Hello Levan, 

I think it just depends on what you are going to do. For example, if you are going to apply Structural Equasion Modelling, factor analysis would be recommended to test whether the well-established factor structure is corresponding with what you have in you sample. Also, sometimes there are numerous factor structures regarding the same psychometric tool. 

On the other hand, if you were using some totals (very strong assumptions need to be met - that every single item contributes to the sum score equally) you could omit factor analysis in my opinion. 

It just depends on what your targets really are. You may also dig into literature and see how others sorted out similar research questions and what methodology was applied.

Best wishes

Dear Witold,

Thank you for answer, highly appreciated.

To be honest, I don't know what is SEM, will look up it in Google for sure. But whatever I did is the following: I used already implemented scales from famous researchers in my research, thus I knew which items where measuring what and under what factors they should be united. That's why I concluded that, coming from the fact that I know what items belongs to which factors (as they have already been used/proved in previous papers), I didn't need to do Factor Analysis. Therefore, I, straight away, calculated the mean of those items under one known factor name and then went with further analysis (Correlation between these factors).

Is this way right?

Regards,

Hello Levan,

I think it is ok. However, it just depends on what you are going to do with your results. If you were aiming at publishing, well... then it could have been noted as limitation in my opinion. In every other case I think it is ok. I was doing some calculations in the exactly same manner as yours.

Let's put it this way - there are different levels and approaches and they may differ for undergrad students and say for post doc researchers. Only when trying to publish you should be ready for harsh scrutiny. 

Using a metaphor - you do not start with learning arpeggios as a guitar beginner, you start with 5 note pentatonic scale. 

So what you are doing is fine, it is just not the only way to approach the task.

Hope it clarifies a bit. 

PS: I have to admit I have seen some published work with serious caveats, too. 

If you have translated the scale into different languages, or if you used it with sample which is different from the sample which the scale was validated, it is the best to at least perform an exploratory factor analysis (EFA). 

Hi.  Let's say that you have a set of scales to measure x, which have been statistically validated.  The  premise of those studies was that they were 'good to go;' you don't need to revalidate them; they'd been validated, no?  Like the doctor using penicillin, you wouldn't have to do a study of whether or not penicillin works; you just administer it to the patient.  *On the other hand*, these days reasonable statistical packages, like Stata, make a confirmatory factor analysis pretty easy.  It involves code like this:

sem (INVOLVE -> s_felt s_work s_imp s_int s_job), stand

and gives pretty easy to read output, which will let you confirm that, in your case, you're getting what you want from the scale.  Remember that all you want is a confirmatory factor analysis that says that all the scale items are measuring one underlying latent variable. 

Here's a good PDF describing how to do CFA in Stata:  http://www.salaswright.com/wp-content/uploads/2012/05/CFA.SEM_using_Stata12.0.pdf

Similar references exist for other statistical languages.  If you are having to do this on the cheap, R (for free download) will do it; you'll need the SEM package, and you can use the R commander package to provide a graphic interface to R.

Hope this helps!

Dear Soon Li Lee,

Thanks for the answer. I am using it on the same language as the previous author.

Hello Mark,

Thanks a lot for such detailed explanation.

The fact here is not about the complexity of the factor analysis process, but other that I will try to explain in the below paragraph. By the way, I am using IBM's SPSS for statistical analysis of my data.

As for now, whatever concerns me is the following. To explain it in an understandable way, I will include some further details, if you allow me to

So, I have a research about the perceptions of specific eco-marketing tools that may lead to consumers trust toward advertisement and then lead to their willingness to buy product. I have the the scales for measuring the consumers' perceptions toward each of these tools (eco-label, eco-packaging, etc..). All of these scales have been adapted from previous researches and been successfully proved (reliable and valid). This scales has their respective ITEMS that has been asked as the questions to consumers. Now, as I have already concluded the questionnaire, I have this items with data (answers). Moreover, I know which of these items belongs to what Scale, thus shortly saying, what they measure.

