RE: st: RE: Error Bars on Histogram or Bar Plots

[Buzz Burhans <wsb2@cornell.edu>]

> Agreed that we agree, mostly. But the premise that "lots of
> people do this, so it is not discussable" is not one I sign
> up to. Lots of people use pie charts, but when I can I try
> to persuade audiences that it's a usually a bad idea. And
> in this detail, and almost all others to do with graphics,
> I stand on the shoulders of giants.
>
> More pertinently, I gave three specific arguments
> against this kind of error bar. I await counter-arguments.
I don't subscribe to the belief that just because others do something it is 
justified either. On the other hand, I don't presume to tell economists, or 
geographers, or anyone else, how their discipline should or shouldn't 
present data. I do recognize that there are conventions that differ in 
different disciplines , and may make more sense to the knowing than to the 
untutored in a particular discipline.

In any case, I don't wish to engage in a debate that has little merit. I 
expressed an opinion, made it clear that is an opinion, and that it is at 
least apparently shared by others in my or related disciplines.  I further 
think that some of the arguments raised in the earlier course of the 
discussion of graphical representations of data were a bit pedantic and 
purist. I am of the opinion that the intent of graphic representations is 
to inform the reader, and I also expect the reader to be intelligent enough 
to invest some thought in making sense of them. I think that the paramount 
issue is remembering that the intent of a graphic should be to inform, in 
many cases a graphic can accomplish this better than text. in as much as 
these issues touch on error bars, in specific, I have the following 
thoughts regarding you arguments:


> 1. Their visual signal conflates two
> distinct issues, estimate relative to
> zero and estimate relative to error
> bar.
Indeed, it does. While I have no interest in debating this, a full 
appreciation of the significance (i.e. meaningfulness, not '"statistical 
significance") of information presented by many graphics requires 
consideration of both of these "distinct issues". Sorry, but the error 
surrounding the statistic is as important a piece of evaluating the 
information as the statistic itself (again, in some / many cases).  In some 
cases graphics without error bars of the type you portrayed may either 
misinform, because without error bars the relative 'statistical" 
significance of the bars is left unaddressed, or may at best leave one 
uninformed. My opinion is that if one is going to represent an estimate to 
me, I would prefer some information about the error at the same time, or I 
do not have enough information to critically asssess the information 
presented.  So I am interested in both the magnitude of the data, i.e. 
relative to 0, and the variance. Only one or the other leaves me either ill 
informed, or as sometimes happens, misinformed when data is presented 
without error information, and that sometimes does 
occur  unfortunately.  In fact, it sometimes occurs intentionally, or in 
some cases it occurs in the company of sloppy or non-rigorous thinking.



> 2. Suppressing the bottom part of the
> error bar obliges the concerned user
> to guess mentally where it ends. That's
> very poor.
True, it imposes the necessity of some mental effort by the reader in order 
to interpret the information as fully as possible. Sometimes, I find I must 
actually read the paper as well !! instead of just looking at the graphics. 
If there is value in graphic representation of data (which I believe there 
is), then it should be as well presented as possible, but even the best 
graphic presentation of the data does not remove the effort a reader must 
make to understand it, nor does it absolve the reader  or the author from 
the need to make an intellectual effort to effectively represent and 
understand the information.  In the case of the half bars, an imperfect but 
genuine representation may require more effort from the reader, but it is 
preferable to providing even less information by omitting the bars altogether.




> 3. Such graphs take a lot of ink to show
> relatively little information.
Au contraire! A point I suppose, of graphic representation, is to 
efficiently represent a substantial amount of information in a fairly 
consolidated manner. While you may prefer to be partially informed, I 
prefer to provide, and I absolutely insist in being provided, more complete 
information that facilitates a more informed assessment of the information 
in the graphic. I think that the fact that so much work published in 
biological sciences journals uses some type of error bar suggests that 
there may be a functional basis for doing so.



> Something like a dot chart seems to me
> far preferable. Cleveland's work is relevant here.
I suppose my ignorance is showing, but I don't know who Cleveland is...and 
I have no quarrel with your liking dot charts. They don't particularly 
appeal to me, and their use is somewhat minimal in the literature I read, 
but that doesn't make them unimportant, or more importantly, 
uninformative.  However, there are many of us for whom the issues we are 
discussing are better , and in my opinion, validly, represented in other 
ways, at least frequently.

Buzz Burhans



*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/