ent in our 90-million-patent dataset.

This involves loading the edge list
(approximately 160 million) of all ci-
tations and looping over citation net-
works until the algorithm converges
under the specified accuracy thresh-
olds. The PageRank offers a more in-
formative metric compared to simple
citation counts. Using the PageRank
approach, two patents with identical
citation counts are not necessarily
equal; the importance of the citing
patents instead reflects the impor-
tance of the focal invention.

Figure 2 shows the average PageRank by decade for selected technology sectors.i The effect we see with ICT patents is clearly reflected in the electrical-engineering sector where there was a distinct increase in importance from the 1970s onward. Moreover, these results seem to suggest the effect is ongoing. Emphasizing the continued effect of electrical engineering, mechanical engineering appears to be decreasing in importance, while chemistry and instruments appear to have remained at a constant (or slightly decreasing) level of importance during the period of study. The other-sectors category, including consumer goods and furniture and games, seem to be gradually increasing in importance, albeit from a more modest starting point.

Table 3 (model 3) lists the regression results of the PageRank analysis against all controls. We find ICT patents received 10% higher PageRanks than other technologies.j Although this appears to be a smaller difference than that of simple citations ( 27.6%), we highlight that the distribution of citation counts and PageRanks differ significantly (see Table 4).

In particular, the distribution of PageRanks is less dispersed than the simple counts, and its coefficient of variation (the standard deviation divided by the mean) highlights this phenomenon, as in Table 4. Simple counts have a coefficient of variation equal to 4.66, while PageRanks had a coefficient of variation of 2. 12, or less than half the dispersion of the citation distribution; that is, a 1% increase in PageRank is equivalent to a 2.2% increase in citation count.k The 10% increase in PageRank ICT patents receive was thus equivalent to a 22% increase in citation counts, only slightly less than the 27.6% we found, confirming the effect.

 

i Although our ICT/non-ICT comparison uses an extended definition of IC T, as in Table 2, we restrict the analysis here to common sectors for clarity.

j The estimated PageRank coefficient is 2.903,

or 10.0% of the mean PageRank of 28.939. k 4.664 divided by 2.118 is 2.202.

further check, we find that the controls have the expected signs and significance, with the stock of patents positively affecting total counts and more popular sectors experiencing higher citation counts.

We now perform the computationally demanding iterative process of measuring the PageRank of each pat-

Figure 2. Yearly average PageRank by technology sector.

5

10

1 5

20

2 5

30 A ve r ag e Ra n k of Pa ge R an k

1950 1960 1970 1980 1990 2000 2010

Chemistry

Electrical engineering Instruments Mechanical engineering Other fields

Table 4. Means, standard deviations, and coefficients of variation.

Mean Standard Deviation Coefficient of Variation

Citations (all publications) 1.473 6.873 4.664 PageRank (x 109) 28. 94 61. 29 2.118

Notes: Summary statistics for citations (across publications) and PageRank variables. Mean and standard deviation are reported in the first two columns. The coefficient of variation, a measure of the dispersion, defined as the ratio of the standard deviation to the mean, is reported in the last column.

Figure 3. Kernel densities for (left) distributions of citations and (right) PageRank.

0 . 1 . 2 . 3 . 4

02468

0 . 1 . 2 . 3 . 4

345678 non-ICT ICT

ke r ne ld en sit y lo g(ci tat io n s)

ke r ne ld en sit y lo g( pa ge r an k)