## wp18276

## Source details

**Canonical URL:** [wp18276](https://www.imf.org/-/media/files/publications/wp/2018/wp18276.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2018/wp18276.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2018/wp18276.pdf.json)

---

### Purpose and contributions
- First paper to examine the channels through which highways affect firm and aggregate productivity growth.
- Two main contributions:
  - Decomposes the change in aggregate productivity resulting from highway connection into four channels: within-firm productivity growth, entry of new firms, reallocation between existing firms, and exit of inefficient firms.
  - Develops a time-varying instrumental variable constructed from engineering and network theory (a least-cost path spanning tree network) to address endogeneity of both the location and timing of highway placement.

### Data, market-access construction, and empirical strategy
- Transportation-cost and network data
  - Road transportation costs estimated between any two prefecture-level cities using price quote data from logistics companies and digitized maps of China’s highway network.
  - Market access estimated for 339 prefecture-level cities in China for the period 1995–2005.
  - Price-quote sources: Arima World Group and Deppon; for each city pair the lower quote is used.
  - Port access measured by transportation cost to the nearest of 9 largest ports (lowest cost chosen).
- Firm-level data and productivity measurement
  - Annual Survey of Industrial Enterprises (ASIE) firm-level data (1998–2007; survey spans 1998–2009) covering SOEs and private firms with annual sales of $650,000 and above.
  - Firm TFP estimated using Olley and Pakes (1996); Levinsohn and Petrin (2003); Ackerberg, Caves, and Frazer (2015); Wooldridge (2009); markups estimated with De Loecker and Warzynski (2012).
- Instrumental variable (time-varying)
  - Constructs a least-cost path spanning-tree network (Kruskal minimum spanning tree using pixel-level construction-costs from terrain ruggedness) as an instrument for actual city market access.
  - Uses Girvan-Newman centrality ranking and actual annual length of highway construction to predict timing of edges being “built.”
  - First-stage: hypothetical market access strongly predicts actual market access; Table 7 first-stage F statistic: 723.433, 489.353, 656.805. log MA (IV) coefficients: 0.967***, 0.922***, 0.960*** (standard errors: (0.035), (0.041), (0.036)).
- Counterfactual construction
  - Baseline counterfactual: removing all highways; compute counterfactual bilateral trade costs τod,t, cities’ counterfactual market access, and resultant city/productivity impacts.
  - Median loss in market access by removing all highways: 62%.

### Key empirical findings and magnitudes
- Aggregate productivity impact (counterfactual removing all highways)
  - Eliminating all highways in China would decrease aggregate productivity by 3.2%.
  - Table 3 / baseline counterfactual (holding city population constant from 2005): Estimated Decrease in Aggregate TFP without Highways: 3.16%.
  - Allowing population redistribution (using 1995 population distribution as proxy for counterfactual without highways): removing all highways would lower aggregate TFP by 3.8%.
  - Additional reported rows:
    - Holding the population distribution from 2000 — 3.12%
    - Holding the population distribution from 1995 — 3.09%
    - Allowing for changes in the distribution of population (holding total population constant) — 3.81%
- Firm-level market-access effects
  - OLS summary: A 1% increase in market access increases firm productivity by 0.05%. A one-standard-deviation increase in market access would boost firm TFP by 5%.
  - Access to international market: A 1% reduction in domestic transportation costs to ports boosted firm TFP in the city by 0.6%.
  - OLS firm-TFP coefficients (selected from Table 2; standard errors in parentheses):
    - 0.134*** (0.023)
    - 0.065*** (0.019)
    - 0.064*** (0.018)
    - 0.069*** (0.019)
    - 0.064*** (0.022)
    - 0.042** (0.020)
    - 0.040** (0.020)
    - 0.046** (0.020)
  - Access-to-port coefficients (Table 2; standard errors in parentheses):
    - 0.087** (0.034) (repeated across specifications)
  - Observations and fit (examples from Table 2):
    - No. of observations: 612484, 598381, 597085, 597085, 594476, 612484, 598381, 597085, 597085, 594476
    - Adjusted. R squared reported: 0.380, 0.655, 0.656, 0.656, 0.656
- IV results and interpretation
  - IV second-stage estimates are slightly smaller than OLS estimates, suggesting OLS upward bias from unobserved forces correlated with cities’ market access expansion.
  - Many robustness checks reported; 2SLS corroborates baseline OLS findings though some coefficients attenuate.
- Prices, marginal costs, and markups
  - Revenue-productivity vs physical-productivity concerns: a 1% increase in market access leads to a 0.12% reduction in firms’ marginal costs (Table 10).
  - Market Access effects on prices and marginal costs (selected Table 10 and Table 11 statistics):
    - Table 10 (Marginal Cost): Market Access coefficients: -0.155***, -0.170***, -0.120** (standard errors: (0.048), (0.049), (0.046)).
    - Table 10 (Price): Market Access coefficients: -0.084*, -0.086*, -0.095** (standard errors: (0.046), (0.047), (0.047)).
    - Table 11 (Market Access on TFPR / revenue productivity): Market Access coefficients: 0.052***, 0.053***, 0.051***, 0.053*** (standard errors: (0.015), (0.014), (0.015), (0.014)).
    - Table 9 (Market Access and markups): Market Access coefficients (OLS/IV): -0.007, -0.002, -0.003, -0.001 (standard errors: (0.020), (0.018), (0.022), (0.019)).
  - Interpretation: firms pass most of cost advantages to consumers; markups show no strong change.
- Transport-cost estimation performance
  - Regression of logistics price quotes on network measures (Table 5) — model fit:
    - Road data + origin/destination fixed effects explains 85% of variation in bilateral transportation costs (Column (2), Table 5).
    - Road data alone explains 69% of variation (Column (1), Table 5).
    - Length of highway coefficients: 0.454***, 0.539***, 1.331***, 1.105*** (standard errors: (0.001), (0.001), (0.007), (0.008)).
    - No. of observations: 97253 (across columns reported); Adjusted. R squared: 0.689, 0.850, 0.869, 0.875.

