stats::line() does not produce correct Tukey line when n mod 6 is 2 or 3

line(1:9+rnorm(9,,1e-1),1:9+rnorm(9,,1e-1))

On 29 May 2017, at 07:21 , GlenB <glnbrntt at gmail.com> wrote:

Tukey divides the points into three groups, not the x and y values

separately.

I'll try to get hold of the book for a direct quote, might take a couple

of days.

Ah well, I can't get it for a week. But the fact that it's often called
Tukey's three group line (try a search on *tukey three group line* and
you'll get plenty of hits) is pretty much a giveaway.


On Mon, May 29, 2017 at 2:19 PM, GlenB <glnbrntt at gmail.com> wrote:

Tukey divides the points into three groups, not the x and y values
separately.

I'll try to get hold of the book for a direct quote, might take a couple
of days.



On Mon, May 29, 2017 at 8:40 AM, Duncan Murdoch <murdoch.duncan at gmail.com>
wrote:

On 27/05/2017 9:28 PM, GlenB wrote:

Bug: stats::line() does not produce correct Tukey line when n mod 6 is 2
or
3

Example: line(1:9,1:9) should have intercept 0 and slope 1 but it gives
intercept -1 and slope 1.2

Trying line(1:i,1:i) across a range of i makes it clear there's a cycle
of
length 6, with four of every six correct.

Bug has been present across many versions.

The machine I just tried it on just now has R3.2.3:

If you look at the source (in src/library/stats/src/line.c), the
explanation is clear:  the x value is chosen as the 1/6 quantile (according
to a particular definition of quantile), and the y value is chosen as the
median of the y values where x is less than or equal to the 1/3 quantile.
Those are different definitions (though I think they would be
asymptotically equivalent under pretty weak assumptions), so it's not
surprising the x value doesn't correspond perfectly to the y value, and the
line ends up "wrong".

So is it a bug?  Well, that depends on Tukey's definition.  I don't have
a copy of his book handy so I can't really say.  Maybe the R function is
doing exactly what Tukey said it should, and that's not a bug.  Or maybe R
is wrong.

Duncan Murdoch

	[[alternative HTML version deleted]]