Multicollinearity problem
in [SAS]
|
From: Hayden Salter on 23 Feb 2010 03:39 Hi, I have 2 questions that are related to each other: 1. I have a set of measurement scales that are highly correlated with each other. I have tried to test the goodness-of-fit of my model using SEM and the Maximum Likelihood method. SEM analyses can be 'adversely' affected by multicollinearity. When I used AMOS to perform the goodness-of-fit test, AMOS gave me an error that my matrix is not definite positive. Is there a way to overcome this problem (i.e. eliminate multicollinearity)? 2. When I used the Unweighted Least Square method instead of the Maximum Likelihood Method, it performed the SEM analysis without an error but the compromise is that you lose the important goodness-of- fit test indices (e.g. RMSEA, CFI, etc.) from being displayed in the results. Is ULS a reliable and renowned method for testing the goodness-of-fit?
From: Dale McLerran on 23 Feb 2010 16:28 --- On Tue, 2/23/10, Hayden Salter <sageman70(a)GOOGLEMAIL.COM> wrote: > From: Hayden Salter <sageman70(a)GOOGLEMAIL.COM> > Subject: Re: Multicollinearity problem > To: SAS-L(a)LISTSERV.UGA.EDU > Date: Tuesday, February 23, 2010, 10:39 AM > Hi Peter, > > Thank you very much for your prompt reply. Do you think that > 'normalising' some of the items of the measurement scales (I did have > some skewed data in some of the scale items) will resolve the > multicollinearity problem? > > Regards, > > Hayden > Normalizing would do nothing to alter collinearity statistics because the normalized variable is perfectly correlated with the originally scaled variable. Dale --------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra(a)NO_SPAMfhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 ---------------------------------------
From: Jeremy Miles on 23 Feb 2010 17:56 You've got a problem in your data that you need to fix. Either you're very, very unlucky, or you've made a mistake somewhere. The most common cause of this error is that you have one variable that is the sum (or other linear combination) of other variables, for example you have 3 test items and a sum of them, or something like a person's age now, a person's age when event X happened, and the time since X happened. It might also be because of pairwise deletion of data. Find out the determinant of the correlation or covariance matrix (if you're unsure how to do this, copy the correlation matrix into excel and use the mdeterm() function). Jeremy P.S. This is the SAS list, you'd be better off on the SPSS list or the SEM list. On 23 February 2010 00:39, Hayden Salter <sageman70(a)googlemail.com> wrote: > Hi, > > I have 2 questions that are related to each other: > > 1. I have a set of measurement scales that are highly correlated with > each other. I have tried to test the goodness-of-fit of my model > using SEM and the Maximum Likelihood method. SEM analyses can be > 'adversely' affected by multicollinearity. When I used AMOS to perform > the goodness-of-fit test, AMOS gave me an error that my matrix is not > definite positive. Is there a way to overcome this problem (i.e. > eliminate multicollinearity)? > > 2. When I used the Unweighted Least Square method instead of the > Maximum Likelihood Method, it performed the SEM analysis without an > error but the compromise is that you lose the important goodness-of- > fit test indices (e.g. RMSEA, CFI, etc.) from being displayed in the > results. Is ULS a reliable and renowned method for testing the > goodness-of-fit? > -- Jeremy Miles Psychology Research Methods Wiki: www.researchmethodsinpsychology.com
From: Hayden Salter on 24 Feb 2010 09:43 Hi Jeremy, I have used factor scores i.e. I add up all the Likert-scale variables making up the factor (measurement scale) and there is no way that one of the variables is the sum of the other variables since each factor has its own constituent items. I will follow your suggestion and identify the 'culprits' of multicollinearity. Once I identify the culprits, what should I do next? Delete the variables or just combine 2 factors into 1. Regards, Hayden On 23 Feb, 22:56, jeremy.mi...(a)GMAIL.COM (Jeremy Miles) wrote: > You've got a problem in your data that you need to fix. Either you're > very, very unlucky, or you've made a mistake somewhere. > > The most common cause of this error is that you have one variable that > is the sum (or other linear combination) of other variables, for > example you have 3 test items and a sum of them, or something like a > person's age now, a person's age when event X happened, and the time > since X happened. It might also be because of pairwise deletion of > data. Find out the determinant of the correlation or covariance > matrix (if you're unsure how to do this, copy the correlation matrix > into excel and use the mdeterm() function). > > Jeremy > > P.S. This is the SAS list, you'd be better off on the SPSS list or > the SEM list. > > On 23 February 2010 00:39, Hayden Salter <sagema...(a)googlemail.com> wrote: > > > > > Hi, > > > I have 2 questions that are related to each other: > > > 1. I have a set of measurement scales that are highly correlated with > > each other. I have tried to test the goodness-of-fit of my model > > using SEM and the Maximum Likelihood method. SEM analyses can be > > 'adversely' affected by multicollinearity. When I used AMOS to perform > > the goodness-of-fit test, AMOS gave me an error that my matrix is not > > definite positive. Is there a way to overcome this problem (i.e. > > eliminate multicollinearity)? > > > 2. When I used the Unweighted Least Square method instead of the > > Maximum Likelihood Method, it performed the SEM analysis without an > > error but the compromise is that you lose the important goodness-of- > > fit test indices (e.g. RMSEA, CFI, etc.) from being displayed in the > > results. Is ULS a reliable and renowned method for testing the > > goodness-of-fit? > > -- > Jeremy Miles > Psychology Research Methods Wiki:www.researchmethodsinpsychology.com
From: Hayden Salter on 24 Feb 2010 10:29
Should I instead combine the scores of the highly correlated variable? On 23 Feb, 21:28, stringplaye...(a)YAHOO.COM (Dale McLerran) wrote: > --- On Tue, 2/23/10, Hayden Salter <sagema...(a)GOOGLEMAIL.COM> wrote: > > > From: Hayden Salter <sagema...(a)GOOGLEMAIL.COM> > > Subject: Re: Multicollinearity problem > > To: SA...(a)LISTSERV.UGA.EDU > > Date: Tuesday, February 23, 2010, 10:39 AM > > Hi Peter, > > > Thank you very much for your prompt reply. Do you think that > > 'normalising' some of the items of the measurement scales (I did have > > some skewed data in some of the scale items) will resolve the > > multicollinearity problem? > > > Regards, > > > Hayden > > Normalizing would do nothing to alter collinearity statistics > because the normalized variable is perfectly correlated with > the originally scaled variable. > > Dale > > --------------------------------------- > Dale McLerran > Fred Hutchinson Cancer Research Center > mailto: dmclerra(a)NO_SPAMfhcrc.org > Ph: (206) 667-2926 > Fax: (206) 667-5977 > --------------------------------------- |