### Decomposition of productivity gains and channels
- Four channels considered (Haltiwanger (1997) decomposition):
  - Within-firm productivity growth among continuing firms.
  - Reallocation of market shares across continuing firms (between incumbents).
  - Entry of new productive firms (extensive and intensive margins).
  - Exit of inefficient firms.
- Main quantitative decomposition results (selected Table 4 coefficients; standard errors in parentheses)
  - Panel (a) OLS — Change in domestic market access: Total effect 0.061** (0.031)
    - Within 0.039*** (0.012)
    - Between 0.014*** (0.005)
    - Entry 0.045** (0.020)
    - Exit 0.009*** (0.002)
  - Panel (b) IV — Change in domestic market access: Total effect 0.080** (0.032)
    - Within 0.007 (0.012)
    - Between 0.021*** (0.006)
    - Entry 0.055** (0.023)
    - Exit 0.011*** (0.003)
  - Change in access to ports (OLS): Total effect 0.139*** (0.027)
    - Within 0.065*** (0.018)
    - Entry 0.097*** (0.020)
    - Exit -0.000 (0.004)
  - Change in access to ports (IV): Total effect 0.134*** (0.027)
    - Within 0.062*** (0.017)
    - Entry 0.095*** (0.020)
    - Exit -0.001 (0.004)
- Empirical conclusions on channels
  - Entry of new and productive firms contributed most to TFP gains.
  - Reallocation among large incumbents also contributed to TFP growth.
  - Exit of inefficient firms made the smallest contribution to TFP gains.
  - Little evidence for within-firm productivity growth as a major channel overall for domestic market access (IV within effect small/insignificant), but reduction in transportation costs to ports promoted within-firm TFP growth.

### Industry heterogeneity and spatial reallocation
- Industry characteristics interacting with market access:
  - Capital intensity (K/L), weight-to-value (iceberg transport cost proxy), and degree of product differentiation predict heterogeneous responses.
- Key heterogeneous-coefficient results (selected tables; standard errors in parentheses)
  - Table 14 (Capital Intensity): K/L*∆Market Access coefficients: 0.642***, 0.644***, 0.462***, 0.463*** (standard errors: (0.136), (0.136), (0.076), (0.078)).
  - Table 15 (Weight-to-Value Ratio): Weight/Value*∆Market Access coefficients: 0.004***, 0.004***, 0.002***, 0.003*** (standard errors: (0.001), (0.001), (0.001), (0.001)).
  - Table 16 (Product Differentiation): ProdDiff*∆Market Access coefficients: -0.008**, -0.009**, -0.007***, -0.009*** (standard errors: (0.004), (0.004), (0.002), (0.002)).
- Quantified implication example
  - A one-standard deviation of a city’s market access would increase annual employment growth in the sector at the 75th percentile of capital intensity by 1.6 percentage points more than in the sector at the 25th percentile (period 2001–2008).

### Identification, robustness, and methodological insights
- Instrument validity and robustness
  - Time-varying least-cost spanning-tree instrument addresses both location and timing endogeneity.
  - Alternative specifications tested: Euclidean-distance spanning tree, inclusion/exclusion of “node” cities, rail/water access controls, nonlinear travel-time terms; results robust.
  - Concern of spatially correlated errors acknowledged; follow-up methods (König and others (2015)) used to correct standard errors.
- Market-access measure and model
  - Market access derived from a multi-city trade model with heterogeneous firms; market-access term ∑j w−γj Yj τγkj informs average firm productivity and aggregate productivity.
  - Numerical solution of a system of 339 equations used to solve MA_i as in Donaldson and Hornbeck extensions.
- Robustness checks (select)
  - Using population of all cities in 1995 so time variation in market access comes only from bilateral transportation costs does not substantially alter results; transportation cost channel slightly larger.
  - Excluding 52 government-targeted “node” cities does not change results.
  - Alternative market access computed numerically (Donaldson and Hornbeck (2016) method) yields very similar regression results.

### Policy implications, caveats, and avenues for future research
- Policy implications
  - National highway network promoted production efficiency, delivered sizable welfare gains to consumers, and led to sectoral reallocation across cities.
  - Transport infrastructure can be a tool to boost aggregate demand in the short run and promote long-run growth; infrastructure investments can yield sizable economic impact when domestic transportation costs are large and misallocation is pervasive.
  - Central policy question highlighted: what is the optimal amount of investment in infrastructure?
- Caveats and limitations
  - Analysis focuses on manufacturing; agricultural and service sectors omitted.
  - Model abstracts from trade in intermediate goods (which accounts for more than 70 percent of total trade); omission may affect magnitudes.
  - Exit decisions may be influenced by economic and political factors not captured here.
  - With restricted labor mobility during the period, empirical gains likely represent a lower bound relative to a full labor-mobility counterfactual.
- Suggested future research
  - Quantitative cost-benefit analysis comparing actual highway construction with counterfactual scenarios.
  - Incorporate inter-sectoral linkages to construct output market access and input market access using input-output tables.
  - Micro-level studies at county or village level to assess effects on poverty and income inequality.

*Italic: Source — wp18276 (PDF chapter/section).*

### References .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .  .

### wp18276 - References

### Purpose and contributions
- First paper to examine the channels through which highways affect firm and aggregate productivity growth.
- Two main contributions:
  - Decomposes the change in aggregate productivity resulting from highway connection into four channels: within-firm productivity growth, entry of new firms, reallocation between existing firms, and exit of inefficient firms.
  - Develops a time-varying instrumental variable constructed from engineering and network theory (a least-cost path spanning tree network) to address endogeneity of both the location and timing of highway placement.

### Data and empirical strategy
- Transportation costs
  - Estimates road transportation costs between any two prefecture-level cities in China using price quote data from logistics companies and digitized maps of China’s highway network.
  - Market access is estimated for 339 prefecture-level cities in China for the period 1995–2005.
- Firm-level productivity
  - Firm TFP is estimated using standard procedures in industrial organization: Olley and Pakes (1996); Levinsohn and Petrin (2003); Ackerberg, Caves, and Frazer (2015); Wooldridge (2009).
- Instrumental variable
  - Constructs a time-varying least-cost path spanning tree network as an instrument for actual city market access.
  - Uses the instrument to perform 2SLS regressions; results confirm baseline OLS findings.
- Counterfactual analysis
  - Chooses the scenario of removing all highways as the baseline counterfactual to evaluate aggregate economic impact.
  - Calculates counterfactual trade costs between all city-pairs and cities’ counterfactual market access in absence of highways.

