orange3-dea 1.1.0

Data Envelopment Analysis add-on for Orange3

Orange3-DEA — Data Envelopment Analysis Add-on

An Orange3 add-on for Data Envelopment Analysis (DEA), inspired by the deaR-Shiny web application (Benítez, Coll-Serrano & Bolós, 2021). This is a full reimplementation of the original prototype with corrected solvers, background execution, and new widgets.

Widgets

Widget	Description
DEA Datasets	Built-in datasets from published papers; panel datasets are sent as a DEA Panel, fuzzy datasets on a dedicated channel. Loads automatically on selection.
DEA Setup	Select the DMU identifier, inputs and outputs. Settings persist by variable name; the Filtered Data output contains exactly the rows used in the analysis.
DEA Model	Choose and run a model. Runs in a background thread (the GUI stays responsive), re-runs automatically on new input, and can attach scale efficiency (CRS/VRS/SE + RTS class).
DEA Results	Sortable tabs: efficiency (with rank and peer counts), slacks, targets (with % change), lambdas, reference sets, cross-efficiency matrix, bootstrap intervals. Sends four Orange Tables.
DEA Plot	Efficiency bars, summary, histogram, peer frequency, and a production-frontier plot (1 input × 1 output) with a zoom/save toolbar.
DEA Malmquist (new)	Malmquist productivity index over a DEA Panel or two DEA Problems: MI = EC × TC, with optional pure/scale decomposition (PEC × SEC).
DEA Window (new)	Window analysis (Charnes et al. 1985): efficiency inside moving windows of consecutive periods, with per-DMU stability summary (mean, SD, min, max).
DEA Context (new)	Context-dependent DEA (Seiford & Zhu 2003): efficiency tiers by frontier peeling, attractiveness and progress scores at a chosen tier distance.
DEA Tests (new)	Banker (1993) RTS tests (exponential and half-normal), plus group-difference tests on efficiency scores: Mann-Whitney, Kolmogorov-Smirnov, Li (1996) with bootstrap p-value, Kruskal-Wallis.
DEA Two-Stage (new)	Simar-Wilson (2007) double bootstrap: truncated regression of DEA scores on environmental covariates with bias correction and valid confidence intervals.
DEA Network (new)	Serial two-stage network DEA: overall = stage 1 × stage 2 (Kao-Hwang 2008) or weighted mean (Chen et al. 2009, CRS/VRS); assign inputs, intermediates and outputs directly from a data table.
DEA Economic (new)	Cost, revenue and profit efficiency with prices; Farrell decomposition (economic = technical × allocative), Nerlovian profit inefficiency.
DEA Conditional (new)	Conditional order-m (Daraio-Simar 2005): kernel-weighted peers in environmental variables; Q = conditional/unconditional diagnoses the environment's effect nonparametrically.
DEA Metafrontier (new)	Group frontiers vs. pooled metafrontier with technology gap ratios (O'Donnell et al. 2008).
DEA Multiverse (new, methodological contribution)	Specification-curve analysis for DEA: runs ~23 defensible specifications (model × orientation × RTS × convexity × frontier type), reports rank distributions with a boxplot, consensus ranking, per-DMU robustness and efficient share, Kendall's W agreement, and an η² decomposition of which methodological choice drives the results. See METHOD_MULTIVERSE.md.
DEA PCA (new)	PCA-DEA (Adler-Golany 2001/2002): principal components per block (standardized, sign-fixed, translated to positive), DEA on components; restores discrimination with many variables. Shows loadings and rank agreement with the full model.
DEA Common Weights (new)	CCA-based common weights (Sinuany-Stern et al. 1994; Friedman & Sinuany-Stern 1997): one weight vector for all DMUs → complete ranking; canonical correlation and Spearman agreement with CCR reported.
DEA Cluster (new)	Discover technology groups by clustering DMUs — on input-output profiles, on the λ (peer) structure ("benchmark communities"), or on multiverse rank profiles; k-means/Ward, automatic k by silhouette; feeds groups straight to DEA Metafrontier.
DEA Map (new)	2-D MDS map of DMUs (Co-Plot spirit, Adler & Raveh 2008) colored by efficiency with efficient units starred; distances from profiles, cross-efficiency appraisal profiles, or multiverse rank profiles; Kruskal stress reported.
DEA Fuzzy (new)	Kao-Liu fuzzy DEA GUI: efficiency intervals per α-cut (table + interval plot); consumes the Fuzzy DEA Problem output of DEA Datasets.
DEA SFA (new)	Stochastic frontier (Cobb-Douglas, half-normal MLE) as a parametric benchmark: elasticities, γ diagnostics, DEA-vs-SFA scatter with Spearman ρ.
DEA Export (new)	Any result table → publication-ready booktabs LaTeX or CSV (preview, clipboard, save).

