Normal Distribution
Normal Distribution
Normal Distribution
2 Normal Distribution
1
2
2
e(x) /2 ,
2
< x < ,
(2.1)
and
x
1
2
2
FX (x) =
e(y) /2 dy
2
1
x
, < x < , < < , > 0, (2.2)
= [1 + er f
2
2
where er f (.) denotes error function, and and are location and scale parameters,
respectively.
(2.3)
and
x
1
2
FX (x) =
et /2 dt, < x < ,
2
x
1
, < x < .
1 + er f
=
2
2
(2.4)
for different values of and 2 , are provided in Figs. 2.1 and 2.2, respectively, by
using Maple 10. The effects of the parameters, and 2 , can easily be seen from
these graphs. Similar plots can be drawn for other values of the parameters. It is
clear from Fig. 2.1 that the graph of the pdf f X (x) of a normal random variable,
X N (, 2 ), is symmetric about mean, , that is f X ( + x) = f X ( x),
< x < .
10
2 Normal Distribution
2.1.3.1 Mode
The mode or modal value is that value of X for which the normal probability density
function f X (x) defined by (2.1) is maximum. Now, differentiating with respect to x
Eq. (2.1), we have
f X
(x) =
(x ) e (x)
3
2 /2 2
which, when equated to 0, easily gives the mode to be x = , which is the mean,
that is, the location parameter of the normal distribution. It can be easily seen that
f X (x) < 0. Consequently, the maximum value of the normal probability density
function f X (x) from (2.1) is easily obtained as f X () = 1 . Since f (x) = 0
2
has one root, the normal probability density function (2.1) is unimodal.
2.1.3.2 Cumulants
The cumulants kr of a random variable X are defined via the cumulant generating
function
11
g(t) =
r =1
kr
tr
, where g(t) = ln E(etX ) .
r!
For some integer r > 0, the r th cumulant of a normal random variable X having the
pdf (2.1) is given by
, when r = 1;
r = 2 , when r = 2;
0, when r > 2
2.1.3.3 Moment Generating Function
The moment generating function of a normal random variable X having the pdf (2.1)
is given by (see, for example, Kendall and Stuart (1958), among others)
M X (t) = E
et X = et +
1
2
t2 2
2.1.3.4 Moments
For some integer r > 0, the r th moment about the mean of a normal random variable
X having the pdf (2.1) is given by
E X
= r =
r (r !)
2 2 [(r/2)!]
, for r even;
0,
(2.5)
for r odd
3
3/2 = 0;
2
4
= 3.
22
12
2 Normal Distribution
Since the coefficient of kurtosis, that is, 2 (X ) = 3, it follows that the normal
distributions are mesokurtic distributions.
2 1
2
(ii) Mean Deviation: 2
eit X = eit
1
2
t2 2
, i=
1.
2.1.3.8 Entropy
For some > 0, entropy of a random variable X having the pdf (2.1) is easily given
by
H X [ f X (x)] = E[ ln( f X (X )]
f X (x) ln [ f X (x)] d x,
=
= ln
2 e
(see, for example, Lazo and Rathie (1978), Jones (1979), Kapur (1993), and
[ f X (x)])
> 0, and
Suhir (1997), among others). It can be easily seen that d(H X d
d 2 (H X [ f X (x)])
d 2
< 0, > 0, . It follows that that the entropy of a random variable X having the normal pdf (2.1) is a monotonic increasing concave function of
> 0, . The possible shape of the entropy for different values of the parameter
is provided below in Fig. 2.3, by using Maple 10. The effects of the parameter
on entropy can easily be seen from the graph. Similar plots can be drawn for others
values of the parameter .
13
entropy
10
12
14
16
18
20
sigma
2.1.4 Percentiles
This section computes the percentiles of the normal distribution, by using Maple 10.
For any p(0 < p < 1), the (100 p)th percentile (also called the quantile of order p)
of N (, 2 ) with the pdf f X (x) is a number z p such that the area under f X (x) to
the left of z p is p. That is, z p is any root of the equation
z p
(z p ) =
f X (u)du = p.
Using the Maple program, the percentiles z p of N (, 2 ) are computed for some
selected values of p for the given values of and , which are provided in Table 2.1,
when = 0 and = 1. Table 2.1 gives the percentile values of z p for p 0.5. For
p < 0.5, use 1 Z1p.
Table 2.1 Percentiles of
N (0, 1)
zp
0.5
0.6
0.7
0.75
0.8
0.9
0.95
0.975
0.99
0.995
0.9975
0.999
0.0000000000
0.2533471031
0.5244005127
0.6744897502
0.8416212336
1.281551566
1.644853627
1.959963985
2.326347874
2.575829304
2.807033768
3.090232306
14
2 Normal Distribution
Dn
2n ,
Dn = ln ln n + ln 4, n = (2 ln n)1/2 , bn (2 ln n)1/2 .