### Key findings and magnitudes
- Aggregate productivity impact (counterfactual removing all highways)
  - Eliminating all highways in China would decrease aggregate productivity by 3.2%.
  - If highways are allowed to influence the distribution of population across cities, the estimated impact on productivity is 3.8%.
- Channels of TFP gains
  - Entry of new and productive firms contributed most to TFP gains.
  - Reallocation among large incumbents contributed to TFP growth.
  - Exit of inefficient firms made the smallest contribution to TFP gains.
  - Little evidence for within-firm productivity growth as a major channel overall; however, a reduction in transportation costs to ports promoted within-firm TFP growth.
- Industry and spatial effects
  - Industries with larger transportation costs and higher capital intensity tend to concentrate in locations with better highway access.
  - Findings are consistent with New Economic Geography models and complement empirical studies on geographic concentration of industries.

### Methodological insights and robustness
- Market access measure
  - Derived from a multi-city trade model with heterogeneous firms; summarizes direct and indirect impacts of transportation costs on city aggregate productivity.
  - Addresses treatment spillover effects where treatment intensity differs across cities despite an integrated national objective.
- Identification and robustness
  - Addresses endogenous highway assignment with a time-varying IV based on a hypothetical least-cost network from engineering and network theory.
  - Performs various robustness checks to test validity of results; 2SLS estimates corroborate OLS findings.

### Policy implications and caveats
- Policy question raised
  - Results highlight the central policy question: what is the optimal amount of investment in infrastructure?
- Caveats and limitations
  - Infrastructure effects depend on channels and time horizon; omission of the agricultural and service sectors and abstraction from trade in intermediate goods may bias results.
  - Exit decisions may be subject to other economic and political factors, which could limit measured contributions from firm exit.
  - The estimated aggregate effect, while sizable, may not be as large as commonly assumed.

### Background facts on China’s highway program
- Policy context
  - In 1992 the Chinese State Council approved construction of the “7-5” network.
  - The NTHS plan outlined construction of 12 trunk highway roads: five longitudinal roads and seven latitudinal roads.
  - Objective: connect all provincial capitals and cities with an urban population above 500,000 through a National Trunk Highway System (NTHS) by 2020.
- Timing and scale
  - China experienced rapid highway expansion in the late 1990s to mid-2000s, moving from virtually no highway to an integrated national highway network.

*Source: wp18276 - References (wp18276.pdf)*

### 2003. The completion of that project marked the beginning of an integrated domestic market

### wp18276 - 2003. The completion of that project marked the beginning of an integrated domestic market

### Highway expansion, market integration, and trade
- The National Trunk Highway System (NTHS) rapid expansion coincided with a drastic expansion of Chinese cities’ access to the international market, especially in coastal areas.
- The majority (76%) of the NTHS is expressways; the rest are highways.
- Inter-provincial trade as a percentage of GDP increased from 82% to 108% for the period from 2002 to 2007.
- The rapid expansion of the highway network is reflected in modal average travel distances: road transportation experienced the largest increase in average travel distance among all transport modes.
- Road transportation gradually became the main mode for inter-city trade after 2000; estimates indicate road transport accounted for nearly 70% of freight value and logistics costs in 2007.
- Average transport distance of highways nearly tripled from 60 kilometers in 2000 to 172 kilometers in 2008; average transport distance of railways and waterways remained at 760 kilometers and 1800 kilometers respectively during the same period.
- Example bilateral iceberg transportation costs (Table 1 excerpts):
  - Yangquan → Taiyuan: 1.04%
  - Chengdu → Zigong: 1.30%
  - Huainan → Lasa: 20.32%
  - Lasa → Kelamayi: 23.70%
  - Shannan → Bayingguole: 28.11%
  - Beijing → Shanghai: 5.54%

### Empirical strategy and identification
- The analysis exploits a national highway “shock” (NTHS construction) and contemporaneous trade liberalization (WTO accession end-2001) while controlling for industry-year fixed effects to separate tariff/industry policy changes from transportation-cost effects.
- Market access is constructed with estimated bilateral iceberg trade costs τod,t and destination GDP Yd, using the baseline market potential formulation:
  - MAo ≈ ∑d τ−θod Yd
- Highway-based least-time-cost routes are used as optimal ground-transport routes; rail and water access are controlled for in regressions and via interaction terms to allow potentially time-varying effects.
- To estimate historical bilateral iceberg trade costs, contemporary price quotes from two large logistics firms are regressed on road-network measures plus origin and destination fixed effects:
  - τo,d,15 = α + β1 hwyo,d,15 + β2 localo,d,15 + feo + fed + εo,d,15
  - Estimated historical τ̂o,d,t = α̂ + β̂1 hwyo,d,t + β̂2 localo,d,t + f̂eo + f̂ed
- Goodness-of-fit of transportation-cost prediction:
  - Model with road data and origin/destination fixed effects explains 85% of variation in bilateral transportation costs (Column (2), Table 5).
  - Road data alone explains 69% of variation (Column (1), Table 5).
  - Adding nonlinear terms and travel time increases R-squared only marginally (Columns (3) and (4), Table 5).
- Use of logistics companies’ published price quotes:
  - Two firms: Arima World Group and Deppon (price quotes used as observed shipping costs).
  - Acknowledged discounts exist (established clients: 10%-20% discount; occasional discounts up to 50%), but quotes are argued to reflect market prices closely; for each city pair the lower of the two companies’ quotes is used as the shipping cost.
- Value-to-weight (industry) adjustments: revenue and quantity data are used to estimate industry-specific value-to-weight ratios to convert price-quote-based costs into iceberg trade-cost equivalents.

