Πανεπιστήμιο Μακεδονίας
A novel multi‐level data envelopment analysis variation for the construction of composite indicators and the measurement of sustainability
Abstract
dc:descriptionIn public decision-making factors such as personal values, cultural background and different individual perspectives play a central role in the policy cycle of design, test, implementation and review. To assist policy makers, analysts have used an array of qualitative and quantitative methods to all steps of the cycle.However, the increasing use of sophisticated methods seems not to be always accompanied by an improvement in the quality of policy making; on the contrary, it seems to attract criticism that is focused on their disadvantages. Furthermore, the rise of Artificial Intelligence (AI) and its expanding use in decision- and/or policy- making, has brought forth the issue of interpretability of algorithms and whether their output can be trusted. Questions such as “which specific feature made the model/algorithm reach the specific decision”, hence issues of transparency and interpretability of the methods, are becoming central issues of the critique on quantitative methods and algorithms.This criticism is not without its merits. The complexity of contemporary problems means that there are issues about which an analyst can only make assumptions due to the existence of deep uncertainty. Moreover, in such complexity, the perception of the analyst may limit the view of the policy cycle under study. As a result, the success of a quantitative method relies on all of the above choices to be exactly “correct”.Sustainable development perfectly encapsulates these issues and in order to achieve it, public policies should have economic, social and environmental dimensions, while taking into account the current technological developments, the cultural context and the value system in which they are applied. Thus, sustainable development is a multi- dimensional concept and from early on arose the need to find an appropriate proxy to measure it; that measure was sustainability.In addition to their multi-dimensional nature, both sustainable development and sustainability have been characterized by different perceptions on how to explicitly define them. Complementary to the lack of a unified definition is also the absence of an official and unified methodological framework. Composite indicators have emerged as a suitable means that allow the proper measurement of sustainability.One of the methods that has been proposed and used in the literature for both the measurement of sustainability and the construction of composite indicators is Data Envelopment Analysis. A literature review was performed in the context of the current thesis for the years 2016-2020 and it was discovered that researchers have made an effort to include parameters that represent the social dimension of sustainability (a feature that was missing the previous years), while more dimensions of sustainability are included in recent studies such as technological innovation and advancement, despite the fact that the three-dimensional construct (economy, environment, society) seems to be the preferred one. Moreover, it was discovered that the choice of inputs and outputs (and intermediate measures) despite commonalities is unique to each research work. In addition, the choice of DEA variation and/or combination with other methodologies implies that the perception of each analyst affects the final result of their work.Apart from those identified gaps, the methodology of DEA itself does not come without its own limitations. First, in its traditional form the efficiency of Decision Making Units is calculated with weights that are most favorable to themselves; i.e. each DMU is evaluated under the most favorable weighting scheme with the purpose of maximizing its own efficiency. As a result, the weights that are chosen for one DMU may be completely different from those selected for another. Moreover, DEA needs to be used in the appropriate context, which means that there is the requirement to decide which parameters will best explain different dimensions of sustainability. This is especially important since, the number of inputs and outputs that can be used is limited by the number of DMUs under evaluation for the measurement to be meaningful, otherwise there would be an increased number of efficient DMUs that would result in inconsistencies.As a result, the purpose of the current thesis is to address all the above gaps and more specifically:1) to propose an alternative version of two-stage Data Envelopment Analysis with a different optimization metric that attempts to intervene on the weights of the inputs, intermediate measures and outputs to better reflect their importance for the DMUs by considering positive and negative deviations in the calculations and limiting the distance of these deviations from the maximum and minimum values.2) to propose a computational framework that will attempt to incorporate different perceptions (meaning different combinations of inputs and outputs) and apply it in the measurement of sustainability of the EU 28 countries.To achieve this objective, the framework will rely on Exploratory Modeling and Analysis (EMA). EMA is a school of thought developed at RAND corporation and promotes the exploratory use of quantitative methods despite methodological limitations, uncertainties and different perceptions (meaning different combinations of inputs, intermediate measures and outputs). Employing an exploratory approach to sustainability measurement could reveal unanticipated implications of the initial assumptions regarding inputs and outputs.The thesis is developed in a series of consecutive steps. First, an alternative two-stage DEA model is introduced that employs positive and negative deviational variables both in the objective function (thus altering the optimization metric) and in the constraints. The model attempts to find the best possible weights for the inputs, intermediate measures and outputs, by minimizing the deviations of both the first and second stage of the model. By minimizing simultaneously the deviations of each stage, the efficiencies of both stages are maximized at the same time and no priority is given into which stage should take precedence. Three lemmas and one theorem are proved, and it is proven that the alternative model has a feasible solution that is optimal.The alternative optimization metric, two-stage proposed DEA model is applied in two case studies: one that calculates the environmental performance of European countries and a second to calculate the agricultural sustainability of European countries.Following the definition of the new model, a new computational framework is defined for the construction of composite indicators. The proposed model is used for the calculation of each sub-indicator that the final indicator will consist of. The calculated sub-indicators are then used as parameters in a Benefit-of-the-Doubt (BoD) model that generates the value of the final index. The computational framework is tested two times in the measurement of