Interpreting First- and Total-Order Sobol Indices

What a Sobol Index Says About a QSP Endpoint

A global sensitivity analysis asks how uncertainty in model inputs contributes to variation in a specified model output. For a QSP model, that output might be a predicted biomarker concentration at a selected time, a response summary, or another defined endpoint. A Sobol index is therefore not a universal property of a parameter: it describes its contribution to variance for a particular endpoint, under particular input distributions.

The first-order and total-order indices answer different questions. A first-order index measures the variance contribution associated with varying one input by itself. A total-order index measures all variance contributions involving that input, including contributions shared through interactions. Comparing them is especially useful when a parameter has little standalone effect but participates in important combined effects.

The conditional expectation in \(S_i\) averages over the other inputs, so the first-order index describes the part of output variance captured by the response to \(X_i\) alone. In contrast, the conditional variance in \(S_{T_i}\) asks how much output variation remains, on average, when all inputs except \(X_i\) are held fixed and \(X_i\) varies. Thus \(S_{T_i}\) includes both the standalone contribution and every interaction contribution involving \(X_i\).

These definitions rely on independent inputs and a finite, positive output variance. Under the functional ANOVA decomposition for independent inputs, the variance is partitioned into nonnegative components \(V_u\) indexed by nonempty subsets \(u\) of inputs. The component \(V_{\{i\}}\) is the first-order contribution; components such as \(V_{\{i,j\}}\) describe interactions. The indices can then be read as follows:

The final expression is the total normalized variance contribution from interactions that include \(X_i\). It does not identify which other inputs form those interactions; a total-order index groups them together. It also does not mean that interactions can always be allocated uniquely to one input in a ranking: an interaction involving two inputs contributes to both of their total-order indices.

How to Read the Indices and Their Gap

Both indices are dimensionless proportions of output variance. For example, \(S_i=0.30\) means that 30% of the output variance is attributable to the first-order component for \(X_i\), given the selected input distributions. It does not mean that changing \(X_i\) changes the endpoint by 30%, nor does it describe a percentage change in the endpoint’s mean.

When \(S_i\) and \(S_{T_i}\) are nearly equal, little of the variance involving \(X_i\) is assigned to interactions. When \(S_{T_i}\) substantially exceeds \(S_i\), interactions involving \(X_i\) make a meaningful contribution. A small \(S_i\) alone is not evidence that an input is unimportant: the total-order index may still be large.

An input is reasonably described as negligible for this analysis only when its total-order contribution is small for the specified output, input distributions, and model domain. This is a scoped conclusion, not a claim that the parameter is biologically irrelevant or unimportant for every prediction. A parameter may matter for another endpoint, another time point, or a different range of plausible inputs. If the output variance is zero, the indices are undefined because their denominator is zero.

Why Rankings Can Be Misleading

A ranking by first-order indices answers which inputs have the largest standalone contributions. A ranking by total-order indices answers which inputs are involved in the largest total contributions, including interactions. Those rankings can differ substantially. Neither ranking is universally correct: the appropriate one depends on whether the question concerns isolated effects or all effects involving each input.

The total-order indices need not sum to one. A variance component for an interaction is included in the total-order index of every input in that interaction. In contrast, the first-order indices sum to at most one under the standard decomposition; the remaining variance is assigned to interactions. A large sum of total-order indices is therefore not a contradiction or a sign that output variance has been counted incorrectly.

Assumptions, Diagnostics, and Alternatives

Before interpreting reported indices, check that the endpoint and input distributions match the scientific question. The same model can yield different indices under different parameter ranges or distributions because the indices quantify variance over those distributions. Check also that the output is a deterministic function of the stated inputs, or that stochastic variation has been handled explicitly. If random simulation noise is mixed into \(Y\) without being modeled as an input or separated from the signal, the standard interpretation may not apply cleanly.

A useful conceptual diagnostic is consistency with the bounds and ordering implied by the definitions: for valid population indices, \(0\leq S_i\leq S_{T_i}\leq 1\). A reported \(S_i>S_{T_i}\), or an index outside this range, signals a problem with estimation, implementation, or interpretation rather than a legitimate population result. In computed analyses, small apparent violations can reflect finite-sample estimation error; they should prompt checking the calculation and its uncertainty rather than being treated as true negative interaction variance.

Compare the two rankings and inspect each gap. Inputs with both indices close to zero have little contribution to output variance over the chosen domain. Inputs with small first-order but larger total-order indices are interaction-dependent and should not be discarded solely because their main effects are small. A large gap flags involvement in interactions but does not identify the interacting partners; a separate interaction analysis is needed for that question.

Sobol indices differ from local derivatives, which describe response to small changes near a chosen parameter value, and from Morris elementary effects, which provide a screening summary across selected changes. Those approaches can be useful when the question or computational budget differs, but they do not have the same variance-share interpretation. Sobol indices also have limitations: they depend on the chosen endpoint and input distributions, standard definitions assume independent inputs, and a ranking does not establish biological causality or parameter identifiability. For correlated inputs, the standard independent-input decomposition and its straightforward attribution should not be assumed.