Generalized additive models with integrated smoothness estimation

Description

Fits a generalized additive model (GAM) to data. The degree of smoothness of model terms is estimated as part of fitting; isotropic or scale invariant smooths of any number of variables are available as model terms; confidence/credible intervals are readily available for any quantity predicted using a fitted model; gam is extendable: i.e. users can add smooths.

Smooth terms are represented using penalized regression splines (or similar smoothers) with smoothing parameters selected by GCV/UBRE or by regression splines with fixed degrees of freedom (mixtures of the two are permitted). Multi-dimensional smooths are available using penalized thin plate regression splines (isotropic) or tensor product splines (when an isotropic smooth is inappropriate). For more on specifying models see gam.models. For more on model selection see gam.selection. For faster fits use the "cr" bases for smooth terms, te smooths for smooths of several variables, and performance iteration for smoothing parameter estimation (see gam.method). For large datasets see warnings.

gam() is not a clone of what S-PLUS provides: the major differences are (i) that by default estimation of the degree of smoothness of model terms is part of model fitting, (ii) a Bayesian approach to variance estimation is employed that makes for easier confidence interval calculation (with good coverage probabilites) and (iii) the facilities for incorporating smooths of more than one variable are different: specifically there are no lo smooths, but instead (a) s terms can have more than one argument, implying an isotropic smooth and (b) te smooths are provided as an effective means for modelling smooth interactions of any number of variables via scale invariant tensor product smooths. If you want a clone of what S-PLUS provides use gam from package gam.

Usage

gam(formula,family=gaussian(),data=list(),weights=NULL,subset=NULL,
    na.action,offset=NULL,control=gam.control(),method=gam.method(),
    scale=0,knots=NULL,sp=NULL,min.sp=NULL,H=NULL,gamma=1,
    fit=TRUE,G=NULL,...)

Arguments

Details

A generalized additive model (GAM) is a generalized linear model (GLM) in which the linear predictor is given by a user specified sum of smooth functions of the covariates plus a conventional parametric component of the linear predictor. A simple example is:

log(E(y_i))=f_1(x_1i)+f_2(x_2i)

where the (independent) response variables y_i~Poi, and f_1 and f_2 are smooth functions of covariates x_1 and x_2. The log is an example of a link function.

If absolutely any smooth functions were allowed in model fitting then maximum likelihood estimation of such models would invariably result in complex overfitting estimates of f_1 and f_2. For this reason the models are usually fit by penalized likelihood maximization, in which the model (negative log) likelihood is modified by the addition of a penalty for each smooth function, penalizing its `wiggliness'. To control the tradeoff between penalizing wiggliness and penalizing badness of fit each penalty is multiplied by an associated smoothing parameter: how to estimate these parameters, and how to practically represent the smooth functions are the main statistical questions introduced by moving from GLMs to GAMs.

The mgcv implementation of gam represents the smooth functions using penalized regression splines, and by default uses basis functions for these splines that are designed to be optimal, given the number basis functions used. The smooth terms can be functions of any number of covariates and the user has some control over how smoothness of the functions is measured.

gam in mgcv solves the smoothing parameter estimation problem by using the Generalized Cross Validation (GCV) criterion

n D/(n - DoF)^2

or an Un-Biased Risk Estimator (UBRE )criterion

D/n + 2 s DoF / n -s

where D is the deviance, n the number of data, s the scale parameter and DoF the effective degrees of freedom of the model. Notice that UBRE is effectively just AIC rescaled, but is only used when s is known. It is also possible to replace D by the Pearson statistic (see gam.method), but this can lead to over smoothing. Smoothing parameters are chosen to minimize the GCV or UBRE score for the model, and the main computational challenge solved by the mgcv package is to do this efficiently and reliably. Various alternative numerical methods are provided: see gam.method.

