Moving from OAT to Sobol Indices for Sensitivity Analysis

Modeling

Many of us default to one-at-a-time (OAT) sensitivity analysis because it is intuitive. If we want to know how temperature affects a reaction, we hold pressure and concentration constant and wiggle the temperature. It is computationally efficient; it provides a clear, linear narrative that is easy to graph and explain. In a perfectly linear system, this approach is entirely sufficient. But consider a scenario where the effect of temperature is dependent on the pressure. If these variables are synergistic, OAT might suggest neither is particularly dominant, or it might miss the peak efficiency window entirely. We might be looking at a narrow slice of a multidimensional surface while assuming the surface is a flat plane. This leads to a false sense of security regarding which parameters actually drive the model. Sobol indices offer a way to move past this by using global sensitivity analysis based on variance decomposition. Instead of local perturbations, the method quantifies how much of the output variance is caused by each input. The implementation generally follows these steps: 1. Define the probability distributions for all input parameters rather than single point values. 2. Generate a quasi-random sample, such as a Sobol sequence, to cover the parameter space more uniformly than standard Monte Carlo sampling. 3. Calculate the first-order index (S_i), which represents the fraction of output variance attributed to variable i alone. 4. Calculate the total-effect index (S_Ti), which accounts for the variance of i plus all its interactions with other variables. The utility is found in the gap between S_i and S_Ti. If the total-effect index is significantly larger than the first-order index, you have a non-linear interaction that OAT would have ignored. It replaces a series of isolated tests with a quantitative map of how uncertainty actually propagates through the system.