Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: RE: confidence intervals histograms

From   "Nick Cox" <>
To   <>
Subject   st: RE: confidence intervals histograms
Date   Tue, 15 Feb 2005 18:42:45 -0000

Histogram bars: sounds tricky. But crudely 
the error on any histogram bin count will
scale as the square root of that count. 
Unfortunately, there is no -xsc(root)- 
option on which bars would be approximately 
the same size. I suggested one some time 
ago to StataCorp. 

I suspect your problem is not literally 
one of drawing a histogram, but just a kind of generic 
fat value bar + thin error bar plot. You can follow
the Roger Newson route, or knit it yourself 
by something like 

. twoway bar ... || rcap ... 

but either way you need to have the variables to hand. 

A search through the archives will find this 
one dividing the community right down the middle. 

Some people are happy with a fat bar + thin bar 
format, even if half the thin bar is hidden beneath 
the fat bar. In my experience it is both common 
and popular in various biological sciences, but less
common and less popular in other fields. 

Others regard it as a very poor format. Stata 
user Paul Seed has used (and may even be 
responsible for) the term "detonator plot". 
At a London users' meeting some years ago 
when it was reported that some people 
wanted this kind of graph wired in as 
an official Stata command, the response 
from the floor was roughly "No. No. Terrible 
idea. Very poor". 

In my view, there are three main criticisms
of this format: 

1. The common practice of burying half an 
error bar obliges the conscientious reader
to try to reconstruct it mentally. That 
is poor graphical etiquette, to make the 
reader guess at what would be shown if
the graph had been properly prepared. 
(In Stata, this is avoidable so long 
as the fat bar has no color or the thin 
bar is drawn on top.) 

2. There is graphical equivocation in 
that fat bars usually have base 0 
and so often look much bigger than error bars. 
Whether the bar value is not 0 is, on 
the other hand, commonly not the scientific issue. 
Thus this design often puts the emphasis
in the wrong place. 

3. This format is often associated 
with a very small amount of information, 
and so a table would often be more 


* poor use does not discredit proper use 

* calling a graph by a derogatory 
name does not make it a poor kind of graph. 


> May I ask if anyone knows if there is an easy way to create 
> graphs with 
> confidence intervals displayed on top of histogram bars?

*   For searches and help try:

© Copyright 1996–2021 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index