error running PLS (singular matrix)

[llucia]

Hi,

I have a model with two formative constructs and seven reflective constructs. One of the formative constructs has 6 formative items and the other one has 4 items. The sample size is 110.
When I try to run the PLS I have the following error:

"A singular matrix occurred during the estimation of the outer weights of a block. Using more indicators could solve the problem"

Can someone tell me what the problem is? I have run the same model with a different sample of 100 and it works.

Thanks in advance

Laura

[mariaemaus]

Hi Illucia, did you find the problem? I came across the same issue..

Kind regards,
Maria

[llucia]

Hi

Yes, I found the problem. There was a error in the dataset. Two variables were the same.

Laura

[anugrah]

hi
I hv that problem when im using SmartPLS version 1.
but it has been solved when im using new version (version 2)

CMIIW

[KatrinF]

Hello,

I get the same error as described above when running the boostrap.

My model consists of 10 latent variables (measured with 26 indicators) and i created 16 moderating effects.

My sample size is only 42.

One independant variable in my model is a dummy variable, so it only consists of 1 and 2... Could this be the problem?

On the other hand, i ran a very similar model bevor, that used a dymmy variable as an independant variable as well and there was no problem (sample size was over 100 in that case).

Does anyone know, if it is plausible that the sample size is the problem in my case or are there other possible errors, i could have done?

Thank you for your help,
Katrin

[christian.nitzl]

Hi Katrin,

the small sample size could be really the problem.

Maybe it will help when you use another weighting scheme e.g. the factor weigthing scheme.

Best regards,

Christian

[Berlin]

Hello,

I have the same problem, bootstrapping doesn't run:
"A singular matrix occurred during the estimation of the outer weights of a block. Using more indicators could solve the problem."

My sample size is 402.

I have a model with nine formative constructs and three reflective constructs. The formative constructs have a different number of formative items and the reflective constructs as well.

I managed to run it once, with sample size 300, but all of the t-Values (outer weights) were round abound 0,0001 which is nearly impossible.

Can anyone help? I couldn't find any mistake in the dataset.

Regards,
Doris

[christian.nitzl]

Hey Doris,

The sample size seems high enough for your model. Did you check if your model is recursive meaning is there no cycling connection in your model? It is not allowed in PLS that there a connection in model which goes back e.g. A->B->C->A and so on.

I hope that helps!

Christian

[Berlin]

Hello Christian

and thanks for your answer. In the meantime I found the mistake.
I forgot to raise the number of "cases", which was 15. So I put it on 402, as well as the sample size and it worked.

Regards,
Doris

[Jennifer Kessler]

Hello,

my issue is somewhat different from the ones reported, but I could not find a fitting thread, so I hope you don't mind if I post the request here.

I have a 2x2 factorial design and wanted to test a potential interaction effect between the two manipulations of my study. I did not find a significant interaction effect and tried to do something like a one-way ANOVA comparing the four groups with each other instead. For doing so, I created six contrasts (e.g. 1/0/0/-1 for comparing group 1 with group 4) and included them as constructs estimating my DVs.

However, when running the PLS algorithm, the singular matrix error occured, so I changed to factor weighting schme instead of path weighting scheme. It worked well and yielded the expected results (group 1 and 4 significantly differed from each other in accordance to path coefficients, the other groups do not differ at all, which I interpreted as an explanation for why interaction is not significant). Still, I am not able to run bootstrapping in order to get the corresponding t-values. Again, the singular matrix error occurs. Why is that? Do I need to conduct another method representing the one-way ANOVA 4-groups comparison I am used to in SPSS? My sample (n=109) was always sufficient in the other models I estimated in PLS. Probably a beginner's question, but I cannot find an answer in literature...

Thanks for every hint!

[Jennifer Kessler]

Since I have to deliver my Master Thesis at the end of the month, help is greatly appreciated :-)

Thanks a lot in advance!!!

[christian.nitzl]

Did you try another sign change option for the bootstrapping?
cn

[Jennifer Kessler]

Hello Christian,

thank you so much for your response.

I just tried the other two bootstrap options, but unfortunately, they did not work either...There was the same error message (singular matrix)

So, the basic procedure for comparing the four groups with each other was right?!?

Thanks again and best regards
Jennifer

[christian.nitzl]

Then it could be that one of your items have to less variance for bootstrapping. That is sometimes happing. To less variance means that for binary items you have to much 0 and to less 1. If I get it right your model could be calculated with PLS but you cannot calculate the s.e.
Best regards
cn

[Jennifer Kessler]

Thanks again.
That might be a potential explanation. Since I have almost equal number of people in each manipulation group, approximately half of the cases is coded 0, one quarter is coded 1 and again one quarter coded -1.

Is there any other way to compare the four groups with each other despite the missing variance?