sustainability of European countries: once with the proposed, alternative, two-stage DEA model and once with the typical two-stage model of Chen et al. (2012).The above calculation of sustainability however is limited by the same notion that was identified in the beginning: since there is no unique, “correct” definition of sustainability, the same indicator can be calculated by using different variations of DEA and/or different combinations of inputs, intermediate measures and outputs.Consequently, there is the need to have an indicator of sustainability that will incorporate all these different perceptions that may arise, where perceptions mean different DEA variation and/or different combination of inputs, intermediate outputs and outputs. The proposed computational framework is based on this principle, and it consists of the following steps:Step 1: Define different perceptions of sustainability and for each perception:a) define how many sub-indicators will be entailed in this perception’s sustainability indexb) define the inputs, intermediate measures and outputs that each sub-indicator will entailc) Repeat for all perceptionsStep 2: Define the variation of DEA that will calculate the value of the sub-indicatorsa) calculate the sub-indicatorsb) calculate the perception’s sustainability index using DEA modelsc) Once all sustainability indices for all perceptions are calculated, calculate the mean value for each country/DMUStep 3: Use machine learning to gain insights into the sustainability of each country under different perceptionsThe proposed computational framework is used with four different versions of two- stage DEA models, and different combinations of inputs, intermediate measures and outputs to the calculation of sustainability of European countries.The final step of the proposed computational framework is to use Machine Learning techniques in the results of the generated computations with the purpose of revealing insights into how the sustainability of countries behaves under different perceptions.Following the logic of EMA, several techniques will be employed in an effort to mitigate intrinsic methodological limitations and find the common, emergent elements that remain robust despite the different methods.The first insights will be revealed by using clustering techniques and more specifically K-Means and Density based clustering (DBSCAN). For the clustering algorithms, the values of the sub-indicators along with those of the sustainability indices under all the computational regimes were used.For the current thesis, three additional techniques were used: Classification and Regression Decision Trees (CART), Random Forests and Boosting Regression.Classification and Regression Decision Trees (CART) since they are not computationally costly, they can be used as communication tools to non-experts and offer deep interpretational capabilities. However, CART trees tend to overfit the data to their training set and are considered weak learners and for that reason two additional ML techniques will be used: Random Forests and boosting regression.Random forests train trees independently using random samples of the available data and the sampling happens with bootstrapping both the sample and the features at every repetition. As a result, they tend to be slower than CART trees, but the generated results are more robust and tend to avoid the pitfalls of overfitting. More specifically, with the random forests 80% of the data will be used for training and the remaining will be used for prediction. Furthermore, for each data row (point) of the remaining data, the contribution of the individual features to the predicted value will be calculated. The average of all the contributions will be plotted in a boxplot to reveal insights on how individual sub-indicators affect the value of the sustainability index.Similarly, boosting regression is also considered a slow learner, but compared to random forests, each tree is generated using information from previous ones. Moreover, the technique will also reveal the relative influence of the individual sub- indicator to the index of sustainability, which could provide further insights into the analysis of the results. Both random forests and boosting regression are more robust than CART trees, but this robustness comes at the detriment of intuitive communication capabilities that are the main characteristic of CART trees. Consequently, the use of all three Machine Learning techniques will limit the methodological weaknesses of each method, while providing results and insights that are robust and independent of the used technique.The final results illustrated that a balance among the performance of various dimensions can be a good policy to achieve sustainable development and when the inclusion of all DEA variations does not alter significantly the mean value of sustainability then the trust in the results increases, thus making them robust.Finally, the blend of DEA with machine learning (applied on the results of DEA for the various scenarios) revealed insights on the areas that policy makers could direct investments to increase sustainability. In addition, the ML applications contributed in the identification of the most important features of sustainability for the various countries something that could have direct implications in the area of EU policy making: for example, countries that share similar features that drive the behavior of sustainability could be grouped together in clusters and policies, laws, regulations etc. could be adapted to those clusters in order to boost the particular features that would increase their sustainability. As a result, policy making has the potential to become customized (adapted to the specifics of each group) without missing its overall and principal theme of pursuing sustainable development. This adaptive and adaptable policy making could greatly be of assistance especially when new countries are negotiating their entry to the Union; based on the features that affect the sustainability of the new countries, they could follow the regulations and laws of the appropriate cluster. Finally, the inclusion of new layers and perceptions renders the algorithms more inclusive and participatory, increasing their transparency, thus improving the trust to the final results.
Degree
thesis:*- Grantor dc:publisher
- Πανεπιστήμιο Μακεδονίας
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsaples, Georgios
- Τσαπλές, Γεώργιος
Subjects
dc:subject × 22- Περιβάλλουσα Ανάλυση Δεδομένων (ΠΑΔ)
- Αειφορία
- Αειφόρος ανάπτυξη
- Σύνθετοι δείκτες
- Ευρωπαϊκές χώρες
- Data envelopment analysis
- Sustainability
- Sustainabile Development
- Composite indicators
- European Countries
- Μαθηματικά
- Φυσικές Επιστήμες
- Μοντελοποίηση και Προσομοίωση
- Περιβαλλοντική Μηχανική
- Επιστήμες Μηχανικού και Τεχνολογία
- Ανανεώσιμη ενέργεια, Αειφορία και Περιβάλλον
- Mathematics
- Natural Sciences
- Modeling and Simulation
- Environmental Engineering
- Engineering and Technology
- Renewable Energy, Sustainability and the Environment
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/52269