2021). Regional Consumption Responses and the Aggregate Fiscal Multiplier. CEPR Discussion Papers 16189, C.E.P.R. Discussion Papers.
- Dupor, W. and McCrory, P. B. (2018). A Cup Runneth Over: Fiscal Policy Spillovers from the 2009 Recovery Act. Economic Journal, 128(611):1476–1508.
- Dupor, W. and Mehkari, M. S. (2016). The 2009 Recovery Act: Stimulus at the extensive and intensive labor margins. European Economic Review, 85(C):208–228.
- Eckert, F., Fort, T. C., Schott, P. K., and Yang, N. J. (2020). Imputing Missing Values in the US Census Bureau’s County Business Patterns. NBER Working Papers 26632, National Bureau of Economic Research, Inc.
- Ferraz, C., Finan, F., and Szerman, D. (2015). Procuring Firm Growth: The Effects of Government Purchases on Firm Dynamics. NBER Working Papers 21219, National Bureau of Economic Research, Inc.
- Galeano, L., Izquierdo, A., Puig, J. P., Vegh, C. A., and Vuletin, G. (2021). Can Automatic Government Spending Be Procyclical? NBER Working Papers 28521, National Bureau of Economic Research, Inc.
- Goldsmith-Pinkham, P., Sorkin, I., and Swift, H. (2020). Bartik instruments: What, when, why, and how. American Economic Review, 110(8):2586–2624.
- Gorodnichenko, Y., Mendoza, E. G., and Tesar, L. L. (2012). The Finnish Great Depression: From Russia with Love. American Economic Review, 102(4):1619–1644.
- Hagedorn, M., Manovskii, I., and Mitman, K. (2019). The Fiscal Multiplier. NBER Working Papers 25571, National Bureau of Economic Research, Inc.
- Ilzetzki, E., Mendoza, E. G., and Végh, C. A. (2013). How big (small?) are fiscal multipliers? Journal of Monetary Economics, 60(2):239–254.
- Isard, W. (1962). Awards of Prime Military Contracts by County, State and Metropolitan Area of the United States, Fiscal Year 1960. Regional Science Research Institute.
- Juarros, P. (2020). Fiscal stimulus, credit frictions and the amplification effects of small firms. Mimeo.
- Lewis, V. and Winkler, R. (2017). Government Spending, Entry, And The Consumption Crowding-In Puzzle. International Economic Review, 58(3):943–972.
- Mountford, A. and Uhlig, H. (2009). What are the effects of fiscal policy shocks? Journal of Applied Econometrics, 24(6):960–992.
- Nakamura, E. and Steinsson, J. (2014). Fiscal Stimulus in a Monetary Union: Evidence from US Regions. American Economic Review, 104(3):753–792.
- Ramey, V. A. and Shapiro, M. D. (1998). Costly capital reallocation and the effects of government spending.
- Ramey, V. A. and Zubairy, S. (2018). Government spending multipliers in good times and in bad: evidence from us historical data. Journal of Political Economy, 126(2):850–901.
- Riera-Crichton, D., Vegh, C. A., and Vuletin, G. (2015). Procyclical and countercyclical fiscal multipliers: Evidence from OECD countries. Journal of International Money and Finance, 52(C):15–31.
- Slavtchev, V. and Wiederhold, S. (2016). Does the Technological Content of Government Demand Matter for Private R&D? Evidence from US States. American Economic Journal: Macroeconomics, 8(2):45–84.
- Suárez-Serrato, J. C. and Wingender, P. (2016). Estimating Local Fiscal Multipliers. NBER Working Papers 22425, National Bureau of Economic Research, Inc.
- Wilson, D. J. (2012). Fiscal spending jobs multipliers: Evidence from the 2009 American Recovery and Reinvestment Act. American Economic Journal: Economic Policy, 4(3):251–82.
- Woodford, M. (2011). Simple Analytics of the Government Expenditure Multiplier. American Economic Journal: Macroeconomics, 3(1):1–35.

