scam (version 1.2-20)

scam.check: Some diagnostics for a fitted scam object

Description

The function takes a fitted scam object produced by scam() and provides diagnostic information about the model fitting procedure and results. It is analogous to the gam.check() function from the mgcv package. By default, the function produces three deviance residual plots and a plot of the response versus fitted values, and prints information about the convergence of the smoothness selection optimization. Alternatively, the function can generate plots for checking the normalized (randomized) quantile residuals of the fitted scam object.

Usage

scam.check(b,type=c("deviance","rquantile","pearson","response"),old.style=FALSE, 
        pch=".", setseed=NULL, rep=0, level=.9, rl.col=3, rep.col="gray80",...)

Arguments

b

a fitted scam object as produced by scam().

old.style

produces qq-norm plots as it was in scam versions < 1.2-15 when set to TRUE.

type

type of residuals, see residuals.scam, used in all plots.

rep, level, rep.col

arguments passed to qq.scam() when old.style is FALSE (default).

rl.col

color for the reference line on the quantile-quantile plot.

pch

plot character to use for the quantile-quantile plot.

setseed

seed to be used in getting randomized quantile residuals.

...

extra graphics parameters to pass to plotting functions.

Author

Natalya Pya based partly on mgcv by Simon N Wood

Details

scam.check produces four standard residual plots, and some convergence diagnostics. By default, the residuals are deviance residuals. The QQ plot in this case is created by a call to qq.scam(), and plots the deviance residuals against approximate theoretical quantilies of the deviance residual distribution, according to the fitted model.

The type of the residuals can be changed to "rquantile", (randomized) quantile residuals as described by Dunn and Smyth (1996). For models with non-normal response distribution, (randomized) quantile residuals may provide a better diagnostic tool. The QQ plot for these residuals is a standard normal QQ plot. Randomization is applied only when the response variable is discrete.

The printed output includes information about the optimization process used for smoothing parameter selection.

References

Wood S.N. (2006) Generalized Additive Models: An Introduction with R. Chapman and Hall/CRC Press.

Dunn, P. K. and Smyth, G. K. (1996) Randomised quantile residuals, J. Comput. Graph. Statist., 5, 236--244

See Also

residuals.scam

Examples

library(scam)
set.seed(2)
n <- 200
x1 <- runif(n)*4-1; x2 <- runif(n)*3-1;
f1 <- function(x) exp(4*x)/(1+exp(4*x)) # monotone increasing smooth
f2 <- function(x) exp(-3*x)/15  # monotone decreasing and convex smooth
f <- f1(x1) + f2(x2)
y <- f+ rnorm(n)*0.2
dat <- data.frame(x1=x1,x2=x2,y=y)
b <- scam(y~ s(x1,bs="mpi")+s(x2,bs="mdcx"),data=dat)
plot(b,pages=1)
scam.check(b)
 
 # \donttest{
## binomial model... 
set.seed(4)
n <- 400
scale <- 3.3
x0 <- runif(n, 0, 1)
f0 <- function(x) 2 * sin(pi * x)
x1 <- runif(n)*4-1
f1 <- function(x) exp(4*x)/(1+exp(4*x)) 
f <- f0(x0) + f1(x1) 
f <- (f-2)*scale
g <- binomial()$linkinv(f)
y <- rbinom(g,1,g)
dat <- data.frame(y=y,x0=x0,x1=x1)

b1 <- scam(y~s(x0)+s(x1,bs="mpi"),family=binomial,data=dat)
scam.check(b1)
## residuals plots using randomized quantile residuals...
scam.check(b1, type="rquantile")

## Residuals vs linear predictor for four realizations of the quantile residuas...
old.par <- par(mfrow=c(2,2))
plot(b1$linear.predictors,residuals(b1, type="rquantile"))
plot(b1$linear.predictors,residuals(b1, type="rquantile"))
plot(b1$linear.predictors,residuals(b1, type="rquantile"))
plot(b1$linear.predictors,residuals(b1, type="rquantile"))
par(old.par)
# }