Data Engineering and BI courses are free this week!
The typical use of this function is to cause EMMs to be computed on a different scale, e.g., the back-transformed scale rather than the linear-predictor scale. In other words, if you want back-transformed results, do you want to average and then back-transform, or back-transform and then average?
regrid(object, transform = c("response", "mu", "unlink", "none", "pass",
links), inv.link.lbl = "response", predict.type,
bias.adjust = get_emm_option("back.bias.adj"), sigma, N.sim,
sim = mvtnorm::rmvnorm, ...)An emmGrid object with the requested changes
An object of class emmGrid
Character, list, or logical value. If "response",
"mu", or TRUE, the inverse transformation is applied to the
estimates in the grid (but if there is both a link function and a response
transformation, "mu" back-transforms only the link part); if
"none" or FALSE, object is re-gridded so that its
bhat slot contains predict(object) and its linfct slot
is the identity. Any internal transformation information is preserved. If
transform = "pass", the object is not re-gridded in any way (this
may be useful in conjunction with N.sim).
If transform is a character value in links (which is the set
of valid arguments for the make.link function, excepting
"identity"), or if transform is a list of the same form as
returned by make.links or make.tran, the results are
formulated as if the response had been transformed with that link function.
Character value. This applies only when transform
is in links, and is used to label the predictions if subsequently summarized
with type = "response".
Character value. If provided, the returned object is
updated with the given type to use by default by summary.emmGrid
(see update.emmGrid). This may be useful if, for example,
when one specifies transform = "log" but desires summaries to be
produced by default on the response scale.
Logical value for whether to adjust for bias in
back-transforming (transform = "response"). This requires a valid value of
sigma to exist in the object or be specified.
Error SD assumed for bias correction (when
transform = "response" and a transformation
is in effect). If not specified,
object@misc$sigma is used, and a warning is issued if it is not found.
Integer value. If specified and object is based on a
frequentist model (i.e., does not have a posterior sample), then a fake
posterior sample is generated using the function sim.
A function of three arguments (no names are assumed).
If N.sim is supplied with a frequentist model, this function is called
with respective arguments N.sim, object@bhat, and object@V.
The default is the multivariate normal distribution.
Ignored.
In cases where the degrees of freedom depended on the linear function being estimated (e.g., Satterthwaite method), the d.f. from the reference grid are saved, and a kind of “containment” method is substituted in the returned object, whereby the calculated d.f. for a new linear function will be the minimum d.f. among those having nonzero coefficients. This is kind of an ad hoc method, and it can over-estimate the degrees of freedom in some cases. An annotation is displayed below any subsequent summary results stating that the degrees-of-freedom method is inherited from the previous method at the time of re-gridding.
The regrid function reparameterizes an existing ref.grid so
that its linfct slot is the identity matrix and its bhat slot
consists of the estimates at the grid points. If transform is
TRUE, the inverse transform is applied to the estimates. Outwardly,
when transform = "response", the result of summary.emmGrid
after applying regrid is identical to the summary of the original
object using type="response". But subsequent EMMs or
contrasts will be conducted on the new scale -- which is
the reason this function exists.
This function may also be used to simulate a sample of regression
coefficients for a frequentist model for subsequent use as though it were a
Bayesian model. To do so, specify a value for N.sim and a sample is
simulated using the function sim. The grid may be further processed in
accordance with the other arguments; or if transform = "pass", it is
simply returned with the only change being the addition of the simulated
sample.
pigs.lm <- lm(log(conc) ~ source + factor(percent), data = pigs)
rg <- ref_grid(pigs.lm)
# This will yield EMMs as GEOMETRIC means of concentrations:
(emm1 <- emmeans(rg, "source", type = "response"))
pairs(emm1) ## We obtain RATIOS
# This will yield EMMs as ARITHMETIC means of concentrations:
(emm2 <- emmeans(regrid(rg, transform = "response"), "source"))
pairs(emm2) ## We obtain DIFFERENCES
# Same result, useful if we hadn't already created 'rg'
# emm2 <- emmeans(pigs.lm, "source", regrid = "response")
# Simulate a sample of regression coefficients
set.seed(2.71828)
rgb <- regrid(rg, N.sim = 200, transform = "pass")
emmeans(rgb, "source", type = "response") ## similar to emm1
Run the code above in your browser using DataLab