Date: Fri, 18 Nov 2005 15:50:24 +0100
Reply-To: Henrik Lolle <lolle@socsci.aau.dk>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Henrik Lolle <lolle@socsci.aau.dk>
Subject: Re: Multicollinearity and Regression
In-Reply-To: <781313109.20051118094704@terra.es>
Content-Type: text/plain; charset="iso-8859-1"; format=flowed
Hi,
John Fox, in his book 'Regression Diagnostics'
(SAGE University Paper), mention a very
substantial meaning of the VIF, or more precisely
the the square root of this measure. The square
root of the VIF tells which factor the standard
error of the estimate of the regression
coefficient is multiplied with in comparison with
the situation where there is no multicollinearity
at all. A VIF of 4 means, for instance, that the
standard error is doubled in comparison with a
situation with no multicollinearity. Or the precision is halfed.
Best,
Henrik Lolle
At 09:47 18-11-2005, Marta García-Granero wrote:
>Hi Siraj,
>
>SUr> Just want to know the acceptable value of VIF, I am using less than 3?
>
>A VIF less than 3 is too strict.
>
>I'm more used to work with Tolerance (VIF=1/Tol). I read once
>somethink about Tol shouldn't be under 0.01 (this means a VIF of 100).
>Anyway, in my (not very large) experience with multiple regression,
>Tol below 0.05 (VIF above 20) can make a model crash.
>
>When studying collinearity, you can also check:
>
>- Correlation among predictors: As a rule of thumb, correlations above
> 0.9 indicate high collinearity.
>
>- Condition index (table "Collinearity diagnostics"). Quoting from the
> SPSS help: "A condition index greater than 15 indicates a possible
> problem and an index greater than 30 suggests a serious problem with
> collinearity".
>
>The following exampe shows a model with serious collinearity problems
>(caused by the use of a non-centered interaction term):
>
>* Systolic blood pressure in new born children *.
>DATA LIST LIST/age weight sbp (3 F8.0).
>BEGIN DATA
>3 135 89
>4 120 90
>3 100 83
>2 105 77
>4 130 92
>5 125 98
>2 125 82
>3 105 85
>5 120 96
>4 90 95
>2 120 80
>3 95 79
>3 120 86
>4 150 97
>3 160 92
>3 125 88
>END DATA.
>
>VAR LABEL age'Age (days)'
> /weight'Birth weight (pounds)'
> /sbp'Systolic Blood Pressure'.
>
>COMPUTE int=age*weight.
>
>REGRESSION
> /DESCRIPTIVES MEAN STDDEV CORR SIG N
> /STATISTICS COEFF OUTS CI R ANOVA COLLIN TOL
> /DEPENDENT sbp
> /METHOD=ENTER age weight int .
>
>I use this example at class, to show my students the efect of
>centering variables before computing interaction terms.
>
>COMPUTE cint=(age-3.3125)*(weight-120.3125).
>
>REGRESSION
> /DESCRIPTIVES MEAN STDDEV CORR SIG N
> /STATISTICS COEFF OUTS CI R ANOVA COLLIN TOL
> /DEPENDENT sbp
> /METHOD=ENTER age weight cint .
>
>HTH
>
>Marta García-Granero
***********************************************************
Henrik Lolle
Associate Professor
Department of Economics, Politics and Public Administration
Aalborg University
Fibigerstraede 1
DK 9220 Aalborg East
http://www.socsci.auc.dk/institut2/dansk/empl/lolle.htm
***********************************************************
| 
