Regression line exceeds confidence band

[Bob Muenchen]

Hi All,

 

I don't recall a regression line in ggplot2 exceeding the shaded confidence band, but that's what happened in the attached plot. Is this normal, and if so, under what conditions does it happen? The smallest x value is a bit of an outlier, so perhaps that's what causes it. Here's the code that did it:

 

# Create a function to apply by group
my_ggplot <- function(x) {
  print(
    ggplot(x, aes(pretest,posttest)) +
      geom_point() + geom_smooth(method = lm) +
      scale_x_continuous(limits = my_xlim) +
      scale_y_continuous(limits = my_ylim) +
      labs(title = unique(x$workshop) )
) }

 

and my_xlim was set to 58 86 to standardize the axes across groups. I'm using ggplot2 0.9.3 and plyr 1.8.

 

Thanks,

Bob

[Ben Bolker]

Because you set the limits with scale_y_continuous, the lowermost
segment of the confidence interval (which would transgress the minimum
y-bound) is getting clipped. This is standard (if sometimes annoying)
behaviour ... you can either use coord_cartesian() to set the limits, or
library(scales); scale_y_continuous(limits = my_ylim, oob = squish) ...

Ben Bolker

> --
> You received this message because you are subscribed to the ggplot2
> mailing list.
> Please provide a reproducible example:
> https://github.com/hadley/devtools/wiki/Reproducibility
>
> To post: email ggp...@googlegroups.com
> To unsubscribe: email ggplot2+u...@googlegroups.com
> More options: http://groups.google.com/group/ggplot2

[Ben Bond-Lamberty]

Try using a single "+coord_cartesian( xlim=my_xlim, ylim=my_ylim )"
instead of those two scale_continuous calls.

Ben

[Bob Muenchen]

Hi Ben,

 

Those were both useful suggestions. The coord_cartesian worked, but it was so precise that it displayed only 1/4 of the extreme point! At first I didn't even realize that the point was there at all. It is almost invisible in the lower left corner. I used a paint package to add an arrow & some text so people would see it. 

 

The squish setting from the scales package did much better, showing the whole point and a bit of axis beyond it. Very nice!  Both plots are attached.

 

From help(squish) I never would have guessed it would have this effect. The examples are rather mysterious. Is the oob argument documented anywhere?

 

 

Thanks!

Bob