|
Abstract:
|
Understanding how a spatio-temporal model is able to capture different structures in the data is an important aspect in the modelling practice. Diagnostic of dependent data could focus on high or low frequency, and could produce very different assessment of the goodness of fit. High frequency diagnostics relies on contrast variance or high frequency behaviour in the periodogram, and could inform on how the model is able to capture local dependence, and where it needs to be improved. However, a high frequency approach fails to capture large scale patterns of scientific significance such as atmospheric teleconnections, and pseudo-realizations from the estimated parameters are not visually similar to the original data set. This creates non negligible communication problems when discussing the results to practitioners outside the statistical community. We show how low frequency diagnostics could result in more intuitive pseudo-realizations, and could be easily explained to the general public with visualization techniques such as movies, or smartphone apps.
|