*Appendix materials below summarize data construction, validation, classification, and imputation methods used in the analysis.*

### Appendix A. Data

### A.1. Military Spending Validation
- Comparison purpose: assess quality of DoD military procurement contract microdata versus previously used datasets (Nakamura and Steinsson (2014), Dupor and Guerrero (2017), Demyanyk et al. (2019)).
- Aggregate comparison: Figure A.1 compares aggregate military spending in nominal terms at the national level by fiscal years; authors conclude their data closely match the trend and level of aggregate spending used in previous works.
- Geographic comparison regression:
  - Specification: Spend_ours_l,t = β Spend_comp_l,t + α_l + δ_t + ε_l,t
  - Interpretation: Spend_ours_l,t is military spending from authors’ data in locality l at time t; Spend_comp_l,t is spending from a comparison dataset; α_l and δ_t are locality and time fixed effects.
- Table A.1: Military Spending Comparisons by Geography — reported estimates (preserving numeric precision):
  - Column (1) Nakamura and Steinsson (2014): β 1.12; 95% C. I. for β (1.02 - 1.21); Observations 2,050; Geographic Unit State; Number Localities 50; Period 1966−2006; Locality FE Yes; Time FE Yes; Within−R^2 0.97.
  - Column (2) Dupor and Guerrero (2017): β 0.94; 95% C. I. for β (0.89 - 0.99); Observations 2,200; Geographic Unit State; Number Localities 50; Period 1966−2009; Locality FE Yes; Time FE Yes; Within−R^2 0.96.
  - Column (3) Demyanyk et al. (2019): β 0.98; 95% C. I. for β (0.89 - 1.07); Observations 10,636; Geographic Unit CBSA; Number Localities 862; Period 2000−2012; Locality FE Yes; Time FE Yes; Within−R^2 0.79.
- Notes and diagnostic findings:
  - Values in brackets report the 95% confidence interval. Standard errors are clustered at the geographic unit level.
  - Largest discrepancy with Nakamura and Steinsson (2014) arises for years between 2000 and 2006; Nakamura and Steinsson (2014) underestimate aggregate spending in that interval.
  - Within−R^2 is over 0.95 for comparisons with Nakamura and Steinsson (2014) and Dupor and Guerrero (2017); lower (0.79) for Demyanyk et al. (2019).
  - Data construction: authors use NARA until 2006; USASpending is used by Demyanyk et al. (2019) starting from 2001. Tests confirm comparability between NARA and USASpending.
  - Conclusion: authors’ data are highly comparable with data used in previous studies.

### A.2. Product Codes Classification
- Dataset includes 3239 distinct 4-digit product codes:
  - 2547 Product Services Codes (PSCs) identifying spending in services.
  - 692 Federal Supply Codes (FSCs) identifying spending in goods.
- PSCs grouped into 24 sub-groups; FSCs grouped into 85 groups.
- Table A.2: List of macro-groups for PSCs and FSCs (preserved labels and codes). Examples (not exhaustive):
  - PSCs: A - Research and Development; B - Special Studies/analysis, Not R&d; C - Architect/engineer Services; D - IT and Telecommunication - Information Technology and Telecommunications; ...; 7A - IT and Telecommunication - Applications; 7B - IT and Telecommunication - Compute; ...; 7K - IT and Telecommunication - Storage.
  - FSCs: 10 - Weapons; 11 - Nuclear Ordnance; 12 - Fire Control Equipment; 13 - Ammunition and Explosives; 14 - Guided Missles; 15 - Aerospace Craft and Structural Components; ...; 99 - Miscellaneous.

