pamela
通常好像都看到的是做的自变量。。。。实在不是很懂这个,自己看的速成的书。
谢谢!
Msmart
Yes, of course. However, you should explain the extracted factors well. We never run a regression without the knowledge of our dv.
pamela
thanks a lot. But what is DV?
I do have some problem in explaining the factors. 'cause i do some extraction about student's job expectations(but after they have got results of their first job). then i think patterns of using web is influential to those expectations. So i use web usage to predict expectations which is not normally seen...
And by the way, can you give some literature support for me about using those extractions as dependent variables?
thank you very much!
anita_jiu
DV-dependent variables, as you spotted at the end of your response.
Factor scores may be used as DV but it depends on what analytical technique(s) you wish to do. Regression, discriminant analysis, logistic regression...? You need to relate your research question with analysis objectives, as subsequent analysis (say, e.g. regression) will usually have different criteria and therefore factor rotation method is a concern.
Msmart
1. Actually, u know, what factor analysis tries to is to find out latent (usually unobservable) structure/variables underlying your observed variables. When your DVs are too many to run a meaningful/explainable multivariate multiple regression, FA techique is employed to extract controllable number of DV. It's a common sense that factors could be used as DVs just like your observed ones, instead of some important innovations documented by literature. However, such an application could be mentioned in any classic multivariate analysis textbook as a basis function of factors. Two of many alternatives are
"Harris, R. J. (1975). A Primer of multivariate statistics. N.Y.: Academic Press."
"Johnson, R. A., & Wichern D. W. (1992). Applied multivariate statistical analysis, Prentice Hall, 1992, 3rd Ed. [ISBN 0-13-041773-4] "
Msmart
2. If you wanna get a great sense of your factors, it's better to dig into the literatures in ur discipline, e.g. labor economics (furthermore, maybe compensation), expectation theory in economics, or behavioral bias in psychology... Actually it is psychologists and sociologists who play with FA frequently, instead of statistician.
3. Some more complicated situation when factors appear in the right side of regression, i.e. as explanatory variabls and do not enter the model dicrectly (or even u dunt know which side or u know both sides), is refered as Structural equation model (including CFA, path analysis), which are less mainstream analysis (let me say) in statistics but essential to psychologists and sociologists. A possible cookbook is
"Hatcher, L. (1994). A Step-by-Step Approach to using the SAS® System for Factor Analysis and Structural Equation Modeling. Cary, N.C.: SAS Institute."
Good luck!
pamela
To anita_jiu:
I want to use FA's factor scores as DV in the regression. But just doubt that sth which are not visible facts but generailized by me can be tested as a DV.
But how do you judge whether to rotate or not? I think there is no such clear line of telling between that?
pamela
To Msmart:
yours are somewhat encouragement for me. Yeah, i really need to better use my own literature. But sometimes you do EFA because you can finds few literature but facts are somewhat observed.
And i also use FA to reduce behavioral phenomenon instead of psychological statements or value statements. And that seems risky.
the rotated results seems very sparsely, but if not rotated they are not easy to explain..how bad it is....
anita_jiu
[quote]引用第6楼pamela于2007-03-15 15:52发表的“”:
To anita_jiu:
I want to use FA's factor scores as DV in the regression. But just doubt that sth which are not visible facts but generailized by me can be tested as a DV.
But how do you judge whether to rotate or not? I think there is no such clear line of telling between that?[/quote]
I don't think that I intended to discourage you from factor rotation; this is not what I meant. In fact, you can only obtain factor scores after the whole analysis is completed including rotation.
What I wanted to say was that if you were not sure about your subsequent analysis objective (as it wasn't shown by the time I replied your previous message/question), then there is a concern regarding which rotation method to be used. I would assume that you know two types of rotation while the selection of either needs to be made based on rationale. This also has an impact on subsequent analysis. For example, if regression is the next step, then factor analysis by an oblimin rotation method is questionable as one 'criterion' regarding regression is that DVs are assumed to be independent.
You could go back to your literature and see what others have done in your field. Sometimes you would find others may have encountered similar situations. You then could evaluate the methods/approaches adopted in a particular situation and perhaps this would be helpful to your research. Good luck.
pamela
To anita_jiu:
You don't disencourage me...hehe.... I just think not knowing enough discourage me.
