The current issue and full text archive of this journal is available at www.emeraldinsight.com/0022-0418.htm JDOC 70,1 52 Received 17 October 2012 Revised 9 January 2013 Accepted 15 January 2013 Assigning publications to multiple subject categories for bibliometric analysis An empirical case study based on percentiles Lutz Bornmann Division for Science and Innovation Studies, Administrative Headquarters of the Max Planck Society, Munich, Germany Abstract Purpose This study is concerned with a problem in measuring citation impact with the aid of percentile data, which arises from the assignment of publications (or the journals in which the publications have appeared) by Thomson Reuters for the Web of Science to more than one subject category. If there is more than one subject category for a publication, it is initially unclear which category is to be used to create the reference set for the calculation of the percentile. This paper seeks to address these issues. Design/methodology/approach In this study the author would like to look at whether the calculation of differences between the citation impact of research institutions is affected by whether the minimum (the maximum percentile), the maximum (the minimum percentile), the mean or the median impact (percentile) for the different subject categories is used. The study is based on a sample of percentile data for three research institutions (n ¼ 4,232). Findings The result of the comparison of citation impact of the three institutions remains very similar for all the calculation methods, but on a different level. Originality/value It is the first study, which investigates how far it makes a difference in the comparison of the citation impact of three different research institutes whether with multiple assignments of subject categories to one publication the minimum, the maximum, the mean or the median inverted percentile is used. An answer to the question is very relevant since different methods are used in practical application. For example, the web-based research evaluation tool InCites uses the minimum percentile. Keywords Subject category, Percentile, Normalized citation impact, InCites Paper type Research paper Journal of Documentation Vol. 70 No. 1, 2014 pp. 52-61 q Emerald Group Publishing Limited 0022-0418 DOI 10.1108/JD-10-2012-0136 1. Introduction Reference sets are used in bibliometrics to make the citation impact of publications (the observed citation impact) with each other comparable. These reference sets are made up of publications from the same publication year, the same field and the same document type as the publications in question. Fields are frequently delimited in bibliometrics with the aid of journal sets (so-called subject categories) which are used in the literature databases Web of Science (Thomson Reuters) and Scopus (Elsevier). The arithmetic mean value of the citations for the publications in the reference set is The author would like to thank Ludo Waltman from the Centre for Science and Technology Studies (CWTS) for the dataset used in this study.
calculated to specify an expected citation impact (Schubert and Braun, 1986). A quotient from the observed and expected citation impact makes it possible to compare the citation impact of publications from different publication years and fields and of different document types. A number of different methods have been developed to calculate the quotients; there is an overview of different methods in Vinkler (2012). There are two significant disadvantages inherent in the calculation of these quotients for the so-called normalisation of citation impact (Bornmann et al., 2011): (1) As a rule, the distribution of citations over publication sets is skewed to the right. The arithmetic mean value calculated for a reference set is therefore determined by a few highly-cited papers. The arithmetic mean as a measure of central tendency is not suitable for skewed data. This is the only reason why, for example, in the Leiden Ranking 2011/2012 the University of Göttingen occupies position 2 in a ranking by citation impact; the relevant mean value for this university turns out to have been strongly influenced by a single extremely highly cited publication (Waltman et al., 2012, p. 2425). (2) The quotient only permits a statement about whether a publication is cited more or less than the average in the reference set. Other attributions which could describe the citation impact of a publication as excellent or outstanding are based on (arbitrary) rules of thumb with no relationship to statistical citation distributions (Leydesdorff et al., 2011). Multiple subject categories 53 Using percentiles (or percentile rank classes) to normalise citation impact can give better comparisons of the impact of publications than normalisation using the arithmetic mean. The percentile provides information about the citation impact the publication in question has had compared to other publications. A percentile is a value below which a certain proportion of observations fall: The higher the percentile for a publication, the more citations it has received compared to publications in the same field and publication year and with the same document type. The percentile for a publication is determined using the distribution of the percentile ranks over all publications: for example, a value of 90 means that the publication in question is among the 10 per