Impact Range Assessment (IRA): An Interpretable Sensitivity Measure for Machine Learning Regression Modelling

arXiv:2602.05239 · stat.ME · Submitted 2026-02-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Impact Range Assessment (IRA)".

Jane: Here's the gist of the paper. Impact Range Assessment (IRA) introduces an interpretability method that quantifies each predictor’s maximal influence by measuring the total potential change in the response variable across its range <ref:2602.05239#pg2> This method provides a simple and intuitive way to interpret and rank the impact of predictors on the response variable, thereby improving model transparency and reliability <ref:2602.05239#pg2>.

Tom: Method Introduction The Impact Range Assessment (IRA) quantifies how much each predictor can influence the response variable when varied across its observed data range in data-driven regression models where predictions are derived from observed data <ref:2602.05239#pg5> The IRA value represents the maximum possible change in the model’s output attributed to a predictor, accounting for the model’s internal sensitivity to that predictor and the distribution of the input data <ref:2602.05239#pg5>.

Jane: The method consists of five steps, starting with selecting a focus predictor <ref:2602.05239#pg5> This process involves interpolating values across the predictor’s range to generate M evenly spaced values <ref:2602.05239#pg5> Next, background observations are generated by drawing K samples randomly from the input dataset <ref:2602.05239#pg5>.

Tom: Modified input sets are then created by duplicating each background observation M times, replacing the focus predictor with the M interpolated points <ref:2602.05239#pg5> Finally, the IRA value is calculated as the average of the ranges between minimum and maximum predicted values across all background observations <ref:2602.05239#pg5>.

Paper summary: Tom: So we're looking at this paper now called "Impact Range Assessment (IRA): An Interpretable Sensitivity Measure for Machine Learning Regression Modelling". The basic idea is that regression models tell us how predictors affect the outcome, but they don't always show us *how much* each predictor actually matters across its whole range.

Jane: Right, so the paper introduces this method called IRA to quantify each predictor’s maximal influence by measuring the total potential change in the response variable across that entire range. That sounds like it could give us a much clearer picture than just looking at a single point in time or one specific region of interest.

Lu: Basically, they claim this method provides a simple and intuitive way to interpret and rank predictors on their impact on the response variable, which should improve model transparency and reliability >

Conclusion: Tom: So we’ve been diving into this paper about Impact Range Assessment, or IRA—it’s basically a new way to measure how much a single input variable actually matters in a prediction model.

Jane: Right, so the title is "Impact Range Assessment (IRA): An Interpretable Sensitivity Measure for Machine Learning Regression Modelling." It sounds like it’s trying to fix that problem where we just see if something is important or not, but we don't really know how much influence it has across all the possible values it could take.

Lu: Exactly. The paper introduces this method to quantify each predictor’s maximal influence by measuring the total potential change in the response variable across its range. It’s about getting that complete picture instead of just a snapshot of one specific area.

Meng: From my side, what I want to know is how practical this is when we’re actually running models at scale. If this method gives us a clear ranking, how does that translate into something we can actually use to decide which features to keep or drop in production?

Tom: Well, the authors show it works on linear regression and random forest models—they found that relevant predictors got higher IRA values than irrelevant ones across both. It seems like a pretty solid way to rank things when you need that transparency.

Jane: But they also flagged some limitations, which is important for us to keep in mind. They said this method isn't great for variables that are discrete or categorical, and the IRA values themselves depend heavily on the dataset you start with.

Lu: That dependence on the dataset is a big caveat because it means you can’t just compare an IRA value from one study to another if they used different input ranges. You have to guide your choice of range based on what you’re actually trying to find out in your specific field.

Meng: So, it's a tool for prioritization, not a universal truth about model performance across every single scenario? That makes sense for production work.

Tom: It definitely offers an intuitive way to assess impact that both technical folks and business people can grasp easily. It’s about making the black box a bit more readable by quantifying the potential change across the whole input space.

Jane: And it’s a step up from just local sensitivity analysis, which only looks at one specific spot. The IRA gives you something much broader, but we still need to watch those dataset dependencies closely as we look at applying this in real-world applications. **(Sound of traffic noise swells slightly)**

Lalam: Based on the literature I’ve processed, the consensus here is that if you need a way to rank predictors based on their estimated influence across the entire input range, IRA offers a clearer measure than local sensitivity analysis because it captures the full potential change. We're fairly sure this approach supports better model interpretability in many contexts.

Jihao You, Dan Tulpan, Jiaojiao Diao, Jennifer L. Ellis

Department of Animal Biosciences, University of Guelph · Department of Integrative Biology, University of Guelph

stat.ME

Submitted: 2026-02-05

Updated: 2026-10-31

Comments: 32 pages, 6 figures. Revised version; resubmitted to MethodsX on October 30, 2026

Code: https://github.com/jyou2025/impact-range-assessment

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: Impact Range Assessment (IRA) introduces an interpretability method that quantifies each predictor’s maximal influence by measuring the total potential change in the response variable across its…

Key concepts

Impact Range Assessment (IRA)
IRA measures the maximum possible change in a model's output caused by a specific predictor when that predictor is varied across its full observed range. It quantifies the total potential influence of a variable, accounting for how sensitive the model is to that input and the distribution of data.
Focus Predictor Selection
The process begins by choosing one specific predictor to analyze in detail. This involves creating evenly spaced values across that predictor's entire range. These interpolated points are then used to modify the input dataset for subsequent steps in the IRA calculation.
Background Observations
These are random samples drawn from the original input dataset. These observations serve as a baseline for comparison when testing how changing the focus predictor affects the model's predictions across different scenarios. They help generate modified input sets.
Model Evaluation
The method is tested on linear and nonlinear models using synthetic datasets. Results showed that relevant predictors consistently produced higher IRA values than irrelevant ones, confirming the method's ability to distinguish important variables in various model complexities.

Terminology

Summary

Impact Range Assessment (IRA) introduces an interpretability method that quantifies each predictor’s maximal influence by measuring the total potential change in the response variable across its range <ref:2602. p. 0

Where the paper says it

ca Keywords Impact Range Assessment; Regression Model; Machine Learning; Sensitivity Analysis; Model interpretation Related research article None Abstract While regression models capture the relationship between predictors and the response variable, they often lack intuitive accompanying methods to understand the influence of predictors on the outcome. To address this, we introduce an interpretability method called Impact Range Assessment (IRA), which quantifies each predictor’s maximal influence by measuring the total potential change in the response variable, across the predictor’s range.

05239#pg2> This method provides a simple and intuitive way to interpret and rank the impact of predictors on the response variable, thereby improving model transparency and reliability <ref:2602. p. 0
Where the paper says it

A case study using a model that predicts pellet quality demonstrated that the IRA provides a simple and intuitive approach to interpret and rank predictor influence, thereby improving model transparency and reliability. • IRA quantifies (for regression models) how much each predictor influences the model’s response variable, across the predictor’s observed range.

05239#pg2>.

Method Introduction

The Impact Range Assessment (IRA) quantifies how much each predictor can influence the response variable when varied across its observed data range in data-driven regression models where predictions are derived from observed data <ref:2602.05239#pg5> The IRA value represents the maximum possible change in the model’s output attributed to a predictor, accounting for the model’s internal sensitivity to that predictor and the distribution of the input data <ref:2602. p. 3

Where the paper says it

The IRA quantifies how much each predictor can influence the response variable when varied across its observed data range in data-driven regression models where predictions are derived from observed data. The IRA value represents the maximum possible change in the model’s output attributed to a predictor, accounting for the model’s internal sensitivity to that predictor and the distribution of the input data.

05239#pg5>. The method consists of five steps, starting with selecting a focus predictor <ref:2602. p. 3
Where the paper says it

The IRA method consists of five steps: 1. Select a focus predictor.

05239#pg5> This process involves interpolating values across the predictor’s range to generate M evenly spaced values <ref:2602. p. 3
Where the paper says it

Interpolate values across the predictor’s range. Generate M evenly spaced values between the predictor’s minimum and maximum values.

05239#pg5> Next, background observations are generated by drawing K samples randomly from the input dataset <ref:2602. p. 8
Where the paper says it

Number of background samples represents the number of background observations randomly drawn from the input dataset. Number of interpolated points represents the number of values generated across the range of each focus predictor.