### A.3. Industry Classification by Labor Intensity
- Procedure:
  - Collect BEA data on value-added and its decomposition—compensation of employees, taxes on production and imports less subsidies, and gross operating surplus—for 66 macro-industries starting from 1997.
  - Compute industry’s average labor share as the average of ratios between compensation of employees and value added.
  - Compute median average labor intensity.
  - Classify an industry as high-labor intensive if its average is above the median; low-labor intensive otherwise.
  - Assign contracts to industries using the first two, three, or four digits of the contractor’s 6-digit NAICs code.
- Table A.3: Industry listings by labor intensity (preserved codes and labels). Examples:
  - Low-Labor Intensive Industries include: 111-112 - Farms; 113-115 - Forestry, fishing, and related activities; 211 - Oil and gas extraction; 22 - Utilities; 311-312 - Food and beverage and tobacco products; 331 - Primary metals; 42 - Wholesale trade; 441 - Motor vehicle and parts dealers; 511 - Publishing industries (includes software); 521-522 - Federal Reserve banks, credit intermediation, and related activities; 524 - Insurance carriers and related activities; 531 - Real Estate; 5411 - Legal services; 562 - Waste management and remediation services; 711-712 - Performing arts, spectator sports, museums, and related activities; 721 - Accommodation; 81 - Other services, except government.
  - High-Labor Intensive Industries include: 213 - Support activities for mining; 23 - Construction; 313-314 - Textile mills and textile product mills; 315, 316 - Apparel and leather and allied products; 321 - Wood products; 323 - Printing and related support activities; 326 - Plastics and rubber products; 332 - Fabricated metal products; 333 - Machinery; 35 - Electrical equipment, appliances, and components; 3364, 3369 - Other transportation equipment; 337 - Furniture and related products; 339 - Miscellaneous manufacturing; 445 - Food and beverage stores; 452 - General merchandise stores; 482 - Rail transportation; 481 - Air transportation; 484 - Truck transportation; 483 - Water transportation; 487-488, 492 - Other transportation and support activities; 485 - Transit and ground passenger transportation; 491 - Federal Government Enterprises; 486 - Pipeline transportation; 493 - Warehousing and storage; 511 - Publishing industries (includes software) appears under low-labor but related IT services (5415) appear under high-labor; 561 - Administrative and support services; 611 - Educational services; 621 - Ambulatory health care services; 622 - Hospitals; 623 - Nursing and residential care facilities; 624 - Social assistance; 713 - Amusements, gambling, and recreation industries; 722 - Food services and drinking places.
- Distributional finding:
  - Out of the 66 industries, half are classified as low-labor intensive and half as high-labor intensive.
  - Contracts in final sample are allocated to all 66 BEA industry groups.
  - DoD spending distribution: about 85% of spending is directed to firms operating in high-labor-intensive industries.

### A.4. Imputation Strategy
- Goals: maximize time series availability for estimating effects of local government spending shocks; address missing industry information and pre-1997 labor-intensity classification.
- Missing industry information (challenge 1):
  - Industry data only available from 1988 for contractors.
  - Use product codes to assign industry to contracts missing contractor industry information.
  - Sequential imputation steps:
    1. Assign the BEA industry that receives the largest share of spending for a specific product code to the contract with missing industry information.
    2. If multiple BEA industries have the same share, assign the finer industry with the highest share of contracts.
    3. If prior steps fail and remaining potential industries all belong to the same labor-intensity group, classify contract as high- or low-labor-intensive without imputing the exact industry.
    4. If none of the above succeed, classify product code as not imputed.
  - Missingness statistics:
    - Industry information missing for about 5% of contracts awarded after 1979; these account for 17% of total spending after 1979.
    - Using imputation strategy, industry cannot be imputed for less than 1% of contracts, corresponding to less than 1% of spending.
    - Vast majority of imputed contracts (accounting for 16% of total spending) are imputed using the first step.
    - Second and third steps yield only minor gains in imputation.
- Pre-1997 labor-intensity classification (challenge 2):
  - Data on value-added components only available from 1997.
  - Apply the classification of a given industry based on data for 1997−2020 to prior years (i.e., classification is held fixed before 1997).
- Assumptions and validation:
  - Assumption 1: share of spending allocated to a product code is a good predictor of the industry that produces the product.
    - Validation: for each product code, compute share of spending covered by the industry assigned as imputed industry.
      - On average, the share is over 70%.
      - For over 75% of product codes, the assigned industry covers over 50% of spending in that product code.
      - Implication: spending for product codes is concentrated in a dominant industry.
  - Assumption 2: classification of industries as high or low labor-intensive is stable over time.
    - Validation: re-classify industries annually and compare to benchmark classification based on full sample.
      - Annual classification matches full-sample classification for more than 91% of cases.
      - Mismatches are distributed across about one-third of industries and are not dominated by a few industries.
      - Apart from a couple of industries, mismatches do not follow a trend suggesting technological shifts.
    - Implication: classification between high- and low-labor-intensive is quite stable over time.

### A.5. Patent Data
- Data source: PatentsView.
- Collection date: end of 2019.
- Coverage: PatentsView database contains the universe of granted patents from the US Patent and Trademark Office (USPTO), starting from 1976 until [text truncated in source].

