CODEWITHSUNDEEP
rggplot2drc
Gaston d.
Gaston d.

Reputation: 1

Discrepancies between geom_smooth fit and predict.drc output in R

I am trying to plot dose-response curves using the drc package using the reduced log-normal fit (LN.2), which should be equivalent to the probit dose-response model. When I use the 'predict' function for both the glm package derived probit model (with dose log10 transformed) and the drc model, the outputs are equivalent as expected.

However, when I use the drc method in geom_smooth, the graphical output is completely different than these predicted values.

require(tidyverse)
require(drc)

df<-tibble(y = c(0.1,0.125,0.275,0.5,0.75, 0.95,1),
           x = c(100, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8))



probit_model <- glm(y ~ log10(x), data = df, family = quasibinomial(link = "probit"))
test_model<-drm(y~x, fct = LN.2(), type = "binomial", data = df)


df <- df %>%
  add_row(x = 10, y = NA) %>%  # To underline that these models are different
  mutate(y_pred = predict(probit_model, newdata = ., type = "response"),
         y_pred_2 = predict(test_model, newdata = data.frame(x = x)))


df %>%
  ggplot(aes(x, y)) +
  geom_point(size = 4) +
  scale_x_log10()+
  geom_line(aes(y = y_pred_2), color = "red", lwd = 1) +
  geom_smooth(formula = y ~ x, color = "blue",
              method = "glm", fullrange = TRUE,
              method.args = list(family = quasibinomial(link = "probit")), se = F)+
  geom_smooth(method = drm, method.args = list(fct = LN.2(), type = "binomial"), se = F, color = "black", linetype = 2)+
  theme_bw()

Here is an example of the behavior.

Upvotes: 0

Views: 238

Answers (1)

Allan Cameron
Allan Cameron

Reputation: 173793

You need to account for the log-transformed x axis values being passed as the input to the predict function. You have already done this with glm, since you are using y ~ x as the formula instead of y ~ log(x), but it will not really work with drm. However, it isn't really necessary to use geom_smooth here, since your model already includes a function to recreate the curve using geom_function

df %>%
  mutate(x = x) %>%
  ggplot(aes(x, y)) +
  geom_point(size = 4) +
  geom_line(aes(y = y_pred_2), color = "red", lwd = 1.5) +
  geom_smooth(formula = y ~ x, color = "skyblue",
              method = "glm", fullrange = TRUE, lwd = 1.5,
              method.args = list(family = quasibinomial(link = "probit")), 
              se = FALSE) +
  geom_function(fun = test_model$curve[[1]], linetype = 2, lwd = 1.5) +
  theme_bw() +
  scale_x_log10()

enter image description here

Note that the two model lines line up perfectly here.

Upvotes: 2

Related Questions