Upload
vivian-benninger
View
218
Download
4
Embed Size (px)
Citation preview
Department of Statistics, Science and Research Branch, Islamic Azad University, Tehran, Irana
Department of Mathematics, Iran University of Science and Technology, Tehran, Iranb
Department of Mathematics, Islamic Azad University, Gahram, Iranc
Introduction
where :
The usual kurtosis measure is
What is kurtosis ?
does not sort the distributions based on the height
Height
We compute for some distributions,
Uniform, Normal, Logistic, Laplace, and Cauchy for the above distributions the kurtosis Measures, ,are 1.8, 3, 4.2, 6,
and respectively.
What is kurtosis ?
Height
As the height of distributions increases,
the order of distributions will be Logistic, Cauchy, Normal, Laplace, and
Uniform respectively.
What is kurtosis ?
Height
Disadvantages of
1- It’s infinite for heavy tail distributions.
2- It doesn’t work well for some distributions such as Ali’s scale contaminated normal distributions.
where is the standard normal distribution function.
3- It can be misleading as a departure from normality.
when
But this sequence converges in distribution to the standard normal distribution as .
is not a sufficient condition for normality.
A Modified Measure of Kurtosis
Where p and q are quantile of order p and q, respectively
Properties of If f(x) > 0 for all x.
P.3. =
Proof:
since
where A(k) = And B(k) =
Lim A(k) = lim B(k) = 0 as
We show that the treatment of is as the same as
Oja (1981) says a location and scale invariant functionl T can be
named a kurtosis measure if whenever G has at leastas much Kurtosis as F according to the definition of relative kurtosis.
T is a kurtosis measure if : 1- It must be location and scale invariant i.e. T(ax+b) = T(x) for a > 0 2- It must preserve one of the orderings. Ordering << were defined in such a way that F<< G means, in some location and scale free, that G has at least as much mass in the center and tails as F i.e. if F<sG then T(F) <= T(G)
Properties of kurtosis measure
Van Zwet (1964) introduced for the class of symmetric distributions an ordering <s
defined by F <s G iff is convex for x > mf. Where mf is the point of symmetric of F. The distributions are ordered by the s ordering.
SIF < 0 iff .742 < x < 2.334 SIF > 0 iff x < .742 or x > 2.334
center
flank
tail
6 = β2(Laplace) > β2(Logistic) = 4.2
By comparing SIF of two distributions the centers of two distributions are approximately equal.
But contamination at the tails of Laplace is more than Logistic.
Contamination at the center has far less influence than that in the extreme values.
We compute SIF for Logistic and Laplace to compare the concentration probability mass in their center, flank and tails.
Let F and G be distribution functions that are symmetric and define
We consider
G =
Shape Parameter of sharply peakedness
The values of are equal to 6 for all values of
Discussion:We discuss about the properties of the usual kurtosis, then we introduce a modified measure of kurtosis and we consider the properties of this kurtosis measure.We show that this kurtosis measure is robust and its treatment is as the same as the usual kurtosis measure. Furthermore for departure from normality the shape of the parameter must be considered, so by noting the sharply peaked parameter the inference of distribution of population will be more correctly.
For < .798 the distribution is shorter than standardized normal distribution. By increasing the distributions willbe more sharper.
References: