comb_facs {emmeans}

Manipulate factors in a reference grid

Description

These functions manipulate the levels of factors comprising a reference grid by combining factor levels, splitting a factor's levels into combinations of newly-defined factors, creating a grouping factor in which factor(s) levels are nested, or permuting the order of levels of a factor

Usage

comb_facs(object, facs, newname = paste(facs, collapse = "."),
  drop = FALSE, ...)

split_fac(object, fac, newfacs, ...)

add_grouping(object, newname, refname, newlevs, ...)

permute_levels(object, fac, pos)

Arguments

object

An object of class emmGrid

facs

Character vector. The names of the factors to combine

newname

Character name of grouping factor to add (different from any existing factor in the grid)

drop

Logical value. If TRUE, any levels of the new factor that are dropped if all occurrences in the newly reconstructed object have weight zero. If FALSE, all levels are retained. (This argument is ignored if there is no .wgt. column in object@grid.)

...

arguments passed to other methods

fac

The name of a factor that is part of the grid in object

newfacs

A named list with the names of new factors and their levels. The names must not already exist in the object, and the product of the lengths of the levels must equal the number of levels of fac.

refname

Character name(s) of the reference factor(s)

newlevs

Character vector or factor of the same length as that of the (combined) levels for refname. The grouping factor newname will have the unique values of newlevs as its levels. The order of levels in newlevs is the same as the order of the level combinations produced by expand.grid applied to the levels of refname – that is, the first factor's levels change the fastest and the last one's vary the slowest.

pos

Integer vector consisting of some permutation of the sequence 1:k, where k is the number of levels of fac. This determines which position each level of fac will occupy after the levels are permuted; thus, if the levels of fac are A,B,C,D, and pos = c(3,1,2,4), then the permuted levels will be B,C,A,D.

Value

A modified object of class emmGrid

The comb_facs function

comb_facs combines the levels of factors into a single factor in the reference grid (similar to interaction). This new factor replaces the factors that comprise it.

Additional note: The choice of whether to drop levels or not can make a profound difference. If the goal is to combine factors for use in joint_tests, we advise against drop = TRUE because that might change the weights used in deriving marginal means. If combining factors in a nested structure, dropping unused cases can considerably reduce the storage required.

The split_fac function

The levels in newfacs are expanded via expand.grid into combinations of levels, and the factor fac is replaced by those factor combinations. Unlike add_grouping, this creates a crossed, rather than a nested structure. Note that the order of factor combinations is systematic with the levels of first factor in newfacs varying the fastest; and those factor combinations are assigned respectively to the levels of fac as displayed in str(object).

The add_grouping function

This function adds a grouping factor to an existing reference grid or other emmGrid object, such that the levels of one or more existing factors (call them the reference factors) are mapped to a smaller number of levels of the new grouping factor. The reference factors are then nested in a new grouping factor named newname, and a new nesting structure refname %in% newname. This facilitates obtaining marginal means of the grouping factor, and contrasts thereof.

Additional notes: By default, the levels of newname will be ordered alphabetically. To dictate a different ordering of levels, supply newlevs as a factor having its levels in the desired order.

When refname specifies more than one factor, this can fundamentally (and permanently) change what is meant by the levels of those individual factors. For instance, in the gwrg example below, there are two levels of wool nested in each prod; and that implies that we now regard these as four different kinds of wool. Similarly, there are five different tensions (L, M, H in prod 1, and L, M in prod 2).

The permute_levels function

This function permutes the levels of fac. The returned object has the same factors, same by variables, but with the levels of fac permuted. The order of the columns in object@grid may be altered.

NOTE: fac must not be nested in another factor. permute_levels throws an error when fac is nested.

NOTE: Permuting the levels of a numeric predictor is tricky. For example, if you want to display the new ordering of levels in emmip(), you must add the arguments style = "factor" and nesting.order = TRUE.

Examples

mtcars.lm <- lm(mpg ~ factor(vs)+factor(cyl)*factor(gear), data = mtcars)
(v.c.g <- ref_grid(mtcars.lm))
(v.cg <- comb_facs(v.c.g, c("cyl", "gear")))
  
# One use is obtaining a single test for the joint contributions of two factors:
joint_tests(v.c.g)

joint_tests(v.cg)

# undo the 'comb_facs' operation:
split_fac(v.cg, "cyl.gear", list(cyl = c(4, 6, 8), gear = 3:5))

IS.glm <- glm(count ~ spray, data = InsectSprays, family = poisson)
IS.emm <- emmeans(IS.glm, "spray")
IS.new <- split_fac(IS.emm, "spray", list(A = 1:2, B = c("low", "med", "hi")))
str(IS.new)

fiber.lm <- lm(strength ~ diameter + machine, data = fiber)
( frg <- ref_grid(fiber.lm) )

# Suppose the machines are two different brands
brands <- factor(c("FiberPro", "FiberPro", "Acme"), levels = c("FiberPro", "Acme"))
( gfrg <- add_grouping(frg, "brand", "machine", brands) )

emmeans(gfrg, "machine")

emmeans(gfrg, "brand")

### More than one reference factor
warp.lm <- lm(breaks ~ wool * tension, data = warpbreaks)
gwrg <- add_grouping(ref_grid(warp.lm), 
    "prod",  c("tension", "wool"),  c(2, 1, 1,  1, 2, 1))
        # level combinations:         LA MA HA  LB MB HB

emmeans(gwrg, ~ wool * tension)   # some NAs due to impossible combinations

emmeans(gwrg, "prod")

str(v.c.g)
str(permute_levels(v.c.g, "cyl", c(2,3,1)))