Calculating percentiles

Ask here about academic and research issues, like designing studies, recruiting participants, choosing statistics, submitting for publication, etc.
Post Reply
User avatar
miriam
Site Admin
Posts: 8158
Joined: Sat Mar 24, 2007 11:20 pm
Location: Bucks
Contact:

Calculating percentiles

Post by miriam »

I'm interested in calculating percentiles to show where a particular individual falls compared to a really large data set. It seems like a simple concept, but in reality turns out to be quite complicated!

If you have a set of data points from 1000 people and you want to calculate where a single individual falls as a percentile, how do you do it? What do you do if a lot of people have "tied" results?

Imagine a data set like:
2
2
3
4
6
7
7
7
7
9
What percentile is a score of 2, or a score of 7?
Miriam

See my blog at http://clinpsyeye.wordpress.com
alexh
Posts: 447
Joined: Sun Sep 23, 2012 7:57 pm

Re: Calculating percentiles

Post by alexh »

I see your paper request is maybe encouraging the use of percentiles for neuropsychological data, but maybe you don't use percentiles for data like this because of the issue you are finding with such a small range? Obviously the classic neuropsychological dataset has a range of around 100 so percentiles make intuitive sense. However, say you effectively only have five clusters, then you might as well call these quintiles? I suppose that doesn't account for the numbers in each 'bin'.
Our language is a necropolis of dead metaphors. Sarbin.
User avatar
miriam
Site Admin
Posts: 8158
Joined: Sat Mar 24, 2007 11:20 pm
Location: Bucks
Contact:

Re: Calculating percentiles

Post by miriam »

Oh, I was just making up a tiny data set to illustrate the issue. The challenge is that sometimes there are clusters of scores in real data. For example, in symptom checklists you might get a large number of 0 scores at the bottom end of the scale. Or on IQ you get tight clusters of scores in the middle of the bell curve. Neuropsych data only appears to have a range of around 100 because it is transformed to fit a pre-assumed normal distribution. Whilst you are right that there are issues with sample size and skew etc, because it is basically a form of ranking, any data that is ordinal can be converted to percentiles.
Miriam

See my blog at http://clinpsyeye.wordpress.com
Post Reply