Source-linked AI summary
Interpreting Economic Complexity
Penny Mealy, J. Doyne Farmer, Alexander Teytelboym
TL;DR
The paper addresses the elusive mathematical and economic interpretation of ECI and PCI, measures that have helped explain development differences but are often linked to export diversity. It shows that they are equivalent to spectral clustering and related dimensionality-reduction measures, yielding similarity-based distances. The measures distinguish specialization by product complexity and reveal informative patterns across countries, US states, and UK regions.
Problem
ECI and PCI have explained GDP per capita and growth, but their precise mathematical and economic interpretations have remained elusive, particularly beyond explanations based on export diversity.
Method
The paper interprets ECI and PCI through spectral clustering, dimensionality reduction, and similarity-based distances on country-product and related regional graphs.
Results
ECI and PCI are equivalent to a spectral-clustering solution and reveal that high-ECI countries tend to specialize in high-PCI products, with analogous informative patterns across US states and UK regions.
Takeaways & Limitations
ECI and PCI capture which types of products or activities countries and regions specialize in, providing information that diversity does not make apparent.
Abstract
from arXiv · showhide
Two network measures known as the Economic Complexity Index (ECI) and Product Complexity Index (PCI) have provided important insights into patterns of economic development. We show that the ECI and PCI are equivalent to a spectral clustering algorithm that partitions a similarity graph into two parts. The measures are also related to various dimensionality reduction methods and can be interpreted as vectors that determine distances between nodes based on their similarity. Our results shed a new light on the ECI's empirical success in explaining cross-country differences in GDP/capita and economic growth, which is often linked to the diversity of country export baskets. In fact, countries with high (low) ECI tend to specialize in high (low) PCI products. We also find that the ECI and PCI uncover economically informative specialization patterns across US states and UK regions.
1 Introduction
The paper reinterprets ECI and PCI as mathematically grounded measures of similarity and specialization rather than proxies for export diversity alone. It shows that these measures connect spectral clustering, dimensionality reduction, and economically informative patterns across countries and regions.
- Core interpretation: ECI and PCI are mathematically equivalent to spectral clustering that partitions a similarity graph into two balanced, internally similar and externally dissimilar components.The equivalence connects the measures to the normalized-cut approximation for graph partitioning.
- Core interpretation: The measures can be interpreted as dimensionality-reduction coordinates that place similar countries or products close together and dissimilar ones farther apart.For export data, ECI orders countries by export similarity, while PCI analogously orders products by their exporters.
- Beyond diversity: ECI is orthogonal to diversity, so it captures information distinct from the number of products a country exports competitively.The paper contrasts this result with earlier descriptions that linked ECI primarily to export-basket diversity.
- Economic specialization: High-ECI countries tend to specialize in high-PCI products, whereas low-ECI countries tend to specialize in low-PCI products.The export baskets of high-ECI countries are also more homogeneous than those of low-ECI countries.
- Regional applications: The framework extends beyond trade data: ECI and PCI reveal specialization patterns in UK local authorities and US states, while diversity is not economically informative there.Regional ECI is strongly correlated with earnings per capita in UK local authorities and with state-level GDP per capita in the United States.
- Construction: The ECI is computed from a binary country-product matrix and the second-largest right eigenvector of a row-stochastic similarity matrix.The matrix encodes country-product competitiveness, while its normalized entries can be interpreted as conditional transition probabilities in a Markov chain.
3 Results
The ECI is mathematically equivalent to spectral clustering and dimensionality reduction, positioning nodes by similarity rather than simply measuring diversity. Across trade and regional data, ECI and PCI reveal specialization patterns and economic relationships that diversity alone does not capture.
- ECI versus diversity: ECI and diversity are mathematically orthogonal, although their empirical correlation varies across countries, UK regions, and US states.In UK local authorities, ECI is negatively correlated with industrial diversity; in US states, it has no significant correlation with occupational diversity.
- Spectral clustering: The ECI is equivalent to spectral clustering that partitions a similarity graph into two balanced components.It approximately minimizes the normalized cut criterion and relates to the normalized Fiedler vector.
- Dimensionality reduction: The ECI places countries on a one-dimensional interval where similar export profiles are close and dissimilar profiles are far apart.This distance is a special case of diffusion-map distance and is closely related to correspondence analysis.
- Regional applications: The ECI correlates with per-capita earnings in UK local authorities and per-capita GDP in US states.This extends economically informative ECI relationships beyond country export data.
- Specialization patterns: High-ECI countries tend to be richer and specialize in high-PCI products, while low-ECI countries specialize in low-PCI products.The same specialization pattern appears across UK industries and US occupations, with urban and professional concentrations associated with higher ECI in the UK.
- ECI versus diversity: Diversity reveals a triangular country-export structure but fails to be economically informative in the UK and US regional examples.Country diversity is positively correlated with per-capita GDP, whereas regional diversity has no positive correlation with earnings or state-level GDP.
4 Discussion
The paper distinguishes ECI and PCI from diversity, showing that they identify economically informative specialization by product or regional industry and occupation type. These measures extend beyond trade data and clarify the distinct roles of complexity and diversity in development.
- 4 Discussion: The paper separates the ECI’s empirical success from diversity, clarifying that the two variables play distinct roles in the development process.Earlier interpretations linked ECI success in explaining GDP per capita and growth to accumulating diverse productive capabilities.
- 4 Discussion: Countries follow a U-shaped diversification pattern, first diversifying and later specializing relatively late in development.Existing studies also report a positive association between export diversification and economic growth, especially among less developed countries.
- 4 Discussion: ECI and PCI reveal the type of exports associated with higher and lower income levels, beyond the number of products countries export.High PCI products tend to involve chemical and machinery exports requiring technologically sophisticated know-how, whereas low PCI products correspond to simpler agricultural products.
