When presenting statistical information using bar charts, when will you need to use Error Bar charts?
$\begingroup$
$\endgroup$
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
Adding error bars to a bar graph is a choice you make as a presenter to communicate more information to your audience. They are useful because they communicate visually how certain you can be, based on your data, of the specific values you are presenting.
In some cases, there is no uncertainty. Imagine you are graphing the number of students in each grade in a school district. These numbers are known so presenting the exact values without error bars makes sense.
If however you are graphing the height of students by gender and you only had time to measure students in one class, you are making statistical inferences about the larger population. In this case, error bars are helpful to communicate the range of likely true values. If the 15 boys in the class you measured averaged 48 inches you could include error bars to show that you are 95% sure that the average of all boys in the district is between say, 46 and 50.
$\begingroup$
$\endgroup$
Here is a heuristic: if the practical importance of the result depends a lot on the value of point estimate, use CIs; if the practical importance of the result depends mainly on the existence/magnitude of effect, consider leaving them out.
Here is why: error bars correct one sort of misunderstanding but invite others. The misunderstanding they correct is obvious: people assume too much precision in point estimate. But the ones they invite can be bad too; these include that all values within the interval are "equally likely"; that "big confidence intervals" are "bad"; & that point estimates w/ overlapping CIs are "not significantly different" from each other. If a particular result is has practical meaning b/c of the point estimate (e.g., likelihood a candidate will win an election or candidates' vote share in election), the former misunderstanding is more consequential -- that is, people will too likely make a mistake in relying on the result if they don't see the imprecision of it. If a particular result, however, is practically significant b/c it discloses an effect that people wouldn't otherwise likely perceive -- consider a small-sample experiment that shows some framing-effect manipulation induces a large change in the valence of subjects' affective reaction to a political candidates' message, where effect is obviously there but scale for measuring it is not that important & CIs are big relative to dimensions of scale b/c of small sample -- then the CI will often add little information, clutter things up, & invite the sort of "a little knowledge is dangerous" types of confusion I mentioned.
Another approach is to try to find some alternative to convey precision of estimate w/o inviting typical misunderstanding of CIs. Some researchers try using graphics w/ multi-color shadings around point estimate, or elongated diamond shapes for pt estimates, to denote probability density of likely "true" values around the point estimate. My understanding is that people who have examined these alternatives conclude that they are confusing too...
BTW, bar graphs are usually pretty rotten way to convey information. They are class Tufte chartjunk. There are lots of better alternatives.
