Learn R Programming

alr3 (version 2.0.5)

pod: Fit partial one-dimensional, or POD models, based on a linear model

Description

A partial one-dimensional model, or a POD model, provides a concise description of a regression model with many predictors and one grouping variable. It requires a nonlinear regression fit.

Usage

## This is a generic function with different arguments depending on the class of the ## first argument. The generic form is
pod (x, ...)
## If the first argument to pod is a formula suitable for defining a linear model,
"pod"(formula, data = sys.parent(), group, subset, weights, na.action, mean.function = c("pod", "common", "parallel", "general"), singular.ok = FALSE, contrasts = NULL, offset, control = nls.control(), ...)
## If the first argument to pod is the result of a linear model fit, the following ## function call is used. All the arguments for pod.formula can also be passed to ## pod.lm.
"pod" (x, group, mean.function, control, ...) ## The following related function require a pod fit as the first argument: "anova"(object,scale=0,test="F",...)
"plot"(x, colors=1:nlevels(x$group), pch=1:nlevels(x$group),key="topleft",identify=FALSE, xlab="Linear Predictor", ylab=as.character(c(formula(x)[[2]])), ...) "plot"(x, colors=1:nlevels(x$group), pch=1:nlevels(x$group), key="topleft", identify=FALSE, xlab="Linear Predictor", ylab=as.character(c(formula(x)[[2]])), ...)

Arguments

formula
A linear regression formula, as used in lm, or the results of a call to pod or a lm fit. See details below.
x
The result of a lm fit from which the formula and the arguments data, subset, weights, na.action, singular.ok, contrasts and offset will be taken.
data
An optional data frame for the data to be used
group
The name of a grouping variable (not quoted) that defines the groups; see details below.
mean.function
Which mean function should be fit? The default is “pod”, that fits the partial one-dimensional mean function. The other options are “common”, which fits a linear model with no grouping effects; “parallel” fits a parallel within-group regression, and “general”, available in pod but not plot.pod, fits a separate coefficeint for each term in the model for each level of the grouping variable.
subset
As in lm, the subset of cases to be used in the analysis
weights
Weights will be used in fitting non-pod models. Since this argument is not supported for nls models, weights are ignored for fitting pod models. If nls is ever updated, then pod models will correctly use the weights.
na.action
At present, only na.omit is likely to work.
singular.ok
The default equal to FALSE is recommended.
contrasts
Same as in lm
offset
Same as in lm
control
A pod model is fit using the nonlinear least squares routine nls. This routine is very sensitive to starting values and other parameters set in the algorithm. This routine selects starting values for you that will often work very well, but in some problems the user may want to change the defaults to nls program using the nls.control function.
...
In pod, other arguments passed to nls, such as control parameters. In pod.anova, there are two additional arguments
object
The result of a call to pod.
scale
Used for test in anova
test
Default is to compute F tests.
colors
Colors for groups in the pod plot
pch
Ploting symbol for the groups in the pod plot
identify
If TRUE, clicking the mouse on a graph will print the case name of the nearest point. This continues until turned off (by pushing the escape key, among other ways of doing this).
key
The default is "topleft", in which case a legend is added to the top left corner of the plot; other choices include "bottomright". If key is a vector of two coordinates, the legend is drawn at the coordinates specified. If key is FALSE, no key is drawn; if TRUE, you can place the key interactively by clicking on the plot.
xlab
Horizontal label, optional
ylab
Vertical label, optional

Value

If mean.function="pod", a list of class pod, with the following components
nls.fit
The usual output from a nls fit.
linear.part
The estimated linear predictor $bhat'x$.
call
The original call to pod
group
The grouping variable
If mean.function has any other value, an object of class c("pod.lm", "lm") is returned. Since the only method for pod.lm objects is a plotting method, except for plotting these objects behave like lm objects.

Details

Suppose we start with a linear mean function specified by y ~ x1 + x2 + x3, where the right-side variables can be any valid R variables, such as transformations, but NOT factors or interactions (if you want to include these, you need to create the dummy variables yourself). The right-hand side variables must also be linearly independent. We also specify a grouping variable z with, say, g levels. Let $Gi$ be a dummy variable with values equal to 1 when $z=i$, and zero otherwise. The pod mean function is then has the nonlinear mean function $$E(y|x,z) = \beta_0 + \beta'x + \sum_{j=2}^g G_j(\theta_{0j} + \theta_{ij}\beta'x)$$ This is a nonlinear mean function that specifies that the response depends on the predictors only through one linear combination, that the dependence is linear, but the slope and intercept may be different for each group.

The pod mean function is easily fit using nls. For example, if z has two levels, a nonlinear mean function that would work is y ~ b0 + b1*x1 + b2*x2 + b3*x3 + G2*(th02 + th12*(b1*x1 + b2*x2 + b3*x3)). Starting values can be determined by fitting the parallel mean function, for this is a linear mean function and easily fit.

The function pod automates this process; you need only input the linear part of the mean function, and pod does the rest.

The function anova.pod is a little different from the generic anova function. It fits and compares all four mean functions that can be fit with the pod function: (1) no group effect; (2) parallel within group method; (3) pod mean function; (4) mean function with all main effects and two-factor interactions with the grouping variable.

The function plot will plot the response versus the fitted linear combination for either the common, parallel or pod models. There is no 2D plot for the general model.

Objects of class pod have methods defined for the generic functions print, summary, coef, anova, deviance, vcov, resid, formula, fitted, df.residual and predict. podresponse returns the values of the response.

References

Cook, R. D. and Weisberg, S. (2004). Partial-one dimensinal models. American Statistician, 58, 110-116.

Weisberg, S. (2005) Applied Linear Regression, third edition. New York: Wiley.

See Also

See Also nls, lm, nls.control

Examples

Run this code
head(ais)
m1 <- pod(LBM ~ Ht + Wt + RCC, data= ais, group= Sex)
anova(m1) # compare four models
plot(m1) # draw the plot
m2 <- update(m1, mean.function="parallel")
plot(m2)

Run the code above in your browser using DataLab