Title: Ottawa, 79 November 2005 http:farmweb.jrc.cec.eu.intci 1
1Weighting issues in the development of a
Composite Indicator Michaela
Saisana michaela.saisana_at_jrc.it European
Commission Joint Research Centre Ispra,
Italy Composite Indicators Workshop Ottawa,
7-9 November 2005
2Prepared with Michela Nardo
Based on Handbook on Constructing Composite
IndicatorsMethodology and User GuideNardo, M.
M. Saisana, A. Saltelli and S. Tarantola
(EC/JRC), A. Hoffman and E. Giovannini (OECD),
OECD Statistics Working Paper JT00188147,
STD/DOC(2005)3.http//www.olis.oecd.org/olis/200
5doc.nsf/LinkTo/std-doc(2005)3
3Step 6. Weighting (and aggregation)
No agreed methodology exists to weight individual
indicators
4Step 6. Weighting (and aggregation)
- Equal weights
- Weights based on statistical models
- Principal component/Factor analysis
- Data envelopment analysis
- Regression approach
- Unobserved components models
- Weights based on opinions participatory methods
- Budget allocation
- Public opinion
- Analytic hierarchy process
- Conjoint analysis
5Equal weights
- Considerations
- Proper normalisation of indicators is needed
- Works well if all dimensions (e.g. economic,
social, environmental, etc.) have the same number
of indicators (e.g. TAI example) - Otherwise, equal weighting implies a higher
weight to the dimension represented by the larger
number of indicators. - Appealing when high correlation of indicators
does not mean redundancy of information in the
composite, i.e. when correlated components
explain different aspects of the picture the
composite aims to capture.
6Weights based on Principal Component Analysis
- Tool to examine thoroughly the relationships
among indicators and statistical dimension of the
dataset. - Two crucial problems
- 1. weights assigned to sub-indicators are based
on correlations which do not necessarily
correspond to the underlying relationships
between the indicators and the phenomena being
measured. - confusion between correlation-redundancy
redundancy implies correlation but the reverse is
not necessarily true - 2. disciplines homogeneity rather than
representing plurality (PCA only applied when
variables are correlated).
7Weights based on Principal Component Analysis
E.g. Environmental Sustainability Index 2005
Conclusions 1. The ESI is a multi-D index and
environmental sustainability is a multi-D
concept. 2. The assumption that if the ESI were
based on a small number of indicators such as the
UNDP - HDI, it would not fully describe all
dimensions of environmental sustainability.
Eigenvalues and scree plot to specify number of
PCs for the 21 indicators
8Weights based on Principal Component Analysis
Environmental Sustainability Index 2005
The higher the loading of an indicator, the more
useful it is for explaining variation in the
direction of the PC.
Conclusions 1. Most indicators load highly on
the first three PCs 2. No indicator has low
loadings on all six PCs, thus none of them is
redundant. 3. Very good interpretation of PCs ?
distinct different aspects of environmental
sustainability captured by each component
9Weights based on Principal Component Analysis
Environmental Sustainability Index 2005
Loadings ? inherent weights
Conclusion Weights based on PCA for the 21
indicators are 1/21. This supports the choice
of equal weights on the indicator level for
calculating the ESI.
10Weights based on Data Envelopment Analysis
Employs linear programming tools to retrieve an
efficiency frontier Two main issues 1.
construction of a benchmark (frontier) 2.
measurement of the distance between countries in
a multi-d framework.
Rearranged from Mahlberg and Obersteiner (2001)
- countries on the frontier (a, b, c) are
classified as the best performing, country d is
the worse performing. - performance indicator d/d
- set of weights of a country depends on its
position with respect to the frontier. - benchmark INTERNAL or EXTERNAL (e.g. Korhonen
et al. 2001, performance indicator of academic
research target in the efficiency frontier
having the most preferred combination of
sub-indicators.
11Table 6.3. BOD approach applied to the TAI
dataset (23 countries). Columns 1 to 8 contain
weights, column 9 displays the countrys
composite indicator.
Weights based on Data Envelopment Analysis
E.g. Technology Achievement Index 2001
The optimal set of weighs (if it exists)
guarantees the best position for the country vis
รก vis all other countries in the sample- any
other weights profile would worsen the relative
position of that country
Notice that 1. weights are country specific 2.
benchmark is country-dependent (no unique
benchmark, unless a country is better-off in all
sub-indicators) 3. indicators must be comparable
(same unit of measurement)
12Weights based on Regression
- Multiple regression models can handle a large
number of indicators. - input indicators (e.g. related to various policy
actions) ? output indicator (e.g. target) - The regression model could
- a. quantify the relative effect of each policy
action on the output - b. be used for forecasting purposes
- In a more general case of multiple output
indicators, canonical correlation analysis
(generalization of multiple regression) could be
applied. -
- However, in any case, there is always the
uncertainty that the relations, captured by the
regression model for a given range of inputs and
output, may not be valid for different ranges.
