Chapter 15 explains a step by step process for building and testing a model. In step 5, we estimate the parameters in a "one at a time fashion" based on all the information sources at our disposal. We expect that many of the parameters will be highly uncertain. But we know that it is often useful to include uncertain parameters, especially if they are needed to close the key feedback loops in the system. Chapter 15 encourages us to make our best estimate and move on to simulating the model. In step 6, we simulate the model to see if it generates the reference mode. In step 7, we test the sensitivity of the model to changes in the uncertain parameters. Chapter 16 illustrates a typical approach to sensitivity testing with the Kaibab deer model. It shows a collection of ten simulations to reveal the importance of three parameters. The ten runs show that the Kaibab model is inclined to generate the overshoot pattern regardless of the particular parameter values adopted in the tests.

The Kaibab chapter illustrates an informal style of sensitivity analysis that is appropriate in your first few iterations through the model building process. This appendix describes a more systematic and comprehensive approach that is useful when you've invested more time and effort in a model. Perhaps the model is gaining credibility within your organization, and you don't want any "unexpected surprises" to pop up when one of simulations calls for an unusual (and untested) combination of parameter values. Perhaps the model has grown to include hundreds of parameters, and you're wondering about the combined effect of their uncertainties. At this stage, we need to move beyond the informal approach described in Chapter 16.

Appendix J describes a formal approach to sensitivity analysis and illustrates the approach with the flower growth model from chapter 6 and the Kaibab model from chapter 16. The examples demonstrate that uncertainty intervals can grow or shrink over time. The examples also demonstrate that the dynamics of uncertainty may be attributed to the feedback structure of the model.