*Italic: References and Appendix materials are drawn from the content unit "wpiea2023052-print-pdf - References."*

### 2019.   These  data  contain  patent-level  information,  including  application  dates,  the  type  of  patent,  the

### Heterogeneous Spending, Heterogeneous Multipliers

### Patent data and sample construction
- Data contain patent-level information: application dates, patent type, inventors’ names, latitude and longitude of inventor addresses, and citation network (number of citations made to and received from other patents).
- Restrict analysis to utility patents (also known as “patents for invention”), which account for about 98% of the universe of patents granted by the USPTO.
- Focus on patents with an application year starting from 1976.
- Identify the year when an innovation occurs as the application year of a patent (the year when the provisional application is considered complete by the USPTO and a filing date is set).
- Final data include only patents for which the DoD has no economic interest to better capture private innovation efforts.
- Restrict analysis to patents that applied by the end of 2006 to address truncation bias (the review process essentially takes 8 years to be fully completed).
- Remove all MSAs with incomplete histories in granted patents or citations.
- USPTO completeness (as reported): “As of 12/31/2012, utility patent data, as distributed by year of application, are approximately 95% complete for utility patent applications filed in 2004, 89% complete for applications filed in 2005, 80% complete for applications filed in 2006, 67% complete for applications filed in 2007, 49% complete for applications filed in 2008, 36% complete for applications filed in 2009, and 19% complete for applications filed in 2010.”

### Adjustments for citation and time biases
- Three major issues affecting over-time comparison of patent statistics:
  - Changes in the propensity to cite (patenting and citing activity accelerated starting in the 1980s).
  - Lifespan of a patent: older patents have more time to accumulate citations.
  - Truncation bias: patents in the last years of the sample are mechanically less cited because records are released at grant dates and review takes time.
- Implement a standard “quasi-structural” approach to correct for the propensity-to-cite bias.
- Measure quality-adjusted innovation rates within a fixed window of five years after the grant year to make citing activities comparable between patents with different lifespans.

### Geolocation and MSA aggregation
- Associate a patent with an MSA using the latitude and longitude of inventor addresses.
- Geolocate coordinates into MSAs using the 2010 TIGER/Line Shapefile constructed by the US Census Bureau.
- When patents have multiple inventors in different MSAs, split patents equally across MSAs according to the number of inventors.
- Restrict geographical aggregation to patents with at least one inventor located in the US mainland.

### Innovation measures
- Construct two measures of innovation quality:
  - Number of patents filed at the USPTO (counts).
  - Patents weighted by the number of adjusted forward citations within a 5-year window after the grant year (quality-weighted measure).

### Additional empirical results — Local Fiscal Multipliers (selected findings and statistics)
- Sample and estimation notes:
  - Balanced panel in many specifications consists of 12,024 observations and includes 334 MSAs for the period 1979−2019. Some specifications use alternate balanced panels (e.g., 8,684 observations and 334 MSAs for 1989−2019; 8,242 observations and 317 MSAs for 1990−2019).
  - Standard errors are clustered by MSA. All regressions include MSA and year-fixed effects.
  - Instruments are computed as specified in the paper equations referenced in the notes.
  - Significance symbols: *, **, and *** represent the 10%, 5%, and 1% significance levels, respectively.

- Table B.1 — Local Fiscal Multiplier: Aggregate Spending (Dependent variable: Employment; Labor Income)
  - Employment coefficients on Spending:
    - tt+1: 0.139*** (0.051), F-stat 8.299
    - tt+2: 0.414*** (0.153), F-stat 9.330
    - tt+3: 0.453*** (0.153), F-stat 10.101
    - tt+4: 0.550** (0.236), F-stat 7.680
    - tt+? (column 5): 0.499*** (0.192), F-stat 10.675
  - Labor Income coefficients on Spending:
    - tt+1: 0.253*** (0.095)
    - tt+2: 0.834*** (0.286)
    - tt+3: 0.873*** (0.285)
    - tt+4: 1.046** (0.424)
    - tt+? (column 5): 0.941*** (0.339)
  - Balanced panel: 12,024 observations, 334 MSAs, period 1979−2019.