Broadly gam works by first constructing basis functions and one or more quadratic penalty coefficient matrices for each smooth term in the model formula, obtaining a model matrix for the strictly parametric part of the model formula, and combining these to obtain a complete model matrix (/design matrix) and a set of penalty matrices for the smooth terms. Some linear identifiability constraints are also obtained at this point. The model is fit using gam.fit, a modification of glm.fit. The GAM penalized likelihood maximization problem is solved by penalized Iteratively Reweighted Least Squares (IRLS) (see e.g. Wood 2000). At each iteration a penalized weighted least squares problem is solved, and the smoothing parameters of that problem are estimated by GCV or UBRE. Eventually both model parameter estimates and smoothing parameter estimates converge. Alternatively the P-IRLS scheme is iterated to convergence for each trial set of smoothing parameters, and GCV or UBRE scores are only evaluated on convergence - optimization is then `outer' to the P-IRLS loop: in this case extra derivatives have to be carried along with the P-IRLS iteration, to facilitate optimization, and gam.fit2 is used.

Five alternative basis-penalty types are built in for representing model smooths, but alternatives can easily be added (see smooth.construct which uses p-splines to illustrate how to add new smooths). The built in alternatives for univariate smooths terms are: a conventional penalized cubic regression spline basis, parameterized in terms of the function values at the knots; a cyclic cubic spline with a similar parameterization and thin plate regression splines. The cubic spline bases are computationally very efficient, but require `knot' locations to be chosen (automatically by default). The thin plate regression splines are optimal low rank smooths which do not have knots, but are more computationally costly to set up. Smooths of several variables can be represented using thin plate regression splines, or tensor products of any available basis including user defined bases (tensor product penalties are obtained automatically form the marginal basis penalties). The t.p.r.s. basis is isotropic, so if this is not appropriate tensor product terms should be used. Tensor product smooths have one penalty and smoothing parameter per marginal basis, which means that the relative scaling of covariates is essentially determined automatically by GCV/UBRE. The t.p.r.s. basis and cubic regression spline bases are both available with either conventional `wiggliness penalties' or penalties augmented with a shrinkage component: the conventional penalties treat some space of functions as `completely smooth' and do not penalize such functions at all; the penalties with extra shrinkage will zero a term altogether for high enough smoothing parameters: gam.selection has an example of the use of such terms.

For any basis the user specifies the dimension of the basis for each smooth term. The dimension of the basis is one more than the maximum degrees of freedom that the term can have, but usually the term will be fitted by penalized maximum likelihood estimation and the actual degrees of freedom will be chosen by GCV. However, the user can choose to fix the degrees of freedom of a term, in which case the actual degrees of freedom will be one less than the basis dimension.

Thin plate regression splines are constructed by starting with the basis for a full thin plate spline and then truncating this basis in an optimal manner, to obtain a low rank smoother. Details are given in Wood (2003). One key advantage of the approach is that it avoids the knot placement problems of conventional regression spline modelling, but it also has the advantage that smooths of lower rank are nested within smooths of higher rank, so that it is legitimate to use conventional hypothesis testing methods to compare models based on pure regression splines. The t.p.r.s. basis can become expensive to calculate for large datasets. In this case the user can supply a reduced set of knots to use in basis construction (see knots, in the argument list), or use tensor products of cheaper bases.

In the case of the cubic regression spline basis, knots of the spline are placed evenly throughout the covariate values to which the term refers: For example, if fitting 101 data with an 11 knot spline of x then there would be a knot at every 10th (ordered) x value. The parameterization used represents the spline in terms of its values at the knots. The values at neighbouring knots are connected by sections of cubic polynomial constrained to be continuous up to and including second derivative at the knots. The resulting curve is a natural cubic spline through the values at the knots (given two extra conditions specifying that the second derivative of the curve should be zero at the two end knots). This parameterization gives the parameters a nice interpretability.

Details of the underlying fitting methods are given in Wood (2000, 2004a).

Value

If fit == FALSE the function returns a list G of items needed to fit a GAM, but doesn't actually fit it.
Otherwise the function returns an object of class "gam" as described in gamObject.

