Morris Screening
A cheap global method that separates influential inputs from negligible ones before spending a full sensitivity budget.
Screening First
When a model has many uncertain inputs, computing full variance-based sensitivity for all of them is expensive. The Morris method is a cheap screening technique that ranks inputs by influence using far fewer runs, so that the negligible ones can be fixed at nominal values and the expensive analysis reserved for the few that matter.
Elementary Effects
Morris works by computing elementary effects: it changes one input at a time by a fixed step, from several different starting points scattered across the input space, and records the resulting change in output. Averaging the magnitude of these effects for an input measures its overall influence; the spread of the effects measures whether the input acts nonlinearly or through interactions.
- A large mean effect: the input is influential and should enter the full analysis.
- A small mean effect: the input can be fixed at nominal, saving budget.
- A large spread with a large mean: the input acts nonlinearly or through interactions, warranting closer study.
Why It Is Efficient
Because it samples from multiple starting points rather than only around one nominal point, Morris is a global method, not a local one, capturing influence across the whole input range. Yet it needs only a modest number of runs, proportional to the number of inputs times the number of starting trajectories, far less than a full Sobol analysis. This makes it the natural first pass for high-dimensional problems.
Its Place in the Workflow
Morris screening is qualitative: it ranks and separates inputs but does not precisely quantify how much variance each explains. The usual workflow runs Morris first to reduce the input set, then applies a quantitative method such as Sobol indices to the survivors. This two-stage approach spends the expensive analysis only where it pays off, and it guards against the waste of computing precise indices for inputs the output barely responds to.