Sie sind auf Seite 1von 18

CHAPTER 6

STANDARDIZED MULTIPLE REGRESSION ANALYSIS

6.1 Introduction

The general form of a standardize multiple linear regression model is:


• Employed to control roundoff errors in normal equations calculations, and
• To permit comparisons of the estimated regression coefficients in common unit.

The least squares method is very sensitive to the rounding of error in the intermediate stage of
calculation.

Roundoff errors in Normal Equations Calculations:


• Roundoff errors tend to enter normal equations calculations primarily when inverse of
𝑋𝑋′𝑋𝑋 is taken.
• The danger of serious roundoff errors in (𝑋𝑋 ′ 𝑋𝑋)−1 when:
 𝑋𝑋′𝑋𝑋 has a determinant that is close to zero
 The elements of 𝑋𝑋′𝑋𝑋 differ substantially in order of magnitude.

• A solution for this condition is to transform the variables and thereby reparameterize
the regression model into the standardized regression model.
• The transformation to obtain the standardized regression model, called the correlation
transformation.

6.2 Correlation Transformation

Objective: To help with controlling roundoff errors.

The correlation transformation is a simple modification of the usual standardization of a


variable. The usual standardization of the response variable Y and the predictor variables
𝑋𝑋1 , 𝑋𝑋2 , … , 𝑋𝑋𝑝𝑝−1 are as follows:

𝑌𝑌𝑖𝑖 − 𝑌𝑌�
(6.1)
𝑠𝑠𝑌𝑌

𝑋𝑋𝑖𝑖𝑖𝑖 − 𝑋𝑋�𝑘𝑘
(𝑘𝑘 = 1, … , 𝑝𝑝 − 1 (6.2)
𝑠𝑠𝑘𝑘

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 1


where;
2
∑(𝑌𝑌𝑖𝑖 − 𝑌𝑌�)2 �∑(𝑋𝑋𝑖𝑖𝑖𝑖 − 𝑋𝑋�𝑘𝑘 )
𝑠𝑠𝑌𝑌 = � 𝑠𝑠𝑘𝑘 =
𝑛𝑛 − 1 𝑛𝑛 − 1

The correlation transformation is a simple function of the standardized variables in equation


(6.1) & (6.2):

1 𝑌𝑌𝑖𝑖 − 𝑌𝑌�
𝑌𝑌𝑖𝑖∗ = � � (6.3)
√𝑛𝑛 − 1 𝑠𝑠𝑌𝑌


1 𝑋𝑋𝑖𝑖𝑖𝑖 − 𝑋𝑋�𝑘𝑘
𝑋𝑋𝑖𝑖𝑖𝑖 = � � (6.4)
√𝑛𝑛 − 1 𝑠𝑠𝑘𝑘

6.3 Standardized Regression Model

Consider multiple linear regression model

𝑦𝑦𝑖𝑖 = 𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥𝑖𝑖1 + 𝛽𝛽2 𝑥𝑥𝑖𝑖2 + ⋯ + 𝛽𝛽𝑝𝑝−1 𝑥𝑥𝑖𝑖,𝑝𝑝−1 + 𝜀𝜀𝑖𝑖

The regression model with transformed variables 𝑌𝑌 ∗ and 𝑋𝑋𝑘𝑘∗ is called a standardized regression
model as follows

𝑦𝑦𝑖𝑖∗ = 𝛽𝛽1∗ 𝑋𝑋𝑖𝑖1


∗ ∗
+ ⋯ + 𝛽𝛽𝑝𝑝−1 ∗
𝑋𝑋𝑖𝑖,𝑝𝑝−1 + 𝜀𝜀𝑖𝑖∗ (6.5)

Why there is no intercept parameter? The least squares or maximum likelihood calculations
always would lead to an estimated intercept term of zero if an intercept parameter were present
in the model.

