problems with SPSS factor analysis - Misc. Stats Topics Discussion

Misc. Stats Topics Discussion > problems with SPSS factor analysis

I am struggling with my dissertation project for my masters. My supervisor has left the country and only gets back 10 days before my final hand in and I have no results. EEK! I think I have vaguely got my head round factor analysis (I am not mathsy at all!) and am attempting to use it in order to validate a questionnaire I have designed. However there is something wrong with my data so that when I try to analysis it a lot of the tables that are meant to be given in the output don't appear (e.g. significant section of the correlation matrix and the KMO, barlets). I don't know why this is happening and think it may be something to do with multicollinarity as at the end of the correlation matrix it says the derminant is .00000 and that the matrix is not positive definite.

I have gone through my results and none of the participants seem to have answered in a weird way (e.g. 3 for everything or 1,2,3). I have also checked the correlation and there are none that correlate at higher than .9 and no variables that correlate higher than .5 on most of the other variables.

I have used PCA to analysis the results (I know that this isnt the ideal FA test to use but this is the only one I can get any results out of even if they are wrong!).

If anyone can help I would be eternally grateful!! I can send you the results in SPSS if this helps to demonstrate what I have tried to explain above.

thanks

hannah

August 8, 2010 |

hannah

Hey Hannah,

Feel free to send me your output and I'll take a look and try to provide some thoughts about what might be going on (statsmakemecryguy@me.com). Thanks!

Jeremy

August 8, 2010 |

Stats Make Me Cry Guy (Jeremy Taylor)

I am developing a questionnaire to use in my research. So I wanted to do a principal component analysis. But It didn't give the tables that are meant to be given in the output; such as significant section of the correlation matrix and the KMO, barlets. Also at the end of the 'Total Variance Explained' table it shows 'The matrix is not positive definite'. At the end of the correlation matrix it says the determinant is .000 and that the matrix is not positive definite.

I have checked my results and none of the participants seem to have answered differently. I have also checked the correlation and there are some that correlate at higher than 0.9. I deleted those variables one-by-one, however it still did not improve the results and appeared as same before. So, now I am up my head and couldn't think what to do.

Could anyone can help me please? I would be heartily grateful!! I can send you the SPSS results which may help to understand what I mean to.

Many thanks.

Dev

February 27, 2011 |

Dev

Hey Dev,

What you are dealing with is a tough issue, because there are a number of conditions that could result in getting such an error message. How big is your sample size? Do you have missing data? If so, how much?

Also, check to make sure that none of your variables weren't accidentally recoded to be constant (e.g. all zeros or all ones...etc). I look forward to your response and I'll do my best to help. If you'd like, you can also send me your dataset and I can take a quick look at it and see if anything seems out of the ordinary. If you want to do that, my email is jtaylo20@me.com.

Thanks.

Jeremy

March 3, 2011 |

Jeremy Taylor

Hi,

I experience the same situation too, but I got the statement 'The matrix is not positive definite' at the end of the Total explained variance.

Could someone share how to resolve this. I would be great full.
Thank you..

March 27, 2011 |

Cris

Sorry for my delay in response, but I've been moving across town, thus somewhat unavailable. As I mentioned above, feel free to send me your output and I'll be happy to see if I can see any reason for the error message. Thanks!

Jeremy

April 5, 2011 |

Jeremy Taylor

Hello, I hope someone can help me with my dilemma for my thesis.

I have a small dataset (n=59) and I have read that if KMO above .6, then a Factorial Analysis may work.

In trying to compute this in SPSS every textbook talks about variables. I suppose each item on a questionnaire is a variable, right?

When SPSS asks for variables, I can send the individual items there, but is the Grouping Variable? I read I can leave this empty.

However, the KMO table does not appear and instead the output says "null null a. This matrix is not positive definite."

What does this mean? I have a small sample, but I was hoping to be able to obtain something from it!

Sorry this may sound silly, but I am trying to check for this before reading loads on the method, as I may not be able to use it due to the small sample. :(

Thanks

January 2, 2012 |

freya

me again...

I have found that when entering all the variables (items = 156 x n=59), the KMO is not calculated...

