Local Polynomial Estimation for Sensitivity Analysis on Models With Correlated Inputs

Sébastien da Veiga, Delphine Sinoquet, Fabrice Gamboa

TL;DR

This paper introduces two novel methods for estimating sensitivity indices in complex models, specifically addressing the challenge of correlated input variables. By employing local polynomial regression, these techniques offer a more robust and computationally efficient alternative to existing methods, particularly when dealing with real-world kinetic models.

What Problem Does This Paper Address?

In many scientific and engineering fields, models are used to predict the behavior of complex systems. These models often have numerous input parameters, and understanding how variations in these inputs affect the model's output is crucial. This is the domain of sensitivity analysis. Traditional methods for calculating sensitivity indices, like variance-based approaches (e.g., Sobol's method, FAST), rely on the assumption that input parameters are independent. However, in many practical scenarios, such as kinetic modeling in chemical processes, input parameters are inherently correlated. Neglecting these correlations can lead to misleading interpretations of sensitivity indices, making it difficult to identify the truly influential parameters. Existing methods that attempt to handle correlated inputs are either computationally expensive, require specific knowledge of input distributions, or are limited in their applicability. This paper aims to bridge this gap by providing a method that is both theoretically sound and practically applicable to models with correlated inputs.

Methods

The core of this paper's contribution lies in its use of local polynomial regression to estimate the conditional moments (expectation and variance) of the model's output with respect to its inputs. The authors propose two estimators, denoted as $\hat{S}_1$ and $\hat{S}_2$, which leverage these estimated conditional moments to compute first-order sensitivity indices.

Here's a breakdown of the approach:

  • Local Polynomial Regression: Instead of fitting a global model to the entire dataset, local polynomial regression fits a polynomial to data points in a local neighborhood around a specific input value. This allows for flexible modeling of complex relationships without strong global assumptions. The paper uses local linear regression for estimating the conditional expectation and a similar approach for the conditional variance.
  • Two-Sample Approach: The methodology requires two distinct samples:
  • A sample of input-output pairs $(X_k, Y_k)$ used to estimate the conditional moments $m(X) = E(Y|X)$ and $\sigma^2(X) = Var(Y|X)$. This sample necessitates running the model.
  • A second sample of input values $(X_j)$ used to compute the sensitivity indices based on the estimated moments. This sample does not require running the model, making it computationally efficient.
  • Estimators $\hat{S}_1$ and $\hat{S}_2$:
    • $\hat{S}_1$ estimates $Var(E(Y|X)) / Var(Y)$.
  • $\hat{S}_2$ estimates $1 - E(Var(Y|X)) / Var(Y)$. Both estimators are derived from the estimated conditional moments.

  • Bandwidth Selection: The performance of local polynomial regression is sensitive to the choice of bandwidth (smoothing parameter). The authors utilize the empirical-bias bandwidth selector (EBBS) for selecting these parameters, which is noted to be efficient on simulated data.

Datasets

The paper does not introduce new benchmark datasets. Instead, it uses:

  • Analytical Examples: Two synthetic models are constructed to demonstrate the theoretical properties and practical performance of the proposed estimators, particularly in scenarios with correlated inputs. These examples are designed to highlight the limitations of existing methods when correlations are ignored.
  • Practical Example (Chemical Field): The method is applied to a kinetic model for the isomerization of normal butane. This real-world case involves a system with multiple outputs and a parameter vector with known correlations, making it a suitable testbed for the proposed techniques.

Key Insights and Ideas

  • The Danger of Ignoring Correlations: The analytical examples starkly illustrate how neglecting input correlations can lead to completely erroneous conclusions about parameter importance. A parameter that appears insignificant when correlations are ignored might be highly influential in reality.
  • Local Polynomials as a Flexible Tool: The paper demonstrates that local polynomial regression is a powerful and adaptable technique for estimating conditional moments, even when input distributions are unknown or complex. This flexibility is key to handling correlated inputs.
  • Computational Efficiency: The two-sample approach is a significant advantage. By separating the model evaluation step from the sensitivity index calculation, the method becomes much more efficient, especially for computationally expensive models.
  • Bridging the Gap: The proposed method offers a practical compromise between methods that require extensive model runs (like some Monte Carlo approaches) and those that rely on strong assumptions about input distributions or analytical tractability of integrals.

Key Results

  • The proposed estimators $\hat{S}_1$ and $\hat{S}_2$ are shown to have good theoretical properties, including asymptotic unbiasedness and consistency.
  • Numerical simulations and the practical example demonstrate that the estimators provide accurate sensitivity indices even with correlated inputs, outperforming methods that ignore these correlations.
  • The method is shown to be computationally efficient, requiring fewer model evaluations than some traditional sampling-based methods. For instance, the authors suggest that $n \approx 100$ and $n' \approx 2000$ (sample sizes for model runs and index calculation, respectively) are often sufficient for accurate estimation.
  • In the chemical process example, the method successfully identified parameters that influence different outputs based on varying operating conditions and feed compositions, providing actionable insights for engineers.

Limitations and Future Work

  • Interpretation of Multidimensional Indices: While the paper addresses the issue of interpreting sensitivity indices with correlated inputs, the authors note that the multidimensional approach (e.g., Jacques' method) can still be limiting when all inputs are correlated, making it difficult to pinpoint individual parameter influences within a correlated block.
  • Theoretical Variance Calculation: The paper mentions that calculating the exact theoretical variance of the $\hat{S}_1$ estimator is complex and requires higher-order moments of the local linear estimator, which is beyond the scope of the current work. However, they suggest that bootstrap methods can be used for practical variance estimation.
  • Future Work: The authors suggest extending their work to develop multi-output sensitivity indices using multivariate nonparametric regression techniques. They also propose that practitioners might benefit from combining their method with other approaches (like Oakley & O'Hagan's Bayesian method) for a more comprehensive analysis.

Summary updated

Abstract

Sensitivity indices when the inputs of a model are not independent are estimated by local polynomial techniques. Two original estimators based on local polynomial smoothers are proposed. Both have good theoretical properties which are exhibited and also illustrated through analytical examples. They are used to carry out a sensitivity analysis on a real case of a kinetic model with correlated parameters.
BibTeX
@article{veiga2009local,
  author = {Sébastien da Veiga and Delphine Sinoquet and Fabrice Gamboa},
  title = {{Local Polynomial Estimation for Sensitivity Analysis on Models With Correlated Inputs}},
  journal = {Technometrics},
  year = {2009},
  doi = {10.1198/tech.2009.08124},
}
Figure