13Weights based on Unobserved Components
- Unobserved components is similar in spirit to
the multiple regression models, it does not need
an explicit value for the dependent variable as
it treats it like another unknown variable to
estimate. - This advantage is counter-balanced by the
inconvenient of the complexity in estimation and
the computational cost
14Weights based on Budget Allocation
Expert 1
Royalties
Internet
Technology exports
Telephones
Electricity
Patents
Schooling years
University Students
allocate 100 points
5 5 5 20 10 5 20 30
Total
5 5 5 20 10 5 20 30
15Weights based on Budget Allocation
Expert 2
Royalties
Internet
Technology exports
Telephones
Electricity
Patents
Schooling years
University Students
allocate 100 points
15 15 10 15 10 5 20 10
Total
5 5 5 20 10 5 20 30
20 20 15 35 20 10 40 40
Based on an internal JRC survey of 21 experts...
16Weights based on Budget Allocation
- Phases
- 1. Selection of experts for the valuation
- 2. Allocation of budget to the sub-indicators
- 3. Calculation of the weights
- 4. Iteration of the budget allocation until
convergence is reached (optional). - Essential to bring together experts that have a
wide spectrum of knowledge, experience and
concerns ?proper weighting system - Example A case study in which 400 German experts
in 1991 were asked to allocate a budget to
several environmental indicators related to an
air pollution problem showed very consistent
results, in spite of the fact that the experts
came from opposing social spheres like the
industrial sector and the environmental sector
(Jesinghaus in Moldan and Billharz, 1997). - Advantages
- Participatory technique
- Limitations
- Design of the survey - choice of experts
(number, background) - Circular thinking
- Up to 8-10 indicators
- Not transferable from one area to another
17Weights based on Analytic Hierarchy Process
USING PAIRWISE COMPARISONS, THE RELATIVE
IMPORTANCE OF ONE CRITERION OVER ANOTHER CAN BE
EXPRESSED
1 EQUAL 3 MODERATE 5 STRONG 7 VERY STRONG
9 EXTREME
18Weights based on Analytic Hierarchy Process
USING PAIRWISE COMPARISONS, THE RELATIVE
IMPORTANCE OF ONE CRITERION OVER ANOTHER CAN BE
EXPRESSED
1 EQUAL 3 MODERATE 5 STRONG 7 VERY STRONG
9 EXTREME
19Weights based on Analytic Hierarchy Process
USING PAIRWISE COMPARISONS, THE RELATIVE
IMPORTANCE OF ONE CRITERION OVER ANOTHER CAN BE
EXPRESSED
1 EQUAL 3 MODERATE 5 STRONG 7 VERY STRONG
9 EXTREME
Weights
solve for the Eigenvector
Inconsistency 17.4
20Weights based on Analytic Hierarchy Process
Based on an internal JRC survey of 18 experts...
Inconsistencies range from 1.1 - 45.5
(desired lt 10-20 )
21Weights based on Analytic Hierarchy Process
- Advantages
- Participatory technique
- Tolerates inconsistency (part of human thinking)
- Pairwise comparisons can be easily handled by
human mind - Limitations
- More time consuming than budget allocation
- (N.(N-1)/2 pairs for comparison)
- Design of the survey - choice of experts
(number, background) - Recommended for less than 10 indicators
- Not transferable from one area to another
22Weights based on BAL AHP
Budget allocation provides weights that are
closer to the average than Analytic Hierarchy
Process
23Weights based on Conjoint Analysis
- Conjoint analysis derives the worth of the
indicators from the worth of a composite, i.e. it
reverses the process of AHP, with which it shares
advantages and disadvantages. - Further complication is the need to specify and
estimate and utility function
24Table 6.6. Weights for the sub-indicators
obtained using 4 different methods equal
weighting (EW), factor analysis (FA), budget
allocation (BAL), and analytic hierarchy process
(AHP)
Weights based on different weighting methods
E.g. Technology Achievement Index 2001
weights
ranks
() For example USA ranks first according to BOD,
second according to EW, FA, and BAL and third
according to AHP.