face (version 0.1-7)

face.sparse: Fast covariance estimation for sparse functional data

Description

The function is to estimate the mean and covariance function from a cluster of functions/longitudinal observations.

Usage

face.sparse(data, newdata = NULL,
             center = TRUE, argvals.new = NULL,
             knots = 7, 
             p = 3, m = 2, lambda = NULL, lambda_mean = NULL,
             search.length = 14, 
             lower = -3, upper = 10, lower2 = -3, upper2 = 5,
             calculate.scores = FALSE,pve=0.99,two_step=FALSE)

Value

newdata

Input

y.pred,mu.pred,Chat.diag.pred, var.error.pred

Predicted/estimated objects at newdata$argvals

Theta

Estimated parameter matrix

argvals.new

Vector of time points to evaluate population parameters

mu.new, Chat.new,Cor.new,Cor.raw.new,Chat.raw.diag.new, var.error.new

Estimated objects at argvals.new

eigenfunctions, eigenvalues

Estimated eigenfunctions (scaled eigenvector) and eigenvalues at argvals.new

mu.hat,var.error.hat

Estimated objects at data$argvals

calculate.scores, rand_eff

if calculate.scores is TRUE (default to FALSE), then predicted scores rand_eff$scores will be calculated.

...

...

Arguments

data

a data frame with three arguments: (1) argvals: observation times; (2) subj: subject indices; (3) y: values of observations. Missing values not allowed.

newdata

of the same strucutre as data; defaults to NULL, then no prediction.

center

logical. If TRUE, then Pspline smoothing of the population mean will be conducted and subtracted from the data before covariance smoothing; if FALSE, then the population mean will be just 0s.

argvals.new

a vector of observation time points to evaluate mean function, covariance function, error variance and etc. If NULL, then 100 equidistant points in the range of data$argvals.

knots

the number of knots for B-spline basis functions to be used; defaults to 7. The resulting number of basis functions is the number of interior knots plus the degree of B-splines.

p

the degrees of B-splines; defaults to 3.

m

the order of differencing penalty; defaults to 2.

lambda

the value of the smoothing parameter for covariance smoothing; defaults to NULL.

lambda_mean

the value of the smoothing parameter for mean smoothing; defaults to NULL.

search.length

the number of equidistant (log scale) smoothing parameters to search; defaults to 14.

lower, upper

bounds for log smoothing parameter for first step of estimation; defaults are -3 and 10, respectively.

lower2, upper2

bounds for log smoothing parameter for second step of estimation; defaults are lower and 5, respectively.

calculate.scores

if TRUE, scores will be calculated.

pve

Defaults 0.99. To select the number of eigenvalues by percentage of variance.

two_step

if TRUE, a two-step estimation procedure will be applied.

Author

Luo Xiao <[e-mail]> and Cai Li <[e-mail]>

Details

This is a generalized version of bivariate P-splines (Eilers and Marx, 2003) for covariance smoothing of sparse functional or longitudinal data. It uses tensor product B-spline basis functions and employes a differencing penalty on the assosciated parameter matrix. The only smoothing parameter in the method is selected by leave-one-subject-out cross validation and is implemented with a fast algorithm.

There are two steps for estimation. During the first step, the objective function to minimize is the penalized least squares on empirical estimates of covariance function. During the second step, the covariance between the empirical estimates (depending on the estimates of covariance function) are accounted and thus a generalized penalized least squares are minimized.

If center is TRUE, then a population mean will be calculated and is smoothed by univariate P-spline smoothing:pspline (Eilers and Marx, 1996). This univariate smoothing uses leave-one-subject-out cross validation to select the smoothing parameter.

The knots are "equally-spaced", the differencing penalty in Eilers and Marx (2003) is used.

If the functional data are observed at the same grid for each function/curve and can be organized into a data matrix, then fpca.face in the package refund should instead be used. fpca.face allows a small percentage (less than 30 percent) of missing data in the data matrix.

References

Luo Xiao, Cai Li, William Checkley and Ciprian Crainiceanu, Fast covariance estimation for sparse functional data, Stat. Comput., tools:::Rd_expr_doi("doi:10.1007/s11222-017-9744-8").

Paul Eilers and Brian Marx, Multivariate calibration with temperature interaction using two-dimensional penalized signal regression, Chemometrics and Intelligent Laboratory Systems 66 (2003), 159-174.

Paul Eilers and Brian Marx, Flexible smoothing with B-splines and penalties, Statist. Sci., 11, 89-121, 1996.