2
3s
1s
f X (x) =
FX (x) =
, where s > 0,
respectively, given by
and
2
3s
1s
1 x
s,
1
2
1
s,
1e
s
x
(2.6)
s
s
1
s
if x
s
(2.7)
, if x >
15
f Z (z) =
s
s
1 e|z|
(2.8)
To describe the shapes of the generalized normal distribution, the plots of the pdf
(2.6), for = 0, = 1, and different values of s, are provided in Fig. 2.4 by using
Maple 10. The effects of the parameters can easily be seen from these graphs. Similar
plots can be drawn for others values of the parameters. It is clear from Fig. 2.4 that
the graph of the pdf f X (x) of the generalized normal random variable is symmetric
about mean, , that is
f X ( + x) = f X ( x), < x < .
2.2.1.2 Mode
It is easy to see that the mode or modal value of x for which the generalized normal
probability density function f X (x) defined by (2.6) is maximum, is given by x = ,
and the maximum value of the generalized normal probability density function (2.6)
16
2 Normal Distribution
is given by f X () =
s
.
2
1s
2.2.1.3 Moments
(i) For some integer r > 0, the r th moment of the generalized standard normal
random variable Z having the pdf (2.8) is given by
r + 1
1 + (1)r
=
s
2
1s
(2.9)
(i) For some integer n > 0, the nth moment and the nth central moment of the
generalized normal random variable X having the pdf (2.6) are respectively given
by the Eqs. (2.10) and (2.11) below:
E
and
Xn
(n )
=
k
n
1 + (1)k
k +s 1
k
k=0
2
1s
n
( n ) 1 + (1)n
n +s 1
n
E (X ) =
2
1s
(2.10)
(2.11)
2
3s
(ii) Variance: V ar (X ) = 2 =
, > 0, s > 0;
1s
3
(iii) Coefficient of Skewness: 1 (X ) = 3/2 = 0;
2
1 5
s
s
4
(iv) Coefficient of Kurtosis: 2 (X ) = 2 =
3 2 , s > 0;
2
17
(v) Median (X ) : ;
2s
(vi) Mean Deviation: E |X | =
, s > 0.
1s
2.2.1.5 Renyi and Shannon Entropies, and Songs Measure of the Shape of the
Generalized Normal Distribution
These are easily obtained as follows, (for details, see, for example, Nadarajah (2005),
among others).
(i) Renyi Entropy: Following Renyi (1961), for some reals > 0,
= 1,
the entropy of the generalized normal random variable X having the pdf (2.6)
is given by
1
R ( ) =
ln
1
+
[ f X (X )] d x
s
ln ( )
=
ln
, > 0, s > 0, > 0,
= 1.
s ( 1)
2
1s
(ii) Shannon Entropy: Following Shannon (1948), the entropy of the generalized
normal random variable X having the pdf (2.6) is given by
H X [ f X (X )] = E[ ln( f X (X )] =
f X (x) ln [ f X (x)] d x,
which is the particular case of Renyi entropy as obtained in (i) above for
1. Thus, in the limit when 1 and using LHospitals rule, Shannon
entropy is easily obtained from the expression for Renyi entropy in (i) above
as follows:
s
1
ln
H X [ f X (X )] =
, > 0, s > 0.
s
2
1s
(iii) Songs Measure of the Shape of a Distribution: Following Song (2001), the
gradient of the Renyi entropy is given by
ln ( )
1
d
1
(2.12)
R ( ) =
R ( ) =
d
s ( 1)
( 1)2
which is related to the log likelihood by
18
2 Normal Distribution
1
R (1) = V ar [ln f (X )] .
2
Thus, in the limit when 1 and using LHospitals rule, Songs measure of the
shape of the distribution of the generalized normal random variable X having the
pdf (2.6) is readily obtained from the Eq. (2.12) as follows:
2 R (1) =
1
,
s
which can be used in comparing the shapes of various densities and measuring
heaviness of tails, similar to the measure of kurtosis.
