# Factor Analysis newbie seeks your help

#### fear_of_gerbils

##### New Member
Hi,

Perhaps unwisely as a newcomer to stats, I have taken on a scale validation project. I’ve searched the literature and can’t find anything so this is the issue:

I’m developing a scale that (hypothetically) will identify, differentiate and measure two different constructs occurring in parental/child relationships.

I’m allowing my participants to redo the scale questions for each child that they have (using Qualtrics), so each parent may complete it once or twice or nine times depending on their fecundity (or patience…)

So I’m going to have less participants than I have responses. Is there any implications for this for my factor analysis? I just can’t find a scale validation study with a similar scenario so worried I’m overlooking something.

I’m thinking of running EFA, then CFA. Hopeful of at least 300 participants and I have <30 items in the scale.

Thanks for any thoughts!

#### spunky

##### Can't make spagetti
I’m allowing my participants to redo the scale questions for each child that they have (using Qualtrics), so each parent may complete it once or twice or nine times depending on their fecundity (or patience…)
So... if I understand this correctly the number of children are "nested" within their parents? So if Parent 1 has 3 children, then she/he will have 3 questionnaire responses (1 per child), if Parent 2 has 2 children, then 2 questionnaires and so on...?

#### fear_of_gerbils

##### New Member
So... if I understand this correctly the number of children are "nested" within their parents? So if Parent 1 has 3 children, then she/he will have 3 questionnaire responses (1 per child), if Parent 2 has 2 children, then 2 questionnaires and so on...?
Yes, that is the case. The theory I'm basing the scale on supposes that there is a qualitative difference in parental-child dyads, and the 2 constructs may or not be present in each separate dyad relationship. So I thought I could treat each completed response individually and separately for the actual scale validation and factor analysis, as I would with a scale validation where a participant has only completed it once. I'll have to think about fatigue effects but can't think of any issues with the FA, unless I'm vastly overlooking something.. which is of course possible! Thanks for reading.

#### spunky

##### Can't make spagetti
I'm vastly overlooking something.
Yes, you're kind of vastly overlooking some very important stuff: the issue that your observations are not independent. The theory of factor analysis (and most linear models) assumes that every person on which you collect data is independent from the next one. That is obviously not the case for you because the responses of children with the same parent will be more correlated among themselves than those of children with different parents.

You will need to use an extension of traditional factor-analytic techniques calledmultilevel factor analysis, which combines features of linear mixed effects models and factor analysis. Otherwise your analysis might be suspect.

If you're doing this for some sort of school project (thesis, final term paper, etc.) you'll probably be OK just doing this and noting it as a limitation towards the end. I mean, strictly speaking, you're still doing it wrong but I'm guessing people who are evaluating you might let it pass. If you're planning to publish this then it might become a problem because, nowadays, a lot of people are aware of that clustered data needs to be handled differently and a reviewer would probably demand you to use the more proper analytic method.

If you really are a "newcomer to stats" I might suggest you do something different, heh.

#### fear_of_gerbils

##### New Member
Yes, you're kind of vastly overlooking some very important stuff: the issue that your observations are not independent. The theory of factor analysis (and most linear models) assumes that every person on which you collect data is independent from the next one. That is obviously not the case for you because the responses of children with the same parent will be more correlated among themselves than those of children with different parents.
You will need to use an extension of traditional factor-analytic techniques called multilevel factor analysis, which combines features of linear mixed effects models and factor analysis. Otherwise your analysis might be suspect.
spunky THANK-YOU this is the answer I have been trying to prise out of my supervisor for weeks, I just had this nagging feeling....

No, I will not be touching multilevel FA with a barge-pole at this point in my limited experience! Plus I only use SPSS and I don't think it even does that...?

And yes it's for a thesis with no expectation of publication. It's posed as a pilot study so I feel a significant re-design and a nice section in limitations/directions for future study coming on.

Thanks again

gerbil_phobe