If I enter all the items under 1 variable (e.g. anxiety) for n=59, then the SPSS gives the KMO. Can I do this individually for each of the pre-arranged variables (e.g. anxiety, depression, cognitive problems)? I thought all the items needed to be included in the same analysis so the SPSS would give the 'true factors', rather than the constructs decided when the questionnaire was developed. Any help please?

January 2, 2012 |

freya

First of all, allow me to apologize for the extremely delayed nature of my response. Unfortunately, my forum notifications were being sent to an inactive email, so I didn't become aware of new posts for a long time.

With respect to your question, if you already have a hypothesis about how the items should be arranged (which "subscale" each belongs on), then I would recommend running a Confirmatory Factor Analysis (CFA), not an Exploratory Factor Analysis (EFA), which sounds like what you are trying to do. CFA's aren't run in SPSS, but instead an SEM software (such as Mplus, AMOS, or LIsrel). They can also be run in R. I hope this helps.

February 23, 2012 |

Jeremy Taylor

I found that my spss cant generate the KMO chart... it is weird because only Q1-Q17 factor can not generate, but the remaining Q 18-33 can generate the KMO chart....

the log of the output file shows:
FACTOR
/VARIABLES Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12 Q13 Q14 Q15 Q16 Q17
/MISSING LISTWISE
/ANALYSIS Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12 Q13 Q14 Q15 Q16 Q17
/PRINT INITIAL CORRELATION KMO EXTRACTION ROTATION
/FORMAT SORT BLANK(.50)
/PLOT ROTATION
/CRITERIA MINEIGEN(1) ITERATE(25)
/EXTRACTION PC
/CRITERIA ITERATE(25)
/ROTATION VARIMAX
/METHOD=CORRELATION.

is that any error of my data?

Thanks for your help

February 28, 2012 |

Sarah

I'm afraid I need a but more info. If you'd like to send your full output and syntax, I can try to review it. Email to statsmakemecryguy@me.com.

February 29, 2012 |

Jeremy Taylor

Hi, i seem to be having the exact same problem as Hannah, where i have done a factor analysis but KMO or Bartlett's is not appearing even though i have selected it? I am looking to do a factor analysis on various traits to find out which are the most relevant. I presume that KMO and Bartlett's is an important determinant to see if the test is reliable and sufficient?

Any help is greatly appreciated,
Regards.

May 17, 2012 |

Mark

In reply to a few posts:

Sarah,

What is your sample size?

Mark,

Since you sent me your output and data, I have a bit more insight into what is likely going on with your situation. I think your problems stem from your sample size, specifically when you select your sample by people that answer Q2 with "1" (which is what was done in your analysis). Below is your syntax:

DATASET ACTIVATE DataSet1.
FACTOR
/VARIABLES Q3_1 Q3_2 Q3_3 Q3_4 Q3_5 Q3_6 Q3_7 Q3_8 Q3_9 Q3_10 Q3_11 Q4_1 Q4_2 Q4_3 Q4_4 Q4_5 Q4_6
Q4_7 Q4_8 Q4_9 Q4_10 Q4_11 Q5_1 Q5_2 Q5_3 Q5_4 Q5_5 Q5_6 Q5_7 Q5_8 Q5_9 Q6_1 Q6_2 Q6_3 Q6_4 Q6_5
Q6_6 Q7_1 Q7_2 Q7_3 Q7_4 Q7_5
/MISSING LISTWISE
/ANALYSIS Q3_1 Q3_2 Q3_3 Q3_4 Q3_5 Q3_6 Q3_7 Q3_8 Q3_9 Q3_10 Q3_11 Q4_1 Q4_2 Q4_3 Q4_4 Q4_5 Q4_6
Q4_7 Q4_8 Q4_9 Q4_10 Q4_11 Q5_1 Q5_2 Q5_3 Q5_4 Q5_5 Q5_6 Q5_7 Q5_8 Q5_9 Q6_1 Q6_2 Q6_3 Q6_4 Q6_5
Q6_6 Q7_1 Q7_2 Q7_3 Q7_4 Q7_5
/SELECT=Q2(1)
/PRINT UNIVARIATE INITIAL CORRELATION KMO EXTRACTION
/FORMAT BLANK(.10)
/PLOT EIGEN
/CRITERIA MINEIGEN(1) ITERATE(25)
/EXTRACTION PC
/ROTATION NOROTATE
/METHOD=CORRELATION.