2 1 12
e
and
FX (x) = er f
x 2
(2.13)
(2.14)
where x , < < , > 0, and er f (.) denotes error function, (for
details on half-normal distribution and its applications, see, for example, Altman
(1993), Chou and Liu (1998), Bland and Altman (1999), McLaughlin (1999), Wiper
et al. (2005), and references therein). Clearly, X = + |Z |, where Z
N (0, 1) has a standard normal distribution. On the other hand, the random variable
X = |Z |follows a negative (general) half- normal distribution. In particular,
if X N 0, 2 , then it is easy to see that the absolute value |X | follows a halfnormal distribution, with its pdf f |X | (x) given by
19
f |X | (x) =
1 x 2
2
e 2 ( ) i f x 0
2
(2.15)
if x < 0
By taking 2 = 2 2 in the Eq. (2.15), more convenient expressions for the pdf and
cdf of the half-normal distribution are obtained as follows
2
x
2 e
if x 0
f |X | (x) =
(2.16)
0
if x < 0
and
F|X | (x) = er f
(2.17)
20
2 Normal Distribution
2.2.3.3 Mode
It is easy to see that the mode or modal value of x for which the half-normal probability
density function f X (x) defined by (2.13) is maximum, is given at x = , and the
maximum value
of the half-normal probability density function (2.13) is given by
f X () =
unimodal.
2
.
2.2.3.4 Moments
(i) kth Moment of the Standardized Half-Normal Random Variable: If the
half-normal random variable X has the pdf given by the Eq. (2.13), then the
H N (0, 1) will
standardized half-normal random variable |Z | = X
1 2
2 e 2 z
2
if z 0
if z < 0
(2.18)
For some integer k > 0, and using the following integral formula (see
Prudnikov et al. Vol. 1, 1986, Eq. 2.3.18.2, p. 346, or Gradshteyn and Ryzhik
21
t 1 e t dt =
, where , Re , Re > 0,
the kth moment of the standardized half-normal random variable Z having the
pdf (2.18) is easily given by
E
Zk
k
1
= 22
k + 1
,
2
(2.19)
where
(.) denotes gamma function.
(ii) Moment of the Half-Normal Random Variable: For some integer n > 0,
the nth moment (about the origin) of the half-normal random variable X having
the pdf (2.13) is easily obtained as
n
n
n
n k k E Z k
= E ( + z ) =
= E X
k
k=0
n
1
n k nk k
k + 1
(2.20)
=
22
k
2
k=0
From the above Eq. (2.20), the first four moments of the half-normal random
variable X are easily given by
2
,
2
2 = E X 2 = 2 + 2
+ 2 ,
2 2
2 3
3
3
2
+ 3 + 2
,
3 = E X
= + 3
1
= E [X ] = +
(2.21)
(2.22)
(2.23)
and
4
= E
X
2 3
2
2 2
+ 6 + 8
3 + 3 4 . (2.24)
= +4
(iii) Central Moment of the Half-Normal Random Variable: For some integer
n > 0, the nth central moment (about the mean 1 = E (X )) of the halfnormal random variable X having the pdf (2.13) can be easily obtained using
the formula
22
2 Normal Distribution
n = E
n
1
n
n n k k
1
=
E X ,
k
k=0
(2.25)
where E X
= k denotes the kth moment, given by the Eq. (2.20), of the
half-normal random variable X having the pdf (2.13).
k
Thus, from the above Eq. (2.25), the first three central moments of the halfnormal random variable X are easily obtained as
2
2
2 ( 2)
,
= 2 1 =
X 1
3
3 = 3 = E X 1
(2.26)
3
2 3 (4 )
,
(2.27)
= 3 31 2 + 2 1 =
2 = E
and
4
2
4
4 = 4 = E X 1
= 4 41 3 + 6 1 2 3 1
4 3 2 4 12
=
.
(2.28)
2
2.2.3.5 Mean, Variance, and Coefficients of Skewness and Kurtosis
These are easily obtained as follows:
2
;
2
(ii) Variance : V ar (X ) = 2 = 2 1
, > 0;
2 (4 )
3
(iii) Coefficient of Skewness : 1 (X ) = 3/2 =
0.9953;
2
( 2)3
8 ( 3)
4
(iv) Coefficient of Kurtosis : 2 (X ) = 2 =
0.7614;
2
( 2)2
(i) Mean : 1 = E (X ) = +
23
2.2.3.6 Median (i.e., 50th Percentile or Second Quartile), and First and Third
Quartiles
These are derived as follows. For any p(0 < p < 1), the (100 p)th percentile
(also called the quantile of order p) of the half-normal distribution, X |,
H N (, ), with the pdf f X (x) given by (2.13), is a number z p such that the area
under f X (x) to the left of z p is p. That is, z p is any root of the equation
z p
F(z p ) =
f X (t)dt = p.
(2.29)
For p = 0.50, we have the 50th percentile, that is, z 0.50 , which is called the median
(or the second quartile) of the half-normal distribution. For p = 0.25 and p = 0.75,
we have the 25th and 75th percentiles respectively.
z 0.50
f X (t)dt =
2 1
z 0.50
2
1 t
e 2
dt.