### Multi-city trade model with heterogeneous firms — analytic findings and predictions
- Preferences and structure:
  - Utility over goods from sectors s ∈ 0,1,...,S given by CES nesting with sectoral expenditure shares βs and elasticities σs.
  - Homogeneous numeraire good 0 produced in every city under constant returns; wage in city k equals wk (price of good 0 normalized to 1).
- Production and firm selection:
  - Firms pay fixed cost fskj and iceberg trade costs τskj; marginal cost expressed in origin labor terms.
  - Optimal price for a firm with productivity φ selling from k to j:
    - pskj(φ) = (σs/(σs − 1)) wk τskj / φ
  - Zero-profit export cutoff productivity φskj is given by:
    - φskj = λs (fskj βs Yj)1/(σs − 1) / (wk τskj Psj)
  - Domestic survival productivity φsk analog similarly with τkk = 1.
- Quantitative implications under Pareto productivity (g(φ) = γ φminγ φ−(γ+1), G(φ) = 1 − (φmin/φ)γ) and entrant mass ∝ wj Lj:
  - Price index, cutoffs, and firm-level export values derived yield a market-potential-like term:
    - φk = κ4 × wk Y1/γk × (∑j w−γj Yj τγkj)1/γ
  - Average firm productivity (TFPk) and aggregate (revenue-weighted) productivity (T̂FPk) are increasing functions of a city’s market access term ∑j w−γj Yj τγkj.
- Main comparative predictions:
  - Cities that can reach large markets (large Yj) cheaply (small τkj) have higher productivity cutoffs and thus higher average firm productivity.
  - Transport infrastructure projects that reduce τkj for a city relative to others should produce a relative increase in that city’s production efficiency.
  - Efficiency gains arise via three channels: entry of new productive firms, exit of inefficient firms, and reallocation across firms. The paper plans to examine the relative importance of these channels in Section 6.

### Data sources, measurement choices, and robustness considerations
- Primary data sources:
  - Digitized maps of China’s highway network (1992–2015) plus local roads, rail, and waterways (vector data compiled from road atlases and digitizations; least-cost-paths constructed for 1995, 2001, and 2005).
  - Annual Survey of Industrial Enterprises (ASIE) firm-level data (1998–2007; survey spans 1998–2009) covering all SOEs and all private firms with annual sales of $650,000 and above; contains identifiers, gross output, total sales, wage bill, employment, fixed capital stock, intermediate inputs, export status, establishment year, ownership, and location.
  - Firm census and economic censuses (1995, 1996, 2001, 2004, 2008) used to construct industry-prefecture aggregates.
  - Price quotes from two large logistics companies used to estimate bilateral transportation costs; port access measured by transportation cost to the nearest of 9 largest ports (lowest cost chosen).
- Construction of market access:
  - Market access computed for 340 prefectures using employment and firm counts for 42 2-digit manufacturing industries.
  - Focus on manufacturing firms (manufactured goods are more tradable); KPMG cited estimate: movement of industrial products accounted for 87% of the value of goods moved in 2006.
  - Use of period 2001–2008 for average growth rates justified by timing of highway network emergence and to avoid post-2008 stimulus confounds.
  - Rail and water access are included in regressions; a rail-access × year dummy term included to capture potential time-varying rail effects.
  - Industry heterogeneity in weight-to-value addressed by estimating industry-specific weight-to-value ratios and constructing industry- and city-specific market access; estimated effects robust to these refinements.
- Firm productivity and markup measurement:
  - TFP estimated using augmented Olley-Pakes (1996) approach (investment proxy preferred in Chinese context due to processing trade prevalence).
  - De Loecker and Warzynski (2012) method used to estimate firm markups.
  - Deflators from Brandt and others (2016); perpetual inventory method used for real capital stock construction.
  - Robustness checks using Ackerberg, Caves, and Frazer (2015) approach reported to yield broadly similar but weaker relationships.

### Key analytical conclusions and empirical implications
- Empirical models and derived theory imply a positive relationship between a city’s market access (proximity to large markets and low transport costs) and:
  - Average firm productivity (TFPk) and revenue-weighted aggregate productivity (T̂FPk).
  - Industry clustering and specialization patterns via differential market access across cities.
- Transportation-cost estimation approach leveraging actual logistics price quotes addresses pitfalls of Euclidean-distance proxies and assumed per-unit cost parameters from other contexts; road-network measures plus city fixed effects explain most cross-sectional variation in observed shipping quotes.
- Policy-relevant interpretation:
  - Highway infrastructure lowering τkj for a city increases that city’s market access and should raise average firm productivity through entry, exit, and reallocation channels.
  - Because labor mobility in China was restricted during the period studied, estimated city-level productivity gains from highways likely represent a lower bound of potential effects under full labor mobility.

*Italic: Source — wp18276 (PDF chapter/section) — "2003. The completion of that project marked the beginning of an integrated domestic market"*

### Section 6 to disentangle gains retained by firms from gains passed on to consumers. Table 6

### Section 6 to disentangle gains retained by firms from gains passed on to consumers. Table 6

### Instrumental Variable Approach
- Identification challenge: endogenous placement and timing of transport infrastructure (highways) may bias OLS estimates upward because planners likely build near cities expected to have higher unobserved growth or with better political connections.
- Instrument construction (time-varying):
  - Step 1: Kruskal minimum spanning tree on least-cost path connections between city centers using remote sensing data on terrain ruggedness to estimate construction cost for each land segment; construct least-cost path spanning-tree network that connects all node cities while minimizing construction costs.
  - Step 2: Girvan-Newman algorithm ranks edges by number of shortest paths traversing the edge to predict optimal timing of construction; use actual annual length of highway construction as an exogenous constraint to determine which least-cost edges are “built” each year.
  - Use the algorithm-generated least-cost path spanning-tree network to compute hypothetical city market access and instrument for actual market access.
- Instrument validity and robustness:
  - Time variation addresses both location and timing endogeneity: location via cost minimization, timing via a “centrality” measure.
  - Concerns discussed:
    - Local construction costs may correlate with potential economic returns (e.g., costly-to-build hilly regions with large benefits). Mitigation: include pre-existing railway and waterway access; alternative Euclidean distance spanning tree (minimizes total distance rather than construction cost) yields similar results.
    - Placed-based policies could correlate with edge centrality. No evidence found that city centrality is strongly correlated with pre-trends.
  - First-stage: hypothetically constructed market access is highly correlated with actual-market-access; F statistic is very large (strong first stage) but may reflect spatial correlation in errors; plan to follow König and others (2015) to correct for spatial correlation.

### Market Access and City Productivity — Firm-Level Regressions
- Baseline specification:
  - logTFP_{i,k,t} = α + δ logMA_{k,t} + (X_{i,k,t})′β + λ_k + μ_{s,t} + ε_{i,k,t}
  - Controls: industry×year fixed effects, city fixed effects, city population, interactions of distance to railway × year and distance to waterway × year, firm markup, firm size, ownership structure, exporter status dummy.
  - Market access measured in three periods: 1995, 2001, 2005; firm TFP observations used in 1998, 2003, 2007 to allow for lags.
- OLS findings:
  - Text summary: A 1% increase in market access increases firm productivity by 0.05%. A one-standard-deviation increase in market access would boost firm TFP by 5%.
  - Access to international market: A 1% reduction in domestic transportation costs to ports boosted firm TFP in the city by 0.6%.
  - Within-firm channel: Including firm fixed effects renders the market-access coefficient insignificant (no strong evidence that market access raised TFP within firms); a reduction in transportation costs to the nearest port led to within-firm productivity growth.
  - Markups: No strong evidence that firm markups respond to market access—implies higher production efficiency translated into lower product prices (both firms and consumers reap benefits).
- Key numeric results from Table 2 (OLS and IV; dependent variable: Firm TFP):
  - Market access coefficients shown (columns labeled (1)–(10) in table snippet):
    - 0.134*** (0.023)
    - 0.065*** (0.019)
    - 0.064*** (0.018)
    - 0.069*** (0.019)
    - 0.064*** (0.022)
    - 0.042** (0.020)
    - 0.040** (0.020)
    - 0.046** (0.020)
  - Access to port coefficients:
    - 0.087** (0.034)
    - 0.087** (0.035)
    - 0.088** (0.034)
    - 0.087** (0.034)
    - 0.087** (0.035)
    - 0.088** (0.034)
  - Exporter dummy coefficients: -0.008 (0.009) (repeated)
  - City population coefficients reported (various): -0.031, 0.027, 0.027, 0.027, -0.031, 0.027, 0.027, 0.027 (standard errors in parentheses: (0.080), (0.070), (0.073), (0.074), (0.080), (0.070), (0.071), (0.073))
  - No. of observations listed: 612484, 598381, 597085, 597085, 594476, 612484, 598381, 597085, 597085, 594476
  - Adjusted. R squared: 0.380, 0.655, 0.656, 0.656, 0.656
  - No. of clusters: 339, 339, 335, 335, 335, 339, 339, 335, 335, 335
  - Notes: Standard errors clustered at the city level. *p<0.10 **p<0.05 ***p<0.01.
- 2SLS results:
  - First stage: hypothetical market access strongly predicts actual market access.
  - IV second-stage estimates are slightly smaller than OLS estimates, suggesting upward bias in OLS from unobserved forces correlated with cities’ market access expansion.

### Counterfactual Impact of Removing Highways
- Counterfactual scenario: removing all highways in China (baseline), holding population fixed (from 2005) in baseline.
- Computation steps:
  - Calculate decline in market access for each city if all highways removed. Median loss in market access by removing all highways is 62%.
  - Estimate decline in productivity per city using estimated effect of market access on productivity.
  - Aggregate city productivity losses weighted by city size to obtain national decline in productivity.
- Counterfactual estimates:
  - Eliminating all highways in China would lower aggregate TFP by 3.2%.
  - Allowing population redistribution (use observed population distribution in 1995 as proxy for counterfactual without highways) yields a larger decline: removing all highways would lower aggregate TFP by 3.8%.
  - Table 3 entry: Baseline counterfactual scenario in absence of highways — Estimated Decrease in Aggregate TFP without Highways: 3.16% (holding city population constant from 2005).
- Caveats and limitations:
  - Model omits trade in intermediate goods (which accounts for more than 70 percent of total trade) — introducing intermediates could raise conceptual and empirical complexities beyond this paper’s scope.
  - Agricultural and services sectors are not captured due to data limitations; calculated TFP gain pertains to manufacturing sector only.
  - Counterfactual exercise intended to provide sense of magnitude; more thorough quantification would require addressing the above issues.
  - With perfect labor mobility, city productivity is not determined by market access in the model—estimated empirical impact should serve as a lower bound.

### Decomposition of Productivity Gains from Highway Access
- Four channels considered for aggregate TFP gains due to market access:
  - Within-firm productivity enhancement among continuing firms.
  - Reallocation of market shares across continuing firms.
  - Entry of productive new firms (which can be further decomposed into extensive margin: number of new firms, and intensive margin: productivity of new firms).
  - Exit of inefficient firms.
- Empirical status:
  - Firm-level regressions established that highways increased city productivity, but they do not identify the relative importance of these four channels; Section 6.2 aggregates TFP growth at industry level and decomposes aggregate TFP growth into these channels (details not included in the supplied excerpt).
  - Section 6.3 aims to estimate extent of revenue-based TFP attributable to changes in markups and changes in physical productivity; robustness checks are presented in Section 6.4 (details not included in the supplied excerpt).

*Source: wp18276 - Section 6 to disentangle gains retained by firms from gains passed on to consumers. Table 6 (PDF).*

### 1. Holding the population distribution from 20003.12%

### 1. Holding the population distribution from 20003.12%

### Counterfactual impact of removing all highways
- Row 1: Holding the population distribution from 2000 — 3.12%
- Row 2: Holding the population distribution from 1995 — 3.09%
- Row 3: Allowing for changes in the distribution of population (holding total population constant) — 3.81%
- Notes: This table shows the estimated impact of removing all highways. Row 1-3 reports the counterfactual impact on aggregate TFP from eliminating all highways in China. In the baseline scenario, population distribution is held fixed from 2005. In Row 2, population distribution is assumed to be from 2000. In Row 3, population distribution is held fixed from 1995. In Row 4, I allow for changes in population distribution over time. The population distribution from 1995 is used as a proxy for the counterfactual population distribution without highways. All results assume that total population is held unchanged.

