piqr: Penalized Quantile Regression Coefficients Modeling

View source: R/qrcmPen.R

piqrR Documentation

Penalized Quantile Regression Coefficients Modeling

Description

This package implements a penalized Frumento and Bottai's (2016) method for quantile regression coefficient modeling (qrcm), in which quantile regression coefficients are described by (flexible) parametric functions of the order of the quantile. This package fits lasso qrcm using pathwise coordinate descent algorithm.

Usage

piqr(formula, formula.p = ~ slp(p, 3), weights, data, s, nlambda=100,
     lambda.min.ratio=ifelse(nobs<nvars, 0.01, 0.0001), lambda,
     tol=1e-6, maxit=100, display=TRUE)

Arguments

formula

a two-sided formula of the form y ~ x1 + x2 + ...: a symbolic description of the quantile regression model.

formula.p

a one-sided formula of the form ~ b1(p, ...) + b2(p, ...) + ..., describing how quantile regression coefficients depend on p, the order of the quantile.

weights

an optional vector of weights to be used in the fitting process.

data

an optional data frame, list or environment containing the variables in formula.

s

an optional 0/1 matrix that permits excluding some model coefficients (see ‘Examples’).

nlambda

the number of lambda values - default is 100.

lambda.min.ratio

Smallest value for lambda, as a fraction of lambda.max. The default depends on the sample size nobs relative to the number of variables nvars. If nobs > nvars, the default is 0.0001, close to zero. If nobs < nvars, the default is 0.01.

lambda

A user supplied lambda sequence.

display

if TRUE something is printed - default is TRUE.

tol

convergence criterion for numerical optimization - default is 1e-6.

maxit

maximum number of iterations - default is 100.

Details

Quantile regression permits modeling conditional quantiles of a response variabile, given a set of covariates. A linear model is used to describe the conditional quantile function:

Q(p | x) = \beta_0(p) + \beta_1(p)x_1 + \beta_2(p)x_2 + \ldots.

The model coefficients \beta(p) describe the effect of covariates on the p-th quantile of the response variable. Usually, one or more quantiles are estimated, corresponding to different values of p.

Assume that each coefficient can be expressed as a parametric function of p of the form:

\beta(p | \theta) = \theta_{0} + \theta_1 b_1(p) + \theta_2 b_2(p) + \ldots

where b_1(p), b_2(p, \ldots) are known functions of p. If q is the dimension of x = (1, x_1, x_2, \ldots) and k is that of b(p) = (1, b_1(p), b_2(p), \ldots), the entire conditional quantile function is described by a q \times k matrix \theta of model parameters.

Users are required to specify two formulas: formula describes the regression model, while formula.p identifies the 'basis' b(p). By default, formula.p = ~ slp(p, k = 3), a 3rd-degree shifted Legendre polynomial (see slp). Any user-defined function b(p, \ldots) can be used, see ‘Examples’.

Estimation of penalized \theta is carried out by minimizing a penalized integrated loss function, corresponding to the integral, over p, of the penalized loss function of standard quantile regression. This motivates the acronym piqr (penalized integrated quantile regression).

See details in iqr

Value

An object of class “piqr”, a list containing the following items:

call

the matched call.

lambda

The actual sequence of lambda values used.

coefficients

a list of estimated model parameters describing the fitted quantile function along the path.

minimum

the value of the minimized integrated loss function for each value of lambda.

dl

a matrix of gradient values along the path.

df

The number of nonzero coefficients for each value of lambda.

seqS

a list containg each matrix s for each value of lambda.

internal

a list containing some initial object.

Note

By expressing quantile regression coefficients as functions of p, a parametric model for the conditional quantile function is specified. The induced PDF and CDF can be used as diagnostic tools. Negative values of PDF indicate quantile crossing, i.e., the conditional quantile function is not monotonically increasing. Null values of PDF indicate observations that lie outside the estimated support of the data, defined by quantiles of order 0 and 1. If null or negative PDF values occur for a relatively large proportion of data, the model is probably misspecified or ill-defined. If the model is correct, the fitted CDF should approximately follow a Uniform(0,1) distribution. This idea is used to implement a goodness-of-fit test, see summary.iqr and test.fit.

The intercept can be excluded from formula, e.g., iqr(y ~ -1 + x). This, however, implies that when x = 0, y is always 0. See example 5 in ‘Examples’. The intercept can also be removed from formula.p. This is recommended if the data are bounded. For example, for strictly positive data, use iqr(y ~ 1, formula.p = -1 + slp(p,3)) to force the smallest quantile to be zero. See example 6 in ‘Examples’.

Author(s)

Gianluca Sottile gianluca.sottile@unipa.it

References

Sottile G, Frumento P, Chiodi M, Bottai M. (2020). A penalized approach to covariate selection through quantile regression coefficient models. Statistical Modelling, 20(4), pp 369-385. doi:10.1177/1471082X19825523.

Frumento, P., and Bottai, M. (2016). Parametric modeling of quantile regression coefficient functions. Biometrics, 72(1), pp 74-84, doi:10.1111/biom.12410.

Friedman, J., Hastie, T. and Tibshirani, R. (2008). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software, Vol. 33(1), pp 1-22 Feb 2010.