(2.30)
Substituting t = u in the Eq. (2.30), using the definition of error function, and
2
solving for z 0.50 , it is easy to see that
2 er f 1 (0.50)
m = Median(X ) = z 0.50 = +
= + ( 2)(0.476936)
+ 0.6745, > 0,
where er f 1 [0.50] = 0.476936 has been obtained by using Mathematica. Note
that the inverse error function is implemented in Mathematica as a Built-in Symbol,
Inverse Erf[s], which gives the inverse error function obtained as the solution for
z in s = er f (z). Further, for details on Error and Inverse Error Functions, see,
for example, Abramowitz and Stegun (1972, pp. 297309), Gradshteyn and Ryzhik
(1980), Prudnikov et al., Vol. 2 (1986), and Weisstein (2007), among others.
24
2 Normal Distribution
+
|x E (X )| f (x)d x,
+
|x M (X )| f (x)d x.
(ii) 2 =
Then
1
2 (1/2)( x )2
dx
|x|
e
2
2
+e
er f ( ) ,
1
=
+
2
2
2
where er f (z) = 0z
2
2 et dt
25
Proof: We have
1
2 (1/2)( x )2
|x |
e
dx
|x ( | 2 (1/2)( x )2
=
dx
e
2 (1/2)u 2
e
|u |
du,
=
0
x
Substituting
= u, and =
2 (1/2)u 2
2 (1/2)u 2
e
e
( u)
du +
(u )
du
=
2
=
er f ( ) + 2e
2
2
er f ( ) + 2e 2
+
2
2
+ e 2 +
er f ( ) .
1
=
2
2
2
1 = E |X 1 |
= 2
|x 1 | f (x)d x
=
0
2
1
1 + e + er f ( 1/2 )
Proof: We have
|x 1 | f (x)d x
1 =
0
2
.
(2.31)
26
2 Normal Distribution
Taking = 1 , we have
2
=
=
.
2
Thus, taking = 1 and =
in the above Lemma, and simplifying, we have
1 = 2
2
1
1 + e + er f ( 1/2 ) ,
2 = E |X m|
=
|x m| f (x)d x
=
0
2
2
k 1 + 2ek + 2k er f (k) ,
(2.32)
where k = er f 1 (0.50).
Proof: We have
|x m| f (x)d x
2 =
0
m
+ 2k
=
=
= 2k
=
2
2
k 1 + 2ek + 2k er f (k) ,
27
2.2.3.10 Renyi and Shannon Entropies, and Songs Measure of the Shape
of the Half-Normal Distribution
These are derived as given below.
(i) Renyi Entropy: Following Renyi (1961), the entropy of the half- normal random variable X having the pdf (2.13) is given by
R ( ) =
1
ln
1
[ f X (X )] d x,
ln ( )
ln
=
2 ( 1)
2 1
, > 0, > 0,
= 1.
(2.33)
(ii) Shannon Entropy: Following Shannon (1948), the entropy of the half-normal
random variable X having the pdf (2.13) is given by
f X (x) ln [ f X (x)] d x,
H X [ f X (X )] = E[ ln( f X (X )] =
0
which is the particular case of Renyi entropy (2.31) for 1. Thus, in the
limit when 1 and using LHospitals rule, Shannon entropy is easily
obtained from the Eq. (2.33) as follows:
1
ln
H X [ f X (X )] = E[ ln( f X (X )] = =
2
2 1
, > 0.
(iii) Songs Measure of the Shape of a Distribution: Following Song (2001), the
gradient of the Renyi entropy is given by
R
d
1
R ( ) =
( ) =
d
2
1
ln ( )
( 1)
( 1)2
(2.34)
28
2 Normal Distribution
R (1) =
1
(< 0),
8
f X (t)dt = p.
(2.35)
Thus, from the Eq. (2.35), using the Maple program, the percentiles z p of the halfnormal distribution, X |, H N (, ) can easily been obtained.
29
f X (x) =
(x 2)
2
+ e
(x 2)
, x 0
(2.36)
0, x < 0
Note that the and 2 are location and scale parameters for the parent normal distribution. However, they are the shape parameters for the folded normal distribution.
Further, equivalently, if x 0, using a hyperbolic cosine function, the pdf f X (x)
of a folded normal distribution can be expressed as
1
f X (x) =
x (x 2 + 2 )
2
2 2
cosh
e
, x 0.
FX (x) =
2
x (y )2
2
(y 2)
2
2
2
e
dy,
+ e
0
x 0, || < , > 0.
(2.37)
Taking z = y
in (2.37), the cdf FX (x) of a folded normal distribution can also
be expressed as
1
FX (x) =
2
(x
)/
12 z 2
+ e
12 z +
2 2
dz,
z 0, || < , > 0,
(2.38)
where and 2 are the mean and the variance of the parent normal distribution. To
describe the shapes of the folded normal distribution, the plots of the pdf (2.36) for
different values of the parameters and are provided in Fig. 2.6 by using Maple
10. The effects of the parameters can easily be seen from these graphs. Similar plots
can be drawn for others values of the parameters.
30
2 Normal Distribution
31
2.2.4.4 Mode
It is easy to see that the mode or modal value of x for which the folded normal
probability density function f X (x) defined by (2.36) is maximum, is given by x = ,
and the maximum value of the folded normal probability density function (2.35) is
given by
2
1
22
.
(2.39)
1 + e
f X () =
2
Clearly, the folded normal probability density function (2.36) is unimodal.
2.2.4.5 Moments
(i) r th Moment of the Folded Normal Random Variable: For some integer
r > 0, a general formula for the r th moment, f (r ) , of the folded normal
random variable X F N , 2 having the pdf (2.36) has been derived
by Elandt (1961), which is presented here. Let =
. Define Ir (a) =
r 1 y 2
1
moment. In particular,
1
I0 (a) =
2
a
1 2
e 2 y dy = 1 (a),
(2.40)
1 2
1 2
Clearly, for r > 0, Ir (a) = 1 a r 1 e 2 a + (r 1) Ir 2 (a).
where (a) =
1
2
= E
=
x f X (x) d x
0
r
r
r
=
r j I j ( ) + (1)r j I j ( ) .
j
(2.41)
j =0
32
2 Normal Distribution
1 2
2
e 2 [1 2I0 ( )]
f (1) = E [X ] = f =
2
1 2
2
e 2 [1 2 ( )] ,
=
2
f (2) = E X 2 = 2f = 2 + 2 ,
f (3) = E X 3 = 2 + 2 2 f 2 [1 2 ( )] ,
and
f (4) = E
X4
= 4 + 62 2 + 3 4 .
(2.42)
(ii) Central Moments of the Folded Normal Random Variable: For some integer
n > 0, the nth central moment (about the mean f (1) = E (X )) of the folded
normal random variable X having the pdf (2.36) can be easily obtained using
the formula
f (n) = E
f (1)
n
n
n r
n
=
E Xr ,
f (1)
r
r =0
(2.43)
where E (X r ) = f (r ) denotes the r th moment, given by the Eq. (2.41), of the
folded normal random variable X . Thus, from the above Eq. (2.43), the first four
central moments of the folded normal random variable X are easily obtained as
follows.
f (1) = 0,
f (2) = 2 + 2 2f ,
3
3
2
12 2
e
,
f (3) = 3 = 2 f f
2
and
4 + 62 2 + 3 4
1 2
8 3
e 2 f + 2 2 3 2 2f 34f .
+
2
(2.44)
f (4) = 4 =
33
(i) Mean: E (X ) = 1 = f =
2
2
1 2
e 2 [1 2 ( )],
[2 ] 2
f (4)
[ f (2) ]2
f X (t)dt = p.
(2.45)
Thus, from the Eq. (2.45), using the Maple program, the percentiles z p of the folded
normal distribution can be computed for some selected values of the parameters.
Note: For the tables of the folded normal cdf FX (x) = P(X x) for different
f
f
> 3.
34
2 Normal Distribution
bution for the random variable X is said to be truncated. We defined the truncated
distributions as follows.
Definition: Let X be a continuous random variable on a probability space
(, S, P), and let T B such that 0 < P {X T } < 1, where B is a - field
on the set of real numbers . Then the conditional distribution P {X x | X T },
defined for any real x, is called the truncated distribution of X . Let f X (x) and FX (x)
denote the probability density function (pdf) and the cumulative distribution function
(cdf), respectively, of the parent random variable X . If the random variable with the
truncated distribution function P {X x | X T } be denoted by Y, then Y has
support T. Then the cumulative distribution function (cdf), say, G (y), and the probability density function (pdf), say, g (y), for the random variable Y are, respectively,
given by
f X (u)du
P {Y y, Y T }
(, y] T
,
G Y (y) = P{Y y | Y T } =
=
P {Y T }
f X (u)du
T
(2.46)
and
gY (y) =
f X (y) ,
f X (u)du
0,
T
y T
Clearly
gY (y) in (2.47) defines a pdf with support T, since
T
f X (y)dy
f X (u)du
(2.47)
y
/ T.
