Professional Documents
Culture Documents
1
=
T
t=1
(x
t
x
t
) (y
t
y
t
)
T
t=1
(x
t
x
t
)
2
,
0
= y
t
1
x
t
2
=
1
T 2
T
t=1
_
y
t
1
x
t
_
2
, Note dof adjustment
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 27 / 51
Example (OLS Using EViews)
Estimate the linear regression equation
y
t
=
0
+
1
x
t
+u
t
u
t
N
_
0,
2
_
using the following annual data from 2003 to 2007 (T = 5)
t : 2003 2004 2005 2006 2007
y
t
: 3 5 10 15 20
x
t
: 2 4 7 6 10
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 28 / 51
Example (OLS Using EViews continued)
The EViews commands are: highlight y and x, double click the shaded
region
Open Equation.../OK
The output is below
Source: lecture1_example1.wf1
The estimates are:
0
= 1.788,
1
= 2.136,
2
= 3.126
2
= 9.772
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 29 / 51
Some extensions of regression models are
1. Dummy variables and seasonality: x
t
=
_
1 : True
0 : False
2. Dynamics: x
t
= fx
t1
, x
t2
, y
t1
, y
t2
, g
3. Nonlinearities: x
t
= flog (w)g , x
t
=
_
w
2
t
_
4. Heteroskedastcity: y
t
=
0
+
1
x
t
+u
t
u
t
N
_
0,
2
t
_
2
t
=
0
+
1
w
t
5. Autocorrelation: y
t
=
0
+
1
x
t
+u
t
u
t
= u
t1
+v
t
v
t
N
_
0,
2
t
_
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 30 / 51
An important assumption of this model is that u
t
is normally
distributed
f (u
t
) =
1
p
2
2
exp
_
u
2
t
2
2
_
, < u
t
<
This implies that y
t
is normal with conditional mean
0
+
1
x
t
and
variance
2
f (y
t
) =
1
p
2
2
exp
_
(y
t
0
1
x
t
)
2
2
2
_
, < y
t
<
or more compactly
y
t
N
_
0
+
1
x
t
,
2
_
with parameters
=
_
0
,
1
,
2
_
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 31 / 51
The linear regression model is just one class of models that
is a subset of the models discussed in this course (say 5%).
The least squares estimator is generally not appropriate to
estimate the parameters of these other models.
An important component of the course is to specify models
where the data are nonnormal.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 32 / 51
Lecture 1 Maximum Likelihood: Basics
Data Characteristics
The assumption that y
t
is normal need not apply to all data sets.
Types of data to consider are:
1. Binary data
2. Skewed data
3. Fat-tailed data
4. Ordered data
5. Count data
6. Duration data
The following examples show that the assumption of normality is just
a very special case.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 33 / 51
Example (Binary Data)
Consider survey data based on a sample of T = 20 people where the
responses are in terms of yes/no answers (ie binary)
y
t
=
_
1 : Yes
0 : No
Here y
t
just takes on one of two values as highlighted by the histogram
0
2
4
6
8
10
12
0.0 0.2 0.4 0.6 0.8 1.0
Series: Y
Sample 1 20
Observations 20
Mean 0.550000
Median 1.000000
Maximum 1.000000
Minimum 0.000000
Std. Dev. 0.510418
Skewness -0.201008
Kurtosis 1.040404
Jarque-Bera 3.334694
Probability 0.188747
Source: lecture1_example2.wf1
Clearly nonnormal.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 34 / 51
Example (Binary Data continued)
One solution for modelling y
t
is to use a Bernoulli distribution
f (y
t
) =
y
t
(1 )
1y
t
, y
t
= 0, 1,
where = fg , is an unknown parameter with the property 0 < < 1.
This parameter represents a probability as
= P (y
t
= 1)
and hence
1 = P (y
t
= 0)
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 35 / 51
Example (Melbourne Age Survey)
Should Connex be allowed to ne train passengers for putting their feet on
seats? [The Age Readerss Poll, 4/2/2009]
Source: Melbourne Age, NEWS, 3/2/2009, p.3
The results of a survey of 1094 people found that
Yes : 55% (1094 0.55 ' 602 people)
No : 45% (1094 0.45 ' 492 people)
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 36 / 51
Example (Skewed Data)
The following gure gives a histogram of annual incomes ($) of 1,526
residents of a country in 2007.
0
40
80
120
160
200
240
280
320
0 20000 40000 60000 80000 100000
Series: Y
Sample 1 1526
Observations 1526
Mean 24584.16
Median 21854.38
Maximum 109712.4
Minimum 4153.896
Std. Dev. 12547.57
Skewness 1.547919
Kurtosis 7.273659
Jarque-Bera 1770.692
Probability 0.000000
Source: lecture1_example3.wf1
The data are all positive and the distribution is skewed to the right.
Clearly nonnormal.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 37 / 51
Example (Skewed Data continued)
Alternatively, by transforming income into natural logarithms
LY = LOG (Y)
the histogram now appears normal
0
20
40
60
80
100
120
140
8.5 9.0 9.5 10.0 10.5 11.0 11.5
Series: LY
Sample 1 1526
Observations 1526
Mean 9.991775
Median 9.992157
Maximum 11.60562
Minimum 8.331802
Std. Dev. 0.488632
Skewness -0.066485
Kurtosis 3.049711
Jarque-Bera 1.281336
Probability 0.526940
Source: lecture1_example3.wf1
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 38 / 51
Example (Skewed Data continued)
This suggests using the lognormal distribution
f (y
t
) =
1
p
2
2
y
t
exp
_
(log y
t
)
2
2
2
_
, y
t
> 0
where =
_
,
2
_
are unknown parameters. An example is given below
with = 10, = 0.5
.00000
.00001
.00002
.00003
.00004
.00005
0 20000 40000 60000 80000 100000 120000
Y
F
Source: lecture1_lognormal.wf1
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 39 / 51
Example (Fat-tailed Data)
The following histogram in (a) gives daily asset returns (%) on the
S&P500 from the 10th of January 1989 to the 31st of July 2007.
