2 Multicollinearity Presented by: Shahram Arsang Isfahan University of Medical Sciences Email: [email protected] April 2014.

2

Multicollinearity

Presented by: Shahram Arsang

Isfahan University of Medical Sciences

Email: [email protected]

April 2014

FOCUS

• Definition of Multicollinearity• Distinguish of Multicollinearity• Remedial measures of Multicollinearity• Example

6 April 2014 3

Multicollinearity

• Definition: predictor variable are highly correlated among themselves

• Example: body fat

• potential harm of collinearity:difficult to infer the separate influence of such explanatory

variables on the response variable.

6 April 2014 4

Problems with Multicollinearity

1. adding or deleting a predictor variable change the regression coefficient.

2. the extra sum of square associated with a predictor varies , depending upon which other predictor variables are already included in the model.

3. the estimated SD of the regression coefficients become large

4. the estimated regression coefficients individually may not be statistically significant even though a definite statistical relations exists between the response variable and the set of predictor variables.

6 April 2014 5

6 April 2014 6

• diagnosis consists of two related but separate elements:

1- detecting the presence of collinear

2-assessing the extent to which these relationships have degraded estimated parameters.

6 April 2014 7

diagnostic Informal diagnostics for multicollinearity

6 April 2014 8

1-large changes in the estimated regression coefficient when a predictor variable is added or deleted, or when an observation is altered or deleted.

2- nonsignificant result in individual tests on the regression coefficient for important predictor variables.

3- estimated regression coefficient with an algebraic sign that is the opposite of that expected from theoretical considerations or prior experience.

4- large coefficient of simple correlation between pairs of predictor variable in the correlation matrix rxx.

5- wide confidence intervals for the regression coefficients representing

important predictor variables .

limitation of informal diagnostics

1. they provide qualitative measurements

2. sometimes the observed behavior may occur without Multicollinearity being present.

Multicollinearity diagnostic methods

• Correlation matrix R (or ) of x`s (absence of high correlations cannot be viewed as evidence of no problem)

• • Variance Inflation Factor (VIF)Weakness:1. unable to reveal the presence of several coexisting near dependencies among

the explanatory variates.

2. meaningful boundary to distinguish between values of VIF

• The technique of Farrar and Glauber (partial correlation)

6 April 2014 10

The technique of Farrar and Glauber• the n*p data matrix X is a sample of size n from

a p-variate Gaussian (normal) distribution

• that is, the partial correlation between Xi and Xj, adjusted for all other X-variates, to investigate the patterns of interdependence in greater detail

6 April 2014 11

Variance inflation factor (VIF)

VIF: how much the variances of the ß are inflated as compared to when the xi`s are not linearly related.

Variance-covariance matrix of the ß and ß*:

Diagnostic uses

severity of multicollinearity:

1. Large value of VIF VIF > 10

2. means of the VIF

: how far the estimated standardized regression coefficient bk* are from the true values βk*.

It can be shown that the expected value of the sum of these squared errors (bk*-βk*)2 is given by :

When no X variable is linearly related to the others in the regression model ;

Sum of (VIF)k ≡ p-1

Provide useful information about the effect of multicollinearity on the sum of the squared errors :

Mean of the VIF values , to be denote by (VIF) :

VIF > 1 indicate of serious Multicollinearity problems.

Body fat example ;

The expected sum of the squared errors in the least squares standardized regression coefficient is nearly 460 times as large as it would be if the x variables were uncorrelated .

Multicollinearity problem ?

Comments

1. reciprocal of the VIF for exclusion x variables:

2. Limitation of VIF: distinguish between several simultaneous multicollinearity

Remedial measures1. Making predictions is not a problem

2. Centered data for x`s

3. Dropping one or more predictors

4. Add some cases that may break the pattern of multicollinearity

5. Use different data sets to estimate different coefficients

6. Principal component analysis

6 April 2014 18

Ridge regression

• By modifying the method of least square to allow biased estimator of the regression coefficients

• Ridge estimators:

by the correlation transform

• Idea is to use a small biasing constant c and find

6 April 2014 19

Ridge regression

• standardized ridge coefficients• C amount of bias in estimator

• C=0 =OLS in standardized form• c>0 `s are biased but more stable than OLS

6 April 2014 20

1( )Rxx yxb r cI r

Ridge regression

• Results:

1. As c increases, bias increases, variance of the betas decreases

2. There always exists a c for which the total MSE for ridge regression is SMALLER than that for OLS.

3. There are no hard and fast ways of finding c.

6 April 2014 21

• Choice of biasing constant c:

1. Ridge trace: Simultaneous plot of the values of the (p-1) estimated ridge standardized regression coefficients for different values of c between 0 and 1.

2. VIF

6 April 2014 22

Choice of biasing constant c

• Smallest value of c where it is deemed that:

1- regression coefficients have steadied itself and

2- VIF is small

6 April 2014 23

Comments:

•

• Limitation of ridge regression:

1-Precision of ridge regression coefficient: Bootstrap

2- choice of c

6 April 2014 24

Use ridge regression to reducing predictor variables:

• Unstable ridge trace with coefficient tending toward zero • Ridge trace is stable but at a very small value• Unstable ridge trace that do not tend toward zero: candidate

6 April 2014 25

VIF - SPSS

6 April 2014 26

Output- Spss

6 April 2014 27

Example 2 – VIF value and remedial measure

6 April 2014 28