# Explaining outliers in laymen terms

I have a presentation coming up where I need to explain what an outlier is to someone with little to no statistical background, as well as explain ways to identify outliers. So I guess what I am looking for help with is

1) How to explain outliers in laymen terms?

2) What are "good" ways to identify outliers?

• (2) is too broad: it takes entire books to answer it. – whuber Oct 22 '15 at 5:41
• @whuber I suppose I meant more so could you name a few standard approaches that an audience could understand and appreciate – user30490 Oct 22 '15 at 5:44
• Will you also deal with the more challenging element 3): "How to deal with outliers in my analysis"? – Michael M Oct 22 '15 at 6:55
• (2) is too broad. See this book for a good introduction. For (1) I usually write: "Observations that are inconsistent with the (multivariate) pattern of the majority of the data." – user603 Nov 3 '15 at 11:14

1) How to explain outliers in laymen terms?

I think this definition on Wikipedia can't get any simpler. An outlier is an observation point that is distant from other observations.

2) What are "good" ways to identify outliers?

One which I use very frequently in my models is the Mahalanobis Distance

Along with that, there are some other tests through which outliers can be identified:

1. Chauvenet's Criterion: Finds the probability of an outlier to be spurious.
2. Grubb's Test: Detect outliers on Univariate data.
• What does "distant from" actually mean? That would seem to be subjective, broad, and open to wildly differing interpretations. :-) – whuber Oct 22 '15 at 5:42
• @whuber You are right. It depends on a lot of things: The problem statement, the analyst's perception, the data, etc, etc. But, here as the OP is explaining the basics to the audience, he shouldn't really worry about that. The word distant from should give the gist of the definition of outliers (at least as a basic definition.). – Dawny33 Oct 22 '15 at 5:47
• That's fine--but I would like to suggest there is value in maintaining a distinction between an intuitive characterization and a definition. You have quoted the former (which is an appropriate response), but not the latter. From recent experiences in teaching graduate stats students, I would also like to mention that "distant from" is remarkably easy to misinterpret, especially when popular software (such as lm.plot in R) automatically flags extreme values regardless of how distant they are from other observations. – whuber Oct 22 '15 at 15:41