PaperlayerView exact revision on arXiv

Source-linked AI summary

Collective Learning in China's Regional Economic Development

Jian Gao, Bogang Jun, Alex "Sandy" Pentland, Tao Zhou, Cesar A. Hidalgo

arXiv:1703.01369v1econ.GNphysics.soc-phq-fin.EC

TL;DR

Causal evidence on how economies learn through industrial diversification is sparse. This paper studies Chinese provincial diversification using a high-speed-rail DID design, finding that neighboring-industry presence is associated with industry emergence and that the learning channels exhibit diminishing returns.

  • Problem

    Causal evidence for inter-industry learning is sparse, leaving how economies learn to diversify incompletely evidenced.

    Source: §1 · p. 5

  • Method

    The paper examines Chinese provincial industrial diversification and uses high-speed-rail introduction in a differences-in-differences design to study inter-regional learning.

    Source: §4 · p. 26

  • Results

    Industry emergence rises with neighboring provinces’ same-industry presence, the learning channels exhibit diminishing returns, and rail-linked provinces show increased industrial similarity and productivity with industry-present neighbors.

    Source: §4 · p. 26

  • Takeaways & Limitations

    The findings provide evidence supporting inter-regional learning theories.

    Source: §4 · p. 26

    The study’s data and geographic resolution are limited.

    Source: §4 · p. 27

Abstract

from arXiv · show

Industrial development is the process by which economies learn how to produce new products and services. But how do economies learn? And who do they learn from? The literature on economic geography and economic development has emphasized two learning channels: inter-industry learning, which involves learning from related industries; and inter-regional learning, which involves learning from neighboring regions. Here we use 25 years of data describing the evolution of China's economy between 1990 and 2015--a period when China multiplied its GDP per capita by a factor of ten--to explore how Chinese provinces diversified their economies. First, we show that the probability that a province will develop a new industry increases with the number of related industries that are already present in that province, a fact that is suggestive of inter-industry learning. Also, we show that the probability that a province will develop an industry increases with the number of neighboring provinces that are developed in that industry, a fact suggestive of inter-regional learning. Moreover, we find that the combination of these two channels exhibit diminishing returns, meaning that the contribution of either of these learning channels is redundant when the other one is present. Finally, we address endogeneity concerns by using the introduction of high-speed rail as an instrument to isolate the effects of inter-regional learning. Our differences-in-differences (DID) analysis reveals that the introduction of high speed-rail increased the industrial similarity of pairs of provinces connected by high-speed rail. Also, industries in provinces that were connected by rail increased their productivity when they were connected by rail to other provinces where that industry was already present. These findings suggest that inter-regional and inter-industry learning played a role in China's great economic expansion.

1. Introduction

The introduction frames how China learned to produce increasingly sophisticated goods during its 1990–2015 expansion, focusing on inter-industry and inter-regional learning. It situates the study against limited causal evidence for these channels and examines evidence consistent with both.

1. Introduction

  • A PPP-adjusted, constant-price measure increased nearly tenfold from USD 1,516 in 1990; if sophisticated-product exports drove China’s expansion, the question is how it learned to produce increasingly sophisticated products.Source: §1 · pp. 2–3
  • Rodrik (2006) argued that unusually sophisticated Chinese exports fueled growth, while economic-complexity research links greater complexity to faster growth among similarly wealthy countries.Source: §1 · pp. 2–3
  • The introduction frames economic learning as drawing on related industries within a region and on neighboring regions across geographic boundaries.It contrasts inter-industry learning, studied more extensively, with the less-studied inter-regional channel.Source: §1 · p. 3
  • Product density, based on products a country is already exporting, is a statistically good predictor of whether that country will develop comparative advantage in a specific product in the future.Swedish manufacturing-plant data showed that the probability an industry would enter a region increases, and its probability of exiting decreases, with the number of related industries.Source: §1 · p. 4
  • The literature has sparse causal evidence supporting both inter-industry and inter-regional learning.The introduction identifies this gap as a limit on the existing evidence.Source: §1 · p. 5
  • A province’s probability of developing a new industry rises with both the number of related industries already present and the number of neighboring provinces where it is present.Source: §1 · p. 5
  • The two learning channels exhibit diminishing returns: when either is sufficiently active, it reduces the marginal contribution of the other.Source: §1 · p. 5
  • DID results show rail-connected province pairs became more industrially similar, while industries gained productivity when connected to provinces where that industry was present.To address endogeneity, the study uses rail introduction as an instrument that affects travel time but not industrial similarity.Source: §1 · p. 6

2. Data

The study combines firm-level data spanning 1990–2015 with province-level industry, productivity, geographic-distance, and macroeconomic measures to examine regional economic development.

