[R] CDF of Sample Quantile

peter dalgaard pdalgd at gmail.com
Tue Mar 8 12:23:34 CET 2011


On Mar 7, 2011, at 19:12 , Bentley Coffey wrote:

> Just to tie up this thread, I wanted to report my solution:
> 
> When (n-1)p is an integer, there is a closed form solution:
> pbinom(j-1,n,...)
> 
> When it is not an integer, its fairly easy to approximate the solution by
> interpolating between the closed-form solutions: fitting log(1 - probability
> from closed form solution) on an orthogonal polynomial in n.  This is a
> _very_ fast and fairly accurate solution.
> 
> Thanks to all who offered their help...


If you have too much time on your hand, Wikipedia has the joint density of two order statistics, from which you could probably proceed to find the marginal density of a linear combination of two neighboring order stats. Just take a large piece of paper and a couple of spare days....  Numerical integration might do the job, with some care.

-p


> 
> On Thu, Feb 17, 2011 at 11:11 PM, Bentley Coffey
> <bentleygcoffey at gmail.com>wrote:
> 
>> 
>> Duncan,
>> 
>> I'm not sure how I missed your message. Sorry. What you describe is what I
>> do when (n-1)p is an integer so that R computes the sample quantile using a
>> single order statistic. My later posting has that exact binomial expression
>> in there as a special case. When (n-1)p is not an integer then R uses a
>> weighted average of 2 order statistics, in which case I'm left with my
>> standing problem...
>> 
>> 
>> On Mon, Feb 14, 2011 at 2:26 PM, Duncan Murdoch <murdoch.duncan at gmail.com>wrote:
>> 
>>> On 14/02/2011 9:58 AM, Bentley Coffey wrote:
>>> 
>>>> I need to calculate the probability that a sample quantile will exceed a
>>>> threshold given the size of the iid sample and the parameters describing
>>>> the
>>>> distribution of each observation (normal, in my case). I can compute the
>>>> probability with brute force simulation: simulate a size N sample, apply
>>>> R's
>>>> quantile() function on it, compare it to the threshold, replicate this
>>>> MANY
>>>> times, and count the number of times the sample quantile exceeded the
>>>> threshold (dividing by the total number of replications yields the
>>>> probability of interest). The problem is that the number of replications
>>>> required to get sufficient precision (3 digits say) is so HUGE that this
>>>> takes FOREVER. I have to perform this task so much in my script
>>>> (searching
>>>> over the sample size and repeated for several different distribution
>>>> parameters) that it takes too many hours to run.
>>>> 
>>>> I've searched for pre-existing code to do this in R and haven't found
>>>> anything. Perhaps I'm missing something. Is anyone aware of an R function
>>>> to
>>>> compute this probability?
>>>> 
>>>> I've tried writing my own code using the fact that R's quantile()
>>>> function
>>>> is a linear combination of 2 order statistics. Basically, I wrote down
>>>> the
>>>> mathematical form of the joint pdf for the 2 order statistics (a function
>>>> of
>>>> the sample size and the distribution parameters) then performed a
>>>> pseudo-Monte Carlo integration (i.e. using Halton Draws rather than R's
>>>> random draws) over the region where the sample quantile exceeds the
>>>> threshold. In theory, this should work and it takes about 1000 times
>>>> fewer
>>>> clock cycles to compute than the Brute Force approach. My problem is that
>>>> there is a significant discrepancy between the results using Brute Force
>>>> and
>>>> using this more efficient approach that I have coded up. I believe that
>>>> the
>>>> problem is numerical error but it could be some programming bug;
>>>> regardless,
>>>> I have been unable to locate the source of this problem and have spent
>>>> over
>>>> 20 hours trying to identify it this weekend. Please, somebody help!!!
>>>> 
>>>> So, again, my question: is there code in R for quickly evaluating the CDF
>>>> of
>>>> a Sample Quantile given the sample size and the parameters governing the
>>>> distribution of each iid point in the sample?
>>>> 
>>> 
>>> I think the answer to your question is no, but I think it's the wrong
>>> question.  Suppose Xm:n is the mth sample quantile in a sample of size n,
>>> and you want to calculate P(Xm:n > x).  Let X be a draw from the underlying
>>> distribution, and suppose P(X > x) = p.  Then the event Xm:n > x
>>> is the same as the event that out of n independent draws of X, at least
>>> n-m+1 are bigger than x:  a binomial probability.  And R can calculate
>>> binomial probabilities using pbinom().
>>> 
>>> Duncan Murdoch
>>> 
>>> 
>> 
> 
> 	[[alternative HTML version deleted]]
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.

-- 
Peter Dalgaard
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Email: pd.mes at cbs.dk  Priv: PDalgd at gmail.com



More information about the R-help mailing list