Upload
gali-halevi
View
375
Download
5
Embed Size (px)
Citation preview
The use of bibliometric indicators
in research assessment: A critical overview
Henk F. Moed
Senior scientific advisor,
Elsevier, Amsterdam, Netherlands
Contents
1 Beyond journal impact factor and H-index
2 University rankings have a limited value
3 Combine indicators and peer review
4 Social sciences deserve special attention
5 Indicators can be manipulated
6 Explore usage-based indicators
Contents
1 Beyond journal impact factor and H-index
2 University rankings have a limited value
3 Combine indicators and peer review
4 Social sciences deserve special attention
5 Indicators can be manipulated
6 Explore usage-based indicators
Journal impact measures are no good predictors of an individual paper’s
actual citation impact
Partly based on International Mathematical Union’s Report ‘Citation Statistics’ (2008)
Normal vs. skewed distributions
0
5
10
15
20
25
30
35
0 15 35 55 75 95 115 135 155 175 195 215Length (cm)
% P
erso
ns
Boys (Meanlength=95 cm)
Players (Meanlength=185 cm)
0
10
20
30
40
50
60
70
80
0 1 2 3 4 5 6 7 8Nr Cites
% P
aper
s
PAMS (JIF=0.43)
TAMS (JIF=0.85)
Adults
Boys Adults
What is the probability that .......
a randomly selected boy is at least as tall as a randomly selected adult?
Av. Length: Boys 85 cm; Adults: 185 cm
Almost zero
62 %a randomly selected PAMS paper is cited at least as often as a randomly selected TAMS paper?
JIF: PAMS: 0.43; TAMS: 0.85
Journal metrics should account for ‘free’ citations (and usage)
Base journal metric
Citations to all docs
# Citable docs
Citable vs. non-citable docs
Citable documents “non-citable” documents
Articles Letters
Reviews Editorials
Discussion papers
The problem of “free” citations - 1
Cites
Docs + + + + +
+ + + + +
The problem of “free” citations - 2
Cites
Docs + +
+ + + + +
“Free” Citations
SNIP corrects for disparities in citation potential among fields
A journal’s ‘Raw’ Citation Impact
‘Topicality’ of its subject field
SNIP: Base concept
SNIP =
How is a field’s ‘topicality’ measured?
Topicality
Citation potential
Length of cited reference lists
Differences in citation potential between fields
Molecular Biology Mathematics
Number of received citations
% Papers
Refe-rencelists
A journal’s raw impact per paper
Citation potential in its subject field
SNIP =
Journal scope, focus
Database coverage
peer reviewed papers only
A field’s frequency & immediacy of citation
Measured relative to database median
Citing papers
Target journalpapers
A journal’s subject field
journal’ssubjectfield
=papers citing the journal
Example 1 : Molec Biol vs. Mathematics
Journal JIF Cit Pot SNIP(= JIF/
Cit Pot)
INVENT MATH1.5
MOLEC CELL13.0
3.80.4
3.2 4.0
Example 2 : Within Mathematics
Journal JIF Cit Pot SNIP(= JIF/
Cit Pot)
Int J Nonlinear Sci & Num Sim
4.2
Commun Partial Different Equat
1.1
2.12.0
0.5 2.1
Example 3 : Social Sci vs. Biol & Med Sci
Journal JIF Cit Pot SNIP
J GERONTOL - A (Biol & Med Sci) 3.7
J GERONTOL - B (Psych & Soc Sci) 2.7
2.0 1.8
2.31.2
Strong points of SNIP
• Takes into account a journal’s scope
• Allows cross-subject comparisons
• Is independent of an a priori subject categorization
• Can be calculated for general journals
• Less potential for gaming
• Accounts for differences between and within journal subject ‘categories’
Institutional research assessment should apply indicators of actual citation impact and adequate benchmarking
Be careful with using the H-Index: Different citation distributions
may have the same value
All three publication lists have a Hirsch Index of 5
30 P110 P2 8 P3 6 P4 5 P5 1 P6 0 P7
30 P110 P2 8 P3 6 P4 5 P5 4 P6 4 P7 4 P8 4 P9
100 P1 70 P2 8 P3 6 P4 5 P5 1 P6 0 P7
H=? H=? H=?5 5 5
123456789
Author 2Author 1 Author 3
Bibliometric indicators are becoming increasingly
‘informative’
Bibliometric indicators more and more....
