Research Radar — 2026-10-27

Methodology Review: Sobol sensitivity indices: separating main effects from interactions

English · About 10 minutes

Main reading checked in full author-hosted journal text; two related readings checked at publisher-abstract level. Verified applications are numerical model benchmarks; omics transfers below are explicitly proposed, not validated.

The intuition

Suppose a computer model returns one number, and several uncertain settings can change that number. Before spending another afternoon refining the model, ask a narrower question: which settings account for variation in this particular output over the range we actually consider plausible? This is not the same as asking which coefficient looks largest, which input has the strongest correlation, or which perturbation should be tested next. The output, input ranges and probability distribution are part of the question, not incidental preprocessing choices.

Variance-based global sensitivity analysis separates variation visible from one input alone from variation involving that input together with others. Sobol indices use an orthogonal decomposition under independent inputs. A first-order index measures a main effect; a total-effect index also includes interactions. Consequently, an input with a modest main effect can still matter greatly. These are properties of a specified model and input distribution, not automatic causal importance scores for measured biological variables. [R2]

A small worked example

Here is an original four-case example, small enough to check without software. Let two independent inputs, \(X_1\) and \(X_2\), each take the values minus one and plus one with equal probability. Define the scalar model output \(Y=X_1+2X_2+3X_1X_2\). In the input order minus/minus, minus/plus, plus/minus and plus/plus, the outputs are zero, minus two, minus four and six. Their mean is zero. Averaging their squared values gives the output variance: \(V=(0+4+16+36)/4=14\). Here \(V\) denotes variance, not a fitted parameter.

Average over the second input while holding the first fixed. The interaction cancels, leaving conditional mean \(X_1\), whose variance is one. Reversing their roles leaves conditional mean \(2X_2\), whose variance is four. The interaction term has variance nine. These three components add to fourteen because the inputs are independent and centered. Thus the first-order indices are \(S_1=1/14\) and \(S_2=4/14\), while the interaction accounts for \(9/14\) of the variance. The symbols \(S_1\) and \(S_2\) denote the first-order fractions for inputs one and two.

Now include every component involving an input. The total-effect fractions are \(T_1=10/14\) and \(T_2=13/14\). Here \(T_1\) and \(T_2\) denote total-effect indices. Their sum is \(23/14\), larger than one, because both totals include the same interaction. This is not an arithmetic error or a division of credit into exclusive shares. For a second check, fix the second input at minus one: the remaining output is \(-2X_1-2\), with variance four. Fix it at plus one: the remaining output is \(4X_1+2\), with variance sixteen. Averaging these variances gives ten, exactly the numerator of \(T_1\).

The necessary mathematics

For a general deterministic function \(f\), write \(Y=f(X_1,\ldots,X_d)\), where \(d\) is the number of mutually independent random inputs and \(X_i\) is input number \(i\). Let \(X_{-i}\) mean the collection of all inputs except number \(i\). Assume the output has finite, nonzero variance \(V=\operatorname{Var}(Y)\). The first-order and total-effect definitions are

Here \(\operatorname{E}\) denotes expectation, \(\operatorname{Var}\) denotes variance, and \(\mid\) denotes conditioning. In the first numerator, average away other inputs before measuring variation across the retained input. In the second, measure variation in that input at each fixed setting of the others, then average those conditional variances. [R1] [R2]

The definitions suggest a computational experiment. Construct paired model evaluations that share all coordinates except one, whose value is independently redrawn. Squared output differences then contain information about the corresponding total effect. Jansen studies variance-estimation designs; Saltelli and colleagues compare estimators and sampling arrangements. Neither an attractive formula nor many simulations alone guarantees a useful estimate. Record the design, budget and repeated-estimate uncertainty rather than reporting a precise-looking rank without numerical diagnostics. [R3] [R1]

Assumptions and failure modes

Independence is a substantive assumption, not a harmless computational convenience. The conventional orthogonal allocation relies on it. If observed inputs are dependent, independently swapping their coordinates changes their joint distribution, and the original interpretation need not survive. The cited main reading explicitly works with independent inputs. Treat dependent-input questions as a different methodological problem rather than quietly applying the same decomposition. [R1] [R2]

Two checks follow directly from the toy example. First, if you decide the second input is always zero instead of equally likely to be minus one or plus one, the model becomes the first input alone: a different input distribution answers a different question. Second, multiplying the entire output by a nonzero constant multiplies all variance components by its square, leaving their fractions unchanged. Changing only the interaction coefficient does not have that invariance. These are useful unit tests for an implementation, not empirical findings.

A stochastic simulator requires another explicit choice: do you want sensitivity of its mean response or sensitivity of individual noisy realizations? Replicating runs can help distinguish those targets, but merely treating random seeds as unexplained noise obscures the question. Also, a model that omits an important mechanism can yield perfectly reproducible indices for the wrong mechanism. As practical safeguards proposed here, check model plausibility separately, report the input distribution, and avoid equating numerical stability with scientific validity.

Verified applications

The verified application in the main reading is evaluation of sensitivity estimators on mathematical model benchmarks with known analytic indices. Its numerical experiments compare sampling strategies and estimator choices across test functions. This is evidence of a scientific-computing application, not evidence that the same rankings identify biological regulators. Jansen also reports small numerical experiments for variance-estimation designs, but only its abstract was inspected here. No omics benchmark or wet-lab validation is claimed for these three readings. [R1] [R3]

Possible transfers — speculative

A possible omics transfer, proposed here rather than validated by these references, is a small sensitivity audit of an already justified simulator. Choose a few controllable inputs, such as simulated sequencing depth or an experimental setting represented by the model, and one predefined output such as recovery of a known synthetic signal. Specify plausible distributions before seeing the rankings. Begin with independent settings only when that assumption matches the intended experiment; do not independently resample columns of a correlated expression matrix and describe the resulting synthetic cells as biologically feasible.

The useful decision might be where to invest model calibration effort, not which gene to perturb. Compare main and total effects, examine a conditional response plot for the largest gap, and repeat the audit under a second defensible input range. If a cheap surrogate replaces an expensive simulator, check its predictions on held-out parameter settings first and disclose that the reported indices concern the surrogate. A future experimental confirmation would need its own design and evidence; this proposed audit supplies neither causal identification nor an automatic laboratory policy.

What to try and what to read next

Spend the first ten minutes enumerating the four outputs and reproducing fourteen, ten and thirteen. Then set the interaction coefficient to zero and predict what will happen before recalculating: the total effects should collapse to the main effects. For a real model, write one sentence defining the output and one defining the input distribution before running anything. Stop if either sentence is vague. The central lesson is that importance can mean acting alone or participating in interactions, and those questions should not share an unlabeled ranking.

Reading list

• [R1] Primary reading: Variance based sensitivity analysis of model output. Design and estimator for the total sensitivity index. Saltelli, Annoni, Azzini, Campolongo, Ratto and Tarantola (2009). Main reading: definitions, sampling construction and numerical experiments. The author-hosted PDF is the published article, not a distinct preprint. Start with sections 1–3 before the estimator comparisons. First available online 2009-09-30; published in Computer Physics Communications volume 181 (2010), pages 259–270. The displayed 2009 date records online-first publication, not the issue year.

• [R2] Related reading: Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates. I. M. Sobol′ (2001). Foundational companion for orthogonal variance decomposition and the distinction between first-order and total contributions. Evidence inspected is publisher abstract and section excerpts.

• [R3] Related reading: Analysis of variance designs for model output. Michiel J. W. Jansen (1999). Estimator-design companion: understand that an index definition and an efficient simulation design are different objects. Abstract-level inspection only; no independent reproduction of its numerical experiments.