2.1. China’s firm data

  • From 1990–2015, listed-firm data cover 31 Chinese provinces and classify industries into 18 sectors and 70 subsectors using registration and financial information.The data include listing and delisting dates, registered addresses, industry categories, yearly revenue, and employee counts.Source: §2.1 · p. 6
  • Industry productivity in a province is measured as total firm revenue divided by total employees in that province-industry.Source: §2.1 · p. 6

2.2. Distance and macroeconomic indicators

  • Interregional proximity is measured by the distance between provincial capital cities, while province-level macroeconomic indicators cover all 31 provinces from 1990–2015.Indicators include GDP per capita, resident population, imports and exports, urban area, and total area; urbanization is measured as the share of urban area.Source: §2.2 · p. 7

3. Results

The results examine how related industries, neighboring provinces, and high-speed rail connections relate to provincial industry development and industrial similarity in China. They also assess productivity patterns associated with rail connections and neighboring provinces’ industry activity.

3.1. Inter-industry learning

  • China’s industry space had a tightly connected manufacturing core and a separate service and information-activities core, while extractive and agricultural industries occupied the periphery.Source: §3.1 · p. 8
  • The probability that an industry develops revealed comparative advantage in a province increases strongly with density of active related industries; five-year adopters had higher density than non-adopters (ANOVA p-value=2.1 × 10−40).Source: §3.1 · p. 12
  • Density of active related industries is a strong, positive, significant predictor of both developing a new industry and keeping an industry, beyond provincial industry count and industry ubiquity.Source: §3.1 · p. 13

3.2. Inter-regional learning

  • Neighboring provinces were significantly more industrially similar than non-neighbors (ANOVA p-value=8.1 × 10−4), and similarity declined with distance (r = −0.32).Source: Figure 5 · p. 15, §3.2 · p. 15
  • The probability of developing revealed comparative advantage increased convexly with active-neighbor density, whose average was higher for province-industry pairs that later developed advantage (ANOVA p-value=1.4 × 10−37).Source: Figure 6 · p. 16, §3.2 · p. 16
  • With controls for provincial industry count and industry ubiquity, active neighboring provinces positively predicted new industry development, but their effect on sustaining revealed comparative advantage was unclear.Source: §3.2 · pp. 17–18
  • For sustaining an industry, the neighboring-province effect was negative before controls and positive after controls, consistent with a tension between learning and competition.Source: §3.2 · p. 18

3.3. Combining inter-industry and inter-regional learning

  • The probability that an industry appears in a province increases with both active neighboring provinces and active related industries, a result robust to other density measures.Source: §3.3 · p. 19
  • A negative, significant interaction between the two learning-channel densities indicates diminishing returns: each channel’s marginal contribution falls when the other is active.Both densities were jointly significant without the interaction, and inter-industry learning had a slightly stronger effect, as indicated by its larger regression coefficient.Source: §3.3 · p. 19
  • The diminishing-returns result persisted under alternative ratio and count measures, although the count-based model had less explanatory power than the density-based model.A negative control using counts of all neighboring provinces and related industries, whether active or not, lost almost all explanatory power, indicating that active channels matter.Source: §3.3 · pp. 19–20

3.4. High-speed rail and inter-regional learning

  • The authors use high-speed rail as an instrument for inter-regional learning because it reduces barriers to that learning but should not affect inter-industry learning.They state that rail does not affect industries’ similarity or productivity within a province, helping isolate inter-regional from inter-industry learning.Source: §3.4 · pp. 20–21
  • The authors report that treatment and control groups had similar industrial-similarity pre-trends before 2005, supporting the DID analysis’s parallel-trends condition.Pairs connected by high-speed rail in 2015 form the treatment group; the analysis focuses on 2004–2014 to capture construction of numerous railroads.Source: §3.4 · pp. 21–22
  • 0.029 was the statistically significant DID estimate for the increase in industrial similarity among province pairs connected by high-speed rail.The event study found no pre-2005 trend in the treatment effect; after rail introduction, the effect increased significantly.Source: §3.4 · p. 22, Figure 8 · p. 23
  • 0.029 was the DID increase in industrial similarity for province pairs after high-speed rail was introduced.The DID coefficient was positive and significant, and robust to controls.Source: §3.4 · pp. 23–24, Table 4 · p. 24
  • Productivity and industrial similarity were neither highly nor consistently correlated over time, allowing productivity to be analyzed separately.Source: §3.4 · p. 24
  • CNY 98,713 (∼USD 15k) was the DID productivity gain for rail-connected province pairs over unconnected pairs.Source: Figure 9 · p. 25, §3.4 · p. 25
  • Industry productivity increased with neighboring provinces’ productivity density, following an increasing and convex relationship.Source: Figure 9 · p. 25, §3.4 · p. 25
  • High-speed rail entry had a positive, significant productivity effect robust to controls, supporting learning when connected provinces had productive firms in that industry.Source: §3.4 · p. 26

4. Conclusion and Discussion

Provincial industry emergence was associated with related industries and neighboring provinces’ industries, while the two learning channels showed diminishing returns. High-speed rail results supported inter-regional learning, though the evidence is limited by the data and does not identify micro-level channels.

Conclusion and Discussion

  • Industry emergence rose with the number of related industries in a province and with the same industry's presence in neighboring provinces; when one learning channel was active, the other contributed less.The authors describe inter-industry and inter-regional learning as substitutes, with learning constrained by the absence of a single opportunity.Source: §4 · p. 26
  • High-speed rail significantly increased industrial similarity between connected province pairs, and productivity rose when a connected province had a productive neighbor in that industry.The DID analysis compared connected neighboring-province pairs with pairs where the industries were present but the provinces did not become rail-connected.Source: §4 · p. 26
  • The study counted new industries only when they had revealed comparative advantage, treating industries without it as absent even if they had many firms.This differs from defining industry presence by a large absolute number of firms.Source: §4 · p. 27
  • The data cover firms listed on China's two major stock markets, biasing the sample toward larger firms and excluding unlisted or foreign-listed firms operating in China.Using provinces as geographic units also limits spatial resolution; some provinces contain more than 100 million people.Source: §4 · p. 27
  • The paper supports collective learning at the macro level but does not identify whether it operates through spin-offs, migrant workers, externalities, labor pooling, or other micro-level channels.The authors leave these micro-level explanations outside the study's scope.Source: §4 · p. 27
  • The findings support viewing economic development as path-dependent learning shaped by related industries and neighbors, with implications for developing countries' planning and development efforts.The authors hope the work encourages research to identify specific learning channels.Source: §4 · p. 27

1. China’s firm data

The study analyzes Chinese listed-firm data from 1990–2015, aggregated into industry categories and filtered to reduce noise. The resulting dataset contains 2,690 firms across 18 sectors and 70 subsectors.

China’s firm data

  • The firm data come from China’s RESSET Financial Research Database, which provides publicly listed firms’ records.The database is provided by Beijing Gildata RESSET Data Tech Co., Ltd.Source: §1 · p. 30
  • The dataset covers 1990–2015 and records listed firms’ registration details, industry, annual revenue, and employee counts.Firm addresses span 31 Chinese provinces.Source: §1 · p. 30
  • Firms are aggregated into sectoral and sub-sectoral levels using the CSRC’s 2011 listed-company classification guidelines.The aggregation also considers the National Economy Industry Classification (GB T4754-2011).Source: §1 · p. 31
  • 2,690 firms remain after filtering, spanning 18 sectoral and 70 sub-sectoral industries.The analysis excludes sub-sectoral industries with three or fewer listed firms to reduce noise.Source: §1 · p. 32

2. Distance, travel time, and macroeconomic indicators

The section operationalizes interregional proximity with three distance measures and three travel-time measures, alongside province-level economic indicators covering 1990–2015 across 31 provinces.

Distance, travel time, and macroeconomic indicators

  • Geographic distance is the geodesic distance between provincial capitals, driving distance the shortest route between them, and neighboring distance the least number of provinces crossed.Driving routes are based on the Google Maps API in 2015.Source: §2 · p. 32
  • Transit time is the shortest high-speed-train journey, with normal-train time used if no high-speed route is available; driving time is the shortest drive between provincial capitals.Driving time is estimated using the Google Maps API in 2015.Source: §2 · pp. 32–33
  • Province-level indicators include GDP per capita, population, imports and exports, urban area, and total area; urbanization is defined as urban area’s share of provincial area.The indicators come from China’s Statistical Yearbooks and cover 31 provinces from 1990 to 2015.Source: §2 · p. 33

3. Representation of industry space

The section represents provincial industries as a weighted bipartite network and constructs an industry space from inter-industry proximity. The resulting proximity matrix shows two large modules and smaller modules, consistent with two density cores.

