SimDec + Global Sensitivity Analysis: Revealing Hidden Effects in Complex Models

Some effects stay hidden — even from the best models.
When model inputs interact in non-linear ways, their combined influence can flip direction, amplify, or disappear altogether. These heterogeneous effects are invisible to traditional sensitivity methods.

That’s what Mariia Kozlova, Robert Moss, Julian Scott Yeomans, and Jef Caers address in their position paper “Uncovering Heterogeneous Effects in Computational Models for Sustainable Decision-Making” (Environmental Modelling & Software, 2024).

🔗 Read the article: https://doi.org/10.1016/j.envsoft.2023.105898

Why Traditional Sensitivity Analysis Falls Short

Most models in sustainability research rely on variance-based sensitivity indices (like Sobol’ indices) to measure how much each input contributes to the output variance.

But there’s a catch:
A single index value can’t tell whether that influence is constant across all regions of the input space.

In other words — a factor might matter only when another variable is high, or might even reverse its effect depending on context. These heterogeneous interactions are widespread in environmental systems, where nonlinear feedbacks rule.

The SimDec Extension: From Sensitivity to Structure

The new framework combines:

1. Global sensitivity analysis (via a simple binning method) — to compute first- and second-order sensitivity indices and identify which variables matter most;
2. Simulation Decomposition (SimDec) — to visualize how combinations of those variables drive the output.

The workflow (see Figure 2, page 4 of the paper) automates the whole process:

1. Compute combined sensitivity indices for each input (including interactions).
2. Select the most influential variables.
3. Divide their ranges into states (e.g., low/medium/high).
4. Form all possible scenarios.
5. Map simulation results to these scenarios.
6. Produce a stacked histogram that decomposes the output distribution by scenario colors.

The result: a visualization that exposes the shape and direction of heterogeneous effects — something that raw sensitivity numbers can never show.

Case Studies: Carbon, Cashflows, and Crying Babies

To demonstrate, the authors apply the method to three models:

• Carbon footprint of shipping pallets: shows that end-of-life treatment (incineration vs. landfill) completely reverses the emission effect of pallet reuse — a change in direction type of heterogeneity.
• Investment profitability model: reveals that the effect of production increases with price — a change in magnitude of interaction.
• Sequential decision-making (the “crying baby” model): illustrates that the importance of reward parameters depends entirely on the solver used — a change in presence of effect.

Each case demonstrates a different type of heterogeneity (see Table 11, page 10):

1. Change in direction,
2. Change in magnitude,
3. Change in presence of effect.

SimDec vs. Classic Sensitivity Tools

As summarized in Table 12 (page 11), one-at-a-time (OAT) and even global sensitivity analyses can detect interactions numerically, but they cannot show what those interactions look like.
SimDec does both — quantifies and visualizes them simultaneously.

Its built-in sensitivity indices rank the key variables, while the stacked-histogram visualization displays their interactions, making the framework both analytical and interpretive.

Why This Matters for Environmental Decision-Making

By integrating SimDec with global sensitivity analysis, researchers can now:

• Identify which factors matter and when,
• Uncover nonlinear and structural heterogeneity,
• Communicate uncertainty to decision-makers through clear, visual evidence.

As Kozlova and colleagues write, SimDec can unify sensitivity and uncertainty analysis — turning complex model behavior into interpretable, actionable insight for sustainable decisions.

Reference:
Kozlova, M., Moss, R. J., Yeomans, J. S., & Caers, J. (2024). Uncovering Heterogeneous Effects in Computational Models for Sustainable Decision-Making. Environmental Modelling & Software, 171, 105898.
https://doi.org/10.1016/j.envsoft.2023.105898

‍