See Also

summary.piqr, plot.piqr, predict.piqr, for summary, plotting, and prediction. gof.piqr to select the best value of the tuning parameter though AIC, BIC, GIC, GCV criteria.

Examples


  ##### Using simulated data in all examples

  ##### Example 1
  set.seed(1234)
  n <- 300
  x1 <- rexp(n)
  x2 <- runif(n, 0, 5)
  x <- cbind(x1,x2)

  b <- function(p){matrix(cbind(1, qnorm(p), slp(p, 2)), nrow=4, byrow=TRUE)}
  theta <- matrix(0, nrow=3, ncol=4); theta[, 1] <- 1; theta[1,2] <- 1; theta[2:3,3] <- 2
  qy <- function(p, theta, b, x){rowSums(x * t(theta %*% b(p)))}

  y <- qy(runif(n), theta, b, cbind(1, x))

  s <- matrix(1, nrow=3, ncol=4); s[1,3:4] <- 0; s[2:3, 2] <- 0
  obj <- piqr(y ~ x1 + x2, formula.p = ~ I(qnorm(p)) + slp(p, 2), s=s, nlambda=50)

  best <- gof.piqr(obj, method="AIC", plot=FALSE)
  best2 <- gof.piqr(obj, method="BIC", plot=FALSE)

  summary(obj, best$posMinLambda)
  summary(obj, best2$posMinLambda)

  ## Not run: 
  ##### other examples
  set.seed(1234)
  n <- 1000
  q <- 5
  k <- 3
  X <- matrix(abs(rnorm(n*q)), n, q)
  rownames(X) <- 1:n
  colnames(X) <- paste0("X", 1:q)
  theta <- matrix(c(3, 1.5, 1, 1,
                    2, 1, 1, 1,
                    0, 0, 0, 0,
                    0, 0, 0, 0,
                    1.5, 1, 1, 1,
                    0, 0, 0, 0),
                  ncol=(k+1), byrow=TRUE)
  rownames(theta) <- c("(intercept)", paste0("X", 1:q))
  colnames(theta) <- c("(intercept)", "slp(p,1)", "slp(p,2)", "slp(p,3)")
  B <- function(p, k){matrix(cbind(1, slp(p, k)), nrow=(k+1), byrow=TRUE)}
  Q <- function(p, theta, B, k, X){rowSums(X * t(theta %*% B(p, k)))}

  pp <- runif(n)
  y <- Q(p=pp, theta=theta, B=B, k=k, X=cbind(1, X))
  m1 <- piqr(y ~ X, formula.p = ~ slp(p, k))
  best1 <- gof.piqr(m1, method="AIC", plot=FALSE)
  best2 <- gof.piqr(m1, method="BIC", plot=FALSE)
  summary(m1, best1$posMinLambda)
  summary(m1, best2$posMinLambda)
  par(mfrow = c(1,3)); plot(m1, xvar="lambda");
                       plot(m1, xvar="objective"); plot(m1, xvar="grad")

  set.seed(1234)
  n <- 1000
  q <- 6
  k <- 4
  # x <- runif(n)
  X <- matrix(abs(rnorm(n*q)), n, q)
  rownames(X) <- 1:n
  colnames(X) <- paste0("X", 1:q)
  theta <- matrix(c(1, 2, 0, 0, 0,
                    2, 0, 1, 0, 0,
                    0, 0, 0, 0, 0,
                    1, 0, 0, 1, -1.2,
                    0, 0, 0, 0, 0,
                    1.5, 0, .5, 0, 0,
                    0, 0, 0, 0, 0),
                  ncol=(k+1), byrow=TRUE)
  rownames(theta) <- c("(intercept)", paste0("X", 1:q))
  colnames(theta) <- c("(intercept)", "qnorm(p)", "p", "log(p)", "log(1-p)")
  B <- function(p, k){matrix(cbind(1, qnorm(p), p, log(p), log(1-p)), nrow=(k+1), byrow=TRUE)}
  Q <- function(p, theta, B, k, X){rowSums(X * t(theta %*% B(p, k)))}

  pp <- runif(n)
  y <- Q(p=pp, theta=theta, B=B, k=k, X=cbind(1, X))
  s <- matrix(1, q+1, k+1); s[2:(q+1), 2] <- 0; s[1, 3:(k+1)] <- 0; s[2:3, 4:5] <- 0
  s[4:5, 3] <- 0; s[6:7, 4:5] <- 0
  m2 <- piqr(y ~ X, formula.p = ~ qnorm(p) + p + I(log(p)) + I(log(1-p)), s=s)
  best1 <- gof.piqr(m2, method="AIC", plot=FALSE)
  best2 <- gof.piqr(m2, method="BIC", plot=FALSE)
  summary(m2, best1$posMinLambda)
  summary(m2, best2$posMinLambda)
  par(mfrow = c(1,3)); plot(m2, xvar="lambda");
                       plot(m2, xvar="objective"); plot(m2, xvar="grad")
  
## End(Not run)
  # see the documentation for 'summary.piqr', and 'plot.piqr'

qrcmNP documentation built on May 29, 2024, 8:29 a.m.