2
$\begingroup$

Recently, I have encountered Hypergeometric function $\,_2 F_1\left(\frac{1}{2},\frac{x+1}{2};\frac{3}{2};-z^2\right)$ in the context of order statistics.
In particular, I am trying to evaluate an integral $n(n-1)\int _{-\infty }^{\infty } (F(d+z)-F(z))^{n-2} f(z) f(d+z) dz $,
where $F(y) = \frac{2 e^y \ _2F_1\left(\frac{1}{2},\frac{x +1}{2};\frac{3}{2};-\frac{e^{2 y}}{x}\right)}{\sqrt{x} B\left(\frac{x}{2},\frac{1}{2}\right)} = I\left({\frac{e^{2 y}}{x+e^{2 y}}};\frac{1}{2},\frac{x}{2} \right)$ to find the sample range of the Sample Range of Fisher's z-distribution with parameters $1$ and $x$.
It is known that this hypergeometric function cannot be expressed in terms of elementary functions if x is not an integer.
That said, is there a way to approximate this integral to find the distribution of the sample range?
Is there anything at all that can be said about the distribution of the sample range for Fisher's z-distribution?
What if, instead the minimum, one wants to use the median of the sample (assuming odd $n$)? Is it possible to use the connection to Beta density to find the density of$X_{(n:n)} - X_{(k:n)}$, where $k<n$, and do not evaluate integrals of the form above directly?

EDIT: The goal is to find (approximate) analytical representation of $X_{(n:n)} - X_{(k:n)}$, so one can assess its dependency on $x$ (degrees of freedom parameter)

$\endgroup$
3
  • 2
    $\begingroup$ You don't need to approximate it: just about any statistical software package will compute it for you using, for instance, an F-Ratio distribution. $\endgroup$
    – whuber
    Commented Sep 6, 2023 at 19:13
  • 1
    $\begingroup$ @whuber Unfortunately, packages like Mathematica cannot find this density analytically, which is the goal here. I am trying to assess the dependency of the range, or, in fact, 𝑋(𝑛:𝑛)βˆ’π‘‹(𝑛/2:𝑛), from x (degrees of freedom parameter ) $\endgroup$ Commented Sep 6, 2023 at 19:18
  • $\begingroup$ @whuber Did I understand your response correctly? Could you please clarify whether you meant numerical or analytical solution? $\endgroup$ Commented Sep 6, 2023 at 19:25

1 Answer 1

1
$\begingroup$

Based on your goal "The goal is to find (approximate) analytical representation of X(n:n)βˆ’X(k:n), so one can assess its dependency on x (degrees of freedom parameter)" and that you are using Mathematica, an approximation to the distribution can be accomplished directly by numerical methods.

Here is the Mathematica code to find the distribution of the range for specified sample sizes and values of the degrees of freedom parameter $x$.

(* PDF of the range of n random samples from a common Fisher Z distribution with parameters 1 and x *)
pdf[d_?NumericQ, n_?IntegerQ, x_?NumericQ] := 
  NIntegrate[PDF[OrderDistribution[{FisherZDistribution[1, x], n}, {1, n}], {x1, x1 + d}],
  {x1, -\[Infinity], \[Infinity]}]

(* Plot pdf of range *)
Plot[pdf[d, 30, 3.75], {d, 0, 10}, PlotRange -> {Automatic, {0, Automatic}}, Frame -> True, 
  FrameLabel -> {"d", "Probability density"}, PlotRange -> All]

PDF of range

(* Mean and variance of range *)
mean = NIntegrate[d pdf[d, 30, 3.75], {d, 0, \[Infinity]}]
(* 5.06407 *)
variance = NIntegrate[(d - mean)^2 pdf[d, 30, 3.75], {d, 0, \[Infinity]}]
(* 1.7515 *)

As a partial check here are simulation results with the same parameters as the above example:

(* Generate 1,000,000 samples of size 30 from a Fisher Z distribution *)
data = RandomVariate[FisherZDistribution[1, 3.75], {1000000, 30}];
(* Calculate ranges *)
range = Differences[MinMax[#]] & /@ data;
Mean[range]
(* 5.06338 *) 
Variance[range]
(* 1.75235 *)

One can investigate the relationships of various summary statistics and the degrees of freedom parameters using the above functions. (Although using simulations to do the same thing isn't much slower.)