(a) Empirical (S&P500) (b) Theoretical (Normal)
Source: lecture1_example4.wf1
The data exhibit fat-tails and a sharp peak relative to the normal
distribution given in (b) (which is simulated returns using the mean and
the standard deviation of Y). Clearly the series is nonnormal.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 40 / 51
Example (Fat-tailed Data continued)
One solution is to use the standardized Student t distribution
f (y
t
) =
_
+ 1
2
_
_
2
( 2)
_
2
_
_
1 +
_
y
t
( 2)
_
2
_
_
+1
2
_
, < y
t
<
where =
_
,
2
,
_
are unknown parameters and
(x) =
_
0
e
s
s
x1
ds
is the gamma function.
This distribution has fat-tails and a sharp peak at . The parameter is
known as the degrees of freedom parameter which controls the fatness
in the tails of the distribution.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 41 / 51
Example (Ordered Data)
The key mechanism of monetary policy in Australia is the target rate.
When the Reserve Bank of Australia conducts monetary policy at the start
of each month, beginning February until December, it changes the target
rate by either a discrete amount (a multiple of 0.25% = 25 basis points),
or not at all
y
t
=
_
_
1.00% : Extreme Easing
0.75% : Strong Easing
0.50% : Moderate Easing
0.25% : Gentle Easing
0.00% : No change
0.25% : Gentle Tightening
0.50% : Moderate Tightening
0.75% : Strong Tightening
1.00% : Extreme Tightening
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 42 / 51
Example (Ordered Data continued)
Here is an example of the RBA cutting its interest rate by 75 basis points.
Source: http://business.theage.com.au/business/
rates-cut-to-525-20081104-5hgv.html
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 43 / 51
Example (Ordered Data continued)
Here is an example of the RBA cutting its interest rate by 100 basis points.
Source: Melbourne Age, 4/2/2009, p.2
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 44 / 51
Example (Ordered Data continued)
A time series plot of the target rate, beginning 1st of October 1993 and
ending 30th of June 2004, is given below in (a). The series follows a step
function. The change in the target policy is given in (b)
Y = D (TARGET)
shows that the change in the target rate is nonnormal.
(a) Target (b) Change in Target
4.0
4.4
4.8
5.2
5.6
6.0
6.4
6.8
7.2
7.6
500 1000 1500 2000 2500
-0.8
-0.4
0.0
0.4
0.8
1.2
500 1000 1500 2000 2500
Source: lecture1_example5.wf1
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 45 / 51
Example (Count Data)
The following table gives data on the number of strikes in the US for the
years 1968 to 1976, T = 9.
t : 1968 1969 1970 1971 1972 1973 1974 1975 1976
y
t
: 8 6 11 3 3 2 18 2 9
Source: Kennan (1985, 5-28), Journal of Econometrics
Clearly assuming that y
t
is normally distributed is inappropriate as the
data are integers and strictly positive. A natural way to model a
dependent variable that measures counts is to assume that y
t
has a
Poisson distribution
f (y
t
) =
y
t
exp ()
y
t
!
, y
t
= 0, 1, 2, , > 0
where = fg is an unknown parameter with the property > 0. This
parameter represents the mean of the distribution.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 46 / 51
Example (Duration Data)
The following histogram gives data on the duration of strikes (in days) in
the US for the years 1968 to 1976, T = 62.
0
4
8
12
16
0 40 80 120 160 200
Series: Y
Sample 1 62
Observations 62
Mean 42.67742
Median 27.00000
Maximum 216.0000
Minimum 1.000000
Std. Dev. 45.84070
Skewness 1.624063
Kurtosis 5.402414
Jarque-Bera 42.16496
Probability 0.000000
Source: lecture1_strikes.wf1
Source: Kennan (1985, 5-28), Journal of Econometrics
The data are nonnormal as (i) the distribution is skewed to the right, (ii)
the data are integers and (iii) the data are strictly positive.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 47 / 51
Example (Duration Data continued)
A natural way to model a dependent variable that measures duration is to
assume that y
t
has an exponential distribution
f (y
t
) = exp [y
t
] , y
t
> 0
with unknown parameter = fg with the property > 0.
The parameter represents the inverse of the mean of the exponential
distribution. This suggests that an estimate of is given by the inverse of
the sample mean (y)
=
1
y
=
T
T
t=1
y
t
We want to show this more formally in this course.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 48 / 51
The aim of maximum likelihood estimation is to use all of the
information in the data to estimate the unknown parameters in .
This requires using the data characteristics and hence knowledge of
the probability density function of y
t
.
Thus the aim is to estimate if the distribution is
1. Bernoulli
2. Poisson
3. Exponential
4. Student t
5. Lognormal
and even
6. Normal
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 49 / 51
Lecture Summary: Key things to know
1. The assumption of normality is used in the regression model, but it
does not apply in all situations.
2. There are other types of models that use alternative types of
distributions. Some examples of other distributions considered in this
course are:
a. Bernoulli
b. Poisson
c. Exponential
d. Student t
e. Lognormal
3. Maximum likelihood estimation provides a general framework to
estimate the unknown parameters for models based on alternative
types of distributions.
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 50 / 51
End of Lecture
Professor Vance L. Martin () Econometrics: ECOM30002 and ECOM90002 Lecture 1 Maximum Likelihood: Basics Semester 1 51 / 51