Custom Transform Coefficients sjPlot::tab_model - Expressing log transformed coefficients as percent change

Question

I fit a log-transformed linear mixed effect model and want to express the coefficients as percent change instead of linear change on the log scale in my summary table with sjPlot

Fake Data:

library(lme4)
library(lmerTest)
library(dplyr)


set.seed(1234)
dat_short <- data.frame(
  dv = c(
    # Add in t1 ctrl 
      rnorm(mean=0.8, sd=0.1, n=6), #Long Healthy
      rnorm(mean=0.7, sd=0.1, n=6), #Lat Healthy
      rnorm(mean=0.6, sd=0.1, n=4),  #Long Damaged
      rnorm(mean=0.5, sd=0.1, n=4),  #Lat Damaged
    # Add in t2 ctrl
    rnorm(mean=0.7, sd=0.1, n=6), #Long Healthy
    rnorm(mean=0.6, sd=0.1, n=6), #Lat Healthy
    rnorm(mean=0.5, sd=0.1, n=4),  #Long Damaged
    rnorm(mean=0.4, sd=0.1, n=4),  #Lat Damaged
    # Add in t1 trt
    rnorm(mean=0.8, sd=0.1, n=6), #Long Healthy
    rnorm(mean=0.7, sd=0.1, n=6), #Lat Healthy
    rnorm(mean=0.6, sd=0.15, n=4), #Long Damaged
    rnorm(mean=0.5, sd=0.15, n=4), #Lat Damaged
    # Add in t2 trt
    rnorm(mean=0.7, sd=0.1, n=6), #Long Healthy
    rnorm(mean=0.6, sd=0.1, n=6), #Lat Healthy
    rnorm(mean=0.65, sd=0.15, n=4),#Long Damaged
    rnorm(mean=0.55, sd=0.15, n=4) #Lat Damaged
  ),
  id=c(rep(c("subj_1", "subj_2"), times=c(40, 40))),
  intervention=c(rep(c("ctrl", "trt"), times=c(40, 40))), 
  timepoint=c(rep(rep(c("t1", "t2"), times=c(20, 20)),2)), 
  direction=c(rep(rep(c("long", "lat", "long", "lat"), times=c(6, 6, 4, 4)),4)),
  region=c(rep(rep(c("healthy", "damaged"), times=c(12, 8)),4))
)  |>
  mutate(dv = case_when(
         id == "subj_1" ~ dv + runif(1, min = 0.01, max = 0.2),
         id == "subj_2" ~ dv))


speed_measures <-data.frame(
  n_speed = c(
    # Add in t1 ctrl 
      round(runif(8, min = 3, max =10),0)
  ),
  id=c(rep(c("subj_1", "subj_2"), times=c(4, 4))),
  timepoint=c(rep(rep(c("t1", "t2"), times=c(2, 2)),2)), 
  direction=c(rep(rep(c("long", "lat"), times=c(1, 1)))
))  

dat_short_combined <- speed_measures |> left_join(dat_short) |> slice_sample(n = 70)

Fit the regression

lmm_1_short <- lmer(dv ~ intervention*timepoint*region + direction + (1|id), data=dat_short)

Summarize it nicely with sjPlot::tab_model, but this is linear change on the log scale, not as intuitive to understand

sjPlot::tab_model(lmm_1_short,
                  show.intercept = FALSE,
                  show.reflvl = TRUE,
                  show.obs = TRUE,
                  df.method = "satterthwaite")

So we can transform the coefficients to percent change:

# Transform regression coefficients to % change
lmm_1_short_perc_summary <- coef(lmm_1_short)$id |>
  summarise(across(.fns = ~100*( exp(.x)-1) ) ) |>
  summarise(across(.fns = ~mean(.x))) |>
  select(-"(Intercept)") |>
  tidyr::pivot_longer(cols = everything(), names_to="Coefficient", values_to = "% Change") |>
  # Add p-values etc. from regression summary
  cbind(coef(summary(lmm_1_short))[-1,-1]) |>
# Compute confidence intervals and transform those to % change as well
  cbind(data.frame(confint(lmm_1_short))[4:11,]) |>
  transmute(Coefficient, across(where(is.numeric), .fns = ~round(.x,3))) |>
    rename("2.5% CI" = `X2.5..`, "97.5% CI" = `X97.5..`) 

row.names(lmm_1_short_perc_summary) <- 1:nrow(lmm_1_short_perc_summary)
lmm_1_short_perc_summary

Which gives this nice summary:

So the goal is, make a pretty sjPlot style regression summary with my log -> % change transformation.

I tried to use the transform function in tab_model, to no avail. I assumed that sjPlot dragged in get_model_data early but really I have no idea.

function_test <- function(arg_1) {
  sjPlot::get_model_data(arg_1, "est") |>
    mutate(estimate = 100*( exp(estimate)-1) )
}

sjPlot::tab_model(lmm_1_short,
                  show.intercept = FALSE,
                  transform = "function_test",
                  show.reflvl = TRUE,
                  show.obs = TRUE,
                  df.method = "satterthwaite")

Error in if (fam.info$is_linear) transform <- NULL else transform <- "exp" : 
  argument is of length zero

Ref: https://cscu.cornell.edu/wp-content/uploads/83_logv.pdf

Have you tried coding your tranformation as a custom function and passing it to sjPlot via the `transform =` argument? — sjp, Sep 13 '22 at 11:16
I had not because I can't really figure out the syntax. Docs say "transform = A character vector, naming a function that will be applied on estimates and confidence intervals. By default, transform will automatically use "exp" as transformation for applicable classes of model (e.g. logistic or poisson regression). Estimates of linear models remain untransformed. Use NULL if you want the raw, non-transformed estimates." What's the "name" of the function I want to apply? — myfatson, Sep 13 '22 at 21:26
I believe you have to save the custom function to the global environment, then pass the name you have given the function — sjp, Sep 14 '22 at 22:19
@sjp I defined a simple function but didn't get far. I'm not really certain what the data structure looks like before it applies the transform, so I made the assumption that it worked off the get_model_data estimates function within the package. Interestingly, I have found that it accepts "exp", "plogis", and "dlogis" as inputs, but it's not documented otherwise. — myfatson, Sep 15 '22 at 03:54

score 0 · Accepted Answer · answered Sep 15 '22 at 03:59

Transform in sjPlot is a single input function, this works:

function_test <- function(arg_1) {
    100*( exp(arg_1)-1) 
}

sjPlot::tab_model(lmm_1_short,
                  show.intercept = FALSE,
                  transform = "function_test",
                  show.reflvl = TRUE,
                  show.obs = TRUE,
                  df.method = "satterthwaite")

Custom Transform Coefficients sjPlot::tab_model - Expressing log transformed coefficients as percent change

1 Answers1