R: P-IRLS GAM estimation with GCV & UBRE derivative calculation

gam.fit2 {mgcv}

R Documentation

P-IRLS GAM estimation with GCV & UBRE derivative calculation

Description

Estimation of GAM smoothing parameters is most stable if optimization of the UBRE or GCV score is outer to the penalized iteratively re-weighted least squares scheme used to estimate the model given smoothing parameters.

These routines estimates a GAM given log smoothing paramaters, and evaluate derivatives of the GCV and UBRE scores of the model with respect to the log smoothing parameters. Calculation of exact derivatives is generally faster than approximating them by finite differencing, as well as generally improving the reliability of GCV/UBRE score minimization.

gam.fit2 evaluates first derivatives, by accumulating them as P-IRLS progresses. gam.fit3 uses a more efficient approach in which the P-IRLS is first run to convergence, and only then are the derivatives evaluated by a separate iteration. gam.fit3 can evaluate second as well as first derivatives.

Not normally called directly, but rather service routines for gam.

Usage

gam.fit2(x, y, sp, S=list(),rS=list(),off, H=NULL, 
         weights = rep(1, nobs), start = NULL, etastart = NULL, 
         mustart = NULL, offset = rep(0, nobs), family = gaussian(), 
         control = gam.control(), intercept = TRUE,deriv=TRUE,
         gamma=1,scale=1,pearson=FALSE,printWarn=TRUE,...)
gam.fit3(x, y, sp, S=list(),rS=list(),off, H=NULL, 
         weights = rep(1, nobs), start = NULL, etastart = NULL, 
         mustart = NULL, offset = rep(0, nobs), family = gaussian(), 
         control = gam.control(), intercept = TRUE,deriv=2,use.svd=TRUE,
         gamma=1,scale=1,printWarn=TRUE,...)

Arguments

`x`	The model matrix for the GAM.
`y`	The response variable.
`sp`	The log smoothing parameters.
`S`	A list of penalty matrices. Typically penalty matrices contain only a smallish square sub-matrix which is non-zero: this is what is actually stored. `off[i]` indicates which parameter is the first one penalized by `S[[i]]`.
`rS`	List of square roots of penalty matrices, each having as few columns as possible, but as many rows as there are parameters.
`off`	`off[i]` indicates which parameter `S[[i]][1,1]` relates to.
`H`	The fixed penalty matrix for the model.
`weights`	prior weights for fitting.
`start`	optional starting parameter guesses.
`etastart`	optional starting values for the linear predictor.
`mustart`	optional starting values for the mean.
`offset`	the model offset
`family`	the family - actually this routine would never be called with `gaussian()`
`control`	control list as returned from `glm.control`
`intercept`	does the model have and intercept, `TRUE` or `FALSE`
`deriv`	Should derivatives of the GCV and UBRE scores be calculated? `TRUE` or `FALSE` for `gam.fit2`, but 0, 1 or 2, indicating the maximum order of differentiation to apply, for `gam.fit3`.
`use.svd`	Should the algorithm use SVD (`TRUE`) or the cheaper QR (`FALSE`) as the second matrix decomposition of the final derivative/coefficient evaluation method? Only used by `gam.fit3`.
`gamma`	The weight given to each degree of freedom in the GCV and UBRE scores can be varied (usually increased) using this parameter.
`scale`	The scale parameter - needed for the UBRE score.
`pearson`	The GCV/UBRE score can be based either on the Pearson statistic or the deviance. The latter is generally to be preferred, as it is less prone to severe undersmoothing. Only used by `gam.fit2`.
`printWarn`	Set to `FALSE` to suppress some warnings. Useful in order to ensure that some warnings are only printed if they apply to the final fitted model, rather than an intermediate used in optimization.
`...`	Other arguments: ignored.

Details

This routine is basically glm.fit with some modifications to allow (i) for quadratic penalties on the log likelihood; (ii) derivatives of the model coefficients with respect to log smoothing parameters to be obtained (by updating alongside the P-IRLS iteration) and (iii) derivatives of the GAM GCV and UBRE scores to be evaluated at convergence.

In addition the routine applies step halving to any step that increases the penalized deviance substantially.

The most costly parts of the calculation are performed by calls to compiled C code (which in turn calls LAPACK routines) in place of the compiled code that would usually perform least squares estimation on the working model in the IRLS iteration.

Estimation of smoothing parameters by optimizing GCV scores obtained at convergence of the P-IRLS iteration was proposed by O'Sullivan et al. (1986), and is here termed `outer' iteration.

Note that use of non-standard families with this routine requires modification of the families as described in fix.family.link.

Author(s)

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

The routine has been modified from glm.fit in R 2.0.1, written by the R core (see glm.fit for further credits).

References

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

http://www.maths.bath.ac.uk/~sw283/