The tenth line of is where problems arise: " /SELECT=Q2(1)"

This command selects your analysis to include only individuals that respond with "1" for question "Q2". The problem is: doing that shrinks your sample size down to only 30 people, which is far too few people when you have approximately 40 variables to analyze.

There is no clear-cut guide for how many people are required per variable (although i've heard anywhere from 5 to 20 used as a rule of thumb), but you should definitely never have more variables than people in your analysis. In your case, I tried running your analysis without that line of code (without selecting that subset of cases) and your sample size is 213, which is enough to allow the model to run correctly (the output then produces KMO and Bartlett's). I hope that is helpful!

Best.

Jeremy

PS: Sarah, might the same be happening to you?

May 17, 2012 |

Jeremy Taylor

Thanks for the reply Jeremy, unfortunately i have to run the factor analysis on each of the brands, so can i still run it on each even though the sample size is small with relevant results, and without producing a bartlett's test of KMO?

May 17, 2012 |

Mark

Unfortunately, I would not trust the results of a factor analysis that has more items than people. You could try reducing the # of items, but you'd have to reduce it by a lot. Alternatively, in other software I think you can run multi-group factor analyses, where it uses the entire sample, but accounts for variability by group (instead of running each group separately), but I don't think SPSS has that capability.

May 17, 2012 |

Jeremy Taylor

Sorry, one final question. Unfortunately i will have to run that analysis as it is what is needed, and due to time constraints i cannot collect anymore data. How would you recommend that i analyze the data given without the KMO or Bartlett's? I simply need to know the relevant traits that exist for each brand. Since you have my output could you attempt to draw a conclusion as to which traits (e.g. Down-to-earth, sincere, etc.) are relevant for retention.

Thanks ever so much, you've been a huge help!

May 17, 2012 |

Mark

I'm afraid it is difficult for me to offer advise on what sounds to be a very complex project, with limited information. However, from what I've heard so far, I'm not actually sure that I'd actually use a factor analysis. From what I understand (which may not be a lot), I think I'd be more likely to use a latent class analysis to explore this issue.

To answer your question more directly, I would not recommend trying to interpret findings of a factor analysis given the small sample sizes of your groups...which is why I might consider a latent class analysis instead.

May 17, 2012 |

Jeremy Taylor

Basically im looking to determine the relevant traits that are associated with a certain brand, e.g. Nokia. So im looking to perform an analysis to find the most relevant traits out of a total of 42 traits. So for example im looking to perform an analysis on Nokia and it could end up that 36 of the 42 traits are relevant to that brand. Can you recommend an analysis to use through SPSS to determine this? Is latent class analysis through SPSS??

Once again, thank you!!

May 17, 2012 |

Mark

Mark,

Latent Class Analysis is definitely where I would start, which would allow you to identify classes of "opinion profiles" for respondents, based on patterns of responses, and then determine if owners of each type of phone are most likely to be in particular profiles...etc. Unfortunately, this isn't something you can do in SPSS. I might recommend using R (perhaps the "mclust" package or "poLCA" package). They are faily straight forward to use.

If you are completely tied to using SPSS, you could try using LCA's ugly cousin: Cluster Analysis. The downside of Cluster Analysis is that it is much less clear cut and empirically based, which can make for ambiguous outcomes. Still, it can provide clusters to compare across phone types, just in a less precise way, compared to LCA.

Here is the cran site for downloading the poLCA and mclust packages:

http://cran.r-project.org/web/packages/poLCA/index.html
http://cran.r-project.org/web/packages/mclust/index.html

Here is a great resource for doing cluster analysis in SPSS, if you decide to go that route:

IBM/SPSS Cluster Analysis Resource

I hope this helps, Mark!

May 17, 2012 |

Jeremy Taylor

Hello, sir i need your assistance regarding Spss. In my Questionnaire there is 11 variables having 31 questions, but the problem is i can not get enough KMO, i.e. 0.3 or 0.4. more over arrangement of questions in the Rotated Component Matrix portion is not in sequence, i tried a lot of ways to arrange them in sequence but i couldn't. even i also attempted to fake or alter the data only to get high KMO and arranged questions in Rotated Component Matrix, but in response i got this statement, 'The matrix is not positive definite'.... plz sir help me regarding this matter.

