Percentile Worksheet Function Discrepancy?

[Selden McCabe]

I've written a program in vb.net to compute percentile rank of a data set.
To check it, I've compared it against EXCEL's PERCENTILE worksheet function.
In some cases, the Excel number is quite a bit off the calculated value.

The source of my calculations is
http://cnx.rice.edu/content/m10805/latest/
which seems pretty straightforward.

Does anyone know what algorithm Excel uses for Percentile?

Or does anyone have any suggestions on how I can track down this
discrepancy?

Thanks!
---Selden McCabe

[Niek Otten]

A bit difficult to say, with just this information.

What source figures did you use? What results did you expect? What did you
get instead? What results did Excel give?

--

Kind Regards,

Niek Otten

Microsoft MVP - Excel

"Selden McCabe" <sel...@msn.com> wrote in message
news:eC4thJh1...@TK2MSFTNGP10.phx.gbl...

[Michael R Middleton]

Selden McCabe -

> Does anyone know what algorithm Excel uses for Percentile? <

Online Help terminology: PERCENTILE(array,k)

Zero percentile (k = 0) is the minimum value (rank = 1) of the data set
array.

100th percentile (k = 1) is the maximum value (rank = n) of the array.

If k is a multiple of 1/(n-1), PERCENTILE returns the appropriate ranked
value (rank = 2 through n-1) of the array.

If k is not a multiple of 1/(n-1), PERCENTILE interpolates (linear between
two adjacent ranked values) to determine the value at the k-th percentile.

- Mike Middleton, www.usfca.edu/~middleton

[Selden McCabe]

Here is the data:
3.46
2.39
2.69
2.68
2.95
3.35
2.29
3.55
3.12
2.72
2.81
3.14
3.4
2.93
3.62
3.4
3.39
3.34
2.9
3.12
2.89

For the 25th percentile, Excel computed 2.81

using the interpolation algorithm
(which you may find at
http://cnx.rice.edu/content/m10805/latest/)

I compute it to be 2.765

Any more thoughts?

Thanks,
---Selden McCabe

[Selden McCabe]

One more detail on my calculation:

I sort the data, (there are 21 items) then compute the rank of the 25th
percentile,
wich is .25 * (21 +1) = 5.5
so the result should be .5 between the 5th and 6th numbers, whcih are 2.72
and 2.81, which would be 2.765.

However the PERCENTILE(a1:a21,.25) function in Excel yields 2.81.

Do my calculations look right? Am I missing something obvious?

Thanks!

[J.E. McGimpsey]

As Michael Middleton told you, XL only interpolates if the desired
percentile (k) is not a multiple of 1/(n-1). Since n = 21, 1/(n-1)
is 5%. 25% is a multiple of 5%, therefore no interpolation is
performed.

Instead the 6th ranked value (1:0%, 2:5%, 3:10%...6:25%) is returned
(2.81).

Note that XL's interpolation method is not the same as that found at
the site you reference, so you shouldn't expect the same results.

In article <OdE626F...@TK2MSFTNGP11.phx.gbl>,

[J.E. McGimpsey]

See my reply to your previous post.

XL's algorithm is based on the lowest value being the zeroth
percentile, so the 25th percentile should be the 6th ranked item,
not interpolated.

In article <e04P$BG2DH...@TK2MSFTNGP11.phx.gbl>,

[Selden McCabe]

Wow! Thank you, both, for your insights!

Not starting the numbering from zero was my main error, along with
multiplying the desired percentile * (N + 1) instead of (N - 1). Now I'm
getting the same results as Excel!

Thanks again!

---Selden

[Jerry W. Lewis]

This may tie things up nicely, even though it is very late for this thread:

In your link, Lane writes that "there is no universally accepted
definition of a percentile."

Hyndman and Fan, 1996, "Sample Quantiles in Statistical Packages", The
American Statistician 50(4):361-365 discuss 9 different definitions and
reference some others. Excel uses Hyndman and Fan's 7th definition,
which considers the min and max to be the 0th and 100th percentiles.
Lane recommends Hyndman and Fan's 6th definition, which considers the
min and max to be the 1/(n+1) and n/(n+1) percentiles.

Both definitions evenly space the percentiles of the remaining
observations between the min and max, and use linear interpolation to
estimate unobserved percentiles. Most of Hyndman and Fan's other
definitions work the same way, but take the percentiles of the min and
max to be somewhere in between the values chosen by Excel and by Lane.

Jerry