| plot.qgcompfit {qgcomp} | R Documentation |
Default plotting method for a qgcompfit object
Description
Plot a quantile g-computation object. For qgcomp.glm.noboot, this function will create a butterfly plot of weights. For qgcomp.glm.boot, this function will create a box plot with smoothed line overlaying that represents a non-parametric fit of a model to the expected outcomes in the population at each quantile of the joint exposures (e.g. '1' represents 'at the first quantile for every exposure')
Usage
## S3 method for class 'qgcompfit'
plot(
x,
suppressprint = FALSE,
pointwisebars = TRUE,
modelfitline = TRUE,
modelband = TRUE,
flexfit = TRUE,
pointwiseref = ceiling(x$q/2),
...
)
## S3 method for class 'qgcompmultfit'
plot(
x,
suppressprint = FALSE,
pointwisebars = TRUE,
modelfitline = TRUE,
modelband = TRUE,
flexfit = TRUE,
pointwiseref = ceiling(x$q/2),
...
)
Arguments
x |
"qgcompfit" object from |
suppressprint |
If TRUE, suppresses the plot, rather than printing it by default (it can be saved as a ggplot2 object (or list of ggplot2 objects if x is from a zero- inflated model) and used programmatically) (default = FALSE) |
pointwisebars |
(boot.gcomp only) If TRUE, adds 95% error bars for pointwise comparisons of E(Y|joint exposure) to the smooth regression line plot |
modelfitline |
(boot.gcomp only) If TRUE, adds fitted (MSM) regression line of E(Y|joint exposure) to the smooth regression line plot |
modelband |
If TRUE, adds 95% prediction bands for E(Y|joint exposure) (the MSM fit) |
flexfit |
(boot.gcomp only) if TRUE, adds flexible interpolation of predictions from underlying (conditional) model |
pointwiseref |
(boot.gcomp only) integer: which category of exposure (from 1 to q) should serve as the referent category for pointwise comparisons? (default=1) |
... |
unused |
Functions
-
plot(qgcompmultfit): Plot method for qgcomp multinomial fits
See Also
qgcomp.glm.noboot, qgcomp.glm.boot, and qgcomp
Examples
set.seed(12)
dat <- data.frame(x1=(x1 <- runif(100)), x2=runif(100), x3=runif(100), z=runif(100),
y=runif(100)+x1+x1^2)
ft <- qgcomp.glm.noboot(y ~ z + x1 + x2 + x3, expnms=c('x1','x2','x3'), data=dat, q=4)
ft
# display weights
plot(ft)
# examining fit
plot(ft$fit, which=1) # residual vs. fitted is not straight line!
## Not run:
# using non-linear outcome model
ft2 <- qgcomp.glm.boot(y ~ z + x1 + x2 + x3 + I(x1*x1), expnms=c('x1','x2','x3'),
data=dat, q=4, B=10)
ft2
plot(ft2$fit, which=1) # much better looking fit diagnostics suggests
# it is better to include interaction term for x
plot(ft2) # the msm predictions don't match up with a smooth estimate
# of the expected outcome, so we should consider a non-linear MSM
# using non-linear marginal structural model
ft3 <- qgcomp.glm.boot(y ~ z + x1 + x2 + x3 + I(x1*x1), expnms=c('x1','x2','x3'),
data=dat, q=4, B=10, degree=2)
# plot(ft3$fit, which=1) - not run - this is identical to ft2 fit
plot(ft3) # the MSM estimates look much closer to the smoothed estimates
# suggesting the non-linear MSM fits the data better and should be used
# for inference about the effect of the exposure
# binary outcomes, logistic model with or without a log-binomial marginal
structural model
dat <- data.frame(y=rbinom(100,1,0.5), x1=runif(100), x2=runif(100), z=runif(100))
fit1 <- qgcomp.glm.boot(y ~ z + x1 + x2, family="binomial", expnms = c('x1', 'x2'),
data=dat, q=9, B=100, rr=FALSE)
fit2 <- qgcomp.glm.boot(y ~ z + x1 + x2, family="binomial", expnms = c('x1', 'x2'),
data=dat, q=9, B=100, rr=TRUE)
plot(fit1)
plot(fit2)
# Using survival data ()
set.seed(50)
N=200
dat <- data.frame(time=(tmg <- pmin(.1,rweibull(N, 10, 0.1))),
d=1.0*(tmg<0.1), x1=runif(N), x2=runif(N), z=runif(N))
expnms=paste0("x", 1:2)
f = survival::Surv(time, d)~x1 + x2
(fit1 <- survival::coxph(f, data = dat))
# non-bootstrap method to get a plot of weights
(obj <- qgcomp.cox.noboot(f, expnms = expnms, data = dat))
plot(obj)
# bootstrap method to get a survival curve
# this plots the expected survival curve for the underlying (conditional) model
# as well as the expected survival curve for the MSM under the following scenarios:
# 1) highest joint exposure category
# 2) lowest joint exposure category
# 3) average across all exposure categories
# differences between the MSM and conditional fit suggest that the MSM is not flexible
# enough to accomodate non-linearities in the underlying fit (or they may simply signal that
# MCSize should be higher). Note that if linearity
# is assumed in the conditional model, the MSM will typically also appear linear and
# will certainly appear linear if no non-exposure covariates are included in the model
# not run (slow when using boot version to proper precision)
(obj2 <- qgcomp.cox.boot(f, expnms = expnms, data = dat, B=10, MCsize=2000))
plot(obj2)
## End(Not run)