[R] Can we get rid of bar charts with error bars?

[Frank E Harrell Jr]

Jim Lemon wrote:
> On 12/04/2009 12:56 AM, Frank E Harrell Jr wrote:
>> Bar charts with error bars are far inferior to dot charts and other 
>> types of displays.  One of many problems is demonstrated if you draw a 
>> bar chart displaying temperature in F then re-draw it on the degrees C 
>> scale.  See http://biostat.mc.vanderbilt.edu/DynamitePlots for much 
>> more information.  The error bars lull us into an assumption that 
>> symmetric confidence intervals are OK, among other things.
>>
>> Frank
>>
> I could promote the "dispersion" function as capable of displaying 
> asymmetric confidence intervals on whatever location indicators you 
> prefer (have a look at the second example for "hierobarp" for example), 
> but the problem is deeper than that. The point of graphic displays is to 
> convey information to someone else, not to tell ourselves what we 
> already know. Do people cheat with pie charts, bar charts, etc? Sure, 
> and we could cheat with dot charts, too. Graphic displays are typically 
> shown to an audience that knows less about the topic than the presenter 
> in the hope that an obscure relationship will be clarified. If I am 
> presenting to colleagues in my field, I will use much more informative 
> and difficult to understand graphics than if I am summarizing the same 
> results to the general public. If both groups leave the respective 
> presentations enlightened, I have done well.
> 
> Jim
> 
> 

Jim I'll just add that in many cases the best graphics are just as easy 
or easier to read and understand than bad graphics.  And dot charts have 
a much less chance of being misleading than bar and pie charts, 
especially if more data characteristics or the data themselves appear.

Frank
-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University