When evaluating the performance of an organization, a factory, or a regional infrastructure system, we often want to know one thing: how well is it being managed? In efficiency analysis, this is known as technical efficiency—the ability to maximize outputs given a set of inputs.
To measure this, researchers traditionally rely on two distinct toolkits:
- Data Envelopment Analysis (DEA): A non-parametric approach that draws a piecewise frontier around the most successful units. It requires no assumptions about the underlying production function, but it is deterministic. It assumes that every deviation from the optimal frontier is caused by pure operational inefficiency.
- Stochastic Frontier Analysis (SFA): A parametric econometric method that constructs a smooth production frontier (Dong; Xu). SFA explicitly accounts for statistical noise and random luck by splitting the deviation from the frontier into random error and systematic inefficiency (Tang et al.; Yuan et al.; Liang et al.). However, it forces researchers to assume a rigid mathematical shape for the frontier (Peng et al.).
The real world, however, presents a messy combination of both challenges. If a regional airport or a farm underperforms, is it due to poor management, a harsh external economic environment, or simply bad luck? Standard DEA fails here because it lumps environmental advantages and random noise together with internal management skill.
To solve this, a hybrid framework emerged: the Three-Stage DEA model. By combining the non-parametric flexibility of DEA with the noise-filtering capabilities of SFA, this methodology mathematically "levels the playing field," allowing researchers to isolate true internal efficiency.
Stage 1: Identifying Initial Slacks
The process begins by running a traditional, deterministic DEA model using raw input and output data (Jia et al. (2020); Zhang et al.). This stage calculates initial efficiency scores and identifies input slacks.
An input slack represents the excess amount of a resource that a unit consumed relative to an efficient peer on the frontier. For example, if an inefficient firm uses 10 more workers than a frontier firm to produce the same output, the input slack for labor is 10.
However, these initial slacks are "polluted." They reflect a combination of three distinct forces:
- External environmental factors: Conditions outside the manager's control (e.g., regional economic status, government subsidies).
- Random noise: Statistical anomalies, measurement errors, or pure luck.
- Managerial inefficiency: True internal operational waste.
Stage 2: Filtering Noise and Environment via SFA
The core innovation of the three-stage framework lies in Stage 2. Instead of treating input slacks as mere error metrics, the framework treats them as dependent variables in an SFA regression (Jia et al. (2021); Yan et al.).
For each input slack, an SFA model is constructed where observable, exogenous environmental variables act as predictors (Wang). The total error term of this regression is broken into two independent components:
$$\text{Total Slack} = f(\text{Environmental Variables}) + v_i + u_i$$
Where:
- $v_i$ represents random noise (typically modeled as a normal distribution).
- $u_i$ represents managerial inefficiency (typically modeled as a truncated normal distribution).
By estimating this regression, the framework quantifies exactly how much of a unit's resource waste was forced upon it by an unfavorable environment, how much was due to bad luck ($v_i$), and how much was due to genuine mismanagement ($u_i$) (An (2023); Feng et al.).
The Input Adjustment
Once the coefficients are known, the raw input data is adjusted to simulate a uniform external environment for all units. The formula increases the inputs of units operating in favorable environments and decreases (or leaves intact) the inputs of those in harsh environments, while simultaneously stripping away the random noise component ($v_i$).
$$X_{ij}^{\text{adjusted}} = X_{ij} + [\max(f(Z_j)) - f(Z_j)] + [\max(v_{ij}) - v_{ij}]$$
This adjustment creates a simulated dataset where every unit is evaluated as if it faced the exact same external difficulties and experienced the exact same luck.
Stage 3: Re-evaluating True Efficiency
In the final stage, the adjusted inputs and the original outputs are fed back into a DEA model (Zhang et al.; An et al. (2025)). Because the external advantages, disadvantages, and statistical noise have been mathematically scrubbed from the input data, the resulting efficiency scores reflect pure internal managerial performance.
[Stage 1: Raw Data] ---> [Standard DEA] ---> [Extract Input Slacks]
|
v
[Stage 2: SFA Regression] <---------------------------+
Decompose Slacks into:
1. Environmental Effects
2. Random Noise (Luck)
3. Managerial Inefficiency
---> [Generate Adjusted Inputs]
|
v
[Stage 3: Adjusted DEA] ---> [True Internal Efficiency Scores]
Why is this three-stage decomposition necessary? As Zhao et al. (2025) demonstrate in healthcare resource allocation, favorable external conditions often mask severe underlying managerial redundancies. Conversely, excellent managers can appear inefficient in standard models simply because they operate in depressed regions or harsh climates.
By separating context from capability, the three-stage framework provides a robust evaluation tool across diverse sectors, including high-tech innovation performance (Zhao et al. (2026)), environmental governance (Li et al.), and regional transportation networks (Jia et al. (2021)).
Key Foundational Work
An Evaluation of Air Transport Sector Operational Efficiency in China based on a Three-Stage DEA Analysis (2020) — contributed 2 concept elements to this direction.
The Influence of Risk Culture on the Performance of International Joint-Venture Securities (2020) — contributed 1 concept elements to this direction.
An Evaluation of the Impact of Ecological Compensation on the Cross-Section Efficiency Using SFA and DEA: A Case Study of Xin’an River Basin (2020) — contributed 1 concept elements to this direction.