05239#pg5>. Modified input sets are then created by duplicating each background observation M times, replacing the focus predictor with the M interpolated points <ref:2602. p. 3
Where the paper says it

Create modified input sets. Duplicate each background observation M times, replacing the focus predictor with the M interpolated points while keeping all other predictors fixed.

05239#pg5> Finally, the IRA value is calculated as the average of the ranges between minimum and maximum predicted values across all background observations <ref:2602. p. 3
Where the paper says it

The range for each background observation, defined as the difference between its minimum and maximum values, is computed. The IRA value for the predictor is the average of these K ranges.

05239#pg5>.

Model Evaluation

The IRA method was evaluated using two benchmark synthetic datasets to assess its ability to capture predictor influence under varying levels of complexity <ref:2602.05239#pg5> For the linear regression model, relevant predictors produced higher IRA values than irrelevant ones <ref:2602. p. 0

Where the paper says it

To address this, we introduce an interpretability method called Impact Range Assessment (IRA), which quantifies each predictor’s maximal influence by measuring the total potential change in the response variable, across the predictor’s range. Validation using synthetic linear and nonlinear datasets demonstrates that relevant predictors produced higher IRA values than irrelevant ones.

05239#pg5>. The single-execution IRA clearly distinguished relevant from irrelevant predictors in the linear regression model (Table 4) <ref:2602. p. 7
Where the paper says it

The large coefficient and intercept values result from applying Box-Cox transformation to PDI (response variable). Method validation For the Linear Regression model, the single-execution IRA clearly distinguished relevant from irrelevant predictors (Table 4).

05239#pg5>. For the nonlinear Random Forest regressor, relevant predictors consistently exhibited higher IRA values compared to the irrelevant ones <ref:2602. p. 8
Where the paper says it

Number of interpolated points represents the number of values generated across the range of each focus predictor. The single-execution IRA for the Random Forest regressor showed that the relevant predictors (𝑋1, 𝑋2, 𝑋3, 𝑋5 and 𝑋6) consistently exhibited higher IRA values compared to the irrelevant ones (𝑋4, 𝑋7 and 𝑋8) (Table 5), aligning with their nonlinear contributions to the outcome.

05239#pg5>. The authors found that increasing the number of interpolated points to 100 or more led to more stable and reliable estimates for the nonlinear model <ref:2602. p. 8
Where the paper says it

For the number of interpolated points, using only 50 slightly underestimated IRA values for predictors involved in nonlinear relationships. Increasing the number of interpolated points to 100 or more led to more stable and reliable estimates.

05239#pg5>.

Robustness and Stability

The robustness of the IRA method was confirmed through repeated evaluations, which produced results closely aligned with those from the single-execution analysis <ref:2602.05239#pg5>. The authors proposed a repeated IRA procedure to generate confidence intervals (CIs) for each predictor by repeating the single-execution analysis several times <ref:2602. p. 4

Where the paper says it

To quantify the uncertainty associated with random background observation resampling and predictor interpolation, a repeated IRA procedure is also proposed to generate confidence intervals (CIs) for each predictor. In this procedure, the single-execution IRA is repeated several times, with each repeat generating an IRA value based on a unique resampling of background observations.

05239#pg5> For the nonlinear dataset, using 100 interpolated points and 200 background observations yielded stable and reliable IRA estimates <ref:2602. p. 8
Where the paper says it

Beyond 100 samples, the changes in IRA values were minimal while computation time increased drastically. In summary, 100 interpolated points and 200 background observations yielded stable and reliable IRA estimates for this nonlinear model (Random Forest regressor).

05239#pg5>. The mean IRA values remained stable when applying the repeated IRA to the Random Forest regressor across different numbers of repeats <ref:2602. p. 9
Where the paper says it

Number of interpolated points represents the number of values generated across the range of each focus predictor. When applying the repeated IRA to the Random Forest regressor across different numbers of repeats, the mean IRA values remained stable (Table 6) and closely matched those obtained from the single-execution IRA (Table 5).

05239#pg5>.

Practical Applications and Limitations

The paper demonstrates that the IRA approach is effective for ranking variables based on their estimated IRA values, which is useful for prioritization <ref:2602.05239#pg5> The method offers practical benefits by providing an intuitive way to assess predictors’ impact accessible to both technical and non-technical users <ref:2602. p. 2

Where the paper says it

The analysis can be performed once as a single-execution procedure for a quick estimation of predictor impact, or repeated to assess the stability of the results. Compared to conventional sensitivity methods, IRA provides a simple and intuitive way to interpret and rank the impact of predictors on the response variable.

