Relying on recent and temporally dispersed science predicts breakthrough inventions

TL;DR

A large corpus of patents is used and features characterizing how patents temporally search in the scientific space are derived, finding that patents that cite scientific papers have more citations and substantially more likely to become breakthroughs.

Abstract

The development of inventions is theorized as a process of searching and recombining existing knowledge components. Previous studies under this theory have examined myriad characteristics of recombined knowledge and their performance implications. One such feature that has received much attention is technological knowledge age. Yet, little is known about how the age of scientific knowledge influences the impact of inventions, despite the widely known catalyzing role of science in the creation of new technologies. Here we use a large corpus of patents and derive features characterizing how patents temporally search in the scientific space. We find that patents that cite scientific papers have more citations and substantially more likely to become breakthroughs. Conditional on searching in the scientific space, referencing more recent papers increases the impact of patents and the likelihood of being breakthroughs. However, this positive effect can be offset if patents cite papers whose ages exhibit a low variance. These effects are consistent across technological fields.

Figures (9)

• Figure 1: The positive effect of search in the scientific space on patent impact. (A) Fractions of patents with patent references (PR) and SNPR over time; (B--C) Mean 10-year forward citations (B) and mean hit probabilities (C) for the groups of patents with and without SNPR; (D--E) Estimated 10-year citations based on the negative binomial regression models without (D) and with (E) considering the interaction between searches in the scientific and technological spaces; (F--G) Estimated hit probabilities based on the logistic regression models without (F) and with (G) considering the interaction between the two searches.
• Figure 2: Temporal search in the scientific space. (A) Patent-level average of $\mu_s$ and $\mu_t$, mean age of cited papers and cited patents for citing patents granted over time. (B) Patent-level average of $cv_s$ and $cv_t$, coefficient of variation of ages of cited papers and patents; (C) Hit probabilities for the four groups of patents based on whether their $\mu_s$ and $cv_s$ are below or above the respective global means over all patents in the decade.
• Figure 3: The effects of temporal search in the scientific space on patent impact. (A--D) Estimated 10-year forward citations based on negative binomial regression models without (A, C) and with (B, D) considering the interaction between temporal searches in the scientific and technological spaces. (E--H) Estimated hit probabilities based on logistic regression models without (E, G) and with (F, H) considering the interaction.
• Figure 4: The effects of temporal search in the scientific space on patent impact, by technological field. (A--B) Impact is based on 10-year forward citations. (C--D) Impact is based on whether patents are hits.
• Figure A1: Regression modeling of temporal search and impact. Only Chemical patents are included.