### Decomposition of TFP gains (methodology and key estimates)
- Decomposition follows Haltiwanger (1997): aggregate TFP changes at industry level are decomposed into four terms (within-firm productivity growth; reallocation between incumbent firms; entry of new firms; exit of incumbents). Equation for industry productivity index:
  - lnTFP_s,k,t = Σ_i θ_s,i,k,t lnTFP_s,i,k,t
- Main regression result summaries:
  - A 1% increase in market access leads to a 0.04% increase in the productivity index (aggregate measure).
  - A 1% reduction in transportation costs to the nearest port increases aggregate TFP by 0.14%.
- Sources of aggregate TFP gains:
  - Entry of productive firms contributed the most.
  - Reallocation among large incumbent firms and exit of inefficient firms also contributed.
  - Little evidence of within-firm productivity improvements from increased domestic market access (OLS significant but IV much smaller and insignificant).
  - For access to ports, within-firm productivity growth was an important source of aggregate TFP growth; little evidence on reallocation across large incumbents.

### Table 4 (selected reported coefficients, OLS and IV)
- Panel (a) OLS — coefficients (standard errors in parentheses)
  - Change in domestic market access: Total effect 0.061** (0.031)
    - Within 0.039*** (0.012)
    - Between 0.014*** (0.005)
    - Entry 0.045** (0.020)
    - Exit 0.009*** (0.002)
  - Change in access to ports: Total effect 0.139*** (0.027)
    - Within 0.065*** (0.018)
    - Between 0.006 (0.008)
    - Entry 0.097*** (0.020)
    - Exit -0.000 (0.004)
  - No. of Observations 64136413641364136413
- Panel (b) IV — coefficients (standard errors in parentheses)
  - Change in domestic market access: Total effect 0.080** (0.032)
    - Within 0.007 (0.012)
    - Between 0.021*** (0.006)
    - Entry 0.055** (0.023)
    - Exit 0.011*** (0.003)
  - Change in access to ports: Total effect 0.134*** (0.027)
    - Within 0.062*** (0.017)
    - Between 0.006 (0.008)
    - Entry 0.095*** (0.020)
    - Exit -0.001 (0.004)
  - No. of Observations 64136413641364136413
- Notes: Regressions include city population, provincial capital dummy, average years of schooling, access to railways, access to waterways and industry fixed effects. Robust standard errors in parentheses. *p<0.10 **p<0.05 ***p<0.01.

### Robustness checks and additional findings
- Revenue productivity versus physical productivity:
  - A 1% increase in market access leads to a 0.12% reduction in firms’ marginal costs (Table 10).
  - Firms pass most of their cost advantages to consumers; markups show no strong change.
  - Implication: market access effects on firm productivity are more than twice as large as TFPR regressions suggest once price effects are considered.
- Isolating transportation costs from GDP growth:
  - Using population of all cities in 1995 so that time variation in market access comes only from bilateral transportation costs does not substantially alter results; transportation cost channel effects are slightly larger than baseline.
- Excluding “node” cities:
  - Excluding 52 government-targeted “node” cities does not change results, reducing concern that results are driven by targeted places.
- Alternative market access estimation:
  - Following Donaldson and Hornbeck (2016) by numerically solving a system of nonlinear equations yields market access measures that produce very similar regression results to the baseline “market potential” measure.

### Spatial reallocation of industries (heterogeneous industry responses)
- Three industry characteristics used to predict heterogeneous responses:
  - Iceberg transportation cost: weight-to-value ratio (median).
  - Capital intensity: median firm capital-to-labor ratio.
  - Degree of product differentiation: trade elasticities/elasticity of substitution.
- Baseline specification:
  - ΔlogY_j,k = α + δ ΔlogMA_k × Ind_j + β X_k,j + μ_k + ν_j + ε_j,k
- Key empirical patterns:
  - Interaction of market access and capital-to-labor ratio is positive: industries with higher capital intensity grew disproportionately faster in cities with large market access.
  - Interpretation example: a one-standard deviation of a city’s market access would increase annual employment growth in the sector at the 75th percentile of capital intensity by 1.6 percentage points more than in the sector at the 25th percentile (period 2001–2008).
  - Interaction of market access and weight-to-value ratio is positive: industries with high transportation costs (high weight-to-value) grew faster in cities with large market access.
  - Product differentiation: industries with a low degree of product differentiation grew disproportionately faster in cities with large market access.

### Discussion, conclusion, and policy implications
- Aggregate findings:
  - National highway network in China promoted production efficiency, delivered sizable welfare gains to consumers, and led to sectoral reallocation between cities.
  - Aggregate TFP growth from reduced transportation costs is attributable primarily to firm entry and resource reallocation.
  - Cities with large market access specialized in industries with low unit cost, higher capital intensity, and low product differentiation.
- Policy relevance:
  - Transport infrastructure can be a tool to boost aggregate demand short run and promote long-run growth; many governments and institutions (World Bank, IMF, Asian Infrastructure Investment Bank, New Development Bank) emphasize infrastructure investment.
  - Findings suggest sizable economic impact of infrastructure investments and market integration when domestic transportation costs are large and misallocation is pervasive.
  - Need to weigh dynamic gains against direct and indirect costs of large-scale infrastructure projects; that evaluation is beyond the scope of the paper.

### Avenues for future research
- Quantitative comparison of aggregate productivity growth under actual highway construction versus counterfactual scenarios, including cost-benefit analysis of infrastructure investments.
- Incorporate inter-sectoral linkages to construct output market access and input market access measures using input-output tables, to distinguish output access from input access effects.
- Micro-level studies at county or village level to assess the role of infrastructure in reducing poverty and affecting income inequality.

