Conceived and designed the experiments: JB. Performed the experiments: JB RC. Analyzed the data: JB AH RC. Wrote the paper: JB HVdS AH. Methodological consultancy: HVdS.
The authors have declared that no competing interests exist.
The impact of scientific publications has traditionally been expressed in terms of citation counts. However, scientific activity has moved online over the past decade. To better capture scientific impact in the digital era, a variety of new impact measures has been proposed on the basis of social network analysis and usage log data. Here we investigate how these new measures relate to each other, and how accurately and completely they express scientific impact.
We performed a principal component analysis of the rankings produced by 39 existing and proposed measures of scholarly impact that were calculated on the basis of both citation and usage log data.
Our results indicate that the notion of scientific impact is a multi-dimensional construct that can not be adequately measured by any single indicator, although some measures are more suitable than others. The commonly used citation Impact Factor is not positioned at the core of this construct, but at its periphery, and should thus be used with caution.
Science is a
A variety of impact measures can be derived from raw citation data. It is however highly common to assess scientific impact in terms of average journal citation rates. In particular, the Thomson Scientific Journal Impact Factor (JIF)
The JIF has achieved a dominant position among measures of scientific impact for two reasons. First, it is published as part of a well-known, commonly available citation database (Thomson Scientific's JCR). Second, it has a simple and intuitive definition. The JIF is now commonly used to measure the impact of journals and by extension the impact of the articles they have published, and by even further extension the authors of these articles, their departments, their universities and even entire countries. However, the JIF has a number of undesirable properties which have been extensively discussed in the literature
The shortcomings of the JIF as a simple citation statistic have led to the introduction of other measures of scientific impact. Modifications of the JIF have been proposed to cover longer periods of time
In addition, the success of Google's method of ranking web pages has inspired numerous measures of journal impact that apply social network analysis
Since scientific literature is now mostly published and accessed online, a number of initiatives have attempted to measure scientific impact from
These developments have led to a plethora of new measures of scientific impact that can be derived from citation or usage log data, and/or rely on distribution statistics or more sophisticated social network analysis. However, which of these measures is most suitable for the measurement of scientific impact? This question is difficult to answer for two reasons. First, impact measures can be calculated for various citation and usage data sets, and it is thus difficult to distinguish the true characteristics of a measure from the peculiarities of the data set from which it was calculated. Second, we do not have a universally accepted, golden standard of impact to calibrate any new measures to. In fact, we do not even have a workable definition of the notion of “scientific impact” itself, unless we revert to the tautology of defining it as the number of citations received by a publication. As most abstract concepts “scientific impact” may be understood and measured in many different ways. The issue thus becomes which impact measures best express its various aspects and interpretations.
Here we report on a Principal Component Analysis (PCA)
The mentioned 39 scientific impact measures were derived from various sources. Our analysis included several existing measures that are published on a yearly basis by Thomson-Reuters and the Scimago project. Other measures were calculated on the basis of existing citation- and usage data. The following sections discuss the methodology by which each of these impact measures was either extracted or derived from various usage and citation sources.
As shown in
Impact measure identifiers refer to
The CDROM version of the 2007 Journal Citation Reports (JCR Science and Social Science Editions) published by Thomson-Reuters Scientific (formerly ISI).
The MESUR project's reference collection of usage log data:
A set of journal rankings published by the Scimago project that are based on Elsevier Scopus citation data:
In the following sections we detail the methodology that was used to retrieve and calculate 39 scientific impact measures from these data sets, and the subsequent analysis of the correlations between the rankings they produced. Throughout the article measures are identified by a unique identifier number that is listed in
Black dots indicate citation-based measures. White dots indicate usage-based measures. The Journal Impact Factor (5) has a blue lining. Measures 23 and 39 excluded.
