📊 statistics

Statistical inference for uncertain single-index models with weather application

This paper proposes and validates two types of uncertain single-index models using semiparametric least-squares estimation with kernel and B-spline methods to effectively handle complex, imprecise data, as demonstrated through simulation studies and a practical weather application.

Original authors: Fuguo Wang, Zhiming LI

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fuguo Wang, Zhiming LI

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the messy reality of the natural world, data is rarely perfect. When scientists measure the weather, track a disease, or monitor a financial market, they often encounter information that is fuzzy, incomplete, or based on human judgment rather than a precise instrument reading. Traditional statistical tools, which rely on the rigid laws of probability, sometimes struggle to make sense of this kind of uncertainty. They can lose important details or produce misleading conclusions when the numbers themselves are not sharp. To handle this, researchers have developed a different mathematical framework called uncertainty theory. Instead of asking how likely an event is to happen based on past frequency, this approach asks how confident an expert is that an event will occur, treating that confidence as a measurable quantity. This shift allows for a more flexible way to model the world when the inputs are vague.

Building on this foundation, a team of researchers at Xinjiang University has created a new way to analyze complex relationships where the data is imprecise. They focused on a specific type of model known as a single-index model. Imagine trying to predict the temperature of a city. You might have five different factors to consider: rain, wind speed, humidity, air pressure, and cloud cover. Instead of trying to figure out a separate, complicated rule for how each of these five factors interacts with the others, a single-index model combines them into one single score. This score acts as a summary, a single number that captures the combined influence of all the variables. The researchers then use this score to predict the outcome, allowing the relationship between the score and the result to be a smooth, flexible curve rather than a rigid straight line. This approach is powerful because it simplifies high-dimensional problems without throwing away the ability to understand what is driving the changes.

The researchers developed two distinct versions of this model to handle different kinds of data. The first version is designed for situations where the input factors, like wind speed or pressure, are known precisely, but the outcome, such as the daily temperature range, is uncertain. The second version is for when both the inputs and the outcomes are fuzzy or imprecise. To make these models work, the team had to invent new methods to find the best possible curve that fits the data. For the first model, they used a technique that looks at the local neighborhood of data points to estimate the shape of the curve, carefully adjusting how much weight to give to nearby points versus distant ones. For the second model, where everything is uncertain, they used a method involving smooth, flexible curves made of connected polynomial pieces. This approach allowed them to enforce a crucial rule: the relationship must be consistent, meaning that as the combined score goes up, the predicted outcome should not suddenly jump up and down in a chaotic way.

To prove their methods worked, the team ran extensive computer simulations. They generated thousands of fake datasets that mimicked the behavior of uncertain variables and tested whether their new models could recover the hidden patterns they had built in. The results were successful; the models accurately identified the underlying relationships and the correct weights for each variable. They also developed a way to check if the models were a good fit by analyzing the leftover errors, or residuals, to ensure they behaved as expected. This process is similar to checking the noise in a recording to see if the microphone is working correctly. The simulations confirmed that their approach could handle the fuzziness of the data without breaking down, providing reliable estimates even when the inputs were not exact numbers.

The true test of their work came when they applied the model to real-world weather data from Lagos, Nigeria. They gathered 535 days of observations, looking at how precipitation, wind, humidity, air pressure, and cloud cover influenced the daily temperature range. By feeding this data into their new model, they discovered that humidity was by far the most dominant factor, carrying a weight of nearly 0.90 compared to the others. This finding makes physical sense for a tropical coastal city, where moisture in the air plays a massive role in regulating heat. The model also revealed a fascinating non-linear pattern. When the combined score of the weather variables was low, the temperature rose as the score increased, reflecting a dry, sunny season where the sun heats the ground directly. However, once the score crossed a certain threshold, the relationship shifted. In the wet season, with high humidity and heavy cloud cover, the temperature actually declined as the score increased. This shift happens because the clouds block the sun and the evaporation of rain cools the air, changing the energy balance of the environment.