Feature Example
Embody ways to put numbers in context
Field-normalized citation measures
Take into account “who” is citing
Citations weighted with impact of citing source
Take into account relationship citing-cited author
Impact outside the own niche; multi-disciplinarity; bridging paradigms
Combine various types of indicators
HR data on personnel (gender, age, funding, ...)
Contents
1 Beyond journal impact factor and H-index
2 University rankings have a limited value
3 Combine indicators and peer review
4 Social sciences deserve special attention
5 Indicators can be manipulated
6 Explore usage-based indicators
University ranking positions are primarily marketing tools,
not research management tools
Research assessment methodologies must take into account… [EC AUBR Expert Group]
1. Inclusive definition of research / output
2. Different types of research and its impacts
3. Differences among research fields
4. Type and mission of institution
5. Proper units of assessment
6. Policy context, purpose and user needs
7. The European dimension
8. Need to be valid, fair and practically feasible
Types of outputs (SSH)
Impacts Publication/text Non-publication
Scientific-scholarly
Journal paper; book chapter; monograph
Research data file; video of experiment
Educational Teaching course book; syllabus
Skilled researchers
Economic Patent Product; process; device; design; image
Cultural Newspaper article; Interviews; events; Performances; exhibits
In institutional research assessment bottom-up approaches must include
data verification by evaluated authors
Top-down institutional analysis
Select an institution’s papers using author affiliations (incl. verification)
Categorize articles intoresearch fields
Calculate indicators
Compare with benchmarks
Bottom-up institutional analysis
Compile a list of researchers
Compile a list of publications per
researcher (incl. verification)
Aggregate researchers into groups, departments, fields, etc.
Calculate indicators;
compare with benchmarks
Metrics provides insight into global or systemic patterns
Gini index of disciplinary specialization
Gini = 0.0
Gini = 0.27
Gini = 0.52
Gini = 0.70
Data for a general, a poly-technical
and a specialized university
Relative Citation Rate (RCR)
The average citation rate of a unit’s papers
÷world citation average in the subfields in
which the unit is active
Corrects for differences in citation practices among fields,
publication years and type of article
Specialized universities perform in their fields of specialization less well than general institutions do
Data: Scopus / Scimagoir (n=1,500)
Data: Scopus / Scimagoir (n=1,500)
Specialized General High
L
ow
No linear correlation between a country’s institutional concentration and its citation impact
Data:Scopus/Scimago
Contents
1 Beyond journal impact factor and H-index
2 University rankings have a limited value
3 Combine indicators and peer review
4 Social sciences deserve special attention
5 Indicators can be manipulated
6 Explore usage-based indicators
Metrics can contribute to keeping the peer review process honest
Case study: A national Research Council
• Proposals evaluated by committees covering a discipline
• Reports from external referees
• Committee members among applicants
Affinity applicants – Committee
0 Applicants are/were not member of any Committee
1 Co-applicant is/was member of a Committee, but not of the one evaluating
2 First applicant is/was member of a Committee, but not of the one evaluating
3 Co-applicant is member of the Committee(s) evaluating the proposal
4 First applicant is member of the Committee(s) evaluating the proposal
For 15 % of applications an applicant is a member of the evaluating Committee (Affinity=3, 4)
0
10
20
30
40
50
60
70
% A
PP
LIC
AT
ION
S
AFFINITY APPLICANTS-COMMITTEE
Projects 63.2 10.2 11.5 5.9 9.1
0 1 2 3 4
Probability to be granted increases with increasing affinity applicants-Committee
30
40
50
60
70
80
% G
RA
NT
ED
AP
PL
ICA
TO
NS
AFFINITY APPLICANTS-COMMITTEE
Projects 37.0 46.9 60.1 62.6 74.0
0 1 2 3 4
Logistic regression analysis: Affinity Applicant-Committee has a significant effect
upon the probability to be granted
MAXIMUM-LIKELIHOOD ANALYSIS-OF-VARIANCE TABLE (N=2,499) Source DF Chi-Square Prob ------------------------------------------------------------- INTERCEPT 1 18.47 0.0000 Publ Impact applicant 3 26.97 0.0000 ** Rel transdisc impact applicant 1 0.29 0.5926 Affinity applicant-committee 2 112.50 0.0000 ** Sum requested 1 45.47 0.0000 ** Institution applicant 4 25.94 0.0000 ** LIKELIHOOD RATIO 199 230.23 0.0638