The calculation of parameter in the standardized regression model as follows:

𝑠𝑠𝑌𝑌
𝛽𝛽𝑘𝑘 = � � 𝛽𝛽𝑘𝑘∗ (6.6)
𝑠𝑠𝑘𝑘

𝛽𝛽0 = 𝑌𝑌� − 𝛽𝛽1 𝑋𝑋�1 − ⋯ − 𝛽𝛽𝑝𝑝 𝑋𝑋�𝑝𝑝−1 (6.6)

6.4 𝑿𝑿′𝑿𝑿 Matrix for Transformed Variables

When the variables have been transformed by the correlation transformation, to study the
inverse matrix and the least squares normal equations, we need to decompose the correlation

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 2


matrix containing all pairwise coefficients among the response and predictor variables into two
matrices:

1. The correlation matrix of the X variables. This matrix is defined as follows:

1 𝑟𝑟12 ⋯ 𝑟𝑟1,𝑝𝑝−1
𝑟𝑟21 1 ⋯ 𝑟𝑟2,𝑝𝑝−1
𝑟𝑟𝑋𝑋𝑋𝑋 =� ⋮ ⋮ � (6.7)
⋱ ⋮
𝑟𝑟𝑝𝑝−1,1 𝑟𝑟𝑝𝑝−1,2 ⋯ 1

2. The vector containing the coefficients of simple correlation between the response
variables Y and each of the X variables:

𝑟𝑟𝑌𝑌1
𝑟𝑟𝑌𝑌2
𝑟𝑟𝑌𝑌𝑌𝑌 =� ⋮ � (6.8)
𝑟𝑟𝑌𝑌,𝑝𝑝−1

The X matrix is:

𝑋𝑋 ∗ ⋯ ∗
𝑋𝑋1,𝑝𝑝−1
⎡ 11 ⎤
𝑋𝑋 ∗ ⋯ ∗
𝑋𝑋2,𝑝𝑝−1
𝑋𝑋 = ⎢ 21 ⎥ (6.9)
⎢ ⋮ ⋱ ⋮ ⎥
∗ ∗
⎣𝑋𝑋𝑛𝑛1 ⋯ 𝑋𝑋𝑛𝑛,𝑝𝑝−1 ⎦

Simply the correlation matrix of the X variables

𝑋𝑋 ′ 𝑋𝑋 = 𝑟𝑟𝑋𝑋𝑋𝑋 (6.10)

Estimated Standardized Regression Coefficients

The least squares normal equations; 𝑋𝑋 ′ 𝑋𝑋𝑋𝑋 = 𝑋𝑋′𝑌𝑌

The least squares estimators; 𝑏𝑏 = (𝑋𝑋′𝑋𝑋)−1 𝑋𝑋′𝑌𝑌

So, the transformation variables X’Y becomes: 𝑋𝑋 ′ 𝑌𝑌 = 𝑟𝑟𝑌𝑌𝑌𝑌

Thus, the least squares normal equations and estimators of the regression coefficients of the
standardized regression model as follows:

𝑟𝑟𝑋𝑋𝑋𝑋 𝑏𝑏 = 𝑟𝑟𝑌𝑌𝑌𝑌 (6.11)

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 3


−1
𝑏𝑏 = 𝑟𝑟𝑋𝑋𝑋𝑋 𝑟𝑟𝑌𝑌𝑌𝑌 (6.12)

𝑏𝑏1∗
𝑏𝑏 ∗
where; 𝑏𝑏 = � 2 �


𝑏𝑏𝑝𝑝−1

For example: When there are two independent variables


𝑦𝑦1∗
∗ ∗ ∗ 𝑟𝑟𝑦𝑦1
𝑥𝑥 𝑥𝑥21 ⋯ 𝑥𝑥𝑛𝑛1 𝑦𝑦2∗
𝑋𝑋 ′ 𝑌𝑌 = � 11∗ ∗ ∗ � � � = �𝑟𝑟 �
𝑥𝑥12 𝑥𝑥22 ⋯ 𝑥𝑥𝑛𝑛2 ⋮ 𝑦𝑦2

