CODEWITHSUNDEEP
remmeans
LucaS
LucaS

Reputation: 1303

emmeans custom contrasts on probability scale

I'm wondering how to obtain a contrast output (for a custom set of contrasts) as a difference in probabilities rather then a difference in log-odds or the odds ratio. I know how to do this for ALL comparisons in emmeans by specifying regrid() prior to a pairs() call, but is there a way to do this for only certain contrasts?

Reproducible code:

data("PimaIndiansDiabetes2", package = "mlbench")
PimaIndiansDiabetes2 <- na.omit(PimaIndiansDiabetes2)

# Convert a couple of numerical variables to binary variables
# Make pregnant binary (0 = 0, ≥ 1 = 1)
PimaIndiansDiabetes2$preg_cat <- factor(ifelse(PimaIndiansDiabetes2$pregnant == 0, "0", "≥1"), levels = c("0", "≥1"))
# Make pressure binary (< 80 = low, ≥ 80 = high)
PimaIndiansDiabetes2$press_cat <- factor(ifelse(PimaIndiansDiabetes2$pressure < 80, "low", "high"), levels = c("low", "high"))

# Logistic model
mod <- glm(diabetes ~ preg_cat + press_cat + glucose, family = binomial, data = PimaIndiansDiabetes2)
summary(mod)

library(emmeans)
# Get means for 4 combinations of pregnancy and pressure categories (setting glucose to 100)
(emm <- emmeans(mod, ~ preg_cat + press_cat, at = list(glucose = 100))) # log-odds
emmeans(mod, ~ preg_cat + press_cat, at = list(glucose = 100), type = "response") # probabilities

# Now say we want to get the 'effect' of preg_count, stratified by press_cat (and keeping glucose = 100)
custom <- list(`preg_count effect for press_cat = low` = c(-1,1,0,0),
               `preg_count effect for press_cat = high` = c(0,0,-1,1))
# Effect on log-odds scale - note they are the same as this scale is additive (linear)
contrast(emm, custom) |> 
  summary(infer = T)
# Effect on probability scale ???? - note they are NOT the same as this scale is multiplicative (non-linear)
contrast(emm, custom) |> 
  regrid() |> 
  summary(infer = T)
# The above doesn't work and only generates output as an odds-ratio

# I can get diffs in probabilities for ALL pairwise comparisons using:
emmeans(mod, ~ preg_cat + press_cat, at = list(glucose = 100)) |> 
  regrid() |> pairs()

# But I only want the comparisons in the custom contrast list

Upvotes: 2

Views: 325

Answers (1)

Russ Lenth
Russ Lenth

Reputation: 6810

You need to regrid before calling contrast().

contrast(regrid(emm), custom)

Upvotes: 2

Related Questions