The goal of igcop is to provide computational tools for the Integrated Gamma (IG) and Integrated Gamma Limit (IGL) copula families.
igcop is not yet available on CRAN, but can be downloaded from this repository using devtools. Just execute this line of code in an R instance, after ensuring you have the devtools R package installed:
devtools::install_github("vincenzocoia/igcop")
The IG copula family is defined by parameters θ > 0 and α > 0, with the IGL copula family obtained with θ → ∞. So, the IGL copula family only has one parameter, α > 0. For a detailed definition, see Coia (2017). Note, however, that this package uses a different parameterization that is more computationally stable.
The IG and IGL copula families are unique in that they are not permutation symmetric. Also, when used in a regression context, the conditional distribution of the response (the 2nd copula variable) has an EVI that increases with the predictor for an IG copula, and reduces a heavy-tailed response to a light-tailed conditional distribution for an IGL copula.
This package piggybacks on the base R syntax for distributions, such as dnorm()
and pexp()
, whose functions adopt the convention:
<prefix><name>
For IG and IGL copulas:
<prefix>
corresponds to one of:
p
for cdf,d
for density (and logd
for log density),q
for quantile (for conditional distributions only), andr
for random number generation (not supported for conditional distributions).<name>
corresponds to the possible names:
ig
and igl
correspond to an IG copula and IGL copula, respectively.condig12
and condigl12
correspond to a conditional distribution of the first variable given the second, of an IG copula and IGL copula respectively.condig21
and condigl21
correspond to a conditional distribution of the second variable given the first, of an IG copula and IGL copula respectively (also available as condig
and condigl
to match the syntax of the CopulaModel package).Here are some examples, starting with evaluating the density of an IG copula at (0.3, 0.6):
Computations are vectorized over both u
and v
(first and second variables), along with the parameter values. Here’s the cdf and density of an IGL copula at different values:
u <- seq(0.1, 0.9, length.out = 9) v <- seq(0.9, 0.5, length.out = 9) pigl(u, v, alpha = 4) #> [1] 0.1000000 0.2000000 0.2999711 0.3988536 0.4888134 0.5508382 0.5683229 #> [8] 0.5447653 0.4998090 digl(0.2, v, alpha = u) #> [1] 0.8522462 0.8230206 0.8471676 0.8915708 0.9458967 1.0058156 1.0691273 #> [8] 1.1345476 1.2012456
It doesn’t make sense to talk about quantiles for a multivariate distribution, so this is only defined for conditional distributions.
Here is an example of a distribution given the first variable (“2 given 1”). Note that the “2 given 1” distributions swap the u
and v
arguments to better align with the conditioning, and you can either explicitly include the 21
suffix or not.
qcondig(v, u, theta = 5, alpha = 3) #> [1] 0.7435415 0.7228302 0.7121613 0.7073784 0.7056649 0.7039164 0.6972994 #> [8] 0.6777041 0.6356285 qcondig21(v, u, theta = 5, alpha = 3) #> [1] 0.7435415 0.7228302 0.7121613 0.7073784 0.7056649 0.7039164 0.6972994 #> [8] 0.6777041 0.6356285
Here is the corresponding “1 given 2” distribution. Since this is less common in regression scenarios, you have to explicitly add the 12
prefix for “1 given 2.”
qcondig12(v, u, theta = 5, alpha = 3) #> [1] 0.8896885 0.8114873 0.7297887 0.6598357 0.6097781 0.5811235 0.5749922 #> [8] 0.5976573 0.6689895
Generating 5 values from an IG copula:
rig(5, theta = 5, alpha = 4) #> # A tibble: 5 x 2 #> u v #> <dbl> <dbl> #> 1 0.474 0.568 #> 2 0.0510 0.0573 #> 3 0.209 0.564 #> 4 0.809 0.636 #> 5 0.115 0.452