Is network DEA the same as analyzing a transport network?

No. It is an efficiency framework for connected production or service stages, though those stages can include transport activities.

Can it use multiple inputs and outputs?

Yes, that is a central reason for using DEA instead of one simple productivity ratio.

Is there one universal overall-score formula?

No. Different network structures and formulations imply different score definitions, decompositions and interpretations.

AI Glossary / Mathematics and optimization / Intermediate

Network Data Envelopment Analysis

Definition

Network Data Envelopment Analysis, or network DEA, evaluates the relative efficiency of organizations while representing their internal stages and the flows connecting those stages. It extends conventional DEA, which often considers only an organization's overall inputs and outputs. Intermediate products become explicit links, making it possible to investigate performance within the production or service process.

What it means

An organization might turn staff time into prepared work, then turn that prepared work into completed services. Two organizations can have the same overall output per staff hour while using very different internal processes. Network DEA represents comparable organizations as decision-making units and their stages as connected subunits. It benchmarks them against an empirical frontier constructed under stated assumptions. The word network refers to the production structure, not necessarily a computer network or neural network. There are multiple network DEA formulations, with different ways to connect and aggregate stage performance.

How it works

Define the stages, external inputs, final outputs and intermediate links. Collect comparable measurements for a set of decision-making units. Specify assumptions such as returns to scale and whether the model emphasizes reducing inputs, expanding outputs or addressing individual slacks. An optimization model constructs benchmarks while respecting its chosen link relationships. Some formulations coordinate common intermediate valuations; others construct linked production possibilities or use slacks-based measures. Overall and stage scores must be interpreted using that specific formulation. Independently benchmarking departments and multiplying their scores is not automatically a valid network DEA analysis, because their proposed targets may not agree on the intermediate flow.

Three practical examples

A manufacturer compares plants where fabrication produces components that assembly turns into finished products, tracking the component flow between stages.
A service organization distinguishes preparation from delivery, so the same final completion rate can be examined through different internal patterns.
A utility study represents generation, transmission and distribution as linked activities, with resources and outputs defined for each stage. Such analysis requires careful treatment of losses and operating conditions rather than a single undifferentiated efficiency ratio.

Think of it this way

Think of evaluating a relay team by examining both its finish time and what happens at each leg and handoff. Looking inside can reveal where the overall result originates. The analogy has limits: DEA builds a mathematical benchmark from several inputs and outputs, and intermediate flows are quantities subject to model constraints rather than simply times added along a race.

Worked example

Consider two hypothetical units with one input, one intermediate quantity and one final output. Unit A uses 100 resource units to produce 80 prepared items and 40 completed items. Unit B uses 100 resource units to produce 100 prepared items and 40 completed items. Both have overall productivity 40/100 = 0.4. Stage-one productivity is 0.8 for A and 1.0 for B. Stage-two productivity is 40/80 = 0.5 for A and 40/100 = 0.4 for B. Under separate, simple constant-returns, single-input single-output stage comparisons, the relative scores would be 0.8 and 1.0 for A, and 1.0 and 0.8 for B. These calculations illustrate contrasting internal patterns. They are not a complete network DEA solution or permission to multiply stage scores in an arbitrary formulation. A linked model must specify feasible intermediate targets and consistent benchmark assumptions before producing an overall network score.

Illustrative case study

In a fictional project, Westhaven Workshops compares preparation and finishing across similar sites. Its initial whole-site ratios make several sites look identical. Intermediate records reveal that some prepare more items but finish a smaller share, while others prepare fewer and finish a larger share. Analysts consult operations staff to distinguish work-in-progress timing from process inefficiency before constructing a linked model. They test alternative assumptions and retain the stage details beside overall scores. This illustration reports no real ranking or measured improvement.

Common misconceptions and mistakes

An efficient unit is efficient relative to the selected data and model, not necessarily perfect in an absolute sense. More intermediate output is not automatically better if downstream stages cannot use it or if it accumulates as unwanted inventory. Do not compare unlike organizations without accounting for material differences in their tasks or environments. Measurement errors and unusual observations can affect a deterministic frontier. A model can identify benchmark gaps, but those gaps alone do not prove managerial failure or explain the causal effect of a proposed intervention.

Business use

Use network DEA when internal process relationships are central to a benchmarking question and reliable stage-level data exist. Agree on the process map with people who operate it. Align reporting periods so outputs from one stage correspond sensibly to inputs of the next. Document orientation, returns-to-scale assumptions, link treatment and any weight restrictions. Test sensitivity to peer selection and variable definitions. Present targets as model-based reference points for investigation, and check whether staffing, quality and operational constraints make the proposed changes feasible.

Check your understanding

Two units both complete 40 items using 100 resource units. Does that establish identical internal efficiency?

Answer: No. Their intermediate quantities and stage performance can differ substantially, as the worked example shows. A network model can examine those differences under explicit assumptions.

Related terms

data envelopment analysis
decision-making unit
production frontier
intermediate product
returns to scale
slacks-based measure

Learn next

draw a two-stage process and label every external and intermediate flow with units and a reporting period. Compare simple stage ratios first, then study how a chosen network DEA formulation enforces consistent link targets. Distinguish an illustrative ratio from an optimized efficiency score.

Sources

Sources: Lewis and Sexton, Network DEA: Efficiency Analysis of Organizations with Complex Internal Structure, author-uploaded primary text, https://www.researchgate.net/publication/223159166_Network_DEA_Efficiency_analysis_of_organizations_with_complex_internal_structure ; Färe and Grosskopf, Network DEA, https://doi.org/10.1016/S0038-0121(99)00012-9 . Relevant Lewis–Sexton introduction and formulation text was accessible; the latter publisher item was available as a search preview, not full text.

Previous: Net reclassification improvement Next: Network Data Model