cent most cited publications; the other 90 per cent of the publications have achieved less impact. A value of 50 represents the median and thus an average citation impact compared to the other publications (from the same field and publication year and with the same document type). Up to now, research in bibliometrics has concerned itself primarily with the calculation of percentiles and their statistical analysis (Bornmann et al., 2013). Some different proposals have been made of appointing publications to percentiles (percentile rank classes) and different suggestions have been published (Bornmann, n.d.b; Leydesdorff et al., 2011; Rousseau, 2012; Schreiber, 2012a, b, Bornmann et al., 2013). Bornmann (n.d.a) proposes statistical methods with which to analyse percentile data meaningfully. This study is concerned with another topic or problem in measuring citation impact with the aid of percentile data, which arises from the assignment of publications (or the journals in which the publications have appeared) by Thomson Reuters for the Web of Science (WoS) to more than one subject category. If there is more than one subject category for a publication, it is initially unclear which category is to be used to create the reference set. According to Herranz and
JDOC 70,1 54 Ruiz-Castillo (2012) this is a significant problem generally encountered in the creation of field-normalised indicators. In the dataset of Herranz and Ruiz-Castillo (2012) where subfields are identified with the 219 Web of Science categories distinguished by Thomson Scientific, 42 per cent of the 3.6 million articles published in 1998-2002 are assigned to two or more, up to a maximum of six subfields. Multiple assignments of subject categories to publications thus affect a little less than half of all the publications, roughly speaking. In InCites (http://incites.thomsonreuters.com/), Thomson Reuters offers one of the few web-based research evaluation tools with which advanced bibliometric citation impact indicators (such as percentiles) can be queried for individual publications. If there is more than one subject category for a publication, Thomson Reuters calculates a percentile for each category and selects the one with the highest citation impact for a publication. Percentiles offered by InCites are thus based on the maximum impact that a publication has achieved. As there are other ways of selecting percentiles, in this study we would like to look at whether the calculation of differences between the citation impact of research institutions is affected by whether the minimum (the maximum percentile), the maximum (the minimum percentile), the mean or the median impact (percentile) for the different subject categories is used. Even though it is obvious for each individual publication which of these percentiles results in a higher or lower citation impact: (1) The varying proportions of publications for institutions with more than one subject category can make it almost impossible to estimate results when publications are aggregated. (2) It is also very difficult to estimate the effect of larger and smaller differences between the percentiles for a publication (assigned by Thomson Reuters to more than one subject category) on an aggregated institution level (e.g. if an average percentile for the institution is calculated). Therefore it is necessary to investigate the different percentiles (minimum, maximum, mean and median) in a real application of an evaluation study. According to Herranz and Ruiz-Castillo (2012) there is also the option of a multiplicative strategy to deal with the problem of assigning publications (journals) to more than one subject category according to which each paper is wholly counted as many times as necessary in the several categories to which it is assigned at each aggregation level. In this way, the space of articles is expanded as much as necessary beyond the initial size. The multiplicative strategy can be interpreted as a weighting scheme according to multidisciplinarity. We will also look at this strategy in terms of the percentile data in this study. 2. Methods 2.1 The dataset used The data for this study is taken from all the publications published by three research institutions (RI1, RI2 and RI3) from 2007 to 2009 in the form of articles or reviews (n ¼ 4,232) (see Table I). The percentiles in this study are calculated in accordance with the InCites approach. The percentile in which the paper ranks in its [subject] category and database year, based on total citations received by the paper. The higher the number [of] citations, the smaller the percentile number. The maximum percentile
value is 100, indicating 0 cites received. Only the article types article, note, and review are used to determine the percentile distribution, and only those same article types receive a percentile value (http://incites.isiknowledge.com/common/help/h_glossary. html). Since in a departure from convention low percentile values mean high citation impact (and vice versa), the percentiles received from InCites are called inverted percentiles. As Table I shows, there are up to six subject categories (and therefore six inverted percentiles) for the publications from the three RIs. Around half to two-thirds of the publications from the RIs were assigned by Thomson Reuters to only one subject category; there are more than three subject categories for only around 6 per cent of the publications. To determine the minimum and maximum inverted percentiles, for publications for which there is more than one subject category, the minimum or maximum from the inverted percentiles was chosen in each case. For the mean and median inverted percentile, the arithmetic mean or the median for the inverted percentiles in each case was calculated. Publications with only one subject category have the same percentile for all the calculation methods. As the proportion of publications in the set with only one subject category is very high at around 60 per cent (see Table I), the statistical analyses have been carried out both for the whole data set and only for the publications for which there are at least two subject categories. The supplementary analysis examines whether the results on the citation impact of the RI s change when the majority of the publications with only one subject category is ignored. A data set is created for the analyses based on the multiplicative strategy in which not the publications, but the subject categories (or the corresponding inverted percentiles) of the individual publications form the units. Publications with more than one subject category are included in the citation impact analysis more than once in this dataset. If there are, for example, six subject categories for a publication, this publication is included with six units in the analysis (see Waltman et al., 2012). Multiple subject categories 55 3. Results 3.1 Results based on the average inverted percentile Table II (A) shows means and medians which were calculated on the basis of differently calculated inverted percentiles for the publications from the three RIs. Viewed through the different calculation methods described above, it can be seen that on average the results are very similar for the RIs. For all the calculation methods, RI2 and RI3 have a slightly higher average citation impact than RI1. For example, the median of the minimum inverted percentile for the publications from RI1 is 45.46, for Number of inverted percentiles RI1 (n ¼ 2,674) RI2 (n ¼ 512) RI3 (n ¼ 1,046) Total (n ¼ 4,232) 1 56.06 55.66 66.35 58.55 2 28.35 24.61 17.88 25.31 3 10.51 9.96 8.13 9.85 4 4.08 3.91 1.82 3.50 5 0.41 2.34 0.48 0.66 6 0.60 3.52 5.35 2.13 Total 100.00 100.00 100.00 100.00 Table I. Percentage for each research institution of the number of inverted percentiles per publication
JDOC 70,1 56 Table II. Citation impact differences between three research institutions (RIs) Research institution RI1 RI2 RI3 Total (A) All publications n 2,674 512 1,046 4,232 Minimum inverted percentile: Mean 48.44 39.13 42.13 45.76 Median 45.46 31.46 33.99 40.65 Maximum inverted percentile: Mean 52.14 43.79 45.07 49.38 Median 51.38 38.52 38.43 46.72 Mean inverted percentile: Mean 50.34 41.49 43.6 47.61 Median 48.2 35.93 35.79 44.38 Median inverted percentile: Mean 50.45 41.52 43.61 47.68 Median 48.39 36.08 35.79 44.27 (B) Only publications with more than one subject category n 1,175 227 352 1,754 Minimum inverted percentile: Mean 46.11 36.28 41.86 43.98 Median 41.94 29.12 33.23 39.27 Maximum inverted percentile: Mean 54.52 46.78 50.59 52.73 Median 55.21 43.29 49.71 52.21 Mean inverted percentile: Mean 50.44 41.58 46.23 48.45 Median 48.83 35.86 42.27 46.14 Median inverted percentile: Mean 50.69 41.66 46.25 48.63 Median 49.23 35.88 42.92 46.19 Notes: If a paper has been assigned by Thomson Reuters to more than one subject category, the minimum inverted percentile, the maximum inverted percentile, the mean inverted percentile or the median inverted percentile is used to calculate the differences in citation impact RI2 and RI3 it is 31.46 and 33.99 respectively. As expected, the highest average inverted percentiles are for the maximum and the lowest for the minimum; mean and median inverted percentiles are between these. The Spearman rank-order correlation between the inverted percentiles based on the different calculation methods is almost 1. This means that the differences between the inverted percentiles can be explained primarily by the fact that a calculation method generally results in higher or lower values than another. How does the picture change when the dataset is restricted to publications with at least two subject categories (and the corresponding percentiles) only? Table II (B) shows the results. It can also be seen here consistent with the results in Table II (A) that for RI1 with each calculation method of the inverted percentiles there is a lower average citation impact than for RI2 and RI3. However, rather more marked differences between RI2 and RI3 than in Table II (A) can be seen here. On average, with all the calculation methods there is a lower citation impact for RI3 than for RI2 (in Table II, A,
