SSGLspr: Do sparse additive regression with splines using spike and...
In jantonelli111/SSGL: Spike and Slab Group Lasso

Description Usage Arguments Value Examples

View source: R/Main.R

This function takes in an outcome and covariates, and fits a semiparametric regression model with additive functions so that the effect of each covariate is modeled with a spline basis representation

1
2
3

SSGLspr(Y, x, xnew = NULL, DF = 2, lambda1 = 1, lambda0seq = seq(1,
  100, by = 1), a = 1, b = dim(x)[2], nFolds = 10,
  updateSigma = TRUE, M = 10, error = 0.001, forceGroups = c())

`Y`	The outcome to be analyzed
`x`	An n by p matrix of covariates
`xnew`	A new n2 by p matrix of covariates at which the outcome will be predicted
`DF`	Degrees of freedom of the splines for each covariate
`lambda1`	Prior parameter for the slab component of the prior
`lambda0seq`	Sequence of lambda0 values to consider
`a`	First hyperparameter for the beta prior denoting the prior probability of being in the slab
`b`	Second hyperparameter for the beta prior denoting the prior probability of being in the slab
`nFolds`	The number of folds to run cross validation on
`M`	Positive number less than p indicating how often to update theta and sigma. There is no need to change this unless trying to optimize computation time
`error`	The l2 norm difference that constitutes when the algorithm has converged
`forceGroups`	A vector containing the indices of any groups you wish to automatically include in the model and not penalize
`UpdateSigma`	True or False indicating whether to estimate the residual variance

A list of values containing the predictions at the observed locations, predictions at the new locations xnew, a vector indicating which covariates have nonzero coefficients, an n by p matrix of predicted f(x) values for each covariate, an n2 by p matrix of predicted f(xnew) values for each covariate, and an estimate of the residual variation in the model.

## Here we generate 200 samples from 100 covariates
n = 200
n2 = 100
G = 100
x = mvtnorm::rmvnorm(n, sigma=diag(G))
xnew = mvtnorm::rmvnorm(n2, sigma=diag(G))
Ynew = 200 + xnew[,1] + xnew[,2] + 0.6*xnew[,2]^2 + rnorm(n2, sd=1)

modSSGLspr = SSGLspr(Y=Y, x=x, xnew = xnew, lambda1=.1, DF=2)
mean((modSSGLspr$predY - Y)^2)
mean((modSSGLspr$predYnew - Ynew)^2)
modSSGLspr$nonzero

plot(xnew[,2], modSSGLspr$fxnew[,2])
points(xnew[,2], xnew[,2] + 0.6*xnew[,2]^2 - 
         mean(xnew[,2] + 0.6*xnew[,2]^2), col=2)
legend("top", c("Estimated", "Truth"),col=1:2, pch=1)