- 4 Discussion: The ECI and PCI reveal similar specialization patterns across richer and poorer UK and US regions, while diversity is not economically informative in these examples.The regional applications use industrial employment concentrations in UK local authorities and occupational employment concentrations in US states.
5 Materials and Methods
The regional applications construct binary region-specialization matrices from location quotients and apply the same eigenvector-based ECI methodology used for trade data. The approach is implemented for UK industries and US occupations.
- 5 Materials and Methods: UK local-authority ECI is constructed from a binary region-industry matrix based on whether each industry’s location quotient exceeds one.The matrix sets W_ri = 1 when LQ_ri > 1 and zero otherwise.
- 5 Materials and Methods: The UK industry-based ECI is calculated from the eigenvector associated with the second-largest eigenvalue of the transformed W matrix.The transformed matrix is constructed analogously to the trade-data matrix used for the original ECI.
- 5 Materials and Methods: US occupation-based ECI is computed analogously from a state-occupation matrix built using occupational location quotients.The data come from the Integrated Public Use Microdata Series, and consistent results are also found using US state-industry data.
S1 Diversity and degree equivalence
The diversity and graph-degree quantities are shown to be equivalent through a similarity matrix constructed from the country-product matrix. Row-stochastic normalization ensures the transformed matrix has the required degree structure.
- S1 Diversity and degree equivalence: The proof compares diversity with graph degree after defining the similarity structure from the country-product matrix.The section introduces both quantities before showing their equivalence.
- S1 Diversity and degree equivalence: The similarity matrix is defined as S = MU⁻¹M′, combining shared products while weighting products by inverse ubiquity.The matrix connects country pairs through their common products and underlies the degree-equivalence argument.
- S1 Diversity and degree equivalence: U⁻¹M′ and D⁻¹M are row-stochastic, so every row of the transformed matrix fM sums to one.This normalization links the matrix representation to the degree structure used in the proof.
- S1 Diversity and degree equivalence: Because each row of fM sums to one, each row of MU⁻¹M′ sums to the corresponding country-diversity value D_ii.The resulting row sums establish the equivalence between the diversity quantity and the graph degree.
S2 Relationship between the ECI and PCI
The paper proves that ECI and PCI are linked through the same spectral structure. In particular, a country’s ECI equals the average PCI of its competitive products, and the two matrices share their eigenvalues.
- S2 Relationship between the ECI and PCI: A country’s ECI equals the average PCI of products in which it has revealed comparative advantage.This proposition directly connects the country-level and product-level complexity measures through the country-product matrix.
- S2 Relationship between the ECI and PCI: The ECI and PCI are defined through corresponding second-eigenvector systems for the country and product matrices.The proof substitutes the ECI eigensystem into the matrix relationship to obtain the PCI eigensystem.
- S2 Relationship between the ECI and PCI: The ECI can be obtained from the PCI using M, and the matrices defining the two measures have identical eigenvalues.This establishes the spectral equivalence connecting the country and product measures.
S3 Interpretation of ECI as a diffusion map
The section places the economic complexity measures in relation to correspondence analysis and kernel principal component analysis.
- The economic complexity measures have a relationship to correspondence analysis.
- These relationships situate the measures within broader methods for analyzing structured data.
- The measures also have a relationship to kernel principal component analysis.
ysis
The diffusion-map interpretation represents countries through random-walk behavior and embeds them in a lower-dimensional space based on export similarity.
- A diffusion map represents complex data in a lower-dimensional Euclidean space by iterating an associated Markov matrix.
- The transition probability f Mij gives the chance that a random walk moves from state i to state j in one step.
- The distance between random walks from two states measures the similarity between their corresponding graph nodes.
- The ECI is the second-largest eigenvector and approximates diffusion distance between stationary random-walk probabilities as t increases.
- For country exports, Figure S1 visualizes countries using the second and third diffusion-map coordinates at t = 1, 5, and 10.
- The diffusion-map representation is also connected to correspondence analysis and kernel principal component analysis.
S4 ECI and PCI Rankings for Regional Data
The supplementary rankings report ECI and PCI positions for UK local authorities, UK industries, US states, and US occupations.
- Together, the rankings cover ECI for regional units and PCI for industries and occupations.
- Table S1 lists the top and bottom 10 UK local authorities ranked by ECI.
- Table S2 lists the top and bottom 10 industries ranked by PCI.
- Table S3 lists the top and bottom 10 US states ranked by ECI.
- Table S4 lists the top and bottom 10 occupations ranked by PCI.
S5 Eigengap Heuristic Analysis
The eigengap analysis finds no clear two-cluster structure in the studied similarity networks, while robustness checks examine how RCA thresholds affect empirical relationships and specialization patterns.
- Eigengap Heuristic Analysis: The export and regional similarity networks do not partition well into two clusters.
- Eigengap Heuristic Analysis: The largest eigenvalue gap lies between the first and second eigenvalues in all three datasets, suggesting one cluster under the eigengap heuristic.
- Eigengap Heuristic Analysis: The eigengap heuristic is limited here because it generally works best when data contain well-pronounced clusters.
- Robustness of empirical results to alternative RCA thresholds: Correlations between ECI and per capita GDP are highest for RCA thresholds between 0.5 and 2, while the ECI–diversity correlation changes with the threshold.
- Robustness of empirical results to alternative RCA thresholds: A lower RCA threshold produces a more triangular specialization pattern because ECI ordering moves closer to diversity ordering.
- Robustness of empirical results to alternative RCA thresholds: Regardless of using per capita or original RCA, threshold 1 gives a strong ECI correlation with per capita GDP, while ECI–diversity correlation decreases as the threshold rises.