ID | Type | Measure | Source | Network parameters | PC1 | PC2 | |
1 | Citation | Scimago Journal Rank | Scimago/Scopus | −0.974 | −8.296 | 0.556^{*} | |
2 | Citation | Immediacy Index | JCR 2007 | 1.659 | −7.046 | 0.508^{*} | |
3 | Citation | Closeness Centrality | JCR 2007 | Undirected, weighted | 0.339 | −6.284 | 0.565^{*} |
4 | Citaton | Cites per doc | Scimago/Scopus | −1.311 | −6.192 | 0.588^{*} | |
5 | Citation | Journal Impact Factor | JCR 2007 | −1.854 | −5.937 | 0.592^{*} | |
6 | Citation | Closeness centrality | JCR 2007 | Undirected, unweighted | −1.388 | −4.827 | 0.619 |
7 | Citation | Out-degree centrality | JCR 2007 | Directed, weighted | −3.191 | −4.215 | 0.642 |
8 | Citation | Out-degree centrality | JCR 2007 | Directed, unweighted | −2.703 | −4.015 | 0.640 |
9 | Citation | Degree Centrality | JCR 2007 | Undirected, weighted | −4.850 | −2.834 | 0.690 |
10 | Citation | Degree Centrality | JCR 2007 | Undirected, unweighted | −4.398 | −2.643 | 0.691 |
11 | Citation | H-Index | Scimago/Scopus | −3.326 | −2.003 | 0.681 | |
12 | Citation | Scimago Total cites | Scimago/Scopus | −4.926 | −1.722 | 0.712 | |
13 | Citation | Journal Cite Probability | JCR 2007 | −5.389 | −1.647 | 0.710 | |
14 | Citation | In-degree centrality | JCR 2007 | Directed, unweighted | −5.302 | −1.429 | 0.717 |
15 | Citation | In-degree centrality | JCR 2007 | Directed, weighted | −5.380 | −1.554 | 0.712 |
16 | Citation | PageRank | JCR 2007 | Directed, unweighted | −4.476 | 0.108 | 0.693 |
17 | Citation | PageRank | JCR 2007 | Undirected, unweighted | −4.929 | 0.731 | 0.726 |
18 | Citation | PageRank | JCR 2007 | Undirected, weighted | −4.160 | 0.864 | 0.696 |
19 | Citation | PageRank | JCR 2007 | Directed, weighted | −3.103 | 0.333 | 0.659 |
20 | Citation | Y-factor | JCR 2007 | Directed, weighted | −2.971 | 0.317 | 0.657 |
21 | Citation | Betweenness centrality | JCR 2007 | Undirected, weighted | −0.462 | 0.872 | 0.643 |
22 | Citation | Betweenness centrality | JCR 2007 | Undirected, unweighted | −0.474 | 1.609 | 0.642 |
23 | / | / | |||||
24 | Usage | Closeness centrality | MESUR 2007 | Undirected, weighted | 3.130 | 2.683 | 0.703 |
25 | Usage | Closeness centrality | MESUR 2007 | Undirected, unweighted | 3.100 | 3.899 | 0.731 |
26 | Usage | Degree centrality | MESUR 2007 | Undirected, unweighted | 3.271 | 3.873 | 0.729 |
27 | Usage | PageRank | MESUR 2007 | Undirected, unweighted | 3.327 | 4.192 | 0.728 |
28 | Usage | PageRank | MESUR 2007 | Directed, unweighted | 3.463 | 4.336 | 0.727 |
29 | Usage | In-degree centrality | MESUR 2007 | Directed, unweighted | 3.463 | 4.015 | 0.728 |
30 | Usage | Out-degree centrality | MESUR 2007 | Directed, unweighted | 3.484 | 3.994 | 0.727 |
31 | Usage | PageRank | MESUR 2007 | Directed, weighted | 3.780 | 4.217 | 0.710 |
32 | Usage | PageRank | MESUR 2007 | Undirected, weighted | 3.813 | 4.223 | 0.710 |
33 | Usage | Betweenness centrality | MESUR 2007 | Undirected, unweighted | 3.988 | 4.271 | 0.699 |
34 | Usage | Betweenness centrality | MESUR 2007 | Undirected, weighted | 3.957 | 3.698 | 0.693 |
35 | Usage | Degree centrality | MESUR 2007 | Undirected, weighted | 5.293 | 3.528 | 0.683 |
36 | Usage | Out-degree centrality | MESUR 2007 | Directed, weighted | 5.302 | 3.518 | 0.683 |
37 | Usage | In-degree centrality | MESUR 2007 | Directed, weighted | 5.286 | 3.531 | 0.683 |
38 | Usage | Journal Use Probability | MESUR 2007 | 8.914 | 1.833 | 0.593 | |
39 | / | / |
The 2007 JCR contains a table listing 4 citation-based impact measures for a set of approximately 7,500 selected journals, namely
(
(
(
In addition, the Scimago project publishes several impact measures that are based on Elsevier's Scopus citation data. We retrieved the following 4 measures from its web site:
(
(
(
(
The Scimago journal rankings were downloaded from their web site in the form of an Excel spreadsheet and loaded into a MySQL database. This added 4 measures of journal impact to our data set bringing the total number of retrieved, existing measures to 8.
