lfbas {locfit} | R Documentation |
By default, Locfit uses a polynomial basis as the fitting functions.
An alternative set of basis functions can be specified through
the basis
argument to locfit.raw
.
lfbas
is a wrapper function used internally in Locfit's processing.
To use the basis
argument, you must write a function
f(x,t)
to evaluate the basis function at a fitting
point t
and data point(s) x
. See below for an example.
Note that the basis function will be called with multiple
data points simultaneously, so should assume x
is a matrix
with d
columns, where d
is the number of independent
variables. The fitting point t
will always be a single point,
in a vector of length d
.
The basis function should return a matrix, with each column being the evaluation of one fitting function.
To ensure that the returned fit estimates the mean function, the
first component of the basis should be 1; the remaining components
should be 0 when x=t
. To help ensure Locfit's interpolation
routines are meaningful, the next d
components should represent
partial derivatives at x=t
. Specifically, df(x,t)/dx[i],
evaluated at x=t
, should be the unit vector with 1 in the
(i+1)st position.
Violation of these rules can be useful, if functionals other than the mean are of interest. But this will require extreme care.
Specifying a user basis results in extensive use of the
call_S
function, which may be slow. Speed may also
be significantly affected by different ways of writing
the basis.
lfbas(dim, indices, tt, ...)
locfit(..., basis)
## Specify a bivariate linear with interaction basis. data(ethanol, package="locfit") my.basis <- function(x, t) { u1 <- x[, 1] - t[1] u2 <- x[, 2] - t[2] cbind(1, u1, u2, u1 * u2) } fit <- locfit(NOx ~ E + C, data=ethanol, scale=0, basis=my.basis) ## With this basis, Locfit's standard interpolation and plot methods ## should be reasonable. plot(fit, get.data=TRUE) ## Estimation of change points. This provides an alternative to using ## left() and right(), and can easily be modified to detecting ## a change in slopes or other parameters. Note that the first ## component is the indicator of x>t, so the coefficient estimates ## the size of the change, assuming the change occurs at t. data(penny, package="locfit") my.basis <- function(x, t) cbind(x > t, 1, x - t) xev <- (1945:1988) + 0.5 fit <- locfit(thickness ~ year, data=penny, alpha=c(0, 10), ev=xev, basis=my.basis) ## The plot will show peaks where change points are most likely. ## in S4, S-Plus 5 etc, ## plot(preplot(fit,where="fitp")^2, type="b") is an alternative. plot(xev, predict(fit, where="fitp")^2, type="b") ## Estimate the mean function using local linear regression, with ## discontinuities at 1958.5 and 1974.5. ## The basis functions must consist of the constant 1, the linear term ## x-t, and indicator functions for two of the three sections. ## Note the care taken to ensure my.basis(t,t) = c(1,0,0,0) for all t. my.basis <- function(x, t) { ret <- NULL if (t<1958.5) ret <- cbind(1, x >= 1958.5, x > 1974.5, x - t) if (t>1974.5) ret <- cbind(1, x <= 1974.5, x < 1958.5, x - t) if (is.null(ret)) ret <- cbind(1, x < 1958.5, x > 1974.5, x - t) ret } fit <- locfit(thickness ~ year, data=penny, alpha=c(0,10), ev=xev, basis=my.basis) plot(preplot(fit,where="fitp", get.data=TRUE))