- Table B.2 — Local Fiscal Multipliers: Spending by Components (Services vs Goods)
  - Dependent variable: Employment — Services Spending coefficients:
    - tt+1: 0.355*** (0.123)
    - tt+2: 0.968*** (0.167)
    - tt+3: 1.191*** (0.199)
    - tt+4: 1.454*** (0.206)
    - col5: 1.208*** (0.217)
  - Dependent variable: Employment — Goods Spending coefficients:
    - tt+1: 0.059 (0.041)
    - tt+2: 0.165 (0.128)
    - tt+3: 0.157 (0.121)
    - tt+4: 0.186 (0.177)
    - col5: 0.174 (0.167)
  - Tests βk_g = βk_s reported with p-values in brackets (e.g., [.0333], [.0003], [.0000], [.0000], [.0001]).
  - Dependent variable: Labor Income — Services Spending coefficients:
    - tt+1: 0.585*** (0.151)
    - tt+2: 1.924*** (0.331)
    - tt+3: 2.424*** (0.403)
    - tt+4: 2.892*** (0.464)
    - col5: 2.448*** (0.510)
  - Dependent variable: Labor Income — Goods Spending coefficients:
    - tt+1: 0.078 (0.064)
    - tt+2: 0.269 (0.189)
    - tt+3: 0.227 (0.173)
    - tt+4: 0.260 (0.249)
    - col5: 0.207 (0.220)

- Table B.3 — Local Fiscal Multipliers: Cumulative Spending by Components (Employment; Labor Income)
  - Employment — Services Spending (cumulative):
    - tt+1: 0.355*** (0.123)
    - tt+2: 0.572*** (0.113)
    - tt+3: 0.561*** (0.093)
    - tt+4: 0.547*** (0.072)
    - col5: 0.449*** (0.060)
  - Employment — Goods Spending (cumulative):
    - tt+1: 0.059 (0.041)
    - tt+2: 0.099 (0.071)
    - tt+3: 0.092 (0.059)
    - tt+4: 0.083 (0.061)
    - col5: 0.073 (0.053)
  - Labor Income — Services Spending (cumulative):
    - tt+1: 0.585*** (0.151)
    - tt+2: 1.086*** (0.198)
    - tt+3: 1.100*** (0.180)
    - tt+4: 1.079*** (0.156)
    - col5: 0.900*** (0.143)
  - Labor Income — Goods Spending (cumulative):
    - tt+1: 0.078 (0.064)
    - tt+2: 0.155 (0.106)
    - tt+3: 0.133 (0.084)
    - tt+4: 0.114 (0.086)
    - col5: 0.091 (0.072)

- Table B.4 — Local Fiscal Multipliers: Outlays by Components
  - Employment — Services Spending:
    - tt+1: 0.283*** (0.089)
    - tt+2: 0.334*** (0.125)
    - tt+3: 0.413* (0.227)
    - tt+4: 0.405 (0.261)
    - col5: 0.423* (0.249)
  - Employment — Goods Spending:
    - tt+1: -0.108 (0.124)
    - tt+2: -0.122 (0.127)
    - tt+3: -0.115 (0.116)
    - tt+4: -0.086 (0.094)
    - col5: -0.104 (0.103)
  - Labor Income — Services Spending:
    - tt+1: 0.865*** (0.292)
    - tt+2: 1.016** (0.410)
    - tt+3: 1.219** (0.564)
    - tt+4: 1.240** (0.615)
    - col5: 1.333** (0.607)
  - Labor Income — Goods Spending:
    - tt+1: -0.263* (0.151)
    - tt+2: -0.310* (0.161)
    - tt+3: -0.315* (0.171)
    - tt+4: -0.261* (0.154)
    - col5: -0.302* (0.166)
  - Balanced panel for these outlays specifications: 8,684 observations, 334 MSAs, period 1989−2019.

- Table B.5 — Local Fiscal Multipliers: Excluding top 1% contracts by value
  - Employment — Services Spending:
    - tt+1: 1.348*** (0.300)
    - tt+2: 2.435*** (0.522)
    - tt+3: 2.759*** (0.651)
    - tt+4: 2.843*** (0.752)
    - col5: 2.497*** (0.643)
  - Labor Income — Services Spending:
    - tt+1: 2.059*** (0.750)
    - tt+2: 4.336*** (1.241)
    - tt+3: 5.044*** (1.393)
    - tt+4: 5.145*** (1.610)
    - col5: 4.523*** (1.425)
  - Balanced panel: 12,024 observations, 334 MSAs, period 1979−2019.

