Categorical Kernel Regression Spline Cross-Validation

Description

krscv computes exhaustive cross-validation directed search for a regression spline estimate of a one (1) dimensional dependent variable on an r-dimensional vector of continuous and nominal/ordinal (factor/ordered) predictors.

Usage

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
krscv(xz,
      y,
      degree.max = 10, 
      segments.max = 10,
      degree.min = 0,
      segments.min = 1, 
      restarts = 0,
      complexity = c("degree-knots","degree","knots"),
      knots = c("quantiles","uniform","auto"),
      basis = c("additive","tensor","glp","auto"),
      cv.func = c("cv.ls","cv.gcv","cv.aic"),
      degree = degree,
      segments = segments,
      tau = NULL,
      weights = NULL,
      singular.ok = FALSE)

Arguments

y

continuous univariate vector

xz

continuous and/or nominal/ordinal (factor/ordered) predictors

degree.max

the maximum degree of the B-spline basis for each of the continuous predictors (default degree.max=10)

segments.max

the maximum segments of the B-spline basis for each of the continuous predictors (default segments.max=10)

degree.min

the minimum degree of the B-spline basis for each of the continuous predictors (default degree.min=0)

segments.min

the minimum segments of the B-spline basis for each of the continuous predictors (default segments.min=1)

restarts

number of times to restart optim from different initial random values (default restarts=0) when searching for optimal bandwidths for the categorical predictors for each unique K combination (i.e.\ degree/segments)

complexity

a character string (default complexity="degree-knots") indicating whether model ‘complexity’ is determined by the degree of the spline or by the number of segments (‘knots’). This option allows the user to use cross-validation to select either the spline degree (number of knots held fixed) or the number of knots (spline degree held fixed) or both the spline degree and number of knots

knots

a character string (default knots="quantiles") specifying where knots are to be placed. ‘quantiles’ specifies knots placed at equally spaced quantiles (equal number of observations lie in each segment) and ‘uniform’ specifies knots placed at equally spaced intervals. If knots="auto", the knot type will be automatically determined by cross-validation

basis

a character string (default basis="additive") indicating whether the additive or tensor product B-spline basis matrix for a multivariate polynomial spline or generalized B-spline polynomial basis should be used. Note this can be automatically determined by cross-validation if cv=TRUE and basis="auto", and is an ‘all or none’ proposition (i.e. interaction terms for all predictors or for no predictors given the nature of ‘tensor products’). Note also that if there is only one predictor this defaults to basis="additive" to avoid unnecessary computation as the spline bases are equivalent in this case

cv.func

a character string (default cv.func="cv.ls") indicating which method to use to select smoothing parameters. cv.gcv specifies generalized cross-validation (Craven and Wahba (1979)), cv.aic specifies expected Kullback-Leibler cross-validation (Hurvich, Simonoff, and Tsai (1998)), and cv.ls specifies least-squares cross-validation

degree

integer/vector specifying the degree of the B-spline basis for each dimension of the continuous x

segments

integer/vector specifying the number of segments of the B-spline basis for each dimension of the continuous x (i.e. number of knots minus one)

tau

if non-null a number in (0,1) denoting the quantile for which a quantile regression spline is to be estimated rather than estimating the conditional mean (default tau=NULL)

weights

an optional vector of weights to be used in the fitting process. Should be ‘NULL’ or a numeric vector. If non-NULL, weighted least squares is used with weights ‘weights’ (that is, minimizing ‘sum(w*e^2)’); otherwise ordinary least squares is used.

singular.ok

a logical value (default singular.ok=FALSE) that, when FALSE, discards singular bases during cross-validation (a check for ill-conditioned bases is performed).

Details

krscv computes exhaustive cross-validation for a regression spline estimate of a one (1) dimensional dependent variable on an r-dimensional vector of continuous and nominal/ordinal (factor/ordered) predictors. The optimal K/lambda combination is returned along with other results (see below for return values). The method uses kernel functions appropriate for categorical (ordinal/nominal) predictors which avoids the loss in efficiency associated with sample-splitting procedures that are typically used when faced with a mix of continuous and nominal/ordinal (factor/ordered) predictors.

For the continuous predictors the regression spline model employs either the additive or tensor product B-spline basis matrix for a multivariate polynomial spline via the B-spline routines in the GNU Scientific Library (http://www.gnu.org/software/gsl/) and the tensor.prod.model.matrix function.

For the discrete predictors the product kernel function is of the ‘Li-Racine’ type (see Li and Racine (2007) for details).

For each unique combination of degree and segment, numerical search for the bandwidth vector lambda is undertaken using optim and the box-constrained L-BFGS-B method (see optim for details). The user may restart the optim algorithm as many times as desired via the restarts argument. The approach ascends from K=0 through degree.max/segments.max and for each value of K searches for the optimal bandwidths for this value of K. After the most complex model has been searched then the optimal K/lambda combination is selected. If any element of the optimal K vector coincides with degree.max/segments.max a warning is produced and the user ought to restart their search with a larger value of degree.max/segments.max.

Value

krscv returns a crscv object. Furthermore, the function summary supports objects of this type. The returned objects have the following components:

K

scalar/vector containing optimal degree(s) of spline or number of segments

K.mat

vector/matrix of values of K evaluated during search

restarts

number of restarts during search, if any

lambda

optimal bandwidths for categorical predictors

lambda.mat

vector/matrix of optimal bandwidths for each degree of spline

cv.func

objective function value at optimum

cv.func.vec

vector of objective function values at each degree of spline or number of segments in K.mat

Author(s)

Jeffrey S. Racine racinej@mcmaster.ca

References

Craven, P. and G. Wahba (1979), “Smoothing Noisy Data With Spline Functions,” Numerische Mathematik, 13, 377-403.

Hurvich, C.M. and J.S. Simonoff and C.L. Tsai (1998), “Smoothing Parameter Selection in Nonparametric Regression Using an Improved Akaike Information Criterion,” Journal of the Royal Statistical Society B, 60, 271-293.

Li, Q. and J.S. Racine (2007), Nonparametric Econometrics: Theory and Practice, Princeton University Press.

Ma, S. and J.S. Racine and L. Yang (under revision), “Spline Regression in the Presence of Categorical Predictors,” Journal of Applied Econometrics.

Ma, S. and J.S. Racine (2013), “Additive Regression Splines with Irrelevant Categorical and Continuous Regressors,” Statistica Sinica, Volume 23, 515-541.

See Also

loess, npregbw,

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
set.seed(42)
## Simulated data
n <- 1000

x <- runif(n)
z <- round(runif(n,min=-0.5,max=1.5))
z.unique <- uniquecombs(as.matrix(z))
ind <-  attr(z.unique,"index")
ind.vals <-  sort(unique(ind))
dgp <- numeric(length=n)
for(i in 1:nrow(z.unique)) {
  zz <- ind == ind.vals[i]
  dgp[zz] <- z[zz]+cos(2*pi*x[zz])
}
y <- dgp + rnorm(n,sd=.1)

xdata <- data.frame(x,z=factor(z))

## Compute the optimal K and lambda, determine optimal number of knots, set
## spline degree for x to 3

cv <- krscv(x=xdata,y=y,complexity="knots",degree=c(3))
summary(cv)