Description Usage Arguments Details Value References See Also Examples
This is the density function and random generation from the horseshoe distribution.
1 2 |
n |
This is the number of draws from the distribution. |
x |
This is a location vector at which to evaluate density. |
lambda |
This vector is a positive-only local parameter lambda. |
tau |
This scalar is a positive-only global parameter tau. |
sigma |
This scalar is a positive-only global parameter sigma. |
log |
Logical. If |
Application: Discrete Scale Mixture
Density: (see below)
Inventor: Carvalho et al. (2008)
Notation 1: theta ~ HS(lambda, tau, sigma)
Notation 2: p(theta) = HS(theta | lambda, tau, sigma)
Parameter 1: local scale lambda > 0
Parameter 2: global scale tau > 0
Parameter 3: global scale sigma > 0
Mean: E(theta)
Variance: var(theta)
Mode: mode(theta)
The horseshoe distribution (Carvalho et al., 2008) is a heavy-tailed discrete mixture distribution that can be considered a variance mixture, and it is in the family of multivariate scale mixtures of normals.
The horseshoe distribution was proposed as a prior distribution, and recommended as a default choice for shrinkage priors in the presence of sparsity. Horseshoe priors are most appropriate in large-p models where dimension reduction is necessary to avoid overly complex models that predict poorly, and also perform well in estimating a sparse covariance matrix via Cholesky decomposition (Carvalho et al., 2009).
When the number of parameters in variable selection is assumed to be sparse, meaning that most elements are zero or nearly zero, a horseshoe prior is a desirable alternative to the Laplace-distributed parameters in the LASSO, or the parameterization in ridge regression. When the true value is far from zero, the horseshoe prior leaves the parameter unshrunk. Yet, the horseshoe prior is accurate in shrinking parameters that are truly zero or near-zero. Parameters near zero are shrunk more than parameters far from zero. Therefore, parameters far from zero experience less shrinkage and are closer to their true values. The horseshoe prior is valuable in discriminating signal from noise.
The horseshoe distribution is the following discrete mixture:
p(theta | lambda) ~ N(0, lambda^2)
p(lambda | tau) ~ HC(tau)
p(tau) ~ HC(sigma)
where lambda is a vector of local shrinkage parameters, and tau and sigma are global shrinkage parameters.
By replacing the Laplace-distributed parameters in LASSO with horseshoe-distributed parameters, the result is called horseshoe regression.
dhs
gives the density and
rhs
generates random deviates.
Carvalho, C.M., Polson, N.G., and Scott, J.G. (2008). "The Horseshoe Estimator for Sparse Signals". Discussion Paper 2008-31. Duke University Department of Statistical Science.
Carvalho, C.M., Polson, N.G., and Scott, J.G. (2009). "Handling Sparsity via the Horseshoe". Journal of Machine Learning Research, 5, p. 73–80.
1 2 3 4 5 6 7 8 9 10 |
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.