*Italic: Source — wp18276 - 1. Holding the population distribution from 20003.12% (excerpt of IMF working paper content).*

### REFERENCES

### REFERENCES

### Key empirical findings from tables and figures
- Market access is a significant determinant of firm outcomes:
  - Table 11: Market Access coefficients: 0.052***, 0.053***, 0.051***, 0.053*** (standard errors: (0.015), (0.014), (0.015), (0.014)).
  - Table 10 (Marginal Cost): Market Access coefficients: -0.155***, -0.170***, -0.120** (standard errors: (0.048), (0.049), (0.046)).
  - Table 10 (Price): Market Access coefficients: -0.084*, -0.086*, -0.095** (standard errors: (0.046), (0.047), (0.047)).
- Access to port shows consistent positive associations:
  - Table 11: Access to port coefficients: 0.075**, 0.076**, 0.075**, 0.076** (standard errors: (0.031), (0.031), (0.031), (0.033)).
  - Table 10: Access to port coefficients: -0.036, -0.022 (standard errors: (0.100), (0.094)).
- First-stage relevance and strength of instruments:
  - Table 7: log MA (IV) coefficients: 0.967***, 0.922***, 0.960*** (standard errors: (0.035), (0.041), (0.036)).
  - Table 7: First-stage F statistic: 723.433, 489.353, 656.805.
  - Table 7: Adjusted. R squared: 0.991, 0.992, 0.982.
- Transport cost estimation (Table 5) – road network variables have large, statistically significant effects on iceberg transport cost:
  - Length of highway coefficients: 0.454***, 0.539***, 1.331***, 1.105*** (standard errors: (0.001), (0.001), (0.007), (0.008)).
  - Length of local road coefficients: 1.092***, 0.712***, 1.215***, 1.199*** (standard errors: (0.005), (0.026), (0.045), (0.045)).
  - Nonlinear terms included: Length of highway (squared) -0.381***, -0.315*** (standard errors: (0.004), (0.004)); Length of local road (squared) -0.938***, -0.780*** (standard errors: (0.056), (0.055)); cubic terms positive and significant.
  - Travel time terms (column (4)): Travel time 0.264*** (0.005); Travel time (squared) -0.022*** (0.001); Travel time (3 squared) 0.001*** (0.000).
  - No. of observations reported across specifications: 97253, 97253, 97253, 97253 (as displayed).
  - Adjusted. R squared: 0.689, 0.850, 0.869, 0.875.
- Market Access and firm markups (Table 9):
  - Market Access coefficients (OLS/IV): -0.007, -0.002, -0.003, -0.001 (standard errors: (0.020), (0.018), (0.022), (0.019)).
  - Access to port: 0.072, 0.072, 0.073, 0.075 (standard errors: (0.046), (0.046), (0.045), (0.045)).
- Heterogeneous industry responses to market access (employment and firm growth):
  - Table 14 (Capital Intensity): K/L*∆Market Access coefficients: 0.642***, 0.644***, 0.462***, 0.463*** (standard errors: (0.136), (0.136), (0.076), (0.078)).
  - Table 15 (Weight-to-Value Ratio): Weight/Value*∆Market Access coefficients: 0.004***, 0.004***, 0.002***, 0.003*** (standard errors: (0.001), (0.001), (0.001), (0.001)).
  - Table 16 (Product Differentiation): ProdDiff*∆Market Access coefficients: -0.008**, -0.009**, -0.007***, -0.009*** (standard errors: (0.004), (0.004), (0.002), (0.002)).

### Main quantitative results and sample sizes
- Table 8 (Firm TFP regressions): No. of observations reported across specifications include 322546, 313209, 312381 (and variants); No. of clusters: 339, 339, 335, 335, 335, 339, 339, 335, 335, 335. Adjusted. R squared values reported include 0.691, 0.822, 0.823, 0.823, 0.823.
- Table 9: No. of observations: 595635, 594476, 594476, 595635, 594476, 594476. Adjusted. R squared reported: 0.84, 0.841, 0.841. No. of clusters: 333, 331, 331, 333, 331, 331.
- Table 10: No. of observations: 197656, 196145, 196143, 197656, 196145, 196143. Adjusted. R squared: 0.389, 0.389, 0.406. No. of clusters: 338, 331, 331, 338, 331, 331.
- Table 12 (excluding “Node” cities): No. of observations: 428739, 418867, 417960, 417960, 416133, 428739, 418867, 417960, 417960, 416133. Adjusted. R squared examples: 0.421, 0.682, 0.693, 0.694, 0.694.
- Table 13 (Alternative TFP estimation using Ackerberg, Caves, and Frazer (2015)): No. of observations across specs include 612259, 609852, 431449, 431066, 612252, 609974, 431454, 431071. Adjusted. R squared: 0.847, 0.847, 0.853, 0.853, 0.853.
- Table 5: No. of observations across specifications reported as 97253 (repeated); Adjusted. R squared: 0.689, 0.850, 0.869, 0.875.

### Methodology and specification notes
- Instrument construction and first-stage:
  - Table 7: Dependent variable is market access estimated with actual highway network data; independent variable is hypothetical market access constructed from least-cost path spanning tree network. Standard errors clustered at the city level. *p<0.10**p<0.05***p<0.01.
- Fixed effects and clustering conventions:
  - Multiple tables indicate inclusion of industry-year FE, city FE or firm FE as appropriate.
  - Standard errors are clustered at the city level in most regressions; some specifications note clustering at province-industry level (employment growth tables).
- Alternative TFP estimation:
  - Table 13 uses Ackerberg, Caves, and Frazer (2015) to estimate firm TFP; results reported for OLS and IV specifications.

### Appendix A — China’s National Trunk Highway System (NTHS) summary
- Policy rationale and financing:
  - Objectives: develop regional economies, boost logistics efficiency, raise living standards, promote trade, increase free flow of commodities, raise competition in the domestic market, ease regional inequalities, and increase employment by connecting different parts of the country through highways.
  - Financing measures introduced in 1984: imposing vehicle purchase tax and raising roadway tolls; government made loans to build highways and repay loans by charging highway tolls.
- Planned network scale and composition (as stated):
  - Total planned transportation system length: 34422 kilometers.
  - Of which: 25765 kilometers are highway, 1479 kilometers are Rate I freeway, and 7178 kilometers are Rate II freeway.
  - Share of total distance by type: the three kinds of express way take up 74.85%, 4.3% and 20.85% of the total distance respectively.