the result is not so unambiguous). The stronger reduction of publications for RI3 than for RI1 and RI2 included in the analysis might be the reason: the lower the case numbers of a sample from the whole data set on which an analysis is based, the less likely it is that the sample can represent the set. As the correlation coefficients show, there are for the reduced dataset similarly high correlations between the inverted percentiles based on the different calculation methods as for the whole dataset which can also be interpreted in a similar way (see above). Multiple subject categories 57 3.2 Results based on the multiplicative strategy (average inverted percentile) The results in Table III are based on a dataset in which the subject categories (or the corresponding inverted percentiles) of the individual publications rather than the publications form the units. Publications with more than one subject category are therefore included more than once in the citation impact analysis. The citation impact differences between the three RIs in Table III (A) correspond approximately to those which are shown in Table II (A) for the mean or median inverted percentile. Table III (B) shows the results of the analysis of the multiplicative strategy which are based on the reduced dataset. There is hardly any difference from the results based on all publications. 3.3 Results based on class 10 per cent publications The proportion of class 10 per cent papers (PP top 10% ) indicator was also examined in a further analysis. The inverted percentiles are categorised in two rank classes for this indicator. As it has now established itself as standard on an institutional level to designate those publications in a set which are among the 10 per cent most cited publications in their subject category (publication year and document type) (P top 10% ) as highly cited publications (Bornmann et al., 2012; Tijssen and Van Leeuwen, 2006; Tijssen et al., 2002; Waltman et al., 2012), the indicator PP top 10% was formed consisting of the proportion of papers with an inverted percentile equal to or smaller than the tenth inverted percentile. The PP top 10% in Table IV is created on the basis of inverted percentiles using the various calculation methods (minimum, maximum, mean and median). Research institution RI1 RI2 RI3 Total (A) All publications n 4,445 938 1,760 7,143 Mean 50.37 40.53 41.13 46.80 Median 49.23 34.28 32.85 42.94 (B) Only publications with more than one subject category n 2,946 653 1,066 4,665 Mean 50.43 40.15 40.37 46.69 Median 49.78 33.66 31.56 42.94 Notes: The results are based on a dataset in which the subject categories (or the corresponding inverted percentiles) of the individual publications rather than the publications form the units for the analysis; Publications with more than one subject category are therefore included more than once in the citation impact analysis Table III. Citation impact differences between three research institutions (RIs)
JDOC 70,1 58 Table IV. PP top 10% for the three research institutions (RIs) Research institution RI1 RI2 RI3 Total (A) All publications n 2,674 512 1,046 4,232 Minimum inverted percentile 13.7 20.7 17.4 15.5 Maximum inverted percentile 11.0 15.6 14.4 12.4 Mean inverted percentile 11.9 17.2 15.6 13.5 Median inverted percentile 11.9 17.6 15.7 13.5 Fractional counted inverted percentile 12.3 17.9 15.9 13.9 (B) Only publications with more than one subject category n 1,175 227 352 1,754 Minimum inverted percentile 14.6 25.1 21.9 17.5 Maximum inverted percentile 8.4 13.7 13.1 10.0 Mean inverted percentile 10.6 17.2 16.5 12.6 Median inverted percentile 10.5 18.1 16.8 12.7 Fractional counted inverted percentile 11.4 18.8 17.3 13.5 Notes: If a publication has been assigned by Thomson Reuters to more than one subject category, the minimum inverted percentile, the maximum inverted percentile, the mean inverted percentile, the median inverted percentile or the fractional counted inverted percentile is used to calculate the differences in citation impact Table IV includes another calculation method, the fractional counted inverted percentile. This method is applied by the Centre for Science and Technology Studies (CWTS, Leiden, The Netherlands) for the most recent edition of the Leiden Ranking (www.leidenranking.com/) (Waltman et al., 2012). If, for example, a single publication belongs to two subject categories and in one category it belongs to the P top 10%, but in the other it does not, in the fractional approach the publication would be counted as P top 10% with a weight of 0.5. If the publication has been assigned by Thomson Reuters to three subject categories and it were to be among the P top 10% in one subject category, the corresponding weight would be 0.33. As the results in Table IV (A) show consistently with the other results that have already been presented, for RI1, for all the calculation methods, there is a lower citation impact than for RI2 and RI3 (that is, a smaller PP top 10% ). Table IV (A) also shows for all the RIs that the minimum or maximum inverted percentile results in the highest or lowest citation impact and that the mean or the median percentile is somewhere between the minimum and maximum. The fractional inverted percentile results in a PP top 10%, which is approximately at the level of the mean or median inverted percentile (the citation impact shown is slightly higher). When we compare the values in Table IV (A) with the values for the calculation of which only publications with at least two subject categories were taken into account (B), more significant changes occur particularly for the minimum and maximum inverted percentiles: The PP top 10% calculated on the basis of the minimum inverted percentile is for all the RIs in the selected data set slightly higher than in the whole set. The reverse is the case for the maximum inverted percentile (as expected). In contrast, there are fewer differences between the tables in the mean and median inverted percentiles. Fractional counted inverted percentiles also exhibit moderate differences.