- Table B.6 — Local Fiscal Multipliers: Pre-determined Shift Share
  - Employment — Services Spending:
    - tt+1: 0.156* (0.080)
    - tt+2: 0.526*** (0.162)
    - tt+3: 0.567*** (0.172)
    - tt+4: 0.717*** (0.226)
    - col5: 0.615*** (0.215)
  - Employment — Goods Spending:
    - tt+1: 0.003 (0.029)
    - tt+2: -0.017 (0.103)
    - tt+3: -0.028 (0.115)
    - tt+4: -0.061 (0.138)
    - col5: -0.056 (0.125)
  - Labor Income — Services Spending:
    - tt+1: 0.432*** (0.152)
    - tt+2: 1.458*** (0.377)
    - tt+3: 1.663*** (0.391)
    - tt+4: 2.109*** (0.495)
    - col5: 1.987*** (0.500)
  - Labor Income — Goods Spending shows negative and sometimes significant coefficients at later horizons.
  - Balanced panel for shift-share instrument: 8,242 observations, 317 MSAs, period 1990−2019; instrument calculated over 1979−1989.

- Table B.7 — Labor Intensity and Employment Concentration Ratios
  - Employment Concentration Ratio coefficients (by largest firms):
    - 4 largest firms: -0.001* (0.000), Observations 198, Number 4-digits Industries 99
    - 8 largest firms: -0.000 (0.001), Observations 204, Number 4-digits Industries 102
    - 20 largest firms: -0.001 (0.001), Observations 204, Number 4-digits Industries 102
    - 50 largest firms: -0.001 (0.001), Observations 204, Number 4-digits Industries 102
  - Balanced panel includes years 2002 and 2017.

- Table B.8 — Local Fiscal Multipliers: Labor Intensity (High vs Low)
  - Employment — Services Spending - High Labor Intensity:
    - tt+1: 0.365** (0.141)
    - tt+2: 0.978*** (0.172)
    - tt+3: 1.183*** (0.257)
    - tt+4: 1.554*** (0.302)
    - col5: 1.276*** (0.307)
  - Employment — Services Spending - Low Labor Intensity:
    - tt+1: 0.016 (0.574)
    - tt+2: -0.164 (0.736)
    - tt+3: -0.259 (0.881)
    - tt+4: -0.839 (0.993)
    - col5: -0.868 (0.951)
  - Employment — Goods Spending - High Labor Intensity:
    - tt+1: 0.073** (0.032)
    - tt+2: 0.283* (0.150)
    - tt+3: 0.302** (0.148)
    - tt+4: 0.537* (0.309)
    - col5: 0.476** (0.240)
  - Employment — Goods Spending - Low Labor Intensity:
    - tt+1: -0.062 (0.074)
    - tt+2: -0.190* (0.113)
    - tt+3: -0.175* (0.090)
    - tt+4: -0.331** (0.148)
    - col5: -0.336** (0.150)
  - Labor Income — Services Spending - High Labor Intensity:
    - tt+1: 0.504*** (0.161)
    - tt+2: 1.828*** (0.345)
    - tt+3: 2.278*** (0.474)
    - tt+4: 2.968*** (0.572)
    - col5: 2.483*** (0.625)
  - Labor Income — Goods Spending - High Labor Intensity:
    - tt+1: 0.114** (0.047)
    - tt+2: 0.495** (0.207)
    - tt+3: 0.464** (0.189)
    - tt+4: 0.795** (0.391)
    - col5: 0.615** (0.286)
  - Labor Income — Goods Spending - Low Labor Intensity shows negative coefficients at several horizons (e.g., tt+3: -0.259* (0.140), tt+4: -0.474** (0.202)).
  - Balanced panel: 12,024 observations, 334 MSAs, period 1979−2019.

*Heterogeneous Spending, Heterogeneous Multipliers — Working Paper No. WP/2023/052*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023052-print-pdf.pdf_