Therefore, First, my question was do I still need to do Factor Analysis to see what item measure what, even though I know it?

And Second, Let's say I am going to do Factor analysis. In this case, should I include all the items from all scales, in one go, during the factor analysis OR should I do Factor Analysis separately for each of the scales' items?

Regards,

Hello Levan,

For example, if one questionnaire is based on 2 factors according to past literature, you test the items in factor analysis taking all the questions from the questionnaire divided across two factors. So in factor analysis you just go with one questionnaire at a time. 

Best wishes

Let me be sure I understand.  You have a set of items that looks like this:

The item is eco-labeled (from Smith's scales on labeling) 

SA A N D SD

The item is eco-packaged (from Jones' scales on packaging) 

SA A N D SD

The item is described as eco-friendly (from Smith's scales on labeling)

SA A N D SD

This is to say that the ITEMS have been validated separately, but the SET of items have not been validated TOGETHER.  Am I right?  

IF that is the case, THEN you need to validate, since your implicit hypothesis is that the items work TOGETHER to measure an underlying factor -- UNLESS you're positing two factors -- ie, Smith's items on labeling measure a labeling factor and are being used separately from Jones' packaging items.  You've made a new scale if you merge two sets of items to create a 'measure' (an underlying factor); you're not creating a new scale if you're measuring two different scales at the same time but not merging the scores.

Hope this helps!

First and foremost, nothing is ever 'proved'.   Reliability, in particular, is an aspect of a measure's application to a sample -- never a population.  If the sample changes, then reliability may change.  You should ALWAYS given some indication of the reliability of the measure with any sample.  As for validity, an adult IQ measure may show consistent validity with college students.  However, no one would accept it as a valid measure of IQ for children below the age of 6 or 7.  In your case, someone made the point of language translation.  That is another excellent example of a measure that might be valid in one sample but not in another sample. 

Mark West describes a more subtle (therefore, easily overlooked) issue that you may have in your particular case.  If you are chopping up measures that have shown construct and criterion-related validity in other research, you will need to demonstrate the same with your reconstructed measure.  There are cases wherein the order that measures are given can influence the outcome.  

If you don't have access to SPSS's SEM software (AMOS), you can come close in the conceptual sense by telling SPSS to extract factors via principal axis factoring rather than using the default (principal components) and rotate the factors using an oblique method. 

First of all, it would be smart if I define what I MEAN by SCALES and ITEMS in aforementioned questions. SCALES are the constructs - for instance, "The perception of eco-label". ITEMS are the questions of these scales that are asked to respondents measure the overall SCALE.

Thank you Mark West, that helped me a lot. Actually, my actual question here was a bit about different concern, but I also have almost same situation whatever you described about merging the scales. Precisely, I have four scales that measures Trust in Advertisement. These four scales are "Affect" "Usefulness", "reliability" and Willing to rely on/use". Thus all of these four scales merged measure the Trust In Advertisement.

However, in my paper, I split these four scales and group it under two groups:

1. "Trust In advertisement" - this will include scales - Affect, usefulness, and Reliability

and

2. "Willing to use advertised product" - that will be measured by "Willingness to rely on/use". I am examining the connection between them, that's why I split them into two parts.

In this case, If I correctly understood your answer, Mark, as these three constructs have never been used before as one merged scale, I have to do factor analysis and in this analysis, I have to include only those ITEMS that belongs to these three SCALES (Affect, Usefulnes, reliability). Is that right?

..As the continuation of the previous question.

As for Factor Analysis, whatever is main concern and confusing part for me is the following: in case I need to do factor analysis for the scales, the question is following:

1. Do I have to do only ONE Factor Analysis for all scales including all of scales' all ITEMS? thus to group them and get grouped them under the factors like in original papers

OR

2. Do I have to do as many Factor Analysis as many Scales I have? I mean, let's say, do one factor Analysis for scale - "perception of eco-label" and in it analyze only those items that were used to measure it?