If we compare the results for the RIs based on average values (Table II, A) with those based on the PP top 10% (Table IV, A) it can be seen that the former is less consistent than the latter. While Table IV (A) shows a higher citation impact for RI2 for all the calculation methods, Table II (A) has one or the other institute with a (slightly) higher impact depending on the calculation method (or the citation impact is practically identical). This result indicates that the indicator PP top 10% is relatively unaffected by the various calculation methods in a comparison of the citation impact of different RIs. Multiple subject categories 59 4. Discussion In evaluative bibliometrics, indicators based on percentiles have proven to be an important alternative to mean-based indicators in field-normalized citation metrics. Compared to the citation impact indicators used for the Leiden Ranking, the percentile indicator PP top 10% is viewed as the most important impact indicator (Waltman et al., 2012, p. 2425). Previous bibliometric research on percentiles had been concerned with the different options for calculating or analysing them. This study has looked at another issue concerning the problem of assigning publications to more than one subject category in the calculation of field-normalised indicators. A sample dataset is used to investigate how far it makes a difference in the comparison of the citation impact of three different RIs whether with multiple assignments of subject categories to one publication the minimum, the maximum, the mean or the median inverted percentile is used. Furthermore, it looks at in how far the use of a multiplicative strategy, in which every publication is taken into account in the analysis more than once (if it has more than one subject category), results in significant deviation in the results. The impact of the calculation method (minimum, maximum, mean or median) on the comparison of citation impact is verified using two indicators: the average inverted percentile and PP top 10%. The results for both indicators show that the minimum results in the highest on average and the maximum in the lowest on average. If a mean value is calculated using the inverted percentiles, this is generally between the minimum and the maximum. The result of the comparison of citation impact of the three RIs remains very similar for all the calculation methods, but on a different level. Thus, the varying proportions of publications for institutions with more than one subject category, and the larger and smaller differences between the percentiles for a publication (see the introduction) have only small effects on the empirical results. In the comparison of the two indicators average inverted percentile and PP top 10% the results show that the latter is less affected by the different calculation methods than the former. As we discussed at the beginning, one of the few available tools (InCites) in which advanced bibliometric indicators can be queried for individual publications, only provides the minimum inverted percentile for a publication. Consequently, as our results show, the citation impact is higher for the institutions in an evaluation study than inverted percentiles which are acquired with another calculation method. Using the minimum inverted percentile calculation method thus results in a citation impact of the institutions being overestimated as a rule. In a comparison of institutions using the InCites method, this does not represent a problem. Problems arise when percentile values from several studies which have calculated the percentiles differently are
JDOC 70,1 60 compared with each other, and the PP top 10% is compared with the expected value of 10 per cent. For a random selection of publications from a database, such as InCites, one would expect that 10 per cent of the publications belong to the 10 per cent of the most cited publications in their field (publication year and document type). Using the minimum inverted percentile calculation method in InCites can be justified in that the best possible citation impact which a publication has achieved is used as a basis for its impact evaluation. However, this does not take into account that the publication has already resulted in a poorer