The researchers did not stop at just finding the pattern; they also tested whether the model was statistically sound. They checked if the small coefficients they found for wind and rain were truly insignificant or just a fluke of the data. Using a specialized test designed for uncertain data, they confirmed that the wind speed factor was indeed significant, even though its weight was small. They also compared their flexible, unknown-curve model against several rigid models that forced the data into specific shapes, like straight lines or simple curves. The flexible model fit the weather data much better, with significantly lower errors, proving that the real-world relationship was too complex for simple formulas. This work demonstrates that by embracing uncertainty rather than fighting it, scientists can build more accurate and insightful models for complex systems, from weather forecasting to any field where human judgment and imperfect measurements play a role.

Technical Summary: Statistical Inference for Uncertain Single-Index Models with Weather Application

1. Problem Statement

The paper addresses the challenge of modeling complex, high-dimensional data where observations are inherently imprecise due to frequency instability and subjective expert judgment. While traditional single-index models (SIMs) effectively mitigate the "curse of dimensionality" for crisp data within a probabilistic framework, they are ill-suited for imprecise data. Applying probability-based SIMs to uncertain data can lead to information loss, estimation bias, and erroneous statistical inferences.

Although uncertainty theory (Liu, 2007, 2010, 2026) provides a robust framework for handling imprecise data through uncertainty measures and distributions, systematic research on single-index models within this theoretical framework remains limited. Existing uncertain regression studies have largely focused on parametric or nonparametric models, leaving a gap in semiparametric single-index modeling that combines dimensionality reduction with the flexibility to handle uncertain explanatory and response variables.

2. Methodology

The authors propose a statistical inference framework for two distinct types of uncertain single-index models, utilizing the profile least-squares principle combined with specific estimation techniques for different data scenarios.

2.1 Model Definitions

The paper defines two models based on the nature of the explanatory vector $x$ :

USIC Model (Uncertain Single-Index Model with Crisp explanatory vector):
$\tilde{y} = g(\beta^T x) + \epsilon$
Here, $x$ is a crisp (precise) vector, while the response $\tilde{y}$ and disturbance $\epsilon$ are uncertain variables.
USIU Model (Uncertain Single-Index Model with Uncertain explanatory vector):
$\tilde{y} = g(\beta^T \tilde{x}) + \epsilon$
Here, both the response $\tilde{y}$ and the explanatory vector $\tilde{x}$ are uncertain variables.

2.2 Estimation Methods

The authors employ a two-step iterative procedure to estimate the unknown link function $g$ and the index coefficients $\beta$ (where $\|\beta\|=1, \beta_1 > 0$ ).

For the USIC Model (Crisp $x$ ):
- Estimator: An Uncertain Nadaraya-Watson (N-W) kernel estimator is used to approximate $g$ .
- Optimization: The minimization of the expected squared error is reduced to minimizing the squared difference between the expected response and the estimated function.
- Kernel Selection: The derivative of the standard uncertain normal distribution is selected as the kernel function due to its favorable theoretical properties.
- Bandwidth Selection: An optimal bandwidth $h$ is determined using V-fold cross-validation, optimized via a Fibonacci search algorithm.
- Residual Analysis: Residuals are defined, and their expectation and variance are estimated to assess model fit.
For the USIU Model (Uncertain $\tilde{x}$ ):
- Estimator: The B-spline method is employed. The N-W estimator is deemed unsuitable here because it lacks an explicit functional form, making it difficult to enforce the strict monotonicity constraints required when both variables are uncertain.
- Monotonicity Constraint: Theorem 2 establishes that if $g$ is strictly increasing, the inverse uncertainty distribution of the error term can be derived explicitly. The B-spline approximation $\hat{g}(t) = \delta^T(t)b$ is constrained such that its derivative remains positive.
- Optimization: The least squares problem is transformed into an integral form involving inverse uncertainty distributions. The L-BFGS algorithm is used to optimize parameters and B-spline coefficients.
- Knot Selection: Knots are placed based on the expected values of the linear combination of uncertain variables, with boundary knots replicated to ensure the curve starts and ends at the boundaries.

