Source-linked AI summary
Unfolding the innovation system for the development of countries: co-evolution of Science, Technology and Production
Emanuele Pugliese, Giulio Cimini, Aurelio Patelli, Andrea Zaccaria, Luciano Pietronero, Andrea Gabrielli
TL;DR
The paper asks how scientific, technological, and production activities interact across countries beyond a presumed one-way innovation chain. It builds a multilayer network to examine those relationships over time and finds that technology is the strongest precursor of scientific and export activity, with predictive signal peaking after about 20 years.
Problem
The paper examines how science, technology, and economic production interact when their relationships may be multidirectional rather than a simple science-to-technology-to-products flow.
Source: pp. 1, 3Method
The authors build a multilayer network linking scientific, technological, and production activities that co-occur significantly across countries, and examine conditional activity probabilities across layers and years.
Source: pp. 1–2, The multilayer space · p. 9Results
Technology is the best precursor for scientific and export activity, with the signal reaching its maximum after around 20 years in both cases.
Source: p. 3Takeaways & Limitations
Current patent portfolios can help forecast the scientific fields and market sectors in which countries may have competitive advantage, including over the long term.
Source: p. 3Combining science, patent, and export data constrains analysis to country-level geography and yearly time intervals, while finer resolutions also reduce statistical power.
Source: Scale of analysis · p. 12
Abstract
from arXiv · showhide
We show that the space in which scientific, technological and economic developments interplay with each other can be mathematically shaped using pioneering multilayer network and complexity techniques. We build the tri-layered network of human activities (scientific production, patenting, and industrial production) and study the interactions among them, also taking into account the possible time delays. Within this construction we can identify which capabilities and prerequisites are needed to be competitive in a given activity, and even measure how much time is needed to transform, for instance, the technological know-how into economic wealth and scientific innovation, being able to make predictions with a very long time horizon. Quite unexpectedly, we find empirical evidence that the naive knowledge flow from science, to patents, to products is not supported by data, being instead technology the best predictor for industrial and scientific production for the next decades.
Science
Scientific output is measured with country-level citation data from SciVal, aggregated across scientific sectors and years. The authors note coverage biases that may affect some comparisons.
Science
- SciVal data, aggregated from Scopus, cover journals, trade publications, books, conference proceedings, and book series.The paper notes that Scopus has broader coverage than Web of Science and a more reliable classification system than Google Scholar, while country-level analyses of these databases usually yield similar results.Source: Science · p. 8
- The data span 1996–2013 and measure citations accrued by scientific publications in each country and sector.The data are aggregated at the country level.Source: Science · p. 8
- Full counting of international coauthorship may favor small, highly internationalized countries, while language coverage may underrepresent anglophone-country output in Social Sciences and Humanities.The authors state these latter sectors are supposedly less connected to patenting and production, and thus do not hinder their analyses.Source: Science · p. 8
Technology
The technology layer uses patent data to represent countries’ technological capabilities. Patent families are the observation units, with their weights distributed across associated countries and technology codes.
Technology
- PATSTAT supplies patent records from almost 100 offices worldwide, with patents classified using the hierarchical International Patent Classification system.PATSTAT records applicant-country origin and groups patents into families linked by shared priorities; the study uses INPADOC extended families.Source: Technology · p. 8
- Each patent family is treated as one observation and its weight is split evenly among the countries and technology codes to which it maps.The technology codes depend on the level of aggregation used.Source: Technology · p. 8
- The country-field-year measure counts patent families, or attributed portions of families, associated with firms located in each country.The measure is defined for a given technological field and year.Source: Technology · p. 8
Products
The products layer uses international export flows as a proxy for economic production. The data classify exports by country, product, and year, with values reported in current US dollars.
Products
- BACI export data, recorded by UN COMTRADE and processed by CEPII, cover approximately 5,000 products classified under the Harmonized System 2007.The database covers import-export flows from 1995 to 2014, and its product codes have six digits arranged hierarchically.Source: Products · p. 8
- Trade flows are measured in thousands of current US dollars, and the resulting country-product matrix records each country's export value for each product and year.The matrix element represents the overall export value for that country-product-year.Source: Products · p. 9
Revealed comparative advantage
The authors identify competitive advantage using revealed comparative advantage (RCA), then convert the underlying activity matrices into binary country-activity indicators. They use a consistent RCA formulation across layers and normalize the science measure for citation-age bias.
