[R] Unintended behaviour (possibly bugs)
Martin Maechler
m@ech|er @end|ng |rom @t@t@m@th@ethz@ch
Thu Jul 2 15:30:05 CEST 2020
>>>>> Alexey Shipunov
>>>>> on Wed, 1 Jul 2020 23:58:04 +0900 writes:
> Dear colleagues,
> There is a new problem with dotchart(), and it is very simple to reproduce.
> Just run example(dotchart).
> On R versions < 4, group labels ("Urban Female" and so on) were
> visible. Now they are not visible.
> If in the dotchart() code, we replace the string
> ===
> goffset <- (max(linch + offset, ginch, na.rm = TRUE) + 1/16)/lheight
> ===
> with the string
> ===
> goffset <- (max(linch + 0.2, ginch, na.rm = TRUE) + 1/16)/lheight
> ===
> everything start to be OK. Probably, the reason that in the code,
> there is another "offset" object and they clash. So if we replace this
> part of code
> ===
> offset <- cumsum(c(0, diff(as.numeric(groups)) != 0))
> y <- seq_len(n) + 2 * offset
> ===
> with
> ===
> offset1 <- cumsum(c(0, diff(as.numeric(groups)) != 0))
> y <- seq_len(n) + 2 * offset1
> ===
> everything will be well again.
Thank you.
I'll have a look *again*, and cautiously consider the above.
Indeed your second patch seems the correct one, distinguishing the
two different offsets that where conflated.
I will commit to R-devel and also to "R 4.0.2 patched" (but note
that no quick R 4.0.3 has been planned).
Note (Alexey knows, almost everbody else probably not):
This has come from another dotchart(* , ylab=.) glitch which
Alexey had reported in February and I had fixed early
March... evidently not fixed quite correctly... and yes, I'm
embarrased.
However I did mention here to have fixed it, on March 12
(--> https://stat.ethz.ch/pipermail/r-help/2020-March/465921.html )
It would have been really great if people would test such
changes, as they were in all pre-releases (R 4.0.0 alpha, beta,
RC), easily available ... and we could have fixed this even
before R 4.0.0 was released more than a month later than my
e-mail above...
Martin
> With best wishes,
> Alexey Shipunov
> пт, 13 мар. 2020 г. в 18:56, Alexey Shipunov <dactylorhiza using gmail.com>:
>>
>> Dear Martin,
>>
>> Great news, thanks!
>>
>> If you wish, please also consider my initial note about help(hist),
>> this is definitely worrying new R users.
>>
>> With best wishes,
>>
>> Alexey
>>
>> пт, 13 мар. 2020 г. в 02:16, Martin Maechler <maechler using stat.math.ethz.ch>:
>> >
>> > >>>>> Alexey Shipunov
>> > >>>>> on Tue, 18 Feb 2020 14:34:48 +0900 writes:
>> >
>> > > Thank you for the detailed explanation. I tend to agree. However, this
>> > > behavior is relatively easy to remediate:
>> >
>> > > This is the piece of the current code:
>> >
>> > > ===
>> > > if (!(is.null(labels) && is.null(glabels))) {
>> > > nmai <- par("mai")
>> > > nmai[2L] <- nmai[4L] + max(linch + goffset, ginch) + 0.1
>> > > par(mai = nmai)
>> > > }
>> > > ===
>> >
>> > > This is my proposal:
>> >
>> > > ===
>> > > yinch <- if (!is.null(ylab)) 0.4 else 0
>> > > if (!(is.null(labels) && is.null(glabels))) {
>> > > nmai <- par("mai")
>> > > nm.2 <- nmai[4L] + max(if(is.null(ylab)) 0 else 0.4) + linch + goffset, ginch) + 0.1
>> > > if (nmai[2L] < nm.2)
>> > > nmai[2L] <- nm.2
>> > > par(mai = nmai)
>> > > }
>> > > ===
>> >
>> > > Then margins and y-axis labels start to work normally. I wonder if
>> > > this (or similar) is possible to introduce into the code?
>> >
>> > > Alexey
>> >
>> > Well, I had looked at this back then (~Feb 18), and now had a
>> > considerable longer look.
>> >
>> > Your suggestion makes sense, but then it needs even more work
>> > to ensure that the 'ylab' y-axis label will be placed properly.
>> >
>> > Of course, Deepayan (author of grid-based 'lattice') is right
>> > that dotchart()s implementation is pretty hackish ... but then
>> > still.
>> >
>> > I have (+-) fixed this in the sources of "R-devel" the
>> > development version of R (which should become R 4.0.0 on April
>> > 24 as was announced today).