3. Representation of industry space

  • A weighted bipartite network links provinces to sub-sectoral industries, with each link weighted by the number of firms in that province-industry pair.The network is G = {P, I, E}, where P is provinces, I is industries, and E is links.Source: §3 · p. 33, Figure S3 · p. 34
  • Industry space combines a connectivity-maximizing spanning network with thresholded high-proximity links, yielding a superposed network of 145 links among 70 sub-sectoral industries.The spanning network has 69 links; the maximum weighted network uses links above φ′ = 0.81 and has 116 links.Source: §3 · pp. 33–34
  • The industry space is laid out with Gephi’s ForceAtlas2 algorithm, with node size and color encoding industry firm counts and sectoral categories and link appearance encoding proximity.The visualization uses 2015 data.Source: §3 · pp. 34–35, Figure S4 · p. 34
  • The hierarchically clustered proximity matrix contains two large modules and smaller modules, supporting the presence of two density cores in industry space.Proximity values have a normal-like distribution with an average around 0.5.Source: §3 · p. 35

4. Robustness check of inter-industry learning

Two robustness checks support inter-industry learning: related-industry density predicts future diversification, and new industries are more likely to emerge where related industries are denser.

4. Robustness check of inter-industry learning

  • Across five-year intervals, provinces with more industries at RCA > 1 had more new industries at RCA > 1 five years later.Source: §4 · p. 36, Figure S6 · p. 36
  • Higher average density among industries at RCA < 1 was strongly positively related to the number of new industries gaining comparative advantage within five years.The density of related industries without current comparative advantage predicted future provincial diversification.Source: §4 · p. 36, Figure S6 · p. 36
  • With time-varying industrial proximity, province-industry pairs that developed comparative advantage had significantly higher related-industry density (ANOVA p-value < 0.01), and entry probability rose convexly with density.The analysis uses five-year intervals.Source: Figure S7 · p. 37, §4 · p. 37

5. Robustness check of inter-regional learning

Robustness checks support the inter-regional learning findings: industrial presence and similarity align with geographic proximity, and a distance-weighted measure of active neighboring provinces is positively associated with later industry development.

5. Robustness check of inter-regional learning

  • Industrial similarity was higher between regions with shorter transit or normal-train times, shorter driving times, or shorter driving distances.The checks report similarity against all four proximity measures.Source: Figure S9 · p. 39, §5 · p. 39
  • Provinces with many firms in an industry tended to neighbor provinces that already had many firms in that industry, supporting the main finding.The spatial check traces industry presence using firm counts; illustrated industries include chemicals, pharmaceuticals, electric machinery, and wholesale.Source: §5 · pp. 38–39, Figure S8 · p. 38
  • ANOVA p-value=7.3×10−36: industry-province pairs that later developed comparative advantage had higher active-neighbor density, and development probability increased convexly with distance-weighted density.The alternative measure weights neighboring provinces by neighboring distance and predicts development over five-year intervals.Source: Figure S10 · p. 40, §5 · p. 40

6. Robustness check of inter-regional and inter-industrial learning

Using alternative ratio measures, the robustness checks find that both related-industry activity and neighboring-province activity are associated with a higher probability of new industries developing.

6. Robustness check of inter-regional and inter-industrial learning

  • The robustness check replaces density with ratios: the proportion of active neighboring provinces and the proportion of active related industries, defined using the 2015 industry space.The figure reports joint and marginal probabilities for these ratios.Source: §6 · p. 40, Figure S11 · p. 41
  • Both ratio measures significantly predict new-industry development, with each showing a strongly increasing, convex relationship with the probability of gaining revealed comparative advantage.These results support the robustness of the earlier findings.Source: Figure S11 · p. 41, §6 · p. 41

7. Robustness check of causal evidence for inter-regional learning

The robustness checks support a positive effect of high-speed rail on inter-regional learning: rail-entry effects on industrial similarity and productivity remain significant after controlling for geographic distance. Connected province pairs also show higher industrial similarity, while productivity rises after rail entry.

7. Robustness check of causal evidence for inter-regional learning

  • Industrial similarity is significantly higher for province pairs connected by high-speed rail (ANOVA p-value=1.2 × 10−18) and increases after the 2005, 2008, and 2012 speed-up campaigns.The 2004–2015 averages for pairs with and without high-speed rail support the robustness of this pattern.Source: §7 · pp. 41–42
  • Despite productivity differences across industries, industrial similarity and productivity have a relatively small correlation across province pairs, supporting consideration of both measures.The passage describes these measures as independent to some extent.Source: §7 · p. 43, Figure S13 · p. 43
  • After 2005 rail entry, productivity treatment coefficients rise significantly, following a similar pre-entry trend for treated and control province pairs.The event study uses 2005 as the baseline; pairs connected by high-speed rail also have significantly higher average productivity over the considered period.Source: §7 · p. 44, Figure S14 · p. 44
  • High-speed rail entry effects on industrial similarity and productivity remain significant when geographic distance is controlled, supporting the robustness of the DID evidence.The regressions use 2004 data before rail entry and 2014 data after entry.Source: §7 · p. 45, Table S5 · p. 45
Loading 1703.01369v1…