𝑦𝑦𝑛𝑛
1 𝑟𝑟
12
𝑋𝑋 ′ 𝑋𝑋 = � � = 𝑟𝑟𝑋𝑋𝑋𝑋
𝑟𝑟21 1

Hence,
1 1 −𝑟𝑟12 𝑟𝑟𝑦𝑦1
𝑏𝑏 = 2 �−𝑟𝑟 � �𝑟𝑟 �
1 − 𝑟𝑟12 21 1 𝑦𝑦2

Thus,
𝑟𝑟𝑌𝑌1 − 𝑟𝑟12 𝑟𝑟𝑦𝑦2 −𝑟𝑟21 𝑟𝑟𝑦𝑦1 + 𝑟𝑟𝑦𝑦2
𝑏𝑏1∗ = 2 and 𝑏𝑏2∗ = 2
1 − 𝑟𝑟12 1 − 𝑟𝑟12

Example 1: From the previous dataset, Dwaine Studios, Inc., operates portrait studios in 21
cities of medium size. These studios specialize in portraits of children.

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 4


We illustrate the calculation of the transformed data for the first case using the means and
standard deviations.

6.5 Multicollinearity

One basic assumptions of the multiple regression model is that there is no exact linear
relationship between any of the independent variables in the model. If such an exact linear
relationship does exist, we called that the independent variables are perfect collinear exists.
Multicollinearity arises when two or more variables (or combinations of variables) are highly
correlated with each other.

When the predictor variables are correlated among themselves, intercorrelation or


multicollinearity among them is said to exist (when correlation among variables is very high).

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 5


Source of multicollinearity:

1. Method of data collection


It is expected that the data is collected over the whole cross-section of variables. It may
happen that the data is collected over a subspace of the explanatory variables where the
variables are linearly dependent. For example, sampling is done only over a limited range
of explanatory variables in the population.

2. Model and population constrain


There may exists some constraints on the model or on the population from where the
sample is drawn. The sample may be generated from that part of population having linear
combinations.

3. Existence of identities or definitional relationships:


There may exist some relationships among the variables which may be due to the definition
of variables or any identity relation among them. For example, if data is collected on the
variables like income, saving and expenditure, then income = saving + expenditure. Such
relationship will not change even when the sample size increases.

4. Imprecise formulation of model


The formulation of the model may unnecessarily be complicated. For example, the
quadratic (or polynomial) terms or cross product terms may appear as explanatory
variables.

5. An over-determined model
Sometimes, due to over enthusiasm, large number of variables are included in the model
to make it more realistic and consequently the number of observations, n, becomes smaller
than the number of explanatory variables, k.

NOTE
• be careful not to apply t tests mechanically without checking for multicollinearity.
• multicollinearity is a data problem, not a misspecification problem.

Effects of Multicollinearity

• Wrong interpretation of the regression coefficients.


• Large variance and covariance for the OLS estimators of the regression parameters.
• Unduly large (in absolute value) estimates of the regression parameters.

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 6


Example 1: Wrong sign problem

Let consider the following data taken from Montgomery et al. (2006).

Y 𝑿𝑿𝟏𝟏 𝑿𝑿𝟐𝟐
1 2 1
5 4 2
3 5 2
8 6 4
5 8 4
3 10 4
10 11 6

When we fit 𝑌𝑌 on 𝑋𝑋1 only we obtain the least squares line as:

𝑌𝑌� = 1.3901 + 0.5493𝑋𝑋1

But this line becomes


𝑌𝑌� = 1.014 − 1.215𝑋𝑋1 + 3.643𝑋𝑋2

The above two fits clearly show that the sign of the coefficient of 𝑋𝑋1 change which is known
as the wrong sign problem.