June 23, 2012 |

Saad

hello i have a question about factor analysis.i would apriciate if someone could answer me. I have 200 questionaires with about 180 variables in each questionaire. i performed a factor analysis and come up with 41 factors. For the purpose of my research this number of factors is very big. So in order to reduce them more i then made a second factor analysis based on this 41 factors i got from the first analysis and come up with 7 factors. This is acceptable number for me. My question is this: what i have done, i mean two times factor analysis, is it possible or is it complete wrong??
If someone could answer back it would be really helpfull for me

September 25, 2012 |

Panos

Panos, what you've done is somewhat unusual, but not necessarily "wrong". It is similar to "parceling" items, which is sometimes done. I would test out the model using a hierarchical confirmatory factor analysis, if I were you... Although your sample size is small for that many items...

October 18, 2012 |

Jeremy Taylor

Sir... I need some reconfirmation from Sir... Is it in the rules of performimg factor analysis,
1) the respondents (N) must more than the variable? in my cases there is 143 respondents while the variable is 176...
2) must be No missing data for all the variable

How can i reduce my variable by using spss? which method should i select to run?

Thanks for your helps...

October 31, 2012 |

Jason Foo

Dear Jason Foo,

I would definitely recommend that you have many more participants then variables when running a factor analysis. In fact, some guidelines indicate you should have at least five participants for each veritable that you are trying to factor analyze (although that is not a concrete rule). In general, if you need to reduce the number of items, you can either exclude some that you believe are least relevant, or you can parcel your items (which is a process of combining several items together that you think are extremely highly correlated and represent similar constructs). I hope this is helpful, Jason!

November 3, 2012 |

Jeremy Taylor

i want to do my research on link between organization ethics and job satisfaction and i cant run factor analysis can you please tell me how can i remove this positive definte thing

March 23, 2013 |

yousaf

Hi yousaf,

I'm afraid we'll need more information to try to help. There are a number of reasons why you could be getting a message that warns that your Matrix is not "positive definate". One possibility is that you simply have insufficient correlation between items to run your factor analysis . To explore this, simply run a correlation matrix and see which item(s) are poorly correlating with other items, those items may need to be removed to allow your model to converge. Another possibility is that your sample size is insufficient, based on your model. Bottom line, there isn't one solution to the "positive definite" problem, there are several solutions that are related to the various reasons you are getting the warning.

March 27, 2013 |

Jeremy Taylor

Hi Jeremy,

I did a questionnaire with 60 items (=variables) with 25 respondents and now I want to run a factor analysis to see if I get the same items as in the original questionnaire (which was tested on a large group of respondents).

However when I'm running a factor analysis I get that the matrix is not positive definite. Is this due to the fact that I have more variables then respondents?

Thanks for your response!

May 20, 2013 |

Leanne

Leanne,

In a way, yes that is why. It isn't specifically because you have "more" variables than respondents, but because your sample size is small in general, especially relative to the number of variables (items) you have. Generally, you want MANY more respondents than items. In fact, I've heard a rule of thumb of 5 to 10 respondents for every item, though this isn't a rule that is always followed.

To run a factor analysis on 60 items, I'd like to see SEVERAL hundred participants in my sample. I hope this helps and sorry to be the bearer of bad news.

May 24, 2013 |

Jeremy Taylor

Hi,
Jeremy
it would be great if you help me with the below data syntax......

Correlation Matrix(a)

a. This matrix is not positive definite.

why it is giving such a results??? i have checked all the entries.....please help

FACTOR
/VARIABLES VAR00001 VAR00002 VAR00003 VAR00004 VAR00005 VAR00006 VAR00007 VAR00008 VAR00009 VAR00010 VAR00011 VAR00012 VAR00013 VAR00014 VAR00015 VAR00016 VAR00017 VAR00018 VAR00019 VAR00020 VAR00021 VAR00022 VAR00023 VAR00024 VAR00025 VAR00026 VAR00027
VAR00028 VAR00029 VAR00030 VAR00031 VAR00032
/MISSING LISTWISE
/ANALYSIS VAR00001 VAR00002 VAR00003 VAR00004 VAR00005 VAR00006 VAR00007 VAR00008 VAR00009 VAR00010 VAR00011 VAR00012 VAR00013 VAR00014 VAR00015 VAR00016 VAR00017 VAR00018 VAR00019 VAR00020 VAR00021 VAR00022 VAR00023 VAR00024 VAR00025 VAR00026 VAR00027
VAR00028 VAR00029 VAR00030 VAR00031 VAR00032
/PRINT INITIAL KMO EXTRACTION ROTATION
/FORMAT SORT
/CRITERIA MINEIGEN(1) ITERATE(25)
/EXTRACTION PC
/CRITERIA ITERATE(25)
/ROTATION VARIMAX
/SAVE REG(ALL)
/METHOD=CORRELATION.

