Singular matrix estat kmo

[Marco Delli]

Dear Stata users,
I am currently trying to practice a bit on STATA.
I decided to make a regression from a database that I constructed by myself making a survey.
This regression explains lnP in function of a lot of variables (27) some are dummies and some perfectly correlated.
Observing the mean of the VIF test higher than 10 I decided to run a ACP.
Everything runs fine. but when I want to run the kmo test the error "correlation matrix is singular".
I would like to know how to fix this problem and why it happens.

Thanks in advance for your help,
Marco

[Carlo Lazzaro]

Marco (as per FAQ, please consider the preference for full given and family names on this forum. Thanks):
- as per FAQ, your chances of getting helpful replies are conditional on posting what you typed and what you got from Stata. Besides, you may want to share an example/excerpt of your dataset (type -search dataex- from within Stata and follow the instructions reported in the help file);
That said:
-a high VIF might be immaterial if it refers to controls and worrisome if it refers to predictors;
- 27 independent variables might be pretty good if you have a sample of 10/15 observations per predictors; otherwise, your results will probably affected by overifitting;
- if with ACP you mean the Italian acronym for PCA (principal component analysis), you may want to take a look at https://stats.stackexchange.com/ques...trix-and-pca.;
- does kmo stay for Kaiser-Mayer-Olkin?

[Marco Delli]

Dear Carlo,

- Here are some informations about my dataset:
The dependent variable is studied throug more than 50 variables, I decided to group them by group in function of their kind of information.
One group is about the profil of the product, the second one about a group of comon characteristics and the last one also.
The sample is constitudes by more than 400 observations grouped by group of 45 or 70 more or less.

- What I try to do, because I am not so experimented, is the following process:

The regression is more or less like the following one:
reg lnP Motiv2 Motiv3 Adv2 Adv3 Shop2 Shop3 Style2 Style3 Impulsive Discount Online Frequency Middel End Indiferent Partner1 Partner2 Partner3 Partner4

For some of them I observe a vif mean really high. Am I true to said that that the coefficient are more volatile due to collinearity ?
And so , I am good to try to reduce this collinearity opting for PCA ?

- I enter some codes:

global profil Motiv2 Motiv3 Adv2 Adv3 Shop2 Shop3 Style2 Style3 Impulsive Discount Online Frequency Middel End Indiferent Partner1 Partner2 Partner3 Partner4
pca $profil
screeplot, yline(1)
pca $profil, comp(6)
predict pc1 pc2 pc3 pc4 pc5 pc6
estat kmo //Matrix singular problem

- My problem is when I run the Kaiser-Mayer-Olkin test in fact. From the link in your previous message, it means that the matrix generated from my variables is singular and so not adapted to the PCA ?

Thank your for your consideration,
Marco

[Carlo Lazzaro]

Marco:
again, your last post is nit the best way for eliciting postive replies (see tha FAQ. Thanks).
That said, your problem probably rests on quasi-extreme multicollinearity: try to rethink about your regression specification, first.