*Italic line removed per instruction: source attribution provided separately.*

### 2020. The plan also stated that the new highway system would link the capital Beijing to

### wp18276 - 2020. The plan also stated that the new highway system would link the capital Beijing to

### Highway plan, scope, and timeline
- The plan stated the new highway system would link the capital Beijing to other provincial cities, connecting 93% of the major cities with other 1 million population and big cities with over 0.5 million population.
- The number of cities that are linked together will exceed 200.
- 0.6 billion people will be affected by the highways.
- The government planned five vertical expressways and seven horizontal expressways to connect the country into a single network.
- The plan was approved by the National People’s Congress in 1992 and established by the Department of Transportation in June, 1993.
- Construction of the highway system can be divided into four time periods: the mid-1980s to 1991, 1992 to 1997, 1998 to 2003, 2003 to 2007.
- It started in the major cities in the mid-1980s. Since the approval of the construction plan in 1992, the project entered the regular phase.
- Until the end of 1997, the total distance of highways in China was 4771 kilometers, of which 70% was the major national highway.
- In 1998 the government decided to build infrastructure and speed up the highway construction project in response to the financial crisis in Asia.
- From 1998 the highway construction project entered the rapid development phase, and in 2001 the total distance of highways in China exceeded 19,000 kilometers, the second longest in the world after the United States.
- From 2003 to 2007 the construction project was further accelerated and was finished in 2007.

### Price quotes and transportation cost data (Appendix B)
- Price quotes were collected from the websites of two logistics companies for any pair of cities in China.
- There are three transport modes that the two companies provide; the fastest mode comes with a higher price. The fastest mode was always chosen to be consistent.
- For every city pair the quotes from the two companies were compared and the lower quote was chosen as the transportation cost between the city pair.

### Revenue productivity, prices and marginal cost (Appendix C) — key decomposition and interpretation
- Concern: If firms with high productivity pass efficiency advantages to consumers via lower prices, regressing revenue productivity produces downward bias.
- Production quantity data at the firm level from 2000 to 2006 were used to estimate firm marginal costs and prices for a subset of firms in the Annual Survey of Industrial Enterprises.
- Decomposition (following Garcia and Voigtländer (2013)):
  - TFPR_it = p_it A_it = μ_it · MC_it(A_it,w_it) · A_it  (equation labeling as in source)
    - where p_it is the output price that firm i charges,
    - A_it is firm i’s physical productivity,
    - μ_it is markup,
    - MC_it is firm i’s marginal cost,
    - w_it is the input price of firm i.
- Interpretation points preserved exactly:
  - Firms with higher physical productivity will have lower marginal cost, and the two effects offset each other.
  - Assuming markups do not change, then the change in revenue productivity is ambiguous.
  - The finding of positive significant effects on revenue productivity suggests that physical productivity might have responded even more to highway access.
  - Input costs of firms are likely to respond to changes in market access:
    - If inputs shipped from other cities becomes cheaper in respond to increase in market access, then revenue TFP will cause even larger downward bias.
    - If labor inputs become more expensive after highway expansion and the effect dominates the cheaper prices of other inputs, then revenue TFP will give us smaller biases.
- Footnotes and methodological notes retained:
  - The firm-level quantity data is used by the National Bureau of Statistics in China to estimate the total quantity of production of major manufacturing products; the data does not cover all firms in the ASIE but is meant to capture all median and large firms.
  - Theoretically, if the effect of labor input is sufficiently large, it is possible that the effect will exactly offset or even dominate physical productivity.

### Bilateral trade and market access derivations (Appendix D)
- Bilateral trade equation (as presented):
  - X^s_{ij} = β^s × Y_i × Y_j / Y × (w_i τ^s_{ij} / Θ^s_j)^{−γ^s} × (f^s_{ij})^{−[γ/(σ^s−1)−1]}.  (equation (24) in source formatting)
- Definition (as in source):
  - Θ^{−γ}_j ≡ ∑_{k=1}^N (Y_k / Y) × (w_k τ_{kj})^{−γ} × f_{kj}^{−[γ/(σ−1)−1]}.
- Consumer Market Access (CMA) defined:
  - Θ^{−γ}_j ≡ CMA_j.  (equation (25))
- Rewritten expression (as in source):
  - X_{ij} = (Y_i / Y) (w_i τ_{ij})^{−γ} CMA^{−1}_j Y_j.  (equation (26))
- Firm Market Access (FMA) defined:
  - FMA_i ≡ ∑_j τ^{−γ}_{ij} CMA^{−1}_j Y_j.  (equation (27))
- CMA expressed in terms of FMA (as in source):
  - CMA_j = ∑_i τ^{−γ}_{ij} FMA^{−1}_i Y_i.  (equation (28))
- Relationship and market access system:
  - Donaldson and Hornbeck (2015) shows FMA_i = ρ CMA_i for some scalar ρ > 0. If MA_i ≡ FMA_i = ρ CMA_i then equation (12) implies:
    - MA_i = κ ∑_j τ^{−γ}_{ij} MA^{−(1+γ)/γ}_j Y_j.  (equation (29) as in source)
  - From the expression above, a city’s market access is the weighted sum of all its neighboring cities’ market access.
  - To solve the market access measure, a system of 339 equations was solved with MatLab.
- Note: An extension to the multi-sector version is noted as an interesting exercise but requires more theoretical and data work.
- Using China’s Input-Output tables for 1997, 2002 and 2007, industry-specific input supply access and output demand access are being constructed (as stated).

### Highway construction costs (Appendix E)
- Estimated construction costs were constructed for each pixel following Faber (2014) using remote sensing data on terrain ruggedness; the darker the pixel the higher the construction costs.
- Hainan province and Taiwan, province of China are excluded from the estimation because they are islands not connected to the mainland by road.

*Source: wp18276 - 2020. The plan also stated that the new highway system would link the capital Beijing to (IMF working paper content provided).*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2018/wp18276.pdf_