a case study on composite indicators - european commission · a case study on composite indicators...
TRANSCRIPT
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
1
Variance-based sensitivity methods
A case study on Composite Indicators
Michaela Saisana
European Commission
Joint Research Centre
Econometrics and Applied Statistics Unit
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
2
Mostly based on …
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
3
Outline
Composite indicators
Motivation and objective of the study
Statistical coherence of two international university rankings
ARWU ranking and THES ranking
Uncertainty analysis
Conclusions
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
4
Composite indicators (1/2)
Composite indicators aggregate individual variables with the aim of capturing
relevant, possible latent dimensions of reality
Most composite indicators are linear, i.e. weighted arithmetic averages. Weights are
understood to reflect variables’ importance in the index (e.g. budget allocation)
Example: university rankings
Very appealing for capturing a university’s multiple missions in a single number
Allow one to situate a given university in the worldwide context
Can lead to misleading and/or simplistic policy conclusions
Caution: Nominal weights are not a measure of variables’ importance
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
5
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
6
Motivation of the study (1/3)
Q: Why nominal weights often do not coincide with the importance of variables?
Example: A dean wants to rank teachers based on ‘hours of teaching’ and
‘number of publications’, adding these two variables up
Assume:
X1 = Hours of teaching
X2 = Number of publications
Y = 0.5 X1 + 0.5 X2
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
7
ρ(x1, x2) =−0.151, σ(x1) = 614, σ(x2) = 116, σ(y) = 162
Estimated 𝑅21= 0.075 and 𝑅2
2 = 0.826
Teachers’ ranking is mostly driven by the number of publications
Motivation of the study (2/3)
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
8
Motivation of the study (3/3)
To obviate this, the dean substitutes the model
with
A professor comes by, looks at the last formula, and complains that publishing is
disregarded in the department …
Y = 0.5 X1 + 0.5 X2
Y = 0.7 X1 + 0.3 X2
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
9
Objective of the study (1/3)
Our suggestion: to assess the quality of a composite indicator using – instead of 𝑅2i
(Pearson product moment correlation coefficient of the regression of y on xi) its
non-parametric equivalent:
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
10
Objective of the study (2/3)
Appealing feature of the first order effect Si (main effect):
it offers a precise definition of importance, that is ‘the expected reduction in variance of the CI that
would be obtained if a variable could be fixed’
it can be used regardless of the degree of correlation between variables
it is model-free, in that it can be applied also in non-linear aggregations
it is not invasive, in that no changes are made to the CI or to the correlation structure of the
indicators
Nominal weights:
Ratios of weights: revealed target relative importance
w1/w2: substitution effect between x1 and x2
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
11
Objective of the study (3/3)
One can hence compare the importance of an indicator as given by the nominal weight
(assigned by developers) with the importance as measured by the first order effect (Si) to test
the index for coherence.
Estimating first order effect:
Non-parametric local linear kernel regression
1) Cross-validation (CV) principle or 2) direct plug-in (DPI) method
Q: What are the limitations of non-parametric kernel regression in a CI context?
Case-studies:
University rankings: ARWU 2008 and THES 2008
Human development: HDI 2009 and HDI 2010
Index of African governance: IAG 2008
Sustainable society index: SSI 2008
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
12
University rankings are used to judge about the performance of university systems, whether
intended or not on by their developers
University rankings: at the forefront of the policy debate
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
13
France:
Creation of 10 centres of HE excellence
The minister of Education set a target to put at least 10
French universities among the top 100 in ARWU by 2012
President has put French standing in these international
ranking at the forefront of the policy debate (Le Monde,
2008).
Italy
No University in the top 100 of the ARWU ranking seen
as failure of the national educational system).
Spain
1 University in the top 200 of the ARWU hailed as a great
national achievement)
University rankings: at the forefront of the policy debate
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
14
An OECD study shows that worldwide university leaders are concerned about
ranking systems with consequences on the strategic and operational decisions they
take to improve their research performance.
(Hazelkorn, 2007)
There over 16,000 HEIs, yet some of the global rankings merely capture the top
100 universities – less than 1%.
(Hazelkorn, 2013)
University rankings: at the forefront of the policy debate
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
15
An extreme impact
What - 2005 THES created a major controversy in Malaysia: country’s top two
universities slipping by almost 100 places compared to 2004.
Why - change in the ranking methodology (not well known fact and of limited comfort)
Impact - Royal commission of inquiry to investigate the matter. A few weeks later, the
Vice-Chancellor of the University of Malaysia stepped down.
University rankings: at the forefront of the policy debate
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
16
University rankings: ARWU 2008
Methodology:
6 indicators; linear aggregation
Best performing institution = 100; score of other institutions calculated as percentage
Weighting scheme chosen by the developers
PROS and CONS
6 « objective » indicators
Focus on research
performance, overlooks
other university missions.
Biased towards hard sciences
intensive institutions
Favours large institutions
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
17
University rankings: THES 2008
Methodology:
6 indicators; linear aggregation
z-score calculated for each indicator
Best performing institution = 100; score of other institutions calculated as %
Weighting scheme chosen by rankers
PROS and CONS Attempt to take into account teaching
quality
Two expert-based indicators: 50% of
total
Subjective indicators: lack of
transparency
Substantial yearly changes in
methodology
Measures research quantity
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
18
Main effects and nominal weights: ARWU 2008
Si’s are more similar to each other than the nominal weights, i.e. ranging between 0.14
and 0.19 (normalized Si’s to unit sum; CV estimates) when weights should either be
0.10 or 0.20
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
19
Main effects and nominal weights: ARWU 2008
Si’s are more similar to each other than the nominal weights, i.e. ranging between 0.14
and 0.19 (normalized Si’s to unit sum; CV estimates) when weights should either be
0.10 or 0.20
2008 ARWU Wi
Wi/Wacademic
performance Si,CV
Si/Sacademic
performance
Si
(unity sum)
Alumni winning Nobel Prize 0.1 1 0.65 0.9 0.14
Staff winning Nobel Prize 0.2 2 0.72 0.9 0.16
Highly cited researchers 0.2 2 0.85 1.1 0.19
Articles in Nature and Science 0.2 2 0.88 1.2 0.19
Articles in Science and Social Sciences Citation Index 0.2 2 0.7 0.9 0.15
Academic performance (size adjusted) 0.1 1 0.76 1.0 0.17
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
20
Main effects and nominal weights: THES 2008
The combined importance of peer-review variables (recruiters and academia) appears larger
than stipulated by developers, indirectly supporting the hypothesis of linguistic bias
The teacher/student ratio, a key variable aimed at capturing the teaching dimension, is much
less important than it should be (normalized Si is 0.09, nominal weight is 0.20)
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
21
Main effects and nominal weights: THES 2008
The combined importance of peer-review variables (recruiters and academia) appears larger
than stipulated by developers, indirectly supporting the hypothesis of linguistic bias
The teacher/student ratio, a key variable aimed at capturing the teaching dimension, is much
less important than it should be (normalized Si is 0.09, nominal weight is 0.20)
Q: Are the Si’s values rescaled to unity sum meaningful?
2008 THES Wi Wi/Wintstudents Si,CV Si/Sintstudents Si (unity sum)
Academic review 0.40 8 0.81 4.8 0.36
Recruiter review 0.10 2 0.54 3.2 0.24
Teacher/student ratio 0.20 4 0.20 1.2 0.09
Citations per faculty 0.20 4 0.41 2.4 0.18
International staff 0.05 1 0.12 0.7 0.05
International students 0.05 1 0.17 1.0 0.08
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
22
Statistical coherence of university rankings
More discrepancy between the nominal weights and the main effects in
THES compared to ARWU
ARWU is less sensitive on the choice of bandwith h than THES
ARWU is more coherent than THES
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
23
1
51
101
151
201
251
301
351
401
451
501Me
dia
n r
an
k (a
nd 9
9%
co
nfid
en
ce in
terv
al) a
cco
un
ting
for
meth
odolo
gic
al u
nce
rtain
ties
Seoul National University
University of Frankfurt
University of Hamburg
University of California-Davis
University of Alaska-
Fairbanks
Hanyang University
54 universities outside the interval (total of 503)
[43 universities in the Top 100]
Uncertainty analysis (ARWU)
Impact of uncertainties on the university rankings
Source: Saisana, D’Hombres,
Saltelli, 2011, Research Policy 40,
165–177
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
24
Uncertainty analysis (THES)
Impact of uncertainties on the university rankings
Source: Saisana, D’Hombres,
Saltelli, 2011, Research Policy 40,
165–177
1
51
101
151
201
251
301
351
401
Me
dia
n r
an
k (a
nd 9
9%
co
nfid
en
ce in
terv
al) a
cco
un
ting
for
meth
odolo
gic
al u
nce
rtain
ties
250 universities outside the interval (total of 400)
[61 universities in the Top 100]
University of California, Santa
Barbara
Stockholm School of Economics
University of st.
Gallen
University of Tokyo
University of
LeichesterUniversity La Sapienza,
Roma
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
25
ARWU versus THES (2008)
1 – Same top10
Universities
2 - Greater variations in
the middle to lower end
of the rankings
3 - Europe is lagging
behind USA
4 – THES favours UK
universities
THES
ranking
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
26
Conclusions
Nominal weights of indicators within a composite do not necessarily coincide with
the importance of a variable
This paper provides a variance-based, non-invasive tool to measure the internal
discrepancy of a composite indicator between nominal weights and effective
importance
Effective tool to:
Avoid associating nominal weights with indicators’ importance
Assess different weighting strategies
Reconsider the aggregation scheme
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
27
1. Paruolo P., Saisana M., Saltelli A., 2013, Ratings and Rankings: voodoo or science?. J Royal Statistical
Society A 176(2).
2. Saisana M., Saltelli A., 2012, JRC audit on the 2012 WJP Rule of Law Index, In Agrast, M., Botero, J.,
Martinez, J., Ponce, A., & Pratt, C. WJP Rule of Law Index® 2012. Washington, D.C.: The World
Justice Project.
3. Saisana M., Philippas D., 2012, Sustainable Society Index (SSI): Taking societies’ pulse along social,
environmental and economic issues, EUR 25578, Joint Research Centre, Publications Office of the
European Union, Italy.
4. Saisana M., D’Hombres B., Saltelli A., 2011, Rickety Numbers: Volatility of university rankings and
policy implications. Research Policy 40, 165–177.
5. Saisana M., Saltelli A., Tarantola S., 2005, Uncertainty and sensitivity analysis techniques as tools
for the analysis and validation of composite indicators. J Royal Statistical Society A 168(2), 307-323.
6. OECD/JRC, 2008, Handbook on Constructing Composite Indicators. Methodology and user Guide, OECD
Publishing, ISBN 978-92-64-04345-9.
References and Related reading
Summer School on Sensitivity Analysis
SAMO2014
Ranco (IT), 24-27/06/2014
28
Forthcoming:
12th JRC Annual Training on Composite Indicators
and Multi-criteria Analysis
22-26 September 2014
http://composite-indicators.jrc.ec.europa.eu