The extremal index \(\theta\) is a
measure of the degree of local dependence in the extremes of a
stationary process. The exdex
package performs frequentist
inference about \(\theta\) using two
types of methodology.
One type (Northrop, 2015) is based on a model that relates the distribution of block maxima to the marginal distribution of the data, leading to a semiparametric maxima estimator. Two versions of this type of estimator are provided, following Northrop, 2015 and Berghaus and Bücher, 2018. A slightly modified version of the latter is also provided. Estimates are produced using both disjoint and sliding block maxima, the latter providing greater precision of estimation. A graphical block size diagnostic is provided.
The other type of methodology uses a model for the distribution of
threshold inter-exceedance times (Ferro and Segers,
2003). Three versions of this type of approach are provided: the
iterated weight least squares approach of Süveges (2007), the
\(K\)-gaps model of Süveges and Davison (2010)
and a similar approach of Holesovsky and Fusek
(2020) that we refer to as D-gaps. For the \(K\)-gaps and \(D\)-gaps models the exdex
package allows missing values in the data, can accommodate independent
subsets of data, such as monthly or seasonal time series from different
years, and can incorporate information from censored inter-exceedance
times. Graphical diagnostics for the threshold level and the respective
tuning parameters \(K\) and \(D\) are provided.
The following code estimates the extremal index using the
semiparametric maxima estimators, for an example dataset containing a
time series of sea surges measured at Newlyn, Cornwall, UK over the
period 1971-1976. The block size of 20 was chosen using a graphical
diagnostic provided by choose_b()
.
library(exdex)
<- spm(newlyn, 20)
theta
theta#>
#> Call:
#> spm(data = newlyn, b = 20)
#>
#> Estimates of the extremal index theta:
#> N2015 BB2018 BB2018b
#> sliding 0.2392 0.3078 0.2578
#> disjoint 0.2350 0.3042 0.2542
summary(theta)
#>
#> Call:
#> spm(data = newlyn, b = 20)
#>
#> Estimate Std. Error Bias adj.
#> N2015, sliding 0.2392 0.01990 0.003317
#> BB2018, sliding 0.3078 0.01642 0.003026
#> BB2018b, sliding 0.2578 0.01642 0.053030
#> N2015, disjoint 0.2350 0.02222 0.003726
#> BB2018, disjoint 0.3042 0.02101 0.003571
#> BB2018b, disjoint 0.2542 0.02101 0.053570
Now we estimate \(\theta\) using the
\(K\)-gaps model. The threshold \(u\) and runs parameter \(K\) were chosen using the graphical
diagnostic provided by choose_uk()
.
<- quantile(newlyn, probs = 0.60)
u <- kgaps(newlyn, u, k = 2)
theta
theta#>
#> Call:
#> kgaps(data = newlyn, u = u, k = 2)
#>
#> Estimate of the extremal index theta:
#> theta
#> 0.1758
summary(theta)
#>
#> Call:
#> kgaps(data = newlyn, u = u, k = 2)
#>
#> Estimate Std. Error
#> theta 0.1758 0.009211
To get the current released version from CRAN:
install.packages("exdex")
See vignette("exdex-vignette", package = "exdex")
for an
overview of the package.