05239#pg5>. Limitations include the fact that the IRA is not well suited for discrete numeric or categorical predictor variables <ref:2602. p. 2
Where the paper says it

There is a need for an intuitive and efficient approach to summarize the influence of individual predictors on the response variable across the predictor space. To address these limitations, we propose a new sensitivity measure named the Impact Range Assessment (IRA), which evaluates the maximum possible change in the response variable when each continuous numeric predictor is varied across its observed range in a regression model.

05239#pg5> Furthermore, IRA values are dataset-dependent and should not be compared across datasets with different predictor ranges for the same regression model <ref:2602. p. 9
Where the paper says it

Thus, using 50 repeats provides an effective balance between estimate stability and computational efficiency for this nonlinear dataset and model. Compared to the single-execution IRA that provides only point estimates, the repeated IRA yields both mean IRA values and 95% confidence intervals, capturing variability from background predictor sampling and providing a more robust measure of the estimated impact for each predictor.

05239#pg5>. The choice of range should therefore be guided by the study objectives and the region of practical or empirical relevance <ref:2602. p. 0
Where the paper says it

A case study using a model that predicts pellet quality demonstrated that the IRA provides a simple and intuitive approach to interpret and rank predictor influence, thereby improving model transparency and reliability. • IRA quantifies (for regression models) how much each predictor influences the model’s response variable, across the predictor’s observed range.

05239#pg5>.

Conclusion

The IRA provides a clearer and more comparable measure of each predictor’s true impact on PDI compared to local sensitivity analysis which only captures effects within a specified region <ref:2602.05239#pg5>. The IRA offers an intuitive way to assess predictors’ impact that is accessible to both technical and non-technical users <ref:2602. p. 2

Where the paper says it

The analysis can be performed once as a single-execution procedure for a quick estimation of predictor impact, or repeated to assess the stability of the results. Compared to conventional sensitivity methods, IRA provides a simple and intuitive way to interpret and rank the impact of predictors on the response variable.

05239#pg5>. This approach can support model interpretability and actionable insights across a range of disciplines <ref:2602. p. 0
Where the paper says it

ca Keywords Impact Range Assessment; Regression Model; Machine Learning; Sensitivity Analysis; Model interpretation Related research article None Abstract While regression models capture the relationship between predictors and the response variable, they often lack intuitive accompanying methods to understand the influence of predictors on the outcome. To address this, we introduce an interpretability method called Impact Range Assessment (IRA), which quantifies each predictor’s maximal influence by measuring the total potential change in the response variable, across the predictor’s range.

05239#pg5>. The IRA provides practical benefits for interpreting data-driven models <ref:2602. p. 3
Where the paper says it

The IRA quantifies how much each predictor can influence the response variable when varied across its observed data range in data-driven regression models where predictions are derived from observed data. The IRA value represents the maximum possible change in the model’s output attributed to a predictor, accounting for the model’s internal sensitivity to that predictor and the distribution of the input data.

05239#pg5>.

References

[1] K. Tyagi, C. Rane, Harshvardhan, M. Manry, Chapter 4 - Regression analysis, in: R. p. 0

Where the paper says it

Ellis1 Affiliations 1 Department of Animal Biosciences, University of Guelph, Guelph, ON, Canada 2 Department of Integrative Biology, University of Guelph, Guelph, ON, Canada Corresponding author’s email address and Twitter handle *jyou03@uoguelph. ca Keywords Impact Range Assessment; Regression Model; Machine Learning; Sensitivity Analysis; Model interpretation Related research article None Abstract While regression models capture the relationship between predictors and the response variable, they often lack intuitive accompanying methods to understand the influence of predictors on the outcome.

Pandey, S.K. Khatri, N. kumar Singh, P. Verma (Eds.), Artifintell Mach Learn EDGE Comput. p. 4
Where the paper says it

Method implementation All data analyses were conducted in Python 3. 17 using scikit-learn 1.

, Academic Press, 2022: pp 53–63 https://doi. p. 1
Where the paper says it

Graphical abstract Specifications table Subject area Mathematics and Statistics More specific subject area Model interpretability approaches for regression models Name of your method Impact Range Assessment (IRA) Name and reference of original method N. Resource availability The Python scripts for the IRA functions and the validation on the synthetic datasets are available at the GitHub link: https://github.

org/10.1016/B978-0-12-824054-

[2] A. Saltelli, M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D Gatelli, M Saisana, S Tarantola Global Sensitivity Analysis: The Primer John Wiley & Sons 2008

[3] B Iooss P Lemaître A Review on Global Sensitivity Analysis Methods in G Dellino C Meloni Eds Uncertain Manag Simul Optim Complex Syst Springer US Boston MA 2015

[4] A Saltelli P Annoni How to avoid a perfunctory sensitivity analysis Environ Model Softw 25 (2010) 1508–1517

[5] H.M Wainwright S Finsterle Y Jung Q Zhou JT Birkholzer Making sense of global sensitivity analyses Comput Geosci 65 (2014) 84–94 https://doi. p. 2

Where the paper says it

Since global sensitivity analysis evaluates the impact of predictors through mathematical formulations, it often becomes computationally intensive and difficult to interpret. In addition to quantitative local and global sensitivity analyses, scatter plots provide a qualitative visual indication of how individual inputs influence the output [8].

org/10.1016/j.cageo.2013.06.006

[6] I M Sobol Distribution of points in a cube and approximate evaluation of integrals USSR Comput Math Math Phys 7 (1967) 86–112

[7] M D Morris Factorial Sampling Plans for Preliminary Computational Experiments Technometrics 33 (1991) 161–174 https://doi.org/10.1080/00401706.

[8] P Zhang A novel feature selection method based on global sensitivity analysis with application in machine learning-based prediction model Appl Soft Comput 85 (2019) 105859 https://doi. p. 2

Where the paper says it

Compared to local sensitivity analysis, global sensitivity analysis provides more robust sensitivity measures and considers nonlinearity by sampling from a wider parameter space [5]. The global sensitivity analysis includes various approaches, such as variancebased methods like the Sobol method [6] and difference-based methods like the Morris method [7].

org/10.1016/j.asoc.2019.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Implement Impact Range Assessment (IRA) to quantify each predictor’s maximal influence by measuring the total potential change in the response variable, across the predictor’s range. This allows AI systems to rank predictors based on their true impact rather than just standardized coefficients, improving model transparency and reliability.

  2. Enable model-agnostic sensitivity analysis for both linear and nonlinear models using IRA, allowing AI to understand how individual predictors influence complex structures that lack interpretable parameters. This provides a method for exploring the full input space beyond local sensitivity analysis regions.

  3. Support fast single-execution and repeated IRA analyses to assess result stability, confirming the method’s robustness. This allows researchers to determine if an IRA value is stable across different sampling methods or execution runs, providing confidence intervals for each predictor's impact.

  4. Use a distribution of IRA values derived from repeated procedures to calculate 95 percent confidence intervals (CIs) for each predictor. This provides a more robust measure of the estimated impact for each predictor, capturing variability from background predictor sampling and offering a more reliable assessment than single point estimates.

  5. Apply IRA to prioritize variables in complex models, as predictors with higher IRA values exert greater influence on the response variable when varied. This helps data scientists focus on the most influential predictors among many variables, providing actionable insights across a range of disciplines.

Abstract

While machine learning regression models capture the relationship between predictors and the response variable, they often lack intuitive methods to help modelers understand how sensitive model predictions are to changes in individual predictors. To address this limitation, we introduce a model-agnostic interpretability method called Impact Range Assessment (IRA), which quantifies the sensitivity of model predictions to changes in each predictor by measuring the range of predicted responses as the predictor varies across its observed range. Validation using synthetic linear and nonlinear datasets demonstrates that relevant predictors produced higher IRA values than irrelevant ones. Moreover, repeated evaluations produced results aligned with those from the single-execution analysis, confirming the method's robustness. For demonstration, two benchmark datasets were used to evaluate IRA across five machine learning regression models, yielding model-specific sensitivity rankings and supporting IRA's applicability as a complementary sensitivity analysis tool alongside feature importance methods such as permutation feature importance and SHapley Additive exPlanations (SHAP).

Related papers