segreg: Fitting segmented/stepmented regression

In segmented: Regression Models with Break-Points / Change-Points Estimation (with Possibly Random Effects)

Fitting segmented/stepmented regression

Description

segreg (stepreg) fits (generalized) linear segmented (stepmented) regression via a symbolic description of the linear predictor. This is an alternative but equivalent function, introduced since version 2.0-0 (segreg) and 2.1-0 (stepreg), to segmented.(g)lm or stepmented.(g)lm.

Usage

segreg(formula, data, subset, weights, na.action, family = lm, control = seg.control(), 
      transf = NULL, contrasts = NULL, model = TRUE, x = FALSE, var.psi = TRUE, ...)
      
stepreg(formula, data, subset, weights, na.action, family = lm, control = seg.control(), 
       transf = NULL, contrasts = NULL, model = TRUE, x = FALSE, var.psi = FALSE, ...)

Arguments

formula	A standard model formula also including one or more 'segmented'/'stepmented' terms via the function seg
data	The possible dataframe where the variables are stored
subset	Possible subset, as in lm or glm
weights	Possible weight vector, see weights in lm or glm
na.action	a function which indicates what happen when the data contain NA values. See na.action in lm or glm.
family	The family specification, similar to family in glm. Default to 'lm' for segmented/stepmented linear models.
control	See seg.control
transf	an optional character string (with "y" as argument) meaning a function to apply to the response variable before fitting
contrasts	see contrasts in glm
model	If TRUE, the model frame is returned.
x	If TRUE, the model matrix is returned.
var.psi	logical, meaning if the standard errors for the breakpoint estimates should be returned in the object fit. If FALSE, the standard errors will be computed by vcov.segmented or summary.segmented. Setting var.psi=FALSE could speed up model estimation for very large datasets. Default to TRUE for segreg and FALSE for stepreg.
...	Ignored

Details

The function allows to fit segmented/stepmented (G)LM regression models using a formula interface. Results will be the same of those coming from the traditional segmented.lm and segmented.glm (or stepmented.lm or stepmented.glm), but there are some additional facilities: i) it is possible to estimate strightforwardly the segmented/stepmented relationships in each level of a categorical variable, see argument by in seg; ii) it is possible to constrain some slopes of the segmented relationship, see argument est or R in seg.

Value

An object of class "segmented" (or "stepmented") which inherits from the class "lm" or "glm" depending on family specification. See segmented.lm.

Warning

Currently for fits returned by segreg, confint.segmented only works if method="delta". Constraints on the mean levels (possibly via argument 'est' of seg) are not yet allowed when calling stepreg.

Note

When the formula includes even a single segmented term with constraints (specified via the argument est in seg()), the relevant coefficients returned do not represent the slope differences as in segmented.lm or segmented.glm. The values depend on the constraints and are not usually interpretable. Use slope the recover the actual slopes of the segmented relationships.

Author(s)

Vito Muggeo

References

Muggeo, V.M.R. (2003) Estimating regression models with unknown break-points. Statistics in Medicine 22, 3055-3071.

See Also

seg, segmented, stepmented

Examples

###########################
#An example using segreg()
###########################

set.seed(10)
x<-1:100
z<-runif(100)
w<-runif(100,-10,-5)
y<-2+1.5*pmax(x-35,0)-1.5*pmax(x-70,0)+10*pmax(z-.5,0)+rnorm(100,0,2)

##the traditional approach
out.lm<-lm(y~x+z+w)
o<-segmented(out.lm, seg.Z=~x+z, psi=list(x=c(30,60),z=.4))

o1<-segreg(y ~ w+seg(x,npsi=2)+seg(z))
all.equal(fitted(o), fitted(o1))

#put some constraints on the slopes
o2<-segreg(y ~ w+seg(x,npsi=2, est=c(0,1,0))+seg(z))
o3<-segreg(y ~ w+seg(x,npsi=2, est=c(0,1,0))+seg(z, est=c(0,1)))

slope(o2)
slope(o3)

##see ?plant for an additional example


###########################
#An example using stepreg()
###########################

### Two stepmented covariates (with 1 and 2 breakpoints)
n=100
x<-1:n/n
z<-runif(n,2,5)
w<-rnorm(n)
mu<- 2+ 1*(x>.6)-2*(z>3)+3*(z>4)
y<- mu + rnorm(n)*.8

os <-stepreg(y~seg(x)+seg(z,2)+w) #also includes 'w' as a possible linear term
os
summary(os)
plot(os, "z", col=2:4) #plot the effect of z

segmented documentation built on Oct. 25, 2024, 5:07 p.m.