gam {mgcv}R Documentation

Generalized Additive Models using penalized regression splines and GCV

Description

Fits the specified generalized additive model (GAM) to data. gam() is not a clone of what S-PLUS provides. Smooth terms are represented using penalized regression splines 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.

Usage


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

Arguments

formula A GAM formula (see also gam.models). This is exactly like the formula for a GLM except that smooth terms can be added to the right hand side of the formula (and a formula of the form y ~ . is not allowed). Smooth terms are specified by expressions of the form:
s(var1,var2,...,k=12,fx=FALSE,bs="tp",by=a.var) where var1, var2, etc. are the covariates which the smooth is a function of and k is the dimension of the basis used to represent the smooth term. If k is not specified then k=10*3^(d-1) is used where d is the number of covariates for this term. fx is used to indicate whether or not this term has a fixed number of degrees of freedom (fx=FALSE to select d.f. by GCV/UBRE). bs indicates the basis to use, with "cr" indicating cubic regression spline, "cc" a cyclic cubic regression spline, and "tp" indicating thin plate regression spline: "cr" and "cc" can only be used with 1-d smooths. by can be used to specify a variable by which the smooth should be multiplied. For example gam(y~z+s(x,by=z)) would specify a model E(y)=f(x)z where f(.) is a smooth function (the formula is y~x+s(x,by=z) rather than y~s(x,by=z) because the smooths are always set up to sum to zero over the covariate values). The by option is particularly useful for models in which different functions of the same variable are required for each level of a factor: see s.
An alternative for specifying smooths of more than one covariate is e.g.:
te(x,z,bs=c("tp","tp"),m=c(2,3),k=c(5,10)) which would specify a tensor product smooth of the two covariates x and z constructed from marginal t.p.r.s. bases of dimension 5 and 10 with marginal penalties of order 2 and 3. Any combination of basis types is possible, as is any number of covariates.
Formulae can involve nested or ``overlapping'' terms such as
y~s(x)+s(z)+s(x,z) or y~s(x,z)+s(z,v): see gam.side.conditions for further details and examples. Note that nesting with te terms is not supported.
family This is a family object specifying the distribution and link to use in fitting etc. See glm and family for more details. The negative binomial families provided by the MASS library can be used, with or without known theta parameter: see gam.neg.bin for details.
data A data frame containing the model response variable and covariates required by the formula. By default the variables are taken from environment(formula): typically the environment from which gam is called.
weights prior weights on the data.
subset an optional vector specifying a subset of observations to be used in the fitting process.
na.action a function which indicates what should happen when the data contain `NA's. The default is set by the `na.action' setting of `options', and is `na.fail' if that is unset. The ``factory-fresh'' default is `na.omit'.
control A list of fit control parameters returned by gam.control.
scale If this is zero then GCV is used for all distributions except Poisson and binomial where UBRE is used with scale parameter assumed to be 1. If this is greater than 1 it is assumed to be the scale parameter/variance and UBRE is used: to use the negative binomial in this case theta must be known. If scale is negative GCV is always used, which means that the scale parameter will be estimated by GCV and the Pearson estimator, or in the case of the negative binomial theta will be estimated in order to force the GCV/Pearson scale estimate to unity (if this is possible). For binomial models in particular, it is probably worth comparing UBRE and GCV results; for ``over-dispersed Poisson'' GCV is probably more appropriate than UBRE.
knots this is an optional list containing user specified knot values to be used for basis construction. For the cr and cc bases the user simply supplies the knots to be used, and there must be the same number as the basis dimension, k, for the smooth concerned. For the tp basis knots has two uses. Firstly, for large datasets the calculation of the tp basis can be time-consuming. The user can retain most of the advantages of the t.p.r.s. approach by supplying a reduced set of covariate values from which to obtain the basis - typically the number of covariate values used will be substantially smaller than the number of data, and substantially larger than the basis dimension, k. The second possibility is to avoid the eigen-decomposition used to find the t.p.r.s. basis altogether and simply use the basis implied by the chosen knots: this will happen if the number of knots supplied matches the basis dimension, k. For a given basis dimension the second option is faster, but gives poorer results (and the user must be quite careful in choosing knot locations). Different terms can use different numbers of knots, unless they share a covariate.
sp A vector of smoothing parameters for each term can be provided here. Smoothing parameters must be supplied in the order that the smooth terms appear in the model formula. With fit method "magic" (see gam.control and magic) then negative elements indicate that the parameter should be estimated, and hence a mixture of fixed and estimated parameters is possible. With fit method "mgcv", if sp is supplied then all its elements must be positive. Note that fx=TRUE in a smooth term over-rides what is supplied here effectively setting the smoothing parameter to zero.
min.sp for fit method "magic" only, lower bounds can be supplied for the smoothing parameters. Note that if this option is used then the smoothing parameters sp, in the returned object, will need to be added to what is supplied here to get the actual smoothing parameters. Lower bounds on the smoothing parameters can sometimes help stabilize otherwise divergent P-IRLS iterations.
H With fit method "magic" a user supplied fixed quadratic penalty on the parameters of the GAM can be supplied, with this as its coefficient matrix. A common use of this term is to add a ridge penalty to the parameters of the GAM in circumstances in which the model is close to un-identifiable on the scale of the linear predictor, but perfectly well defined on the response scale.
gamma It is sometimes useful to inflate the model degrees of freedom in the GCV or UBRE score by a constant multiplier. This allows such a multiplier to be supplied if fit method is "magic".
fit If this argument is TRUE then gam sets up the model and fits it, but if it is FALSE then the model is set up and an object G is returned which is the output from gam.setup plus some extra items required to complete the GAM fitting process.
G Usually NULL, but may contain the object returned by a previous call to gam with fit=FALSE, in which case all other arguments are ignored except for gamma, family, control and fit.
... further arguments for passing on e.g. to gam.fit