In
The 2007 JCR contains a table that lists the number of citations that point from one journal to another. The number of citations is separated according to the publication year of both the origin and target of the citation. For example, from this table we could infer that 20 citations point from articles published in “Physica Review A” in 2006 to articles published in “Physica Review B” in 2004 and 2005. Each such data data point can thus be described as the n-tuple
We attempted to ensure that our citation network conformed to the definition of the Journal Impact Factor rankings published in the 2007 JCR. We therefore extracted citations from the JCR that originated in 2006 publications and pointed to 2004 and 2005 publications. The resulting citation network contained 897,608 connections between 7,388 journals, resulting in a network density of 1.6% (ratio of non-zero connections over all possible non-reflexive connections). This citation network was represented as a 7,338×7,338 matrix labeled C whose entries
In
In short, the MESUR project's reference collection of usage log data consists of log files recorded by a variety of scholarly web portals (including some of the world's most significant publishers and aggregators) who donated their usage log data to the MESUR project in the course of 2006–2007. All MESUR usage log data consisted of a list of temporally sorted “requests”. For each individual request the following data fields were recorded: (1) date/time of the request, (2) session identifier, (3) article identifier, and (4) request type. The session identifier grouped requests issued by the same (anonymous) user, from the same client, within the same session. This allowed the reconstruction of user “clickstreams”, i.e. the sequences of requests by individual users within a session. Since each article for this investigation is assumed to be published in a journal, we can derive journal clickstreams from article clickstreams.
Over all clickstreams we can thus determine the transition probability
This analysis was applied to the MESUR reference data set, i.e. 346,312,045 user interactions recorded by the web portals operated by Thomson Scientific (Web of Science), Elsevier (Scopus), JSTOR, Ingenta, University of Texas (9 campuses, 6 health institutions), and California State University (23 campuses) between March 1st 2006 and February 1st 2007. To ensure that all subsequent metrics were calculated over the same set of journals, the resulting set of journal transition probabilities were trimmed to 7,575 journals for which a JIF could be retrieved from the 2007 JCR. All usage transition probabilities combined thus resulted in the 7,575×7,575 matrix labeled
Four classes of social network measures were applied to both the citation and usage network represented respectively by matrix
(
(
(
(
The definitions of each of the measures in these classes were varied according to the following network factors: (1) Weighted vs. unweighted connections, i.e. measures can be calculated by assuming that each non-zero connection valued 1 vs. taken into account the actual weight of the connection, (2) Directed vs. undirected connections, i.e. some measures can be calculated to take into account the directionality of journal relations or not, and finally (3) Citation vs. usage network data, i.e. any of these measure variations can be calculated for either the citation or the usage network.
These factors result in 2^{3} = 8 variations for each the above listed 4 classes of social network measures, i.e. 32 variants. However, not all permutations make equal sense. For example, in the case of Betweenness Centrality we calculated only two of these variants that both ignored connection directionality (irrelevant for betweenness) but one took into account connection weights (weighted geodesics) and another ignored connections weights (all connections weighted >0). Each of these variants were however calculated for the citation and usage-network. The final list of social network measures thus to some degree reflect our judgment on which of these permutations were meaningful.