The future of research assessment exercises lies in the intelligent
combination of metrics and peer review
Intelligent combination of ‘metrics’ and peer review
• Policy makers let the type of peer review depend upon the outcomes of a bibliometric study
• Peer committees use citation analysis for initial rankings and explicitly justify why their final judgments deviate
• Metrics are used to assess peer review processes
Contents
1 Beyond journal impact factor and H-index
2 University rankings have a limited value
3 Combine indicators and peer review
4 Social sciences deserve special attention
5 Indicators can be manipulated
6 Explore usage-based indicators
CINon-CI
Non-CI CI
Citing/Source
Cited/Target
? %? %
Coverage of journal-based citation index (CI)
CINon-CI
Non-CI CI
Citing/Source
Cited/Target
± 80%± 20%
Science
CINon-CI
Non-CI CI
Citing/Source
Cited/Target
± 20%± 80%
Humanities
CI coverage by field
EXCELLENT(>80%)
GOOD (60-80%)
FAIR(40-60%)
MODERATE(<40%)
Biochem & Mol Biol
Appl Phys & Chem
Mathematics Other Soc Sci
Biol Sci – Humans
Biol Sci – Anim & Plants
EconomicsHumanities & Arts
ChemistryPsychol & Psychiat
Engineering
Clin Medicine Geosciences
Phys & Astron
Soc Sci ~ Medicine
JournalsBooks,
proceedings
Options for creating a comprehensive database of
research outputs in social sciences & humanities
Option Example/Case
1 Combine existing SSH bibliographies CSA-Illumina
2 Create new SSH databases Iberian Citation Index
3 Expand existing citation indexes WoS, Scopus
4 Explore Google Scholar; Book Search
5 Combine output registration systems MAETIS (NL)
6 Citation index from repositories Book Citation Index Project
7 Electronic Library Catalogues WorldCat
Contents
1 Beyond journal impact factor and H-index
2 University rankings have a limited value
3 Combine indicators and peer review
4 Social sciences deserve special attention
5 Indicators can be manipulated
6 Explore usage-based indicators
Users and producers of metrics should be alert on ‘manipulation’
Effects of editorial self-citations upon journal impact factors
[Reedijk & Moed, J. Doc., 2008]
• Editorial self-citations: A journal editor cites in his editorials papers published in his own journal
• Focus on ‘consequences’ rather than ‘motives’
Case: ISI/JCR Impact Factor of a Gerontology Journal (published in the journal itself)
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
2000 2001 2002 2003 2004
IMPACT FACTOR YEAR
CIT
ES
PE
R 'C
ITA
BL
E' D
OC
Decomposition of the IF of a Gerontology journal
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
2000 2001 2002 2003 2004
IMPACT FACTOR YEAR
CIT
ES
PE
R 'C
ITA
BL
E' D
OC
Editorial self citations
Free citations
One can identify and correct for the following types of
strategic editorial behavior
• Publish ‘non-citable’ items• Publish more reviews• Publish ‘top’ papers in January• Publish ‘topical’ papers (with high short term
impact) • Cite your journal in your own editorials• Excessive journal self-citing
Contents
1 Beyond journal impact factor and H-index
2 University rankings have a limited value
3 Combine indicators and peer review
4 Social sciences deserve special attention
5 Indicators can be manipulated
6 Explore usage-based indicators
Analogy Model
Formal use Informal use
(Collections of) publishing authors
(Collections of) users
Citing a document Retrieving the full text of a document
Article User session
Author’s institutional affiliation
User’s account name
Number of times cited Number of times retrieved as full text
Age distribution downloads vs. citations[Tetrahedron Lett, ScienceDirect; Moed, JASIST, 2005]
0
4
8
12
16
20
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
AGE (MONTHS)
%
SD USESCITATIONSDownloads
Citations%
Age (months)
Ageing downloads vs. citations: Two factor vs. single factor model
0.01
0.1
1
10
100
0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 64 68 72 76 80 84 88
Age (months)
Use
s
DownloadsObserved
DownloadsComputed
DownloadsSingularPoints
CitationsObserved
CitationsComputed
%
Age (months)
Downloads
Citations
More downloads more citations
or
More citations more downloads?
Citations lead to downloads[Moed, J. Am Soc Inf Sci Techn, 2005]
1
10
100
1000
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
AGE PAPER A (MONTHS)
DO
WN
LO
AD
S
A
B (B cites A)
C (C cites A and B)
PaperA published
Paper B published;it cites A
Download of A increases
Paper C published;it cites A and B
Downloads and citations relate to distinct phases in
scientific information processing
.... but (many) more cases must be studied
Thank you for your attention!