>> >
>> > Now, things like this (extended) example work nicely :
>> >
>> > op <- par(xaxs = "i") # 0 -- 100\%
>> > dotchart(t(VADeaths), xlim = c(0,100), bg = "skyblue",
>> > main = "Death Rates in Virginia - 1940", xlab = "rate [ % ]",
>> > ylab = "Grouping: Age x Urbanity . Gender")
>> > par(op)
>> >
>> >
>> > Thank you, Alexey, for your report and bug fix suggestion!
>> >
>> > Best regards,
>> >
>> > Martin Maechler
>> > ETH Zurich and R Core team
>> >
>> >
>> >
>> > > ........... 17:37, Deepayan Sarkar <deepayan.sarkar using gmail.com>:
>> > >>
>> > >> On Mon, Feb 17, 2020 at 10:24 AM Rui Barradas <ruipbarradas using sapo.pt> wrot=
>> > > e:
>> > >> >
>> > >> > Hello,
>> > >> >
>> > >> > Yes, this is definitely a bug.
>> > >>
>> > >> I would argue that the only bug here is that the documentation doesn't
>> > >> say that 'ylab' may not behave as expected.
>> > >>
>> > >> dotchart() is mainly designed for 2-way tables (see the VADeaths
>> > >> example), but it's implementation is really pretty hackish because it
>> > >> has to work within the limited traditional graphics framework. The
>> > >> main problem is that dot plots want to put horizontal y-axis labels
>> > >> (usually derived from factor levels), which are often longer than the
>> > >> default margins, so the margins are modified. Unfortunately they are
>> > >> only re-set on exit, and so the ylab that is plotted inside dotchart()
>> > >> may be clipped. Traditionally, Cleveland dot plots don't have a y-axis
>> > >> label; it's assumed that the factor levels are sufficient (and for
>> > >> 2-way tables, there would be two variables, so there is no sensible
>> > >> default).
>> > >>
>> > >> I doubt that dotchart() is worth fixing (except to maybe disallow
>> > >> ylab). If you want flexibility, use modern grid-based alternatives
>> > >> such as lattice::dotplot() or ggplot2.
>> > >>
>> > >> -Deepayan
>> > >>
>> > >> > Even the matrix plot is puzzling, with a "1" as top row sort-of-label
>> > >> > but no grid line. I'm trying to follow the source code of dotchart but
>> > >> > am yet to understand exactly what it does to decide the margins setting=
>> > > s.
>> > >> >
>> > >> > if (!(is.null(labels) && is.null(glabels))) {
>> > >> > nmai <- par("mai")
>> > >> > nmai[2L] <- nmai[4L] + max(linch + goffset, ginch) +
>> > >> > 0.1
>> > >> > par(mai = nmai)
>> > >> > }
>> > >> >
>> > >> > This should be moved to r-devel?
>> > >> >
>> > >> > Rui Barradas
>> > >> >
>> > >> > 03:33 de 17/02/20, Alexey Shipunov escreveu:
>> > >> > > John and Rui, thanks!
>> > >> > >
>> > >> > > However, if we use the proper object, the problem still persists:
>> > >> > >
>> > >> > > dotchart(c("3"=1, "2"=2, "1"=3), ylab="Ylab") # ylab is invisible
>> > >> > > dotchart(c("aa"=1, "b"=2, "cc"=3), ylab="Ylab") # ylab is partly visible (!!!)
>> > >> > > dotchart(c("aaa"=1, "bbb"=2, "ccc"=3), ylab="Ylab") # ylab is well visible
>> > >> > >
>> > >> > > If the object is matrix, ylab is visible:
>> > >> > >
>> > >> > > dotchart(matrix(1:3, dimnames=list(c("aa","bb","cc"), NULL)), ylab="Ylab")
>> > >> > >
>> > >> > > But the ?dotchart explicitly says that "x: either a vector or matrix
>> > >> > > of numeric values" and then "labels: a vector of labels for each
>> > >> > > point. For vectors the default is to use "names(x) = ...".
>> > >> > >
>> > >> > > So this is likely a bug. Do you agree?
>> > >> > >
>> > >> > > Alexey
>> > >> > >
>> > >> > > ..... 01:55, Rui Barradas <ruipbarradas using sapo.pt>:
>> > >> > >>
>> > >> > >> Hello,
>> > >> > >>
>> > >> > >> I believe you are wrong, the error is not in dotchart, it's in your
>> > >> > >> code. You assume that to plot an object of class "table" is the same as
>> > >> > >> to plot an object of class "numeric".