T
gY (y)dy
bers. In particular, if the values of Y below a specified value a are excluded from the
distribution, then the remaining values of Y in the population have a distribution with
the pdf given by g L (y; a) = 1 f XF(y)
, a y < , and the distribution is
X (a)
said to be left truncated at a. Conversely, if the values of Y above a specified value a
are excluded from the distribution, then the remaining values of Y in the population
have a distribution with the pdf given by g R (y; a) = Ff XX (y)
(a) , 0 y a,
and the distribution is said to be right truncated at a. Further, if Y has a support
T = [a1 , a2 ], where < a1 < a2 < , then the conditional distribution
of Y, given that a1 y a2 , is called a doubly truncated distribution with the
cdf, say, G (y), and the pdf, say, g (y), respectively, given by
G Y (y) =
and
gY (y) =
f X (y)
FX (a2 ) FX (a1 ) ,
0,
y [a1 , a2 ]
y
/ [a1 , a2 ] .
(2.48)
(2.49)
35
The truncated distribution for a continuous random variable is one of the important
research topics both from the theoretical and applications point of view. It arises
in many probabilistic modeling problems of biology, crystallography, economics,
engineering, forecasting, genetics, hydrology, insurance, lifetime data analysis,
management, medicine, order statistics, physics, production research, psychology,
reliability, quality engineering, survival analysis, etc, when sampling is carried out
from an incomplete population data. For details on the properties and estimation of
parameters of truncated distributions, and their applications to the statistical analysis
of truncated data, see, for example, Hald (1952), Chapman (1956), Hausman and
Wise (1977), Thomopoulos (1980), Patel and Read (1982), Levy (1982), Sugiura
and Gomi (1985), Schneider (1986), Kimber and Jeynes (1987), Kececioglu (1991),
Cohen (1991), Andersen et al. (1993), Johnson et al. (1994), Klugman et al. (1998),
Rohatgi and Saleh (2001), Balakrishnan and Nevzorov (2003), David and Nagaraja
(2003), Lawless (2003), Jawitz (2004), Greene (2005), Nadarajah and Kotz (2006a),
Maksay and Stoica (2006) and Nadarajah and Kotz (2007) and references therein.
The truncated distributions of a normally distributed random variable, their properties and applications have been extensively studied by many researchers, among
them Bliss (1935 for the probit model which is used to model the choice probability
of a binary outcome), Hald (1952), Tobin (1958) for the probit model which is used
to model censored data), Shah and Jaiswal (1966), Hausman and Wise (1977), Thomopoulos (1980), Patel and Read (1982), Levy (1982), Sugiura and Gomi (1985),
Schneider (1986), Kimber and Jeynes (1987), Cohen (1959, 1991), Johnson et al.
(1994), Barr and Sherrill (1999), Johnson (2001), David and Nagaraja (2003), Jawitz
(2004), Nadarajah and Kotz (2007), and Olive (2007), are notable. In what follows,
we present the pdf, moment generating function (mgf), mean, variance and other
properties of the truncated normal distribution most of which is discussed in Patel
and Read (1982), Johnson et al. (1994), Rohatgi and Saleh (2001), and Olive (2007).
Definition: Let X N , 2 be a normally distributed random variable with
the mean and the variance 2 . Let us consider a random variable Y which represents
the truncated distribution of X over a support T = [a, b], where < a < b <
. Then the conditional distribution of Y, given that a y b, is called a
doubly truncated normal distribution with the pdf, say, gY (y), given by
gY (y) =
y
1
b
a
0,
y [a, b]
(2.50)
y
/ [a, b]
where (.) and (.) are the pdf and cdf of the standard normal distribution, respectively. If a = , then the we have a (singly) truncated normal distribution from
above, (that is, right truncated). On the other hand, if b = , then the we have
a (singly) truncated normal distribution from below, (that is, left truncated). The
following are some examples of the truncated normal distributions.
36
2 Normal Distribution
2 (y) , < y 0
,
0,
y > 0
(2.51)
where (.) is the pdf of the standard normal distribution. The shape of right truncated
normal pdf gY (y) in (2.51) is illustrated in the following Fig. (2.7).
37
M (t) = E etY |Y [a, b]
a
b
t
2
2
t
= et + 2
b a
(2.52)
(B) Mean, Second Moment and Variance: Using the expression for the mgf (2.52),
these are easily given by
Mean = E (Y |Y [a, b]) = M (t)t = 0
a
b
= +
b a
(i)
(2.53)
Particular Cases:
(I) If b in (2.52), then we have
E (Y |Y > a) = + h,
where h =
a
.
E (Y |Y < b) =
b
V (Y ) = [1 + ]
2
(ii)
Second Moment = E Y 2 |Y [a, b] = M (t)t = 0
= 2 {E (Y |Y [a, b])} 2
a
b
= 2 + 2
b a
(2.54)
38
2 Normal Distribution
a a
b b
+ 2 1 +
b a
(2.55)
and
(iii)
(
)
E Y 2 |Y [a, b]
{E (Y |Y [a, b])}2
a a
b
b
= 2 {1 +
a
b
b
a 2
b a
(2.56)
39
(1982), and Johnson and Kotz (1994), and references therein. In what follows,
we present briefly the pdf, cdf, mean, variance and other properties of the inverse
Gaussian distribution (IGD).
Definition: The pdf of the Inverse Gaussian distribution (IGD) with parameters
and is given by
f (x, , ) =
2 x 3
1/2
where is location parameter and is a shape parameter. The mean and variance of
this distribution are and 3 / respectively. To describe the shapes of the inverse
Gaussian distribution, the plots of the pdf (2.57), for = 1 and = 1, 3, 5 are
provided in Fig. 2.8 by using Maple 10. The effects of the parameters can easily be
seen from these graphs. Similar plots can be drawn for others values of the parameters.
Properties of IGD:
Let x1 , x2 , ..., xn be a random sample of size n from the inverse Gaussian distribution
(1.1). The maximum likelihood estimators (MLEs) for and are respectively given
by
40
2 Normal Distribution
= x =
i=1
xi /n, =
n
1
n
1
, where V =
.
V
xi
x
i=1
41
1
2
1 2
e 2 x and (x) =
x
(2.58)
sity function and cumulative distribution function of the standard normal distribution
respectively.
42
2 Normal Distribution
t 2
2 [1 + i h (t)] , t
(g) The characteristic function
ofx X 2is given by (t) = e
y
2
e 2 dy and h (x) = h (x) for x 0.
, where h (x) =
43
a skew-normal distribution
X = Y
, where 0, > 0, we obtain
with parameters (, , ), denoted by Y S N , 2 , , if its probability
density function is given by
(y )
y
, < y < ,
(2.59)
where (y) and (y) denote the probability density function and cumulative
distribution function of the normal distribution respectively, and 0, >
0 and < < are referred as the location, the scale and the shape
parameters respectively. Some characteristic values of the random variable Y are
as follows:
f Y (y; , , ) = 2
44
2 Normal Distribution
2
I. Mean: E (Y ) = +
2
2 2
II. Variance: V ar (Y ) =
4
[E (X )]3
III. Skewness: 1 =
3
2
[V ar (X )] 2
IV. Kurtosis: 2 = 2 ( 3)
[E (X )]4
[V ar (X )]2
45
a probability value of obtaining data which have a larger value of the test statistic than
the value observed, is determined, which is known as the p-value. Smaller p-values
(for example, less than 0.01) indicate a poor fit of the distribution. Higher values of p
(close to one) correspond to a good fit of the distribution. We consider the following
parametric and non-parametric goodness-of-fit tests
(Oi E i )2
=
Ei
(2.60)
i =1
46
2 Normal Distribution
0, x < y1 ,
i
Fn (x) =
, yi x < yi + 1 , i = 1, 2, . . . , n 1,
n
1, yn x.
(2.61)
Clearly,
Fn (x) =
1
[Number of Observations x] .
n
where
Dn+ =
max
i = 1, 2, ..., n
and
Dn
max
i = 1, 2, ..., n
i
Fn (yi )
n
i 1
Fn (yi )
.
n
+
0.12
. The
critical value of Dn is calculated by the formula d / n + 0.11
n
null and alternative hypotheses being tested are, respectively, given by:
H0 : The data follow the specified continuous distribution;
H1 : The data do not follow the specified continuous distribution.
The null hypothesis (H0 ) is rejected at the chosen significance level, say, , if the
Kolmogorov-Smirnov test statistic, Dn , is greater than the critical value calculated
by the above formula.
A2 = A2n =
47
n
*
+
1
Fractiles of the distribution of A2n for = 0.15, 0.10, 0.05, and 0.01, denoted by
a , are given in Table 11.2 on p. 400 of Blischke and Murthy (2000). The null and
alternative hypotheses being tested are, respectively, given by:
H0 : The data follow the specified continuous distribution
H1 : The data do not follow the specified continuous distribution.
The null hypothesis (H0 ) is rejected if the A-D test statistic, A2n , is greater than the
above tabulated constant a (also known as the critical value for A-D test analysis) at
one of the chosen significance levels, = 0.15, 0.10, 0.05, and 0.01. As pointed
out in Blischke and Murthy (2000), the critical value a does not depend on n, and
has been found to be a very good approximation in samples as small as n = 3.
