This is an important issue in the application of uncertainty estimation methods in that a variety of dependencies will have an effect on the resulting prediction bounds. There will be dependencies between different model components and different model parameters in producing good model performance. There will be dependencies between errors in boundary condition data, model structural error and observations that will have an effect on calibrated values of model parameters. There will be spatial and temporal dependencies in model residuals that may be difficult to represent by some simple error structure (but which if ignored are known to lead to bias in probabilistic inference of parameter values, even for linear systems). It may be very difficult, even with experience of the application of a particular model, to specify the expected dependencies for a new application. It may be possible to elucidate some of these dependencies a posteriori as part of the calibration process; others may be difficult such as the interaction between input error, model structural error and effective parameter values when the true nature of the errors in inputs or model structure cannot be fully known given the information available. It is clearly also an issue in trying to provide prior estimates of effective parameters in applications (like the prediction of future change) where no conditioning data is available to allow evaluation of a posteriori distributions and dependencies. What is important is that all model applications should be aware of the potential effects of such dependencies.