WARNINGS

If fit method "mgcv" is selected, the code does not check for rank deficiency of the model matrix that may result from lack of identifiability between the parametric and smooth components of the model.

You must have more unique combinations of covariates than the model has total parameters. (Total parameters is sum of basis dimensions plus sum of non-spline terms less the number of spline terms).

Automatic smoothing parameter selection is not likely to work well when fitting models to very few response data.

With large datasets (more than a few thousand data) the "tp" basis gets very slow to use: use the knots argument as discussed above and shown in the examples. Alternatively, for 1-d smooths you can use the "cr" basis and for multi-dimensional smooths use te smooths.

For data with many zeroes clustered together in the covariate space it is quite easy to set up GAMs which suffer from identifiability problems, particularly when using Poisson or binomial families. The problem is that with e.g. log or logit links, mean value zero corresponds to an infinite range on the linear predictor scale. Some regularization is possible in such cases: see gam.control for details.

Author(s)

Simon N. Wood simon.wood@r-project.org

Front end design inspired by the S function of the same name based on the work of Hastie and Tibshirani (1990). Underlying methods owe much to the work of Wahba (e.g. 1990) and Gu (e.g. 2002).

References

Key References on this implementation:

Wood, S.N. (2000) Modelling and Smoothing Parameter Estimation with Multiple Quadratic Penalties. J.R.Statist.Soc.B 62(2):413-428

Wood, S.N. (2003) Thin plate regression splines. J.R.Statist.Soc.B 65(1):95-114

Wood, S.N. (2004a) Stable and efficient multiple smoothing parameter estimation for generalized additive models. J. Amer. Statist. Ass. 99:673-686

Wood, S.N. (2004b) On confidence intervals for GAMs based on penalized regression splines. Technical Report 04-12 Department of Statistics, University of Glasgow.

Wood, S.N. (2004c) Low rank scale invariant tensor product smooths for generalized additive mixed models. Technical Report 04-13 Department of Statistics, University of Glasgow.

Key Reference on GAMs and related models:

Hastie (1993) in Chambers and Hastie (1993) Statistical Models in S. Chapman and Hall.

Hastie and Tibshirani (1990) Generalized Additive Models. Chapman and Hall.

Wahba (1990) Spline Models of Observational Data. SIAM

Background References:

Green and Silverman (1994) Nonparametric Regression and Generalized Linear Models. Chapman and Hall.

Gu and Wahba (1991) Minimizing GCV/GML scores with multiple smoothing parameters via the Newton method. SIAM J. Sci. Statist. Comput. 12:383-398

Gu (2002) Smoothing Spline ANOVA Models, Springer.

O'Sullivan, Yandall and Raynor (1986) Automatic smoothing of regression functions in generalized linear models. J. Am. Statist.Ass. 81:96-103

Wood (2001) mgcv:GAMs and Generalized Ridge Regression for R. R News 1(2):20-25

Wood and Augustin (2002) GAMs with integrated model selection using penalized regression splines and applications to environmental modelling. Ecological Modelling 157:157-177

http://www.stats.gla.ac.uk/~simon/

See Also

gamObject, gam.models, s, predict.gam, plot.gam, summary.gam, gam.side, gam.selection,mgcv, gam.control gam.check, gam.neg.bin, magic,vis.gam

