Re: The D'Agostino Pearson k^2 test implemented in mathematica / variance of difference sign test
- To: mathgroup at smc.vnet.net
- Subject: [mg65727] Re: The D'Agostino Pearson k^2 test implemented in mathematica / variance of difference sign test
- From: Maxim <m.r at inbox.ru>
- Date: Sun, 16 Apr 2006 03:48:55 -0400 (EDT)
- Sender: owner-wri-mathgroup at wolfram.com
[This post has been delayed due to email problems - moderator]
Suppose that we have n independent identically distributed random variables {u[1], ..., u[n]} and P[u[i] == u[j]] == 0 for i != j. We form another sequence {xi[1] = Boole[u[1] > u[2]], ..., xi[n - 1] = Boole[u[n - 1] > u[n]]} and we're looking for the variance of the sum of xi[i]:
D[N[n]] == Variance[Sum[xi[i], {i, n - 1}]] ==
Variance[Sum[xi[i], {i, n - 2}] + xi[n - 1]] ==
Variance[Sum[xi[i], {i, n - 2}]] + Variance[xi[n - 1]] +
2*Covariance[Sum[xi[i], {i, n - 2}], xi[n - 1]] ==
D[N[n - 1]] + 1/4 + 2*Sum[Covariance[xi[i], xi[n - 1]], {i, n - 2}]
For any pair of adjacent elements we have
Covariance[xi[1], xi[2]] ==
P[xi[1] == 1 && xi[2] == 1] - P[xi[1] == 1]*P[xi[2] == 1] ==
P[u[1] > u[2] > u[3]] - P[u[1] > u[2]]*P[u[2] > u[3]] ==
1/6 - 1/4 == -1/12
because all permutations of {u[1], ..., u[n]} are equally probable. For any non-adjacent elements Covariance[xi[i], xi[j]] == 0. Therefore,
D[N[n]] == D[N[n - 1]] + 1/4 + 2*(-1/12), D[N[2]] = 1/4
and D[N[n]] == (n + 1)/12 if n >= 2.
Here is a check for n = 6:
In[1]:= n = 6;
Lvalfreq = {First@ #, Length@ #}& /@ Split@ Sort@
(Count[Sign[Most@ # - Rest@ #], 1]& /@
Permutations@ Range@ n)
{Lval, Lp} = {Lvalfreq[[All, 1]], Lvalfreq[[All, 2]]/n!};
mu = Lval.Lp
sigma = ((Lval - mu)^2).Lp
Out[2]= {{0, 1}, {1, 57}, {2, 302}, {3, 302}, {4, 57}, {5, 1}}
Out[4]= 5/2
Out[5]= 7/12
And a numerical test:
In[6]:= Lcnt = Array[
Count[Sign[Most@ # - Rest@ #]&@ Array[Random[]&, n], 1]&,
10^5];
{Mean@ Lcnt, Variance@ Lcnt} - {mu, sigma} // N
Out[7]= {0.00262, 0.0033856695}
Maxim Rytin
m.r at inbox.ru
On Sat, 11 Mar 2006 11:47:33 +0000 (UTC), Darren Glosemeyer <darreng at wolfram.com> wrote:
>
> For the variance quoted on the TimeSeries page, I initially thought the
> same thing you did. Assuming the signs are independent and there are
> equal
> probabilities of getting positive and negative signs (and 0 probability
> of
> getting a 0 difference), the statistic would follow
> BinomialDistribution[n-1, 1/2], which would have a variance of
> (n-1)/4. Simulations give a variance that appears to be (n+1)/12 (which
> would still indicate a typo in the TimeSeries documentation). I haven't
> figured out why this should be the variance yet. My best guess is that
> the
> assumption of independence is not valid given the differencing and as a
> result the distribution is something other than
> BinomialDistribution[n-1, 1/2].
>
>
> Darren Glosemeyer
> Wolfram Research
>
>
> At 05:15 AM 3/10/2006 -0500, john.hawkin at gmail.com wrote:
>> Hello,
>>
>> I have two questions.
>>
>> 1. Are there any resources of .nb files available on the internet
>> where I might find an implementation of the D'Agostino Pearson k^2 test
>> for normal variates?
>>
>> 2. In the mathematica time series package (an add-on), the
>> "difference-sign" test of residuals is mentioned (url:
>> http://documents.wolfram.com/applications/timeseries/UsersGuidetoTimeSeries/1.6.2.html).
>> It says that the variance of this test is (n+1) / 2. However, it
>> would seem to me that a simple calculation gives a variance of (n-1)/4.
>> It goes as follows:
>>
>> If the series is differenced once, then the number of positive and
>> negative values in the difference should be approximately equal. If Xi
>> denotes the sign of each value in the differenced series, then
>> Mean(Xi) = 0.5(1) + 0.5(0) = 0.5
>> Var(Xi) = Expectation( (Xi - Mean(Xi))^2 )
>> = Expectation( Xi^2 -Xi + 0.25 )
>> = 0.5 - 0.5 + 0.25
>> = 0.25
>>
>> And assuming independence of each sign from the others, the total
>> variance should be the sum of the individual variances, up to n-1 for n
>> data points (since there are only n-1 changes in sign), thus
>>
>> Variance = (n-1) / 4
>>
>> There is an equivalent problem in Lemon's "Stochastic Physics" about
>> coin flips, for which the answer is listed, without proof, as (n-1)/8.
>> Because of these three conficting results I am wondering if I have made
>> an error in my calculation, and if anyone can find one please let me
>> know.
>>
>> Thank you very much,
>>
>> -John Hawkin
>