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 or an Un-Biased Risk Estimator criterion (UBRE) which is in practice an approximation to AIC. 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. Two alternative numerical methods are provided, see mgcv, magic and gam.control.

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.

The fitting approach just described, in which the smoothing parameters are estimated for each approximating linear model of the IRLS process was suggested by Chong Gu (see, e.g. Gu 2002), and is very computationally efficient. However, because the approach neglects the dependence of the iterative weights on the smoothing parameters, it is usually possible to find smoothing parameters which actually yield slightly lower GCV/UBRE score estimates than those resulting from this `performance iteration'. gam therefore also allows the user to `improve' the smoothing parameter estimates, by using O'Sullivan's (1986) suggested method, in which for each trial set of smoothing parameters the IRLS is iterated to convergence before the UBRE/GCV score is evaluated. This requires much less efficient minimisation of the power iteration based on nlm, and is therefore quite slow.

Three alternative bases 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. Multivariate terms 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.

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 "mgcv" GCV/UBRE minimization method are given in Wood (2000): the basis of the approach is to alternate efficient global optimization with respect to one overall smoothing parameter with Newton updates of a set of relative smoothing parameters for each smooth term. The Newton updates are backed up by steepest descent, since the GCV/UBRE score functions are not positive definite everywhere.

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" which has the following elements:

boundary did parameters end up at boundary of parameter space?
call the matched call (allows update to be used with gam objects, for example).
coefficients the coefficients of the fitted model. Parametric coefficients are first, followed by coefficients for each spline term in turn.
converged indicates whether or not the iterative fitting method converged.
deviance (unpenalized)
df.null null degrees of freedom.
edf estimated degrees of freedom for each coefficient.
family family object specifying distribution and link used.
fit.method The underlying multiple GCV/UBRE method used: "magic" for the new more stable method, "mgcv" for the Wood (2000) method.
fitted.values fitted model predictions of expected value for each datum.
formula the model formula.
full.formula the model formula with each smooth term fully expanded and with option arguments given explicitly (i.e. not with reference to other variables) - useful for later prediction from the model.
gcv.ubre The minimized GCV or UBRE score.
hat array of elements from the leading diagonal of the `hat' (or `influence') matrix. Same length as response data vector.
iter number of iterations of P-IRLS taken to get convergence.
linear.predictor fitted model prediction of link function of expected value for each datum.
method Either "GCV" or "UBRE", depending on smoothing parameter selection method used (or appropriate, if none used).
mgcv.conv A list of convergence diagnostics relating to smoothing parameter estimation. Differs for method "magic" and "mgcv". Here is the "mgcv" version:
score
corresponding to edf, an array of GCV or UBRE scores for the model given the final estimated relative smoothing parameters.

g
the gradient of the GCV/UBRE score w.r.t. the relative smoothing parameters at termination.

h
the second derivatives corresponding to g above - i.e. the leading diagonal of the Hessian.

e
the eigen-values of the Hessian. All non-negative indicates a positive definite Hessian.

iter
the number of iterations taken.

in.ok
TRUE if the second smoothing parameter guess improved the GCV/UBRE score.

step.fail
TRUE if the algorithm terminated by failing to improve the GCV/UBRE score rather than by `converging'. Not necessarily a problem, but check the above derivative information quite carefully.

In the case of "magic" the items are:
full.rank
The apparent rank of the problem given the model matrix and constraints.

rank
The numerical rank of the problem.

fully.converged
TRUE is multiple GCV/UBRE converged by meeting convergence criteria. FALSE if method stopped with a steepest descent step failure.

hess.pos.def
Was the hessian of the GCV/UBRE score positive definite at smoothing parameter estimation convergence?

iter
How many iterations were required to find the smoothing parameters?

score.calls
and how many times did the GCV/UBRE score have to be evaluated?

rms.grad
root mean square of the gradient of the GCV/UBRE score at convergence.

min.edf Minimum possible degrees of freedom for whole model.
model model frame containing all variables needed in original model fit.
nsdf number of parametric, non-smooth, model terms including the intercept.
null.deviance deviance for single parameter model.
prior.weights prior weights on observations.
residuals the deviance residuals for the fitted model.
sig2 estimated or supplied variance/scale parameter.
smooth list of smooth objects, containing the basis information for each term in the model formula in the order in which they appear. These smooth objects are what gets returned by the smooth.construct objects.
sp smoothing parameter for each smooth.
Vp estimated covariance matrix for the parameters. This is a Bayesian posterior covariance matrix that results from adopting a particular Bayesian model of the smoothing process.
weights final weights used in IRLS iteration.
y response data.

WARNINGS

The "mgcv" 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@stats.gla.ac.uk

References

Key References:

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. (in press) Stable and efficient multiple smoothing parameter estimation for generalized additive models. J. Amer. Statist. Ass.

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.

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

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

Wahba (1990) Spline Models of Observational Data. SIAM

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

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

Examples

library(mgcv)
set.seed(0) 
n<-400
sig2<-4
x0 <- runif(n, 0, 1)
x1 <- runif(n, 0, 1)
x2 <- runif(n, 0, 1)
x3 <- runif(n, 0, 1)
pi <- asin(1) * 2
f <- 2 * sin(pi * x0)
f <- f + exp(2 * x1) - 3.75887
f <- f+0.2*x2^11*(10*(1-x2))^6+10*(10*x2)^3*(1-x2)^10-1.396
e <- rnorm(n, 0, sqrt(abs(sig2)))
y <- f + e
b<-gam(y~s(x0)+s(x1)+s(x2)+s(x3))
summary(b)
plot(b,pages=1)
# 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(0.5,41)) 
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/5)
y<-rpois(rep(1,n),g)
b2<-gam(y~s(x0)+s(x1)+s(x2)+s(x3),family=poisson)
plot(b2,pages=1)
# 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")
[Package Contents]