For a possible explanation we plot two explanatory variables against one another and it shows
a linear relationship with a correlation coefficient 0.945 with p-value 0.000. This finding
confirms our suspicion that multicollinearity causes wrong sign problem.

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 7


Example 2: Let us consider dataset from Kutner et al. (2008).

Important Conclusion

• When predictor variables are correlated, the regression coefficient of any one variable
depend on which other predictor var. are included and which ones are left out in the
model.
• Regression coefficients doesn't reflect the effect of individual predictor var. on the
response, but only a marginal or partial effect, given whatever other correlated predictor
variables are included in the model.

Another important is related to the error sums of squares (ESS). In general, when two or more
predictor variables are uncorrelated, the marginal contribution of one predictor variable in
reducing the error sum of squares when the other predictor variables are in the model is exactly
same as when this predictor variable is in the model alone.

𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 |𝑋𝑋2 ) = 𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 ) (6.13)

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 8


From table,
𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 |𝑋𝑋2 ) = 𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋2 ) − 𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 , 𝑋𝑋2 )
= 248.750 − 17.625 = 231.15

𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 ) = 231.125

Otherwise, when predictor variables are correlated, the marginal contribution of any one
predictor variable in reducing the error sum of squares varies, depending on which other
variables are already in the regression model, just as for regression coefficients.

𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 |𝑋𝑋2 ) ≠ 𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 ) (6.14)

From previous Body Fat example, the ESS for 𝑋𝑋1

𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 ) = 352.27

𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 |𝑋𝑋2 ) = 3.47

• 𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 |𝑋𝑋2 ) so small compared with 𝑆𝑆𝑆𝑆𝑆𝑆(𝑋𝑋1 ) because 𝑋𝑋1 and 𝑋𝑋2 are highly correlated
with each other and with response variable.

Thus, when 𝑋𝑋2 already in the model, marginal contribution of 𝑋𝑋1 in reducing SSE is
comparatively small because 𝑋𝑋2 contain much of the same information as 𝑋𝑋1.

6.6 Indications of Multicollinearity

High Correlation Values

Relatively high simple correlation between one or more pairs of explanatory variables.
Calculate regression coefficients between all explanatory variables and test the maximum (in
absolute value) correlation coefficient by the statistic

𝑟𝑟𝑖𝑖𝑖𝑖 √𝑛𝑛 − 2
𝑡𝑡 = ~𝑡𝑡𝑛𝑛−2 (6.15)
�1 − 𝑟𝑟𝑖𝑖𝑖𝑖2

There is an evidence of multicollinearity at the 5% level of significance if

|𝑡𝑡| > 𝑡𝑡0.975, 𝑛𝑛−2

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 9


Variance Inflation Factors (VIF)

VIF commonly used diagnostic method to identify multicollinearity. The VIF is the ratio of
variance in a model with multiple terms, where to measures how much the variance of an
estimated regression coefficient increases if your predictors are correlated (multicollinear).

The largest VIF value among all variables an indicator of multicollinearity. The rule of thumb:

𝑉𝑉𝑉𝑉𝑉𝑉 < 5 No Multicollinearity


5 ≤ 𝑉𝑉𝑉𝑉𝑉𝑉 ≤ 10 Moderate Multicollinearity
𝑉𝑉𝑉𝑉𝑉𝑉 > 10 Severe Multicollinearity

The VIF is the most popular method to identify multicollinearity and it is given by:

1
𝑉𝑉𝑉𝑉𝑉𝑉𝑖𝑖 = (6.16)
1 − 𝑅𝑅𝑗𝑗2

where; 𝑅𝑅𝑗𝑗2 denotes the coefficient of determination obtained when 𝑋𝑋𝑗𝑗 is regressed on other 𝑋𝑋𝑝𝑝−1
variables in the model.

One or more large VIFs indicate the presence of multicollinearity in the data.

