CODEWITHSUNDEEP
rggplot2confidence-interval
maxjohnson
maxjohnson

Reputation: 29

Coefficient plot of two regressions

I am trying to create a coefficient plot using ggplot that combines the results of two different regressions. The coefficients, standard errors and confidence interval bounds of each regression are stored as a data frame of the form:

## regression1: 
   var     betas       crse       upper       lower
1  x1  0.517251974 0.58176862  0.37751553  0.05698842
2  x2 -0.260210445 0.03521915 -0.12118217 -0.25923872
3  x3  0.680752318 0.08844444  0.75410023  0.40740441
4  x4  0.663395004 0.05090350  0.26316403  0.06362598
5  x5 -0.551992451 0.03289870 -0.08751219 -0.21647271

## regression2:
   var     betas       crse       upper       lower
1  x1 -0.343254719 0.05498965  0.01451302 -0.20104246
2  x2  0.126434568 0.02243139  0.17040108  0.08247165
3  x3 -0.178460203 0.06215729 -0.05663415 -0.30028625
4  x4  0.301058265 0.03737595  0.37431378  0.22780275
5  x5 -0.054594805 0.02037967 -0.01465139 -0.09453822

My approach was to combine both regressions into one data frame using:

combined <- rbind(regression1, regression2)

And then I use ggplot:

ggplot(combined, aes(x=var, y=betas)) +
  geom_point(aes(x=var, y=betas), 
             color="red", 
             shape=15) + 
  geom_errorbar(aes(ymin=lower, ymax=upper), 
                width=.25, 
                size=.65)

However, the lines of the two models with the coefficient / confidence intervals overlap each other in the plot and it is not easy to distinguish which one is which. Is there a way to separate the lines so that they can be distinguished? Perhaps I am using the wrong approach and I should not rbind the two plots.

Upvotes: 1

Views: 490

Answers (3)

jay.sf
jay.sf

Reputation: 73134

You may use matplot. However, your regression seems to be flawed, since betas don't lie between the CI's.

matplot(regression1[, c(2, 4:5)], type='l', lty=c(1, 2, 2), col=1)
matlines(regression2[, c(2, 4:5)], type='l', lty=c(1, 2, 2), col=2)
legend('topright', legend=paste('r', 1:2), text.col=1:2, bty='n')

enter image description here

Upvotes: 2

Ben Bolker
Ben Bolker

Reputation: 226522

@user2602640's answer gets the essential points but there are a couple of additional shortcuts you can use.

  • dplyr::bind_rows() will automatically set up an "id" column (you need to name the arguments, which is a little annoying; otherwise the id reverts to numeric indices)
  • if you're OK with lines without end caps for your error bars (recommended by the minimalist/Tufte school of quantitative graphics anyway) then geom_pointrange() will do both the points and the CIs at the same time (perhaps there's an analogous combined-points-and-errorbars geom out there somewhere in the extended ggplot-verse?)
  • you may be interested in the broom and dotwhisker packages, which can further streamline similar tasks (e.g. if you have two regressions dotwhisker::dwplot(r1, r2) will automatically construct the combined coefficient plot)

The chosen width (0.25) for position_dodge() is an aesthetic choice you may want to adjust (ggplot dodges geoms like boxplots automatically, but considers points zero-width, so you need to specify an amount of displacement)

library(tidyverse)
combined <- bind_rows(reg1=reg1, reg2=reg2, .id="model")
ggplot(combined, aes(x=var, y=betas, ymin=lower, ymax=upper, colour=model)) +
   geom_pointrange(position=position_dodge(width=0.25))

data setup

reg1 <- read.table(header=TRUE, text="
   var     betas       crse       upper       lower
 x1  0.517251974 0.58176862  0.37751553  0.05698842
 x2 -0.260210445 0.03521915 -0.12118217 -0.25923872
 x3  0.680752318 0.08844444  0.75410023  0.40740441
 x4  0.663395004 0.05090350  0.26316403  0.06362598
 x5 -0.551992451 0.03289870 -0.08751219 -0.21647271
")

reg2 <- read.table(header=TRUE, text="
   var     betas       crse       upper       lower
 x1 -0.343254719 0.05498965  0.01451302 -0.20104246
 x2  0.126434568 0.02243139  0.17040108  0.08247165
 x3 -0.178460203 0.06215729 -0.05663415 -0.30028625
 x4  0.301058265 0.03737595  0.37431378  0.22780275
 x5 -0.054594805 0.02037967 -0.01465139 -0.09453822
")

Upvotes: 1

user2602640
user2602640

Reputation: 816

You just need to create a new variable to track the two regressions and allow the colour of the points / errorbars change with that variable. The dodging allows to separate the two regressions sideways. So:

regression1$Reg <- "Regression 1"
regression2$Reg <- "Regression 2"

combined <- rbind(regression1, regression2)

ggplot(combined, aes(x=var, y=betas)) +
  geom_point(aes(x=var, y=betas, colour = Reg), 
             shape=15, position = position_dodge(width = 0.25)) + 
  geom_errorbar(aes(ymin=lower, ymax=upper, colour = Reg), 
                width=.25, position = position_dodge(width = 0.25), 
                size=.65)

Upvotes: 2

Related Questions