Exact distributions

For (very) small values of $n$ (such as $n=2$), exact formulas for the pdf's, means, and variances can be obtained.

(* Fisher's z distribution functions *)
Fdz = 2 Exp[d + z] Hypergeometric2F1[1/2, (x + 1)/2, 
     3/2, -Exp[d + z]^2/x]/(Sqrt[x] Beta[x/2, 1/2]);
Fz = 2 Exp[z] Hypergeometric2F1[1/2, (x + 1)/2, 
     3/2, -Exp[z]^2/x]/(Sqrt[x] Beta[x/2, 1/2]);

(* Associated density functions *)
fz = D[Fz, z];
fdz = fz /. z -> d + z;

(* Integrand for n=2 *)
integrand = 2 fz fdz // FullSimplify;

(* Find a table of pdf's for x = 1,2,...,10 *)
pdf = FullSimplify[
   Table[{x, Integrate[integrand, {z, -\[Infinity], \[Infinity]}, Assumptions -> d > 0]},
   {x, 1, 10}], Assumptions -> z > 0 && d > 0]// TrigToExp // Together;

TableForm[pdf[[{1, 3, 5, 7, 9}]], TableHeadings -> {None, {"x", "PDF for n=2"}}]

PDF's for odd values of x

TableForm[pdf[[{2, 4, 6, 8, 10}]], TableHeadings -> {None, {"x", "PDF for n=2"}}]

PDF's of even values of x

Means, variances, and medians can be obtained (although only numerically for medians):

means = Table[{x, Integrate[d pdf[[x, 2]], {d, 0, \[Infinity]}]}, {x, Length[pdf]}];
variances = Table[{x, Integrate[(d - means[[x, 2]])^2 pdf[[x, 2]], {d, 0, \[Infinity]}]},
  {x, Length[pdf]}];
medians = Table[
  {x, d /. FindRoot[CDF[ProbabilityDistribution[pdf[[x, 2]], {d, 0, \[Infinity]}], d] == 1/2,
  {d, means[[x, 2]]}]}, {x, Length[means]}] // Chop;
TableForm[Transpose[{Range[10], means[[All, 2]], variances[[All, 2]], medians[[All, 2]]}], 
  TableHeadings -> {None, {"x", "Mean", "Variance", "Median"}}]

Table of means, variances, and medians

$\endgroup$
5
  • $\begingroup$ Thank you so much for this thorough explanation! That was very helpful! $\endgroup$ Commented Sep 7, 2023 at 21:32
  • $\begingroup$ What sample sizes are you considering? I ask because there are closed formulas for the pdf for small values of $n$ and integer values of $x$. $\endgroup$
    – JimB
    Commented Sep 8, 2023 at 4:08
  • $\begingroup$ I am actually trying to solve the reverse problem: Given a sample of size $𝑛$ and the distance between the maximum and the median (quantities that are actually observable), what can we tell about about the degrees of freedom parameter $\nu$? So, essentially, I am trying to find the law of $\nu$ given the distance and the same size. If there were an explicit closed form pdf, it would've been a straightforward MLE / Bayesian Inference type problem. Could you please tell me where are I can find these formulas for the pdf for small values of $𝑛$? That would certainly make things easier. $\endgroup$ Commented Sep 8, 2023 at 5:14
  • $\begingroup$ Maybe I was too loose with the phrase "there are". Specifically one can "find" the pdf's using similar functions in Mathematica. I'll include an example shortly. $\endgroup$
    – JimB
    Commented Sep 8, 2023 at 17:18
  • $\begingroup$ Thank you so much! $\endgroup$ Commented Sep 8, 2023 at 17:31

Your Answer

By clicking β€œPost Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.