6.7 Corrections for Multicollinearity

As remedial measures of multicollinearity problem we can take the following steps:

• Collection of more data


Additional data may help in reducing the sampling variance of the estimates. The data need
to be collected such that it helps in breaking up the multicollinearity in the data.

• Dropping off unimportant variables


If possible, identify the variables which seems to causing multicollinearity. The process of
omitting the variables way be carried out on the basis of some kind of ordering of
explanatory variables, e.g., those variables can be deleted first which have smaller value
of t-ratio. In another example, suppose the experimenter is not interested in all the
parameters.

• Use Ridge Regression


• Use Principal Component Regression

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 10


Example 3: The blood pressure data in which researchers observed the following data on 20
individuals with high blood pressure:

The Blood Pressure Dataset

Result:

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 11


6.8 Autocorrelated Errors

Recall that one of the assumptions when building a linear regression model is that the errors
are independent. One of the basic assumptions in linear regression model is that the random
error components or disturbances are identically and independently distributed. So, in the
model 𝑦𝑦 = 𝑋𝑋𝑋𝑋 + 𝜀𝜀, it is assumed that

𝜀𝜀𝑖𝑖𝑖𝑖 ~𝑁𝑁(0, 𝜎𝜎 2 ), and 𝜎𝜎�𝜀𝜀𝑖𝑖 , 𝜀𝜀𝑗𝑗 � = 0, 𝑖𝑖 ≠ 𝑗𝑗, 𝑖𝑖, 𝑗𝑗 = 1,2, … , 𝑛𝑛

The assumption is that errors are uncorrelated. In real practice, independent error is violated,
where errors are frequently correlated positively over time. When autocorrelation is present,
some or all off diagonal elements in 𝐸𝐸(𝜀𝜀𝜀𝜀′) are nonzero.

Error terms corrected over time are said to be autocorrelated or serially correlated. The
assumption of independent not always true when data are obtained in a time sequence, where
error terms are correlated with the previous error.

Sometimes the study and explanatory variables have a natural sequence order over time, i.e.,
the data is collected with respect to time. Such data is termed as time series data. The
disturbance terms in time series data are serially correlated.

Major cause of positively autocorrelated error terms in business and economic involving time
series data, is the omission of one or more variable which has an effect on y but were not
included in model.

Example 4:
Suppose that every morning you watch a weather report featuring your favourite meteorologist
on the local morning news. You use his forecasts to decide what clothes to wear.
− You notice that in the summer months, once you get outside, you are always hot. After
a while, you realize that the temperature predicted by the meteorologist is always higher
than the actual temperature.
− Then, during the rainy, you realize the temperature he predicts is always lower than the
actual temperature. Something is wrong with the meteorologist’s method.
− Since his winter forecasts are always too high, why doesn’t he start lowering his
predictions? When rainy rolls around, then his predictions are always too low. Why
doesn’t he realize this and raise his rainy forecasts?

This is the essence of autocorrelation: The errors follow a pattern, showing that something is
wrong with the regression model.

Example 5:
Suppose that we wish to regress annual sales of a soft drink concentrate against the annual
advertising expenditures for that product. Now, the growth in population over the period of

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 12


time used in the study will also influence the product sales. If population size is not included
in the model, this may cause the errors in the model to be autocorrelated because population
size is correlated with product sales.

When error terms are autocorrelated, some issues arise when using ordinary least squares.
These problems are:
1. The OLS estimated regression coefficients are still unbiased but they are longer having
minimum variance property and may be quite inefficient.
2. The MSE may seriously under estimate the variance of the error terms.
3. The standard error of the regression coefficients may seriously underestimate the true
standard deviation of the estimated regression coefficients. This will result in a
conclusion that the estimate is good but actually it is not.
4. Statistical intervals and inference procedures are no longer strictly applicable.

Detecting the presence of autocorrelation


Since problem of autocorrelation is very serious – it’s important to detect them. There are two
general approaches for dealing with the detecting of autocorrelation.
1. Plot residual against time

2. Formal statistical test- Durbin Watson

This test is based on the assumption that the errors in the regression model are generated
by a first-order autoregressive error model.