Example workflows are in examples/ (basic analysis, panel productivity, undesirable outputs, research/multiverse); see USER_GUIDE.md for recommended analysis paths and CHANGELOG.md for version history.

Implemented models

Category	Model	Notes
Radial	CCR / BCC	Two-phase (slacks maximized at the optimal radial factor), so weakly efficient DMUs are identified correctly; CRS/VRS/NIRS/NDRS
Non-radial	Additive	Weights: ones / MIP / RAM; score = total slack (0 = efficient), RAM score in [0, 1]
Non-radial	SBM (Tone 2001)	Non-oriented, input- and output-oriented
Non-radial	Russell	Per-input/per-output factors
Directional	DDF (Chambers et al. 1996)	β = 0 means efficient
Ranking	Radial super-efficiency	Infeasible LPs under VRS reported explicitly
Ranking	Super-SBM (Tone 2002)	Only SBM-efficient DMUs re-scored
Ranking	Cross-efficiency	Arbitrary / benevolent / aggressive secondary goals (Doyle & Green 1994)
Non-convex	FDH	Enumeration
Inference	Bootstrap (Simar & Wilson 1998)	Smoothed bootstrap with reflection, bias correction, CIs, reproducible seed
Panel	Malmquist (Färe et al. 1994)	EC/TC + optional PEC/SEC decomposition; consistent for both orientations
Fuzzy	Kao & Liu alpha-cuts	Correct per-DMU best/worst scenarios (API only)
Extras	Scale efficiency	SE = TE(CRS)/TE(VRS) with IRS/DRS/CRS classification
Environmental	Seiford-Zhu (2002)	Bad outputs translated and treated as goods (VRS); mark undesirable outputs in DEA Setup
Environmental	Environmental DDF (Chung et al. 1997)	Weak disposability: expands good outputs, contracts bad outputs
Weights	Multiplier form + assurance regions	Ratio bounds on input/output weights, text syntax a / b in [lo, hi]; lambdas recovered from LP duals
Robust	Order-m (Cazals-Florens-Simar 2002)	Partial frontier, insensitive to outliers; Monte Carlo with seed
Panel	Window analysis	Moving-window pooled frontiers
Benchmarking	Context-dependent DEA	Tier stratification, attractiveness/progress
Inference	Banker (1993) RTS tests	Exponential F(2n,2n) and half-normal F(n,n) variants
Inference	Li (1996) test	Kernel density equality with bootstrap p-value; plus MW/KS/KW
Inference	Simar-Wilson (2007) Algorithm 2	Double-bootstrap truncated regression (L1 bias correction, L2 coefficient CIs)
Network	Kao-Hwang (2008), Chen et al. (2009)	Two-stage serial network with exact multiplicative/additive decomposition
Economic	Cost / revenue / profit efficiency	Farrell decomposition with per-DMU or common prices
Panel	Global Malmquist (Pastor-Lovell 2005)	Circular index over a pooled global frontier (TC column = best-practice gap change)
Panel	Malmquist-Luenberger (Chung et al. 1997)	Green productivity with weak disposability of bad outputs
Panel	Bootstrap CIs for Malmquist	Pairs bootstrap over the panel (approximate; seeded)
Robust	Order-α (Daouia-Simar 2005)	Quantile frontier; α = 1 reproduces FDH; in DEA Model
Robust	Conditional order-m (Daraio-Simar 2005)	Kernel-weighted peers in Z; Q diagnostics
Screening	Wilson (1993) outliers	Leave-one-out super-efficiency flags; in DEA Tests
Groups	Metafrontier + TGR	TE_meta = TE_group × TGR
Meta	DEA Multiverse	Specification-space analysis: consensus ranks, robustness, Kendall's W, η² factor influence (see METHOD_MULTIVERSE.md)
Multivariate	PCA-DEA (Adler-Golany)	Block-wise PCA with translation; discrimination diagnostics
Multivariate	NMF-DEA	Non-negative factorization per block: no translation needed, archetypal profiles; same widget as PCA-DEA
Multivariate	CCA common weights	First canonical pair as common weights; complete ranking
Multivariate	Cluster-DEA	Profiles / peer-structure / multiverse-rank clustering with silhouette and auto-k; pipeline to metafrontier
Multivariate	MDS efficiency map	Classical MDS with stress-1; multiverse-rank distances are a novel option
Inference	Tobit second stage	Censored MLE with SEs and p-values, offered alongside Simar-Wilson in DEA Two-Stage (with the methodological caveat)

What was fixed relative to the prototype

• Radial model now runs a second phase, so "efficient" means strongly (Pareto) efficient; targets are computed from the phase-2 lambdas.
• Additive model no longer reports a meaningless 0/1 score; NIRS/NDRS constraints are no longer silently dropped (additive, SBM, super-SBM).
• DDF had an inverted sign on the input direction; fixed.
• "Additive-min" (which always returned 1.0 by construction) was removed.
• Cross-efficiency now solves the Doyle–Green secondary-goal LP instead of relying on arbitrary epsilon bounds.
• Malmquist handles the output orientation via proper distance functions (1/φ), so MI > 1 always means productivity growth.
• Bootstrap is a genuine Simar–Wilson smoothed bootstrap (was: naive resampling), with a random seed for reproducibility.
• Kao–Liu fuzzy intervals evaluate each DMU at its own best/worst scenario (was: one common scenario for all DMUs — incorrect).
• The DEA Setup widget's Filtered Data output was not filtered; fixed, along with settings restore bugs and index-based DMU selection.
• Long runs no longer freeze the GUI (background threads in DEA Model and DEA Malmquist).

Installation

pip install -e .          # from this directory

Then launch Orange Canvas — the DEA category appears in the toolbox.

Typical workflow

File / DEA Datasets → DEA Setup → DEA Model → DEA Results → DEA Plot
                    ↘ (DEA Panel) → DEA Malmquist

Programmatic use

from orangecontrib.dea.core import DEAProblem, run_dea

problem = DEAProblem(dmu_names, input_names, output_names, X, Y)
res = run_dea(problem, model="Basic Radial (CCR/BCC)",
              orientation="input", rts="VRS")
print(res.efficiency, res.efficient, res.reference_set(0))

Tests

pytest orangecontrib/dea/tests

65 tests cover closed-form CCR scores, model orderings (CCR ≤ NIRS ≤ BCC, SBM ≤ radial, Russell ≤ radial, FDH ≥ VRS, order-m ≥ FDH), weak-efficiency detection, super-efficiency, cross-efficiency consistency, Malmquist identities, bootstrap reproducibility, fuzzy interval nesting, multiplier/envelopment equivalence, weight-restriction monotonicity, weak-disposability targets, window/context properties, Li-test power and size, recovery of a known covariate effect in the Simar-Wilson two-stage model, the Kao-Hwang product identity, network ≤ black-box efficiency, Farrell decompositions, global-Malmquist circularity and identity, Malmquist-Luenberger identity, order-α = FDH at α = 1 with monotonicity, conditional order-m reproducibility, outlier flagging, and TGR ≤ 1 with TE_meta = TE_group × TGR.

References

• Charnes, Cooper & Rhodes (1978). EJOR 2(6), 429–444.
• Banker, Charnes & Cooper (1984). Management Science 30(9), 1078–1092.
• Tone (2001, 2002). EJOR 130(3), 498–509; 143(1), 32–41.
• Doyle & Green (1994). JORS 45(5), 567–578.
• Färe, Grosskopf, Norris & Zhang (1994). AER 84(1), 66–83.
• Simar & Wilson (1998). Management Science 44(1), 49–61.
• Simar & Wilson (2007). Journal of Econometrics 136(1), 31–64.
• Kao & Liu (2000). Fuzzy Sets and Systems 113(3), 427–437.
• Seiford & Zhu (2002). EJOR 142(1), 16–20. (undesirable outputs)
• Seiford & Zhu (2003). EJOR 151(2), 411–420. (context-dependent)
• Chung, Färe & Grosskopf (1997). J. Environ. Manage. 51(3), 229–240.
• Cazals, Florens & Simar (2002). J. Econometrics 106(1), 1–25.
• Banker (1993). Management Science 39(10), 1265–1273.
• Li (1996). Econometric Reviews 15(3), 261–274.
• Charnes, Clark, Cooper & Golany (1985). Annals of OR 2, 95–112. (window)
• Benítez, Coll-Serrano & Bolós (2021). Sustainability 13(12), 6774.

License

MIT