Chapter 8
Statistical Analysis of
GCM Output
by Claude Frankignoul
8.1 Introduction
In general circulation model (GCM) studies, statistical methods are needed
for a number of purposes: to validate a model with observations, to identify
its response to anomalous boundary conditions or its sensitivity to changes
in model formulation, and to determine its predictive skill. The first two
problems are discussed in this Chapter 8 and in the next Chapter 9, while
the evaluation of forecasts is discussed in Chapter 10.
In model validation studies, the question is whether a GCM is consiste nt with reality, for example its mean state (climate) and its variability
(weather), which are themselves estimated from the observations. One may
also ask whether a change in a parameterization improves the model climate, or whether a model is more rea/istic than another one. Although visual comparisons can reveal obvious differences, they become less effective as
the model's fidelity increases, and they are inadequate for separating the effect of model inadequacies from that of the observational data uncertainties.
Hence, quantitative measures of agreement based on statistical techniques
are needed.
In sensitivity or response studies, models are integrated twice in different
conditions (e.g., boundary conditions) to document the model response to
the prescribed change. The question is whether there is a significant climatic
impact, and possibly wh ether the model response is consistent with a response
Statistical Analysis of
GCM Output
by Claude Frankignoul
8.1 Introduction
In general circulation model (GCM) studies, statistical methods are needed
for a number of purposes: to validate a model with observations, to identify
its response to anomalous boundary conditions or its sensitivity to changes
in model formulation, and to determine its predictive skill. The first two
problems are discussed in this Chapter 8 and in the next Chapter 9, while
the evaluation of forecasts is discussed in Chapter 10.
In model validation studies, the question is whether a GCM is consiste nt with reality, for example its mean state (climate) and its variability
(weather), which are themselves estimated from the observations. One may
also ask whether a change in a parameterization improves the model climate, or whether a model is more rea/istic than another one. Although visual comparisons can reveal obvious differences, they become less effective as
the model's fidelity increases, and they are inadequate for separating the effect of model inadequacies from that of the observational data uncertainties.
Hence, quantitative measures of agreement based on statistical techniques
are needed.
In sensitivity or response studies, models are integrated twice in different
conditions (e.g., boundary conditions) to document the model response to
the prescribed change. The question is whether there is a significant climatic
impact, and possibly wh ether the model response is consistent with a response