Simon N. Wood, P-splines with derivative based penalties and tensor product smoothing of unevenly distributed data, Stat. Comput., tools:::Rd_expr_doi("10.1007/s11222-016-9666-x").

See Also

fpca.face and fpca.sc in refund

Examples

if (FALSE) {
##########################
#### CD4 data example
##########################

require(refund)
data(cd4)
n <- nrow(cd4)
Tt <- ncol(cd4)

id <- rep(1:n,each=Tt)
t <- rep(-18:42,times=n)
y <- as.vector(t(cd4))
sel <- which(is.na(y))


## organize data and apply FACEs
data <- data.frame(y=log(y[-sel]),
argvals = t[-sel],
subj = id[-sel])
data <- data[data$y>4.5,]
fit_face <- face.sparse(data,argvals.new=(-20:40))

## set calculate.scores to TRUE if want to get scores
fit_face <- face.sparse(data,argvals.new=(-20:40),calculate.scores=TRUE)
scores <- fit_face$rand_eff$scores

data.h <- data
tnew <- fit_face$argvals.new

## scatter plots
Xlab <- "Months since seroconversion"
Ylab <- "log (CD4 count)"
par(mfrow=c(1,1),mar = c(4.5,4.5,3,2))
id <- data.h$subj
uid <- unique(id)
plot(data.h$argvals,data.h$y,
type = "n", ylim = c(4.5,8),
xlab = Xlab, ylab = Ylab,
cex.lab = 1.25,cex.axis=1.25,cex.main = 1.25)

for(i in 1:10){
seq <- which(id==uid[i])
lines(data.h$argvals[seq],data.h$y[seq],lty=1,col="gray",lwd=1,type="l")
#points(data.h$argvals[seq],data.h$y[seq],col=1,lty=1,pch=1)
}

Sample <- seq(10,50,by=10)
for(i in Sample){
seq <- which(id==uid[i])
lines(data.h$argvals[seq],data.h$y[seq],lty=1,col="black",lwd=1,type="l")
}
lines(tnew,fit_face$mu.new,lwd=2,lty=2,col="red")

## plots of variance/correlation functions

Cov <- fit_face$Chat.new
Cov_diag <- diag(Cov)
Cor <- fit_face$Cor.new

par(mfrow=c(1,2),mar=c(4.5,4.1,3,4.5))


plot(tnew,Cov_diag,type="l",
xlab = Xlab, ylab="",main= "CD4: variance function",
#ylim = c(0.8,1.5),
cex.axis=1.25,cex.lab=1.25,cex.main=1.25,lwd=2)

require(fields)
image.plot(tnew,tnew,Cor,
xlab=Xlab, ylab = Xlab,
main = "CD4: correlation function",
cex.axis=1.25,cex.lab=1.25,cex.main=1.25,
axis.args = list(at = c(0,0.2,0.4,0.6,0.8,1.0)),
legend.shrink=0.75,legend.line=-1.5)


## prediction of several subjects

par(mfrow=c(2,2),mar=c(4.5,4.5,3,2))
Sample <- c(30,40,50,60)
for(i in 1:4){
sel <- which(id==uid[Sample[i]])
dati <- data.h[sel,]

seq <- -20:40
k <- length(seq)
dati_pred <- data.frame(y = rep(NA,nrow(dati) + k ),
argvals = c(rep(NA,nrow(dati)),seq),
subj=rep(dati$subj[1],nrow(dati) + k )
)

dati_pred[1:nrow(dati),] <- dati
yhat2 <- predict(fit_face,dati_pred)

data3 <- dati
Ylim <- range(c(data3$y,yhat2$y.pred))

plot(data3$argvals,data3$y,xlab=Xlab,ylab=Ylab, main = paste("Male ",i,sep=""),
ylim = c(4,8.5),
cex.lab=1.25,cex.axis = 1.25,cex.main = 1.25,pch=1,xlim=c(-20,40))

Ord <- nrow(dati) + 1:k
lines(dati_pred$argvals[Ord],yhat2$y.pred[Ord],col="red",lwd=2)
lines(dati_pred$argvals[Ord],
yhat2$y.pred[Ord] - 1.96*yhat2$se.pred[Ord], col="red",lwd=1,lty=2)
lines(dati_pred$argvals[Ord],
yhat2$y.pred[Ord] + 1.96*yhat2$se.pred[Ord], col="red",lwd=1,lty=2)

lines(tnew,fit_face$mu.new,lty=3,col="black",lwd=2)
legend("bottomleft",c("mean","prediction"),lty=c(3,1),col=1:2,lwd=2,bty="n")
}
}