November 27, 2014 |

ajay

hi there,

for some reason, the KMO is not displayed in my output. i definitely have too many cases in relation to how many variables i have, but even if i don't include all the variables in the PCA, the KMO is not displayed. what other reason that i could potentially fix could be behind this? i'd be very grateful for your help..

ira

February 18, 2015 |

Ira

Hi, There
i am trying to extract components from my data set and its a very small dataset (n=6).the result says the correlation matrix is not positive definite and the KMO and Barallett test results are not displayed in the output which have been given as input. Kindly do help me with this issue.

August 9, 2015 |

GOPINATH

Probably you have very correlated variables!!

April 10, 2016 |

HackGen10

I need your guidance my KMO is coming out to be .92 which is very high and there are no cross loadings...my guide told me to reduce the value of KMO but how to do that....

March 15, 2017 |

Priya

Less Your sample size.... to reduce the value of KMO...!!!

April 13, 2017 |

Hammad Zahid

Correlation Matrixa a This matrix is not positive. i try do KMO but cant get the result. anyone plz help me.

June 10, 2017 |

nathan

Hi,
Can anyone help? I am having the same issues as Hanna who posted on August 8, 2010...

"However there is something wrong with my data so that when I try to analysis it a lot of the tables that are meant to be given in the output don't appear (e.g. significant section of the correlation matrix and the KMO, barlets). I don't know why this is happening and think it may be something to do with multicollinarity as at the end of the correlation matrix it says the derminant is .00000 and that the matrix is not positive definite.

I have used PCA to analysis the results (I know that this isnt the ideal FA test to use but this is the only one I can get any results out of even if they are wrong!).

If anyone can help I would be eternally grateful!! I can send you the results in SPSS if this helps to demonstrate what I have tried to explain above."

thanks
Kudirat

August 6, 2017 |

Kudirat

Hi,
Can anyone help? I am having the same issues as Hanna who posted on August 8, 2010...

I have used PCA to analysis the results (I know that this isnt the ideal FA test to use but this is the only one I can get any results out of even if they are wrong!).

If anyone can help I would be eternally grateful!! I can send you the results in SPSS if this helps to demonstrate what I have tried to explain above."

thanks
Kudirat

August 6, 2017 |

Kudirat

Hi,
Can anyone help? I am having the same issues as Hanna who posted on August 8, 2010...

I have used PCA to analysis the results (I know that this isnt the ideal FA test to use but this is the only one I can get any results out of even if they are wrong!).

If anyone can help I would be eternally grateful!! I can send you the results in SPSS if this helps to demonstrate what I have tried to explain above."

thanks
Kudirat

August 6, 2017 |

Kudirat

Hi,
Can anyone help? I am having the same issues as Hanna who posted on August 8, 2010...

I have used PCA to analysis the results (I know that this isnt the ideal FA test to use but this is the only one I can get any results out of even if they are wrong!).

If anyone can help I would be eternally grateful!! I can send you the results in SPSS if this helps to demonstrate what I have tried to explain above."

thanks
Kudirat

August 6, 2017 |

Kudirat

Hi,
Can anyone help? I am having the same issues as Hanna who posted on August 8, 2010...

I have used PCA to analysis the results (I know that this isnt the ideal FA test to use but this is the only one I can get any results out of even if they are wrong!).

If anyone can help I would be eternally grateful!! I can send you the results in SPSS if this helps to demonstrate what I have tried to explain above."

thanks
Kudirat

August 6, 2017 |

Kudirat

STATS MAKE ME CRY

Misc. Stats Topics Discussion > problems with SPSS factor analysis

Post Your Reply