[S] Re: barcharts, piecharts, and dotplots - ease up!

eric.gibson@pharma.Novartis.com
Fri, 13 Nov 1998 11:27:25 -0500


Dear Colleagues,

Suppose a physician comes into your office and
requests a barplot for an upcoming meeting that is
30 minutes from now.
Suppose also that he/she has a set of 8 numbers representing
the average percent change from baseline in the
primary efficacy varaible for all patients by
treatment and week (visit).
A side by side display of bars across the 4
visits is helpful for showing a trend over time,
and this is what the physician has seen in the medical
literature for decades, therefore this is what he/she requests.

Week 4 6 8 12
Drug 72 68 74 78
Placebo 23 18 12 12

I was just in this position two weeks ago.
This is what I gave her (the physician).

PLOT A (What she requested)

data<-c(72,23,68,18,74,12,78,12)
xvals<-barplot(data, beside=T, space=c(.25,0), width=2.0, angle=c(35,135),
density=c(10,10), col=1, ylab="Percent Reduction", xlab="Weeks",
ylim=c(0,100) )
axis(side=1, at=apply(matrix(xvals,nrow=2), 2, mean),
labels=c("4","8","12","16"),ticks=T)
legend(locator(1), c("Drug","Placebo"), angle=c(35,135), density=c(10,10),
col=1)
text(xvals,data+3,signif(data,3))

PLOT B (a single alternative)

data<-c(72,23,68,18,74,12,78,12)
drug<-c(72,68,74,78)
placebo<-c(23,18,12,12)
week<-c(4,8,12,16)
weeks<-c(4,4,8,8,12,12,16,16)
plot(weeks, data, type="n", xlab="Week", ylab="Percent Reduction",
ylim=c(0,100), xlim=c(4,16), xaxt="n")
axis(side=1, at=c(4,8,12,16), labels=c("4","8","12","16"), tck=1, lty=2)
axis(side=2, tck=1, lty=2)
points(week,drug,pch=1,type="o",lty=1,lwd=1)
points(week,placebo,pch=2,type="o",lty=1,lwd=1)
legend(4.5, 95, legend=c("Drug","Placebo"), marks=c(1,2), bty="o", ncol=1)

When she came back a few minutes later,
I handed her the two plots, and then said
"by the way I did an extra graph because
I thought you might want to have two options
to chose from. Go with whatever you prefer."

* I could have included many alternative plots,
such as a dotplot, but I kept it simple.
* I could have bored her to death with a long monologue
about how barplots aren't really the ideal way to display
data, but I did not.
* I could have pulled out Cleveland's book (an excellent book)
and given her a lecture about the misuse and abuse
of things like piecharts and barplots, but I didn't.
* I could have talked down to her in a belittling fashion
as though she were somehow statistically illiterate
(which she isn't), but I didn't.

I simply smiled and gave her what she asked for along with
a single, simple alternative, and told her to go with
what she feels is best.

This approach does not always work. Sometimes the physician
elects to use the barplot or the piechart anyway,
but I don't think Biblical Armageddon will result because
a pie chart is used.
The majority (albeit a narrow one) of the time, this approach
works well and the better graphical display wins the day because
the physician is convinced by comparing the two graphs.

If I had a dollar for every time I heard a statistician lecture
a scientist about piecharts and barcharts I could retire young
You can attract more flies with honey, than with vinegar.
If statisticians wonder why piecharts and barcharts are still
so pervasive, they have only themselves to blame.
Use more honey.

Best wishes,
Eric Gibson, Ph.D.
Novartis Pharmaceuticals
The opinions expressed herein are those of Eric Gibson and
in no way represent Novartis Pharmaceuticals.
eric.gibson@pharma.novartis.com

-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news