In addition to the existing measures and the social network measure, we calculated, a number of measures that did not fit any the above outlined classes, namely
(
(
(
(
In total, we calculated 32 citation- and usage-based impact measures; 16 social network measures on the basis of matrix
The set of selected measures was intended to capture the major classes of statistics and social network measures presently proposed as alternatives to the JIF. In summary, the set of all measures can be categorized in 4 major classes. First,
Spearman rank-order correlations were then calculated for each pair of journal rankings. Because
A sample of matrix R for 10 selected measures is shown below. For example, the Spearman rank-order correlation between the Citation H-index and Usage PageRank is 0.66. The IDs listed in
Not all pair-wise correlations were statistically significant. Two measures in particular lacked significant correlations (
The resulting PCA components were ranked according to the degree by which they explain the variances in
PC1 | PC2 | PC3 | PC4 | PC5 | |
Proportion of Variance | 66.1% | 17.3% | 9.2% | 4.8% | 0.9% |
Cumulative Proportion | 66.1% | 83.4% | 92.6% | 97.4% | 98.3% |
We projected all measures unto the first two components, PC1 and PC2, to create a 2-dimensional map of measures. A varimax rotation was applied to the measure loadings to arrive at a structure that was more amenable to interpretation. The measure loadings for each component are listed in
To cross-validate the PCA results, a hierarchical cluster analysis (single linkage, euclidean distances over
The map in
A complete linkage hierarchical cluster analysis based on the Euclidean distances of the measure
Cluster | Measures | Interpretation |
1 | 38 | Journal Use Probability |
2 | 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37 | Usage measures |
3 | 1, 2, 3, 4, 5 | JIF, SJR, Cites per Document measures |
4 | 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 | Total Citation rates and distributions |
5 | 16, 17, 18, 19, 20, 21, 22 | Citation Betweenness and PageRank |
The pattern of clusters indicate that some measures express a more distinct aspect of scientific impact and will thus be farther removed from all other measures.
To interprete the meaning of PC1 and PC2 we need to investigate the distribution of measures along either axis of the map in
PC1 clearly separates usage measures from citation measures. On the positive end of PC1, we find a sharply demarcated cluster of all usage measures, with the exception of the Journal Use Probability (ID 38) which sits isolated on the extreme positive end of PC1. On the negative end of PC1, we find most citation measures. Surprisingly, some citation measures are positioned close to the cluster of usage measures in terms of their PC1 coordinates. Citation Closeness (ID 3) and in particular Citation Immediacy Index (ID 2) are located on the positive end of PC1, i.e. closest to the usage measures. Citation Betweenness Centrality (IDs 21 and 22) are also positioned closely to the cluster of usage measures according to PC1.
This particular distribution of citation measures along PC1 points to an interesting, alternative interpretation of PC1 simply separating the usage from the citation measures. In the center, we find Citation Immediacy Index (ID 2) positioned close to the cluster of usage measures in terms of its PC1 coordinates. The Citation Immediacy Index is intended to be a “rapid” indicator of scientific impact since it is based on same-year citations. Its proximity to the usage measures according to PC1 may thus indicate that the usage measures are equally rapid indicators, if not more so. The assumption that usage measures are “Rapid” indicators of scientific impact is furthermore warranted for the following reasons. First, usage log data is generally considered a more “rapid” indicator of scientific impact than citation data, since usage log data is nearly immediately affected by changes in scientific habits and interests whereas citation data is subject to extensive publication delays. It has in fact been shown that present usage rates predict future citation rates
PC2 separates citation statistics such as Scimago Total Cites (ID12), JIF (
Consequently, the PCA results could be interpreted in terms of a separation of measures along two dimensions: “Rapid” vs. “Delayed” (PC1) and “Popularity” vs. “Prestige” (PC2). Surprisingly, most usage-based measures would then fall in the “Rapid, “Prestige” quadrant, approximated in this aspect only by two Citation Betweenness Centrality measures. The majority of citation-based measures can then be classified as “Delayed”, but with the social network measures being indicative of aspects of “Prestige” and the normalized citation measures such as the JIF, Scimago Journal Rank (ID 1) and Cites per Doc indicative of journal “Popularity”. We also note that the Scimago Journal Rank is positioned among measures such as the JIF and Cites per Doc. This indicates it too expresses “Delayed” “Popularity”, in spite of the fact that SJR rankings originate from 2007 citation data and that the SJR has been explicitly defined to “transfer(s) (of) prestige from a journal to another one” (
Another interesting aspect of the distribution of measures along PC1 and PC2 relates to the determination of a “consensus” view of scientific impact. The
The presented results pertain to what we believe to be the largest and most thorough survey of usage- and citation based measures of scientific impact. Nevertheless, a number of issues need to be addressed in future research efforts.