The generalized simple linear regression model for one predictor variables when the random
error terms follow a first-order autoregressive, process is:

𝑌𝑌𝑡𝑡 = 𝛽𝛽0 + 𝛽𝛽1 𝑋𝑋𝑡𝑡 + 𝜀𝜀𝑡𝑡 (6.17)


𝜀𝜀𝑡𝑡 = 𝜌𝜌𝜌𝜌𝑡𝑡−1 + 𝑢𝑢𝑡𝑡
where:
𝜌𝜌 is a parameter ∋ |𝜌𝜌| < 1 𝑢𝑢𝑡𝑡 are independent 𝑁𝑁(0, 𝜎𝜎 2 )

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 13


Each error term in above model (6.17) consists of a fraction of the previous error term (when
𝜌𝜌 > 0) plus a new disturbance term 𝑢𝑢𝑡𝑡 . The parameter ρ is called the autocorrelation parameter.

Suppose we transform the response variable so that 𝑦𝑦𝑡𝑡′ = 𝑦𝑦𝑡𝑡 − 𝜌𝜌𝜌𝜌𝑡𝑡−1. Substitute in this
expression for 𝑦𝑦𝑡𝑡 and 𝑦𝑦𝑡𝑡−1 according to the regression becomes:

𝑦𝑦𝑡𝑡′ = (𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥𝑡𝑡 + 𝜀𝜀𝑡𝑡 ) − 𝜌𝜌(𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥𝑡𝑡−1 + 𝜀𝜀𝑡𝑡−1 )

= 𝛽𝛽0 (1 − 𝜌𝜌) + 𝛽𝛽1 (𝑥𝑥𝑡𝑡 − 𝜌𝜌𝜌𝜌𝑡𝑡−1 ) + 𝜀𝜀𝑡𝑡 − 𝜌𝜌𝜌𝜌𝑡𝑡−1

𝑦𝑦𝑡𝑡′ = 𝛽𝛽0 (1 − 𝜌𝜌) + 𝛽𝛽1 (𝑥𝑥𝑡𝑡 − 𝜌𝜌𝜌𝜌𝑡𝑡−1 ) + 𝑢𝑢𝑡𝑡 (6.18)

where 𝑢𝑢𝑡𝑡 independence disturbance term.

Now, we write the model (6.18) becomes:

𝑦𝑦𝑡𝑡′ = 𝛽𝛽0′ + 𝛽𝛽1′ 𝑥𝑥𝑡𝑡′ + 𝑢𝑢𝑡𝑡 (6.19)

The reparameterized model (6.19) cannot be used directly because the new regressor and
response variables 𝑥𝑥𝑡𝑡′ and 𝑦𝑦𝑡𝑡′ are functions of the unknown parameter ρ.

The autoregressive error process can be viewed as a regression through the origin.

𝜀𝜀𝑡𝑡 = 𝜌𝜌𝜌𝜌𝑡𝑡−1 + 𝑢𝑢𝑡𝑡 (6.20)

Properties of Error Terms

The error terms 𝜀𝜀𝑡𝑡 still have mean 0 and constant variance:

𝐸𝐸(𝜀𝜀𝑡𝑡 ) = 0 (6.21)

𝜎𝜎 2
𝜎𝜎 2 (𝜀𝜀𝑡𝑡 ) = (6.21)
1 − 𝜌𝜌2

The covariance (a measure of the relationship between two variables) between adjacent error
terms is:
𝜎𝜎 2
𝜎𝜎(𝜀𝜀𝑡𝑡 , 𝜀𝜀𝑡𝑡−1 ) = 𝜌𝜌 � � (6.22)
1 − 𝜌𝜌2

which implies the coefficient of correlation between adjacent error terms is:

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 14


𝜎𝜎 2
𝜌𝜌 � �
1 − 𝜌𝜌2
𝜌𝜌(𝜀𝜀𝑡𝑡 , 𝜀𝜀𝑡𝑡−1 ) = = 𝜌𝜌 (6.23)
𝜎𝜎 2 𝜎𝜎 2
� �
1 − 𝜌𝜌2 1 − 𝜌𝜌2

which is the autocorrelation parameter we introduced above. Thus, the autocorrelation


parameter ρ is the coefficient of correlation between adjacent error terms.

