was fitted to the data, where gt has an ARMA structure (Shumway and Stoffer, 2011). The arima function in R
(http://www.r-project.org/) can be used to fit a regression model of the form (1). This model was used as it can accurately
represent residual autocorrelation in regression, and hence can provide a realistic stochastic representation of a time series
suitable for simulation, as well as an unbiased assessment of the statistical significance of regression coefficients compared
to those obtained from ordinary least squares regression (Maddala, 2001).
The model was selected according to a Bayesian Information Criterion (BIC) using the auto.arima function in R (Hyndman
and Khandakar, 2008) with default parameters and a search restricted to stationary models, and the entire time series from
1882 to 2010 was used. We start with the simple (stationary in this case) and only move to the complex (non-stationary) if
the performance of the simpler model was found to be wanting. The selected model has autoregressive terms of order 1, and
moving average terms of orders 1 and 2 (Table 2, Model A). However, the partial autocorrelation plot indicated strong autocorrelation
for a lag of 24 months. Therefore, moving average terms of lags 12 and 24, were added to the model (Table 2,
Model B). The residual for this model is of the form: