Virtually all attempts to date to detect an anthropogenic signal in
the observed climate system have excluded from consideration uncertainties in the model predicted field, as well as uncertainties in the observations themselves. The current work attempts to first estimate, then assess, the impact on detection, of both type of uncertainties.
The uncertainties in the model predicted anthropogenic signals are
estimated in several ways. One approach uses the CMIP2 data sets. These simulations come from 9 different GCMs, each forced by a 1%/year increase in CO2. The differences between the model signals can be taken as a measure of the uncertainty in the estimates from a true (or mean) signal. The results suggest the average signal from the 9 models is a fairly good approximation to the observed global temperature change observed over the last 50 years. This in turn suggests the main uncertainty in model signals is associated with internal variability produced by the models' themselves. Various statistics from these runs are used to construct, via a Monte
Carlo approach, a sampling distribution associated the sampling
uncertainties. Claims of detection from each model are evaluated against this distribution.
Two different models have produced a set of five simulations that
include the effects of both CO2 and sulfate aerosols. The differences
between these simulations are again used to develop a sampling distributiondue to uncertainty associated with inter- and intra-model differences. The same set of detections metrics used in the CO2 only runs are again estimated and compared with their sampling distribution to assess detection metrics for significance.
The uncertainties in the observations are estimated from the
sampling errors associated with construction of observed fields of global temperature change. The statistics of these errors were reconstructed from the standard error and decorrelation scale length obtained in Jones et al., 1997. The result was the establishment, via Monte Carlo methods, of 1000 different estimates of the error fields. Each was physically different, although they each had the same basic statistical properties. These different realizations of observational uncertainty were used as above to construct a sampling distribution associated with observational error. The various model anthropogenic signals were tested to see if they could fall within these distributions. Statement of signal significance will be made from these tests