2.3 Statistical Inference

Hypothesis Testing: The paper utilizes uncertain hypothesis tests (Ye and Liu, 2022, 2023) to:
1. Assess the adequacy of the disturbance term's distribution (e.g., testing if residuals follow a specific normal uncertain distribution).
2. Perform significance tests on individual coefficients (e.g., testing $H_0: \beta_k = 0$ ) to determine if a variable contributes significantly to the model.
Prediction: Forecasting involves calculating the expected value of the response and constructing confidence intervals based on the inverse uncertainty distribution of the predicted variable.

3. Key Contributions

The paper outlines three primary contributions:

Model Establishment: It establishes two types of uncertain single-index models (USIC and USIU) to provide a flexible tool for nonlinear modeling of high-dimensional imprecise data.
Estimation and Testing Procedures: It proposes specific estimation methods (Uncertain N-W kernel for crisp inputs; B-spline with monotonicity constraints for uncertain inputs) and derives explicit forms for inverse distributions. It further develops hypothesis testing procedures for residual analysis and coefficient significance.
Validation and Application: It validates the models through numerical simulations and demonstrates their practical effectiveness using a real-world weather dataset from Lagos, Nigeria.

4. Results

4.1 Numerical Simulations

Example 1 (USIC): Using 500 simulated samples with crisp inputs and uncertain outputs, the proposed method successfully recovered the true index coefficients ( $\beta \approx (0.2, -0.4, 0.9)^T$ ) and the link function. The residual analysis confirmed that the disturbance term followed the assumed normal uncertain distribution, and the model provided a good fit.
Example 2 (USIU): Using 50 samples with fully uncertain inputs and outputs, the B-spline method successfully estimated the link function and coefficients. The monotonicity constraint was satisfied, and the hypothesis test confirmed the model's validity. The method effectively handled the complexity of uncertain explanatory variables.

4.2 Weather Application (Lagos, Nigeria)

The model was applied to a dataset of 535 daily weather observations (Jan 2023 – July 2024) to predict the daily temperature range ( $y$ ) based on precipitation, wind speed, humidity, sea-level pressure, and cloud cover ( $x_1$ to $x_5$ ).

Coefficient Estimation: The estimated index coefficients identified humidity ( $\beta_3 = 0.8952$ ) as the dominant factor, followed by sea-level pressure and cloud cover.
Nonlinear Dynamics: The estimated link function revealed a unimodal pattern. Below a threshold ( $z \approx -1.3$ ), temperature rises with the index (dry season, radiation-dominated). Above the threshold, temperature declines (wet season, latent heat-dominated).
Significance Testing: A significance test on the wind speed coefficient ( $\beta_2 \approx 0.0381$ ) concluded that despite being close to zero, it is statistically significant in the model context.
Model Comparison: When compared against four parametric models (identity, quadratic, logarithmic, exponential), the proposed uncertain single-index model with an unknown link function achieved the lowest residual variance ( $\sigma^2 = 0.9765$ ), demonstrating superior flexibility and fit.

5. Significance and Claims

The paper claims that its proposed framework offers a novel and powerful tool for handling high-dimensional imprecise data where traditional probabilistic models fail. By integrating uncertainty theory with semiparametric single-index structures, the authors provide a systematic approach that:

Retains the interpretability of parametric models while capturing complex nonlinear relationships.
Effectively manages the "curse of dimensionality" in uncertain environments.
Provides rigorous statistical inference tools (residual analysis, hypothesis testing) specifically tailored for uncertain data.

The authors maintain a modest tone regarding future work, suggesting that the current framework should be extended to uncertain partially linear single-index models and uncertain varying-coefficient single-index models to further enhance practical applicability. They also note the need for further comparative studies of estimation methods across various scenarios.