An Epsilon Half Normal Slash Distribution and Its Applications to Nonnegative Measurements

We introduce a new class of the slash distribution using the epsilon half normal distribution. The newly defined model extends the slashed half normal distribution and has more kurtosis than the ordinary half normal distribution. We study the characterization and properties including moments and some measures based on moments of this distribution. A simulation is conducted to investigate asymptotically the bias properties of the estimators for the parameters. We illustrate its use on a real data set by using maximum likelihood estimation.


Introduction
The epsilon half normal distribution, proposed by Castro et al. [1], is widely used for nonnegative data modeling, for instance, to consider the lifetime process under fatigue.We say that a random variable X has an epsilon half normal distribution with parameters 0   and , denoted by , if its density function is given by, for , where   .


denotes the standard normal density function.When , the epsilon half normal distribution reduces to the half normal distribution investigated in [2][3][4].0   Castro et al. [1] provided mathematical properties of the epsilon half normal distribution and discussed some inferential aspects related to the maximum likelihood estimation.
On the other hand, a random variable has a standard slash distribution with parameter , introduced in [5], if can be represented as where and are independent.It generalizes normality and has been much studied in the statistical literature.
For the limit case , yields the standard normal distribution q     SL q Z .Let , the canonical slash distribution follows, see [6].It is well known that the standard slash density has heavier tails than those of the normal distribution and has larger kurtosis.It has been very popular in robust statistical analysis and studied by some authors.

q 
The general properties of this canonical slash distribution were studied in [5,7].Kafadar [8] investigated the maximum likelihood estimates of the location and scale parameters.
Gómez et al. [9] replaced standard normal random variable Z by an elliptical distribution and defined a new family of slash distributions.They studied its general properties of the resulting families, including their moments.Genc [10] proposed the univariate slash by a scale mixtured exponential power distribution and investigated asymptotically the bias properties of the estimators.Wang et al. [11] introduced the multivariate skew version of this distribution and examined its properties and inferences.They substituted the standard normal random variable Z by a skew normal distribution studied in [12] to define a skew extension of the slash distribution.
Olmos et al. [3] introduced the slashed half normal distribution by a scale mixtured half normal distribution and showed that the resulting distribution has more kurtosis than the ordinary half normal distribution.Since the epsilon half normal distribution is an extension of the half normal distribution, it is naturally to define a slash distribution based on it in which skewness  , EHN   and thick tailed situations may exist.It leads to a new model on nonnegative measurements with more flexible asymmetry and kurtosis parameters.
The paper is organized as follows: in Section 2, we introduce the new slash distribution and study its relevant properties, including the stochastic representation etc.In Section 3 we discuss the inference, moments and maximum likelihood estimation for the parameters.Simulation studies are performed to investigate the behaviors of estimators in Section 4. In Section 5, we give a real illustrative application and report the results.Section 6 concludes our work.

Stochastic Representation Definition 2.1 A random variable
has an epsilon half normal slash distribution if it can be represented as the ratio where defined in (1) and , ,q  HNS  , , .We denote it as .
. Then, the density function of is given by, for where 0   , , .0 1    0 q  Proof.From (1), the joint probability density function of X and U is given by, for , 0, 0 Using the transformation: the joint probability density function of Y and U is given by, for , 0, 0 The marginal density function of is given by After changing the variable into 1 q t u  , the dendity function will be obtained as stated.

Remark 2.3 If
which is the density function for the slashed half normal distribution studied by [3].As , The limit case of the epsilon half normal slash distribution is the epsilon half normal distribution.For 1 q    , the canonical case follows.
Figure 1 shows some plots of the density function of the epsilon half normal slash distribution with various parameters.
The cumulative distribution function of the epsilon half normal slash distribution is given as follows.For , where is the cumulative distribution function for the standard normal random variable.
where 0 Proposition 2.4 shows that the epsilon half normal slash distribution can be represented as a scale mixture of an epsilon half normal distribution and uniform distribution.The result provides another way besides the definition (3) to generate random numbers from the epsilon half normal slash distribution Proposition 2.7 Let , then the moment generating function of Y is given by, for  , ,

Moments and Measures Based on Moments
 (8) In this section, we derive the moment generating function,the k-th moment and some measures based on the moments.
where 0 , then the moment generating function of X is given by Proof.From Proposition 2.4 and using properties of the conditional expectation, we have and the result follows.
Proposition 2.8 Let

 
, The following results are immediate consequences of (10) Proof.From the stochastic representation defined in (3) and the results in (9), the claim follows in a straightforward manner.
For the standardized skewness and .The skewness and kurtosis coefficients of are given by where BB q q q q q q q q q q q q q q q Remark 2.12 As , the skewness coefficient converges.
and the kurtosis coefficient converges as well 12 1 10 5 π 48 1 3 π 48 1 64 1 1 6 π which are the corresponding skewness and kurtosis coefficients for the epsilon half normal distribution
Figure 2 shows the skewness and kurtosis coefficients with various parameters for the model.The skewness and kurtosis coefficients decrease as increases.The parameter  , ,

EHNS q
   q  does not affect the two coefficents.

Maximum Likelihood Inference
In this section, we consider the maximum likelihood estimation about the parameters of the model defined in (4).For , where .

 
, ,q Suppose 1 2 is a random sample of size from the epsilon half normal slash distribution , , , n y y y  n   , , EHNS q   .Then the log-likelihood function can be written as The estimates of the parameters maximize the likelihood function.By taking the partial derivatives of the log-likelihood function with respect to , ,q   respectively and equalizing the obtained expressions to zero, we obtain the following maximum likelihood estimating equations.
The maximum likelihood estimating equations above are not in a simple form.In general, there are no implicit expression for the estimates.The estimates can be obtained through some numerical procedures such as Newton-Raphson method.Many programs provide routines to solve such maximum likelihood estimating equations.
In this paper, all the computations are performed using software R. The MLE estimators are computed by the optim function which uses L-BFGS-B method.In the following section, a simulation is conducted to illustrate the behavior of the MLE.
For asymptotic inference of , the Fisher information matrix I θ plays a key role.It is well known that its inverse is the asymptotic variance matrix of the maximum likelihood estimators.For the case of a single observation   1 n  , we take the second order derivatives of the log-likelihood function in (14) and the Fisher information matrix is defined as is the maximum likelihood estimator of , we have θ Proof.It follows directly by the large sample theory for maximum likelihood estimators and the Fisher information matrix given above.

Data Generation
In this section, we present how to generate the random numbers from the epsilon half normal slash distribution .

 
, , EHNS q   Proposition 4.1 Let , Z V be two independent random variables, where , where 0 The density function of X is

Behavior of MLE
In this section, we perform a simulation study to illus trate the behavior of the MLE estimators for parameters  , and . q It is known that as the sample size increases, the distribution of the MLE tends to the normal distribution with mean   , ,q   and covariance matrix equal to the inverse of the Fisher information matrix.However, the log-likelihood function given in ( 14) is a complex expression.It is not generally possible to derive the Fisher information matrix.Thus, the theoretical properties (asymptotically normal, unbiased etc) of the MLE estimators are not easily derived.We study the properties of the estimators numerically.
We first generate 500 samples of size 50 n  and 100 n  from the

 
, , EHNS q   distribution for fixed parameters.The estimators are computed by the optim function which uses L-BFGS-B method in software R. The empirical means and standard deviations(SD) of the estimators are presented in Table 1.
It can be seen from Table 1 that the parameters are well estimated and the estimates are asymptotically un-

Real Data Illustration
In this section, we consider the s life of fatigue fracture of Kevlar 49/epoxy that are subject to the pressure at the 90% level.The data set has been previously studied in [1, 3,13,14].Table 2 summarizes descriptive statistics of the data set where 1 b and 2 b are sample asymmetry and kurtosis coefficients, respectively.This data set indicates non negative asymmetry.
We fit the data set with the half normal, the epsilon half normal and the epsilon half normal slash distributions using maximum likelihood method.The results are reported in Table 3.The usual Akaike information criterion (AIC) and Bayesian information criterion (BIC) to measure of the goodness of fit are also computed.tion is conducted and demonstrates the good perto thank the anonymous review-d epsilon half normal, the half normal distribution etc. Probabilistic and inferential properties are derived.A imula s formance of the maximum likelihood estimators.We apply the model to a real dataset and the results demonstrate that the proposed model is very useful and flexible for non negative data.

Figure 1 .
Figure 1.The density function of

Figure 2 .
Figure 2. The plot for the skewness k is the number of parameters in the distribution and is the maximized value of the likelihood function.e results indicate that EHNS model fits best.Figures 3(a) and (b) display the fitted models using the MLE estimates.

LFigure 3 .
Figure 3. Models fitted for the life of fatigue fracture data set.a) Histogram and fitted curves; b) Empirical and fitte CDF.

Table 1 . Empirical means and SD for the MLE estimators of  , and .
 q

Maximum likelihood parameter estimates(with T (SD)) of the HN, EHN and EHNS models for the stress- rupture data.
iased.The empirical mean square errors decrease as tress-rupture data set, the b sample size increases as expected.