As already noted above, for theoretical reasons we may expect a linear relationship between the radiative forcing covariates
and global mean temperature. To explore this issue in greater detail we fitted non-parametric curves, as recommended
by Fox and Weisberg (2011), to the scatter plots of each covariate against the mean global temperature anomaly (Fig. 4).
These figures showed only a slight degree of non-linearity. However, they do not account for dependencies in the data,
and so we refitted Model A using the gamm function from the mgcv library in R. We specified the same residual autocorrelation
structure when fitting this model, but instead of linear functions we used non-parametric splines for all 4 covariates.
Note that it is not possible to fit the other models with the gamm function because of their complicated autocorrelation
structure, but this should only have minor influence on the curvature of the splines. The four splines were estimated to
be exactly linear and coincided very closely to the arima function estimates (Fig. 4). Hence, it was concluded that a multiple
linear regression was an accurate representation for modelling the mean global temperature anomaly.
Also, as already noted, the correlation between the covariates is very small for the data from 1950 onwards, thus one
would not expect collinearity to be an issue. To examine whether this is indeed the case, each covariate was removed from
Model B, one at a time, and the remaining parameters were re-estimated. In all cases the regression parameters only changed
by a small amount. For example, when f(eCO2) was removed, the new regression parameter estimates for SOI, TSI and VOLRF
were 0.013, 0.008 and 0.10, respectively, all within the 95% confidence intervals of the original estimates. Also note that the
regression parameter estimates for Model E, fitted with data from 1950 onwards, are very similar to Model B. Thus we can
safely interpret the regression coefficient as mostly representing the effect of the variable they are associated with.