First, although an attempt was made to establish a representative sample of existing and plausible scientific impact measures, several other conceivable impact measures could have been included in this analysis. For example, the HITS algorithm has been successfully applied to web page rankings. Like Google's PageRank it could be calculated for our citation and usage journal networks. Other possible measures that should be considered for inclusion include the Eigenfactor.org measures, and various information-theoretical indexes. The addition of more measures may furthermore enable statistical significance to be achieved on the correlations with now-removed measures such as Citation Half-Life and the Usage Impact Factor, so that they could be included on the generated PCA map of measures.
Second, we projected measure correlations onto a space spanned by the 2 highest-ranked components, the first of which seems to make a rather superficial distinction between usage- and citation-derived impact measures and the second of which seems to make a meaningful distinction between “degree” and “quality” of endorsement. Future analysis should focus on including additional components, different combinations of lower-valued components and even the smallest-valued components to determine whether they reveal additional useful distinctions. In addition, non-linear dimensionality reduction methods could be leveraged to reveal non-linear patterns of measure correlations.
Third, a significant number of the measures surveyed in this article have been standard tools for decades in social network analysis, but they are not in common use in the domain of scientific impact assessment. To increase the “face-validity” of these rankings, all have been made available to the public on the MESUR web site and can be freely explored and interacted with by users at the following URL:
Fourth, the implemented MESUR services can be enhanced to support the development of novel measures by allowing users to submit their own rankings which can then automatically be placed in the context of existing measures. Such a service could foster the free and open exchange of scientific impact measures by allowing the public to evaluate where any newly proposed measure can be positioned among existing measures. If the measure is deemed to similar to existing measures, it need not be developed. If however, it covers a part of the measure space that was previously unsampled, the new measure may make a significant contribution and could therefore be considered for wider adoption by those involved in scientific assessment.
Our results indicate that scientific impact is a multi-dimensional construct. The component loadings of a PCA indicate that 92% of the variances between the correlations of journal rankings produced by 37 impact measures can be explained by the first 3 components. To surpass the 95% limit, a 4-component model would have to be adopted.
A projection of measure correlations onto the first 2 components (83.4%) nevertheless reveals a number of useful distinctions. We found that the most salient distinction is made by PC1 which separates usage from citation measures with the exception of Citation Betweenness centrality and Citation Immediacy. The position of the latter and the time periods for which usage was recorded suggests an interpretation of PC1 as making a distinction between measures that provide a “rapid” vs “delayed” view of scientific impact.
PC2 seems to separate measures that express Popularity from those that express Prestige. Four general clusters of impact measures can be superimposed on this projection: (1) usage measures, (2) a group of distinctive yet dispersed measures expressing per document citation popularity, (3) measures based on total citation rates and distributions, and (4) finally a set of citation social network measures. These 4 clusters along with the PCA components allows us to quantitatively interpret the landscape of presently available impact measures and determine which aspects of scientific impact they represent. Future research will focus on determining whether these distinctions are stable across a greater variety of measures as well other usage and citation data sets.
Four more general conclusions can be drawn from these results; each has significant implications for the developing science of scientific assessment.
First, the set of usage measures is more strongly correlated (average Spearman rank-order correlation = 0.93, incl. Usage Probability) than the set of citation measures (average Spearman rank-order correlation = 0.65). This indicates a greater reliability of usage measures calculated from the same usage log data than between citation measures calculated from the same citation data. This effect is possibly caused by the significantly greater density of the usage matrix
Second, if our interpretation of PC2 is correct, usage-based measures are actually
Third, some citation measures are more closely related to their usage counterparts than they are to other citation measures such as the JIF. For example, the Spearman rank-order correlation between Citation Betweenness Centrality and Usage Betweenness Centrality is 0.747. In comparison, the Spearman rank-order correlation between the JIF and Citation Betweenness Centrality is only 0.52. This indicates that contrary to what would be expected, usage impact measures can be closer to a “consensus ranking” of journals than some common citation measures.
Fourth, and related, when we rank measures according to their average correlation to all other measures
The ranking data produced to support the discussed Principal Component Analysis is available upon request from the corresponding author with the exception of those that have been obtained under proprietary licenses.
(0.06 MB PDF)