W =
i =1
ai (yn i + 1 yi )2
n
i =1
,
(xi x)2
where x denotes the sample mean, and, for the observed sample size n 50, the
coefficients ai , i = 1, . . . , k, where k is approximately n2 , are available in Table
A16, pp. 550552, of Conover (1999). For the observed sample size n > 50, the
interested readers are referred to DAgostino (1971) and Shapiro and Francia (1972)
(Fig. 2.12).
For the Shapiro-Wilk test, the null and alternative hypotheses being are, respectively, given by:
H0 : F (x) is a normal distribution with unspecified mean and variance
H1 : F (x) is non-normal.
The null hypothesis (H0 ) is rejected at one of the chosen significance levels if the
Shapiro-Wilk test statistic, W , is less than the quantile as given by Table A 17,
48
2 Normal Distribution
Probability Density Function
0.4
0.36
0.32
0.28
f(x)
0.24
0.2
0.16
0.12
0.08
0.04
0
120
130
140
150
160
170
180
190
200
210
220
230
x
Histogram
Normal
pp. 552553, of Conover (1999). The p- value for the Shapiro-Wilk test may be
calculated by following the procedure on p. 451 of Conover (1999).
Note: The Shapiro-Wilk test statistic may be calculated using the computer softwares
such as R, Maple, Minitab, SAS, and StatXact, among others.
2.3.5 Applications
In order to examine the applications of the above tests of normality, we consider the
following example of weights of a random sample of 40 adult men (Source: Biostatistics for the Biological and Health Sciences, Mario F Triola, Publisher: Pearson,
2005).
Example: We consider the weights of a random sample of 40 adult men as given
below:
{169.1, 144.2, 179.3, 175.8, 152.6, 166.8, 135.0, 201.5, 175.2, 139.0, 156.3, 186.6,
191.1, 151.3, 209.4, 237.1, 176.7, 220.6, 166.1, 137.4, 164.2, 162.4, 151.8, 144.1,
204.6, 193.8, 172.9, 161.9, 174.8, 169.8, 213.3, 198.0, 173.3, 214.5, 137.1, 119.5,
189.1, 164.7, 170.1, 151.0}.
49
Value
Percentile
Value
Sample size
Range
Mean
Variance
Standard deviation
Coefficient of variation
Standard error
Skewness
Excess Kurtosis
40
117.6
172.55
693.12
26.327
0.15258
4.1627
0.37037
0.16642
Min
5%
10 %
25 % (Q1)
50 % (Median)
75 % (Q3)
90 %
95 %
Max
119.5
135.11
137.56
152.0
169.95
190.6
212.91
220.29
237.1
P-value
K-S test
A-D test
Chi-Squared test
Shapiro-Wilk test
0.112
0.306
2.712
0.967
0.652
0.552
0.844
0.379
Do not reject
Do not reject
Do not reject
Do not reject
H0
H0
H0
H0
Using the software EasyFit, the descriptive statistics are computed in the Table 2.2
below. The frequency histogram of the weights of 40 adult men is drawn in Fig. 2.12.
The goodness of fit (or GOF) tests, as discussed above, are applied to test the compatibility of our example of weights of the random sample of 40 adult men with our
hypothesized theoretical probability distribution, that is, normal distribution, using
various software such as EasyFit, Maple, and Minitab. The results are summarized
in the Table 2.2 below. The chosen significance level is = 0.05. The null and
alternative hypotheses being tested are, respectively, given by:
H0 : The data follow the normal distribution;
H1 : The data do not follow the normal distribution.
It is obvious from Table 2.3 is that the normal distribution seems to be an appropriate
model for the weights of 40 adult men considered here. Since the sample size is large
enough, all tests are valid for this example. In this section, we have discussed various
tests of normality to test the suitability of a random sample with a theoretical probability distribution function. In particular, we have applied to test the applicability of
normal distribution to a random sample of the weights of 40 adult men. It is hoped
that this study may be helpful to apply these goodness of fit (or GOF) tests to other
examples also.
50
2 Normal Distribution
2.4 Summary
The different forms of normal distributions and their various properties are discussed
in this chapter. The entropy of a random variable having the normal distribution has
been given. The expressions for the characteristic function of a normal distribution are provided. Some goodness of fit tests for testing the normality along with
applications is given. By using Maple 10, various graphs have been plotted. As a
motivation, different forms of normal distributions (folded and half normal etc.) and
their properties have also been provided.
http://www.springer.com/978-94-6239-060-7