6.9 Durbin-Watson Test for Autocorrelation

We shall now consider a test of the null hypothesis of that no serial correlation is present
(𝜌𝜌 = 0). By far the most popular test autocorrelation (serial correlation) is the Durbin-Watson
test.

This test involves the calculation of a test statistics based on the OLS residuals. The statistic
is defined as
∑𝑛𝑛𝑡𝑡=2(𝑒𝑒𝑡𝑡 − 𝑒𝑒𝑡𝑡−1 )2
𝐷𝐷 = (6.24)
∑𝑛𝑛𝑡𝑡=1 𝑒𝑒𝑡𝑡2

When successive values of 𝑒𝑒̂𝑡𝑡 are close to each other, the DW statistic will be low, indicating
the presence of positive serial correlation. The summary of decision rule of the Durbin-Watson
test is:

Value of DW Result
0 < 𝐷𝐷𝐷𝐷 < 𝑑𝑑𝐿𝐿 Reject null hypothesis; positive serial correlation present
𝑑𝑑𝐿𝐿 < 𝐷𝐷𝐷𝐷 < 𝑑𝑑𝑈𝑈 Indeterminate
𝑑𝑑𝑈𝑈 < 𝐷𝐷𝐷𝐷 < 2 Accept null hypothesis
2 < 𝐷𝐷𝐷𝐷 < 4 − 𝑑𝑑𝑈𝑈 Accept null hypothesis
4 − 𝑑𝑑𝑈𝑈 < 𝐷𝐷𝐷𝐷 < 4 − 𝑑𝑑𝐿𝐿 Indeterminate
4 − 𝑑𝑑𝐿𝐿 < 𝐷𝐷𝐷𝐷 < 4 Reject null hypothesis; negative serial correlation present

Values of 𝑑𝑑𝐿𝐿 and 𝑑𝑑𝑈𝑈 are available from the Durbin-Watson table.

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 15


Example 6: Consumer Expenditure data given by Chartterjee and Hadi, 2006.

Consumer Money Stock Consumer Money Stock


Expenditure Expenditure
214.6 159.3 238.7 173.9
217.7 161.2 243.2 176.1
219.6 162.8 249.4 178.0
227.2 164.6 254.3 179.1
230.9 165.9 260.9 180.2
233.3 167.9 263.3 181.2
234.1 168.3 265.6 181.6
232.3 169.7 268.2 182.5
233.7 170.5 270.4 183.3
236.5 171.6 275.6 184.3

Time series plot of residuals indicates that positive autocorrelation is present in the data. for
this data, we obtain:

8195.21
𝐷𝐷𝐷𝐷 = = 0.328
7587.92

At the 5% level of significance, the critical values corresponding to 𝑛𝑛 = 20 are 𝑑𝑑𝐿𝐿 = 1.20 and
𝑑𝑑𝑈𝑈 = 1.41. Since the observed value of d is less than 𝑑𝑑𝐿𝐿 , we reject the null hypothesis and
conclude that positive autocorrelation is present in our data.

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 16


6.10 Corrections for Autocorrelation

When autocorrelated error terms are found to be present, then one of the first remedial measures
should be to investigate the omission of a key predictor variable. If such a predictor does not
aid in reducing/eliminating autocorrelation of the error terms, then certain transformations on
the variables can be performed.

