Performs a slightly inefficient but numerically stable version of fast
backward elimination on factors, using a method based on Lawless and Singhal
(1978).
This method uses the fitted complete model and computes approximate Wald
statistics by computing conditional (restricted) maximum likelihood estimates
assuming multivariate normality of estimates.
fastbw
deletes factors, not columns of the design matrix. Factors requiring multiple d.f. will be retained or dropped as a group.
The function prints the deletion statistics for each variable in
turn, and prints approximate parameter estimates for the model after
deleting variables. The approximation is better when the number of
factors deleted is not large. For ols
, the approximation is exact for
regression coefficients, and standard errors are only off by a factor
equal to the ratio of the mean squared error estimate for the reduced
model to the original mean squared error estimate for the full model.
If the fit was from ols
, fastbw
will compute the usual \(R^2\)
statistic for each model.
fastbw(fit, rule=c("aic", "p"),
type=c("residual", "individual", "total"), sls=.05, aics=0, eps=1e-9,
k.aic=2, force=NULL)# S3 method for fastbw
print(x, digits=4, estimates=TRUE, ...)
a list with an attribute kept
if bw=TRUE
, and the
following components:
matrix of statistics with rows in order of deletion.
names of factors kept in final model.
the subscripts of factors kept in the final model
opposite of factors.kept
.
column numbers in design matrix corresponding to parameters kept in the final model.
opposite of parms.kept
.
vector of approximate coefficients of reduced model.
approximate covariance matrix for reduced model.
matrix of coefficients of all models. Rows correspond to the successive models examined and columns correspond to the coefficients in the full model. For variables not in a particular sub-model (row), the coefficients are zero.
fit object with Varcov(fit)
defined (e.g., from ols
, lrm
, cph
, psm
, glmD
)
Stopping rule. Defaults to "aic"
for Akaike's information criterion. Use
rule="p"
to use \(P\)-values
Type of statistic on which to base the stopping rule. Default is
"residual"
for
the pooled residual chi-square. Use type="individual"
to use Wald
chi-square of individual factors.
Significance level for staying in a model if rule="p"
. Default is .05.
For rule="aic"
,
variables are deleted until the chi-square - k.aic
times d.f. would rise above aics
.
Default aics
is zero to use the ordinary AIC. Set aics
to say 10000
to see all variables deleted in order of descending importance.
Singularity criterion, default is 1E-9
.
multiplier to compute AIC, default is 2. To use BIC, set k.aic
equal
to \(\log(n)\), where \(n\) is the effective sample size (number of events
for survival models).
a vector of integers specifying parameters forced to be in the model, not counting intercept(s)
result of fastbw
number of significant digits to print
set to FALSE
to suppress printing table of
approximate coefficients, SEs, etc., after variable deletions
ignored
Frank Harrell
Department of Biostatistics, Vanderbilt University
fh@fharrell.com
Lawless, J. F. and Singhal, K. (1978): Efficient screening of nonnormal regression models. Biometrics 34:318--327.
rms
, ols
, lrm
,
cph
, psm
, validate
,
solvet
, rmsMisc
if (FALSE) {
fastbw(fit, optional.arguments) # print results
z <- fastbw(fit, optional.args) # typically used in simulations
lm.fit(X[,z$parms.kept], Y) # least squares fit of reduced model
}
Run the code above in your browser using DataLab