Multiple extreme weather event models
To quantify the correlation between the counts of the various stress
events we used the single event models to detrend the data for each
event metric and create a transformed data set which does not exhibit
spatio-temporal variability. Using the transformed data, the dependency
across the various event metrics was quantified using correlation. The
single event Poisson models were used to transform the original datays,t (for each stress) to the scale of a Gaussian
random variable with mean zero and variance one. At that scale, all
spatial and temporal variability has been factored out and the sample
correlations between the transformed counts for each event are estimates
of the dependency between each event.
A modified simulation technique was employed to sample from the
predictive distribution of the counts for each event, allowing for the
correlation between them. Firstly, we generated random samples of the
data at the detrended scale, respecting the correlation between the
event metrics at this scale. Then, we transformed these samples back to
original scale of the data to obtain a set of simulated counts in each
grid cell and year, thus maintaining both the spatio-temporal
variability in each event but also the correlation between event
metrics.
The thresholds were set as the sample mean of each event metric across
all grid cells and years. The joint probability that the annual mean
count of two or more event categories exceeds a particular threshold was
then determined. Comparison of differences in these probabilities
between 1961–1988 and 1989–2016 lie in the region between -1 and 1,
and conveys information about whether the risk of two or more stress
events occurring within one year has increased. Significant changes are
ones that are above 0.05 or below -0.05.