Setting y range with bars

[Winston Chang]

Hi everyone - 

Someone just asked me a question about changing the y-min on a bar graph. When you use ylim() or scale_y_continuous(limits=xxx), of course, the bars disappear. The usual answer is to use coord_cartesian(ylim=xxx) instead.

I discovered that you can also use scale_y_continuous(limits=xxx, oob=rescale_none). Is this considered an OK way to do it? I find that the coord/scale limits thing is not very intuitive for ggplot2 users (including me sometimes!)

dat <- data.frame(x=c("A","B"), y=c(10,12))

p <- ggplot(data=dat, aes(x=x, y=y)) + geom_bar()

# Normal version

p

# No visible bars

p + scale_y_continuous(limits=c(5, 15))

# Visible bars with oob=rescale_none

p + scale_y_continuous(limits=c(5, 15), oob=rescale_none)

# Visible bars with coord_cartesian

p + coord_cartesian(ylim=c(5,15))

-Winston

[Hadley Wickham]

On Mon, Apr 23, 2012 at 12:38 AM, Winston Chang <winsto...@gmail.com> wrote:
> Hi everyone -
>
> Someone just asked me a question about changing the y-min on a bar graph.
> When you use ylim() or scale_y_continuous(limits=xxx), of course, the bars
> disappear. The usual answer is to use coord_cartesian(ylim=xxx) instead.
>
> I discovered that you can also use scale_y_continuous(limits=xxx,
> oob=rescale_none). Is this considered an OK way to do it? I find that the
> coord/scale limits thing is not very intuitive for ggplot2 users (including
> me sometimes!)

I think using squish would be canonical.

And this is a valid use, but it does create a deceiving graph.

Hadley


--
Assistant Professor / Dobelman Family Junior Chair
Department of Statistics / Rice University
http://had.co.nz/

[Winston Chang]

> Someone just asked me a question about changing the y-min on a bar graph.
> When you use ylim() or scale_y_continuous(limits=xxx), of course, the bars
> disappear. The usual answer is to use coord_cartesian(ylim=xxx) instead.
>
> I discovered that you can also use scale_y_continuous(limits=xxx,
> oob=rescale_none). Is this considered an OK way to do it? I find that the
> coord/scale limits thing is not very intuitive for ggplot2 users (including
> me sometimes!)

I think using squish would be canonical.

And this is a valid use, but it does create a deceiving graph.

I'm just wondering about how, in general, one should go about zooming in on a graph -- by using scale limits with oob=rescale_none, or with coord transforms. Both work but I think from a user's point of view, modifying the scale is less confusing.

Here's a simple example where squish can have bad consequences:

# Plot with a single segment

dat <- data.frame(x=1, xend=5, y=1, yend=2)

p <- ggplot(dat, aes(x=x, xend=xend, y=y, yend=yend)) + geom_segment()

p

# Shrink x range with default oob=censor: no segment

p + scale_x_continuous(limits=c(1,2))

# With squish, line changes slope

library(scales)

p + scale_x_continuous(limits=c(1,2), oob=squish)

# With rescale_none, it looks OK

p + scale_x_continuous(limits=c(1,2), oob=rescale_none)

# With coord_cartesian, it looks OK

p + coord_cartesian(xlim=c(1,2))

-Winston

[Hadley Wickham]

Using the coord is definitely the best way. But it might be worth considering if changing scale limits should work like coords by default. 

Hadley

--
Sent from my iPhone

[Winston Chang]

In my opinion, it makes sense for the default behavior for scale_x_continuous to NOT drop out-of-range values (oob=rescale_none). I think in most cases, if users want that data to be clipped, they'd rather do it themselves, or explicitly specify the value of oob.

The current default behavior of scale_colour_continuous seems good to me.

[Hadley Wickham]

> In my opinion, it makes sense for the default behavior for
> scale_x_continuous to NOT drop out-of-range values (oob=rescale_none). I
> think in most cases, if users want that data to be clipped, they'd rather do
> it themselves, or explicitly specify the value of oob.

I think that makes sense too. The current behaviour is confusing, and
it's rarely desired. We might also think about having an explicit
out-of-bounds flag for non-positions scales.

Could you please file an issue and assign it to 0.10? We can discuss
potential changes in more depth when we get closer to that release.

Thanks!