Cochrane-Orcutt Procedure (Cochrane and Orcutt, 1949)

This procedure involves a series of iterations, each of which produces a better estimate of than
does the previous one.

Step 1: The OLS method is used to estimate regression parameters and the correlation
coefficient is estimated by
∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡 𝜀𝜀̂𝑡𝑡−1
𝜌𝜌� = (6.25)
�∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡2 �∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡−1
2

Step 2: The estimated value of ρ is used to perform the generalized differences

∗ ∗
𝑌𝑌𝑡𝑡∗ = 𝑌𝑌𝑡𝑡 − 𝜌𝜌𝑌𝑌𝑡𝑡−1 , 𝑋𝑋1𝑡𝑡 = 𝑋𝑋1𝑡𝑡 − 𝜌𝜌𝑋𝑋1𝑡𝑡−1 , . . . 𝑋𝑋𝑘𝑘𝑘𝑘 = 𝑋𝑋𝑘𝑘𝑘𝑘 − 𝜌𝜌𝑋𝑋𝑘𝑘𝑘𝑘−1

We estimate regression parameters from the transformed model:

∗ ∗ ∗
𝑌𝑌𝑡𝑡∗ = 𝛽𝛽0 (1 − 𝜌𝜌) + 𝛽𝛽1 𝑋𝑋1𝑡𝑡 + 𝛽𝛽2 𝑋𝑋2𝑡𝑡 + ⋯ + 𝛽𝛽𝑘𝑘 𝑋𝑋𝑘𝑘𝑘𝑘 + 𝑢𝑢𝑡𝑡

calculate a new set of residuals as

𝜀𝜀̂𝑡𝑡 = 𝑌𝑌𝑡𝑡 − 𝛽𝛽̂1 − 𝛽𝛽̂2 𝑋𝑋2𝑡𝑡 − 𝛽𝛽̂3 𝑋𝑋3𝑡𝑡 − ⋯ − 𝛽𝛽̂𝑘𝑘 𝑋𝑋𝑘𝑘𝑘𝑘

Step 3: Recalculate the correlation coefficient

∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡 𝜀𝜀̂𝑡𝑡−1


𝜌𝜌� = (6.26)
�∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡2 �∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡−1
2

and so on …

The iterative procedure can be carried on for as many steps as desired. Standard procedure is
to stop the iterations when the new estimate of ρ differs from the previous one by less than 0.01
or 0.005.

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 17


Example 7: For the Consumer Expenditure Data, we have;

∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡 𝜀𝜀̂𝑡𝑡−1


𝜌𝜌� = = 0.751
�∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡2 �∑𝑇𝑇𝑡𝑡=2 𝜀𝜀̂𝑡𝑡−1
2

Let us define 𝑌𝑌𝑡𝑡∗ = 𝑌𝑌𝑡𝑡 − 𝜌𝜌�𝑌𝑌𝑡𝑡−1 , 𝑋𝑋𝑡𝑡∗ = 𝑋𝑋1𝑡𝑡 − 𝜌𝜌�𝑋𝑋𝑡𝑡−1

Now a fitting 𝑌𝑌𝑡𝑡∗ on 𝑋𝑋𝑡𝑡∗ gives 𝜀𝜀̂𝑡𝑡 = 𝑌𝑌𝑡𝑡∗ − 𝑌𝑌�𝑡𝑡∗

For this revised residuals, Durbin-Watson statistic is 1.43.

From the table given by Durbin-Watson (𝑛𝑛 = 19, 𝑝𝑝 = 1), we obtain 𝑑𝑑𝐿𝐿 = 1.18 and 𝑑𝑑𝑈𝑈 = 1.40.

Since the D-W statistic lies between 𝑑𝑑𝑈𝑈 and 2, we may accept the null hypothesis and conclude
that of there is no evidence of autocorrelation in this data.

Regression Analysis | Chapter 6 | MTH5404| Mustafa, M. S (UPM) 18

Das könnte Ihnen auch gefallen