Revealed comparative advantage
- RCA compares an activity's weight in a country's activity basket with its weight in total world activity; RCA > 1 indicates comparative advantage.The authors binarize the raw activity matrices using this threshold.Source: Revealed comparative advantage · p. 9
- The same RCA formulation is used across science, technology, and economic activities for consistency, with science RCA normalized to address the citation-age bias toward older papers.The paper describes RCA-like metrics as common in studies of these activities.Source: Revealed comparative advantage · p. 9
The multilayer space
The multilayer space links activities in different layers through countries, using transition probabilities to represent how information moves between activities over time. Its construction assumes that transfer proceeds in two probabilistic steps and that the second step is independent of the starting activity.
The multilayer space
- Information moves first from an activity to countries with comparative advantage, then from those countries to activities in the destination layer.The first step assigns equal probabilities to countries active in the source activity; the second assigns probabilities according to each country’s destination-layer diversification.Source: The multilayer space · p. 9
- The Assist matrix represents information flow from an activity in one layer and year to activities in another layer and year.Its transition probability is composed by passing through countries, and the time lag sets the process’s time direction.Source: The multilayer space · pp. 9–10
- The transfer probabilities encode effort inversely related to the number of countries active in the source activity and the number of destination activities in each country.The authors motivate the latter relation by countries’ finite, fixed resources for activity transfer.Source: The multilayer space · p. 10
- The model assumes that transfer effort toward a destination activity is independent of the starting activity and year, making the jump a Markov process.When the destination year precedes the source year, the same transition probability is interpreted in the opposite information-flow direction.Source: The multilayer space · p. 10
Hypothesis testing for the multilayer space
The authors test Assist-matrix links against a degree-constrained random-network null model. The null distributions support significance assessment, while the choice of significance threshold must be matched to the number of sampled realizations.
Hypothesis testing for the multilayer space
- The Bipartite Configuration Model randomizes network connections while constraining the ensemble-average country and activity degrees to their observed values.This preserves diversification and ubiquity while removing connectivity-pattern signal beyond the degree values.Source: Hypothesis testing for the multilayer… · p. 10
- The null model chooses the maximum-entropy distribution over network configurations subject to the degree constraints.Its connection probabilities depend on country and activity Lagrange multipliers, and links are independent random variables in this model.Source: Hypothesis testing for the multilayer… · pp. 10–11
- The authors sample null Assist matrices and use their distributions to assign p-values and assess whether observed links are statistically significant.The required ensemble size depends on the threshold: a 99% threshold needs at least 100 realizations, and a 99.9% threshold at least 1000.Source: Hypothesis testing for the multilayer… · p. 12
- For the time-lag signal-to-noise analysis, the authors define Φ as the average percentage of significant connections across years using a 95% threshold.When no signal is found, the expected value is approximately 5%.Source: Hypothesis testing for the multilayer… · p. 12
Scale of analysis
The analysis can vary its activity, geographical, and temporal aggregation. Resolution choices shape which capabilities can be identified and are constrained by data availability and statistical power.
Scale of analysis
- Analysis scale varies across activity detail, geographical units, and temporal intervals, and the chosen scale can affect which capabilities are identified.Country-level requirements may differ from local ones; for example, broad capabilities may include education, while local requirements may involve infrastructure or climate.Source: Scale of analysis · p. 12
- A joint analysis of scientific papers, patents, and exports cannot use a geographical resolution finer than countries or a time interval shorter than one year.Exports are recorded by customs and are not readily available at finer geographical resolutions.Source: Scale of analysis · p. 12
- Finer activity or time resolution reduces statistical power, so specific-activity analyses require pooling years, while fine-time analyses use aggregated fields or counts of significant fields.Pooling can sum yearly matrices or stack yearly observations as separate rows.Source: Scale of analysis · p. 12