Thanks for answers in advance,

Regards,

Dear Levan,

I hope it answers both of your questions. You can do this two-way.

Either you can follow the factor structure suggested by the original scale design (e.g. DASS-21 scale has 3 factor structure suggested by Lovibond and Lovibond).

Sometimes there are competing alternative factor structures suggested fro the same scale (hence for DASS-21 you have several factor structures stemming from 1 to 4 factor structures).

I would rather follow the original factor structure suggested by the researchers who designed that and surely tested their new scale among more than on sample to check its psychometric properties including factor structure.

Hope it helps a bit.

I think it is as often as not that you will NOT get the same factor structure as the scale author has advertised.  However, that is not necessarily an indictment of the measure. Again, many of these things can depend on the sample (size, homogeneity, etc.)  I think most of us would be inclined to do what Witold suggests.  Start with the theory as given and work from there.  With SPSS PAF, you can force the program to derive 3 factors, or four, or one...however many the scale says you should have.  You can't tell the items where to fall, like you can with AMOS, but you can force the program to give you the same number of factors that theory says there should be and see if the items fall into place as expected.  If they don't, you at least have some direction with which to investigate further.  

I just read back through this thread and realized I missed some of your question.  It appears that what you really want to do is a 2nd-order factor analysis.  That is, items load on the original measure's factors, and you are theorizing that those original factors actually load on newer, 2nd-order factors.  If this is correct, I don't think you are going to be able to do this in a single step without some sort of structural equation modeling software (AMOS is the SPSS product, M-Plus, EQS, Lisrel, etc).  I think your only other option would be to get factor scores for the original scales (call these the 1st-order factors) either by following the original measure author's instructions for obtaining such scores, or by having SPSS create factors scores via PAF.  Then, use those scores as the items in another factor analysis.  The danger there is you might need several scores to load on each factor.  You may not have enough data (the 1st-order factor scores acting as items for the 2nd-order factoring) to complete a robust secondary analysis.  I say all of this reluctantly.  I am not sure I would try a 2nd-order analysis without SEM software.

Here is a second-order CFA in Stata: 

. sem (Phys -> phyab1 phyab2 phyab3 phyab4)

> (Appear -> appear1 appear2 appear3 appear4)

> (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)

> (Par -> parrel1 parrel2 parrel3 parrel4)

> (Nonacad -> Phys Appear Peer Par)

Nonacad is the second-order factor.  This is the  Marsh, H. W. and Hocevar, D., (1985) self-concept model.

That's why I'd argue for Stata or R as an OK choice for this problem. 

Many widely used scales were evaluated on samples from specific populations that may differ from your sampled population. What is reliable on White College Students may not be reliable on Latina mothers, etc. 

Mark West shows the code in Stata for a 2nd order factor analysis. An advantage of Stata compared to SPSS plus AMOS is that Stata does not charge extra for modules the way SPSS does. Also, if your university has a grad plan with Stata the price of Stata is much less than SPSS and you own it--not a one year lease. Of course, R is absolutely free.

 R can be sort of brutal, but (to my mind) Stata is the easiest of the statistical packages, with its excellent documentation -- especially if you use a good intro guide like Dr. Acock's :)  And the SEM package is easier than MPLUS or (for sure) LISREL, both of which I've used.

In alternative you can do a 2nd order factor analysis in AMOS. It is a graphical program easy to manage.

Hello..I've already asked similar questions related to what I'm doing at the moment. Based on many colleagues' answers, I think that you should go for a Confirmatory Factor Analysis (CFA), which is hard to do with SPSS, but much easier with AMOS. In this regard, you can find a book online by Green and Salkind in PDF format. It has a very useful chapter on factor analysis.

Also, this YouTube presentation might be very good:

https://www.youtube.com/watch?v=MCYmyzRZnIY