Plot vs ggplot2 in R and how to extract fit parame

Question

I have the following data in a data.frame called t

   DayNum   MeanVolume    StdDev    StdErr
1      13   207.0500  41.00045  5.125057
2      15   142.7625  27.87236  3.484045
3      18    77.5500  19.43928  2.429910
4      21    66.3750  20.56403  2.570504
5      26    67.0500  29.01576  3.626970
6      29    66.4750  25.94537  3.243171
7      33    76.9625  25.31374  3.164218
8      36    91.2875  37.01719  4.627149
9      40   102.0500  29.39898  3.674872
10     43   100.8250  24.22830  3.028538
11     47   120.5125  28.80592  3.600740
12     50   147.8875  35.82894  4.478617
13     54   126.7875  45.43204  5.679004
14     57   139.8500  56.01117  7.001397
15     60   179.1375  69.64526  8.705658
16     64   149.7625  39.10265  4.887831
17     68   229.5250 121.08411 15.135514
18     71   236.5125  76.23146  9.528933
19     75   243.2750 101.69474 12.711842
20     78   331.6750 141.25344 17.656680
21     82   348.2875 122.86359 15.357948
22     85   353.7750 187.24641 23.405801
23     89   385.4000 154.05826 19.257283
24     92   500.9875 263.43714 32.929642
25     95   570.2250 301.82686 37.728358
26     98   692.2250 344.71226 43.089032
27    102   692.8000 283.94120 35.492650
28    105   759.2000 399.19323 49.899153
29    109   898.2375 444.94289 55.617861
30    112   920.1000 515.79597 64.474496

I am trying to fit x = DayNum to y = MeanVolume in t.

Here is what I did:

Fit to data

model<-lm(log(t$MeanVolume) ~ t$DayNum, data=t)

Plot data

plot(MeanVolume~DayNum, data=t, ylab="Mean Volume (mm3)", xlim=c(0,120), ylim=c(0,1000))
arrows(t$DayNum, t$MeanVolume-t$StdErr, t$DayNum, t$MeanVolume+t$StdErr, length=0.01, angle=90, code=3)

Create fit data

t$pred<-exp(predict(model))

Plot fit

lines(t$DayNum,t$pred,col="blue")

On the other hand, if I use ggplot2 to do this by using

ggplot(data = t, mapping = aes(x = DayNum, y=MeanVolume)) + 
  geom_line() + 
  geom_point(size=3, color="blue") + 
  geom_smooth(method="glm", method.args=list(family=gaussian(link="log"))) +
  labs(x="Days", y="Mean Volume (mm3)", title="Data") +
  geom_errorbar(aes(ymin = MeanVolume - StdErr, ymax = MeanVolume + StdErr), width=.2)

I get the following plot

As you can see the fitted curve in the ggplot case is better than in the plot case. Why? Also I would like to fit parameters such as intercept and the slope of the exponential fit line. How can I extract them from ggplot call?

Answer 1

lm with log transformed y is not the same as glm with gaussian error distribution and log link (as to why check link in the comment by @Lyngbakr)

gz <- read.table("somet.txt")
gz <- as.data.frame(gz)
model_lm <- lm(log(MeanVolume) ~ DayNum, data = gz)
model_glm <- glm(MeanVolume ~ DayNum, data = gz, family = gaussian(link = "log"))
pred_lm <- exp(predict(model_lm))
pred_glm <- predict(model_glm, type = "response")

plot(MeanVolume ~ DayNum, data = gz, ylab = "Mean Volume (mm3)", xlim = c(0,120), ylim = c(0,1000))
arrows(gz$DayNum, gz$MeanVolume - gz$StdErr, gz$DayNum, gz$MeanVolume + gz$StdErr, length = 0.01, angle = 90, code = 3)

lines(gz$DayNum, pred_lm, col = "blue")
lines(gz$DayNum, pred_glm, col = "red")

legend("topleft", col = c("blue", "red"), lty = 1, legend = c("lm", "glm"))

as for the second part of the question:

library(ggplot2)
p = ggplot(data = gz, mapping = aes(x = DayNum, y=MeanVolume)) + 
  geom_line() + 
  geom_point(size = 3, color="blue") + 
  geom_smooth(method = "glm", method.args = list(family = gaussian(link = "log"))) +
  labs(x = "Days", y = "Mean Volume (mm3)", title = "Data") +
  geom_errorbar(aes(ymin = MeanVolume - StdErr, ymax = MeanVolume + StdErr), width=.2)

to extract the data from a ggplot one can use:

build = ggplot_build(p)

the data for the curve are in build$data[[3]]

p +  geom_line(data = build$data[[3]], aes(x = x, y = y), lty = 2, color = "red", size = 1.5)

This data is the same as data in pred_glm - well its a bit more dense (more data points). As far as I am aware there is no method to extract the coefficients from the ggplot just the predictions, but you can always build the glm model as described above.