impact (if it has been assigned to more than one subject category by Thomson Reuters). For the Leiden Ranking, the fractional counted inverted percentile is used which, as our results show, leads to results very similar to those with the mean or median of the percentiles. Against the background of the results in this study we would like to recommend using the mean or median from the percentiles in evaluation studies. Both reflect the actual citation impact of a publication best, in that they take account of every citation impact achieved in a field and lead to similar results as the fractional counted inverted percentile or the analysis of the data using the multiplicative strategy. References Bornmann, L. (n.d.a), How to analyze percentile impact data meaningfully in bibliometrics? The statistical analysis of distributions, percentile rank classes and top-cited papers, Journal of the American Society for Information Science and Technology, in press. Bornmann, L. (n.d.b), The problem of percentile rank scores used with small reference sets, Journal of the American Society of Information Science and Technology, in press. Bornmann, L., De Moya Anegón, F. and Leydesdorff, L. (2012), The new excellence indicator in the world report of the SCImago Institutions Rankings 2011, Journal of Informetrics, Vol. 6, pp. 333-335. Bornmann, L., Leydesdorff, L. and Mutz, R. (2013), The use of percentiles and percentile rank classes in the analysis of bibliometric data: opportunities and limits, Journal of Informetrics, Vol. 7, pp. 158-165. Bornmann, L., Mutz, R., Marx, W., Schier, H. and Daniel, H.-D. (2011), A multilevel modelling approach to investigating the predictive validity of editorial decisions: do the editors of a high-profile journal select manuscripts that are highly cited after publication?, Journal of the Royal Statistical Society Series A (Statistics in Society), Vol. 174, pp. 857-879. Herranz, N. and Ruiz-Castillo, J. (2012), Sub-field normalization in the multiplicative case: highand low-impact citation indicators, Research Evaluation, Vol. 21, pp. 113-125. Leydesdorff, L., Bornmann, L., Mutz, R. and Opthof, T. (2011), Turning the tables in citation analysis one more time: principles for comparing sets of documents, Journal of the American Society for Information Science and Technology, Vol. 62, pp. 1370-1381. Rousseau, R. (2012), Basic properties of both percentile rank scores and the I3 indicator, Journal of the American Society for Information Science and Technology, Vol. 63, pp. 416-420. Schreiber, M. (2012a), Inconsistencies of recently proposed citation impact indicators and how to avoid them, available at: http://arxiv.org/abs/1202.3861 (accessed 20 February 2013). Schreiber, M. (2012b), Uncertainties and ambiguities in percentiles and how to avoid them, available at: http://arxiv.org/abs/1205.3588 (accessed 21 May 2013). Schubert, A. and Braun, T. (1986), Relative indicators and relational charts for comparative assessment of publication output and citation impact, Scientometrics, Vol. 9, pp. 281-291.
Tijssen, R. and Van Leeuwen, T. (2006), Centres of research excellence and science indicators. Can excellence be captured in numbers?, in Glänzel, W. (Ed.), Ninth International Conference on Science and Technology Indicators, Katholieke Universiteit Leuven, Leuven. Tijssen, R., Visser, M. and Van Leeuwen, T. (2002), Benchmarking international scientific excellence: are highly cited research papers an appropriate frame of reference?, Scientometrics, Vol. 54, pp. 381-397. Vinkler, P. (2012), The case of scientometricians with the absolute relative impact indicator, Journal of Informetrics, Vol. 6, pp. 254-264. Waltman, L., Calero-Medina, C., Kosten, J., Noyons, E.C.M., Tijssen, R.J.W., Van Eck, N.J., Van Leeuwen, T.N., Van Raan, A.F.J., Visser, M.S. and Wouters, P. (2012), The Leiden Ranking 2011/2012: data collection, indicators, and interpretation, Journal of the American Society for Information Science and Technology, Vol. 63, pp. 2419-2432. Multiple subject categories 61 Corresponding author Lutz Bornmann can be contacted at: bornmann@gv.mpg.de To purchase reprints of this article please e-mail: reprints@emeraldinsight.com Or visit our web site for further details: www.emeraldinsight.com/reprints