Examples

library(mgcv)
set.seed(0) 
n<-400
sig<-2
x0 <- runif(n, 0, 1)
x1 <- runif(n, 0, 1)
x2 <- runif(n, 0, 1)
x3 <- runif(n, 0, 1)
f0 <- function(x) 2 * sin(pi * x)
f1 <- function(x) exp(2 * x)
f2 <- function(x) 0.2*x^11*(10*(1-x))^6+10*(10*x)^3*(1-x)^10
f3 <- function(x) 0*x
f <- f0(x0) + f1(x1) + f2(x2)
e <- rnorm(n, 0, sig)
y <- f + e
b<-gam(y~s(x0)+s(x1)+s(x2)+s(x3))
summary(b)
plot(b,pages=1,residuals=TRUE)
# same fit in two parts .....
G<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),fit=FALSE)
b<-gam(G=G)
# an extra ridge penalty (useful with convergence problems) ....
bp<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),H=diag(0.5,37)) 
print(b);print(bp);rm(bp)
# set the smoothing parameter for the first term, estimate rest ...
bp<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),sp=c(0.01,-1,-1,-1))
plot(bp,pages=1);rm(bp)
# set lower bounds on smoothing parameters ....
bp<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),min.sp=c(0.001,0.01,0,10)) 
print(b);print(bp);rm(bp)

# now a GAM with 3df regression spline term & 2 penalized terms
b0<-gam(y~s(x0,k=4,fx=TRUE,bs="tp")+s(x1,k=12)+s(x2,k=15))
plot(b0,pages=1)
# now fit a 2-d term to x0,x1
b1<-gam(y~s(x0,x1)+s(x2)+s(x3))
par(mfrow=c(2,2))
plot(b1)
par(mfrow=c(1,1))

# now simulate poisson data
g<-exp(f/4)
y<-rpois(rep(1,n),g)
b2<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),family=poisson)
plot(b2,pages=1)
# repeat fit using performance iteration
gm <- gam.method(gam="perf.magic")
b3<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),family=poisson,method=gm)
plot(b3,pages=1)

# a binary example 
g <- (f-5)/3
g <- binomial()$linkinv(g)
y <- rbinom(g,1,g)
lr.fit <- gam(y~s(x0)+s(x1)+s(x2)+s(x3),family=binomial)
## plot model components with truth overlaid in red
op <- par(mfrow=c(2,2))
for (k in 1:4) {
  plot(lr.fit,residuals=TRUE,select=k)
  xx <- sort(eval(parse(text=paste("x",k-1,sep=""))))
  ff <- eval(parse(text=paste("f",k-1,"(xx)",sep="")))
  lines(xx,(ff-mean(ff))/3,col=2)
}
par(op)
anova(lr.fit)
lr.fit1 <- gam(y~s(x0)+s(x1)+s(x2),family=binomial)
lr.fit2 <- gam(y~s(x1)+s(x2),family=binomial)
AIC(lr.fit,lr.fit1,lr.fit2)

# and a pretty 2-d smoothing example....
test1<-function(x,z,sx=0.3,sz=0.4)  
{ (pi**sx*sz)*(1.2*exp(-(x-0.2)^2/sx^2-(z-0.3)^2/sz^2)+
  0.8*exp(-(x-0.7)^2/sx^2-(z-0.8)^2/sz^2))
}
n<-500
old.par<-par(mfrow=c(2,2))
x<-runif(n);z<-runif(n);
xs<-seq(0,1,length=30);zs<-seq(0,1,length=30)
pr<-data.frame(x=rep(xs,30),z=rep(zs,rep(30,30)))
truth<-matrix(test1(pr$x,pr$z),30,30)
contour(xs,zs,truth)
y<-test1(x,z)+rnorm(n)*0.1
b4<-gam(y~s(x,z))
fit1<-matrix(predict.gam(b4,pr,se=FALSE),30,30)
contour(xs,zs,fit1)
persp(xs,zs,truth)
vis.gam(b4)
par(old.par)
# very large dataset example using knots
n<-10000
x<-runif(n);z<-runif(n);
y<-test1(x,z)+rnorm(n)
ind<-sample(1:n,1000,replace=FALSE)
b5<-gam(y~s(x,z,k=50),knots=list(x=x[ind],z=z[ind]))
vis.gam(b5)
# and a pure "knot based" spline of the same data
b6<-gam(y~s(x,z,k=100),knots=list(x= rep((1:10-0.5)/10,10),
        z=rep((1:10-0.5)/10,rep(10,10))))
vis.gam(b6,color="heat")