>> > >> > >>
>> > >> > >> Inline.
>> > >> > >>
>> > >> > >> =C3=80s 12:21 de 16/02/20, Alexey Shipunov escreveu:
>> > >> > >>> Dear list,
>> > >> > >>>
>> > >> > >>> I have been advised to share these with R-help instead of filling the
>> > >> > >>> bug report:
>> > >> > >>>
>> > >> > >>> 1) dotchart() does not allow to see the left axis title ('ylab') and
>> > >> > >>> cannot change the left margin (outer margin 2) of the plot
>> > >> > >>>
>> > >> > >>> The code:
>> > >> > >>>
>> > >> > >>> aa <- table(c(1, 1, 1, 2, 2, 3))
>> > >> > >>> dotchart(aa, ylab="Ylab") # does not show 'ylab'
>> > >> > >>
>> > >> > >> You are right, it does *not* show 'ylab' but the user is warned.
>> > >> > >>
>> > >> > >>
>> > >> > >> aa <- table(c(1, 1, 1, 2, 2, 3))
>> > >> > >> dotchart(aa, ylab = "Ylab") # does show 'ylab'
>> > >> > >> #Warning message:
>> > >> > >> #In dotchart(aa, ylab = "Ylab") :
>> > >> > >> # 'x' is neither a vector nor a matrix: using as.numeric(x)
>> > >> > >>
>> > >> > >>
>> > >> > >> My code:
>> > >> > >>
>> > >> > >>
>> > >> > >> (mar <- par("mar")) # new R session
>> > >> > >> #[1] 5.1 4.1 4.1 2.1 # the left margin is 4.1
>> > >> > >>
>> > >> > >> aa <- as.numeric(table(c(1, 1, 1, 2, 2, 3)))
>> > >> > >>
>> > >> > >> dotchart(aa, ylab = "Ylab") # It does show 'ylab'
>> > >> > >> old.par <- par(mar = mar + c(0, 5, 0, 0))
>> > >> > >> par("mar")
>> > >> > >> #[1] 5.1 9.1 4.1 2.1
>> > >> > >>
>> > >> > >> dotchart(aa, ylab = "Ylab") # The left margin is now 9.1, much bigger
>> > >> > >>
>> > >> > >> par(old.par) # It does change the left margin
>> > >> > >> dotchart(aa, ylab = "Ylab") # but only when a new graph is plotted.
>> > >> > >>
>> > >> > >>
>> > >> > >>
>> > >> > >>> old.par <- par(mar=c(1, 10, 1, 1)) ; dotchart(aa, ylab="Ylab") ;
>> > >> > >>> par(old.par) # does not change left margin
>> > >> > >>>
>> > >> > >>> Possible solution:
>> > >> > >>>
>> > >> > >>> I researched the problem and think that the dotchart() code will need
>> > >> > >>> few corrections. If there is an interest, I can post it here; or you
>> > >> > >>> can look at the code of shipunov::Dotchart1() function.
>> > >> > >>>
>> > >> > >>> 2) example(hist) includes two "wrong" and "extreme" examples which
>> > >> > >>> slow down and even crash R on some systems; this make it unsuitable
>> > >> > >>> for demonstration in the class and strikes beginners in R who just
>> > >> > >>> want to understand how hist() works. Actually, I did it last week (I
>> > >> > >>> was not aware of these examples), and in the class two computers hang,
>> > >> > >>> and many others were extremely slow.
>> > >> > >>>
>> > >> > >>> The code:
>> > >> > >>>
>> > >> > >>> example(hist)
>> > >> > >>>
>> > >> > >>> Possible solution:
>> > >> > >>>
>> > >> > >>> If R maintainers will enclose parts of "hist" example in \dontrun{},
>> > >> > >>> this will allow to see the code but in the same time will not strike
>> > >> > >>> beginners in R who just
>> > >> > >>> want to understand how hist() works. They will still be possible to
>> > >> > >>> run with example(..., run.dontrun=TRUE).
>> > >> > >>
>> > >> > >> Agree, it's annoying. Sometimes there's a Warning section after the
>> > >> > >> Details section. Maybe such a section could get users' attention to
>> > >> > >> those examples? At least it wouldn't hurt...
>> > >> > >>
>> > >> > >>
>> > >> > >> Hope this helps,
>> > >> > >>
>> > >> > >> Rui Barradas
>> > >> > >>
>> > >> > >>>
>> > >> > >>> With best wishes,
>> > >> > >>>
>> > >> > >>> Alexey Shipunov
>> > >> > >>> ______________________________________________
More information about the R-help
mailing list