Yeah, I know the rotation method.
And what confused me is just that for behaviroal elements, whether we can also use FA to reduce data. It seems it is more applied to psychological statements,etc. statements of that kind are used to "construct" sth that maybe exist or maybe not, but behaviroal statements are actually existing, i am afraid by rotating, i will lose some of them. I don't know what if I lose them...
Msmart
I dunt know whether I got your points correctly, however, according to my limit knowledge,
1. No matter what kind of data we are analyzing by EFA, we must have some kind of beliefs there are somethings in common in those varables (even we are developing a new theory, we need support from related theory or more basic disciplines). Observability or unobservability doesn't matter (then you might argue existence or nonexistence...), common or uniqueness does. So we extract commonness (factors) and drop (not lose, right? :-)) uniqueness (randomness);
2. For rotation. Rotation is just for a simpler structure and better explanation. However, there are always differrent sorts of commonness, right? Then things get easier. If we believe in equal-weight commonness, we rotate factor scores by quartimax. If we have some knowledge in hierarchical commonness, varimax is employed to make difference. If we think there are several uncorrelated commonness, orthogonal rotation is natural, otherwise we turn to oblique one.
3. EFA is not merely for dimension reduction (but PCA is more on this side...). Actually, it aims at finding out latent structure (unobservable, reasonable, interpretable, and so on. ) based on your different assumptions for this structure (equal-weight, hierarchical, uncorrelated...)
Msmart
In summary, dunt worry about that. Existence or nonexistence doesn't matter. Some kind of commonness in those variables making sence in your discipline does.
P.S., your question is quite typical for EFA or PCA. For a great EFA analyst in some discipline, to be acquaint with those EFA techniques is just the first half. Knowledge in the story itself finishes the whole picture. Go ahead, you can do it.
[quote]引用第9楼pamela于2007-03-15 17:52发表的“”:
And what confused me is just that for behaviroal elements, whether we can also use FA to reduce data. It seems it is more applied to psychological statements,etc. statements of that kind are used to "construct" sth that maybe exist or maybe not, but behaviroal statements are actually existing, i am afraid by rotating, i will lose some of them. I don't know what if I lose them...
.......[/quote]
pamela
thanks a lot.
After some clearance, i do find out that PCA is more suitable than EFA in my case. now things come to finding out how to use spss to do that.
And the book you recommend, "a primer" is too mathematical in its nature, then i gave up reading...:(.. time is pressing though..
anita_jiu
Your philosophy position should guide you to determine and justify a suitable approach of methology in your research. The question is not to do with whether or not you would lose X or Y; rather, I think PCA tells you the relative importance among some variables. You then tell the others what you have found.
Behavioural measurement and perception constructs - again, this leads me to the previous question: what philosophy position do you hold? Positivism, post-positvism, phenomenology, constructivism, or...? Well of course, this question depends on the criteria of your research. I suppose you perhaps would not need to address this question formally but it has impacts on your project explictly. This question is crutial to social science researchers (in fact, to many other researchers) who are interested in measuring people's perception, attitude, behaviour as it fundemantally affects the way how a particular research is designed, structured, and undertaken. This then results in a seriels of questions, e.g. what analysis objective(s) should be set up, what analysis techniques should be applied in order to gain the objectives (e.g. in your case, whether PCA or FA is more appropriate to use) ...
Have fun with SPSS!
Msmart
I agree with you.
[quote]引用第13楼anita_jiu于2007-03-16 18:21发表的“”:
Your philosophy position should guide you to determine and justify a suitable approach of methology in your research. The question is not to do with whether or not you would lose X or Y; rather, I think PCA tells you the relative importance among some variables. You then tell the others what you have found.
Behavioural measurement and perception constructs - again, this leads me to the previous question: what philosophy position do you hold? Positivism, post-positvism, phenomenology, constructivism, or...? Well of course, this question depends on the criteria of your research. I suppose you perhaps would not need to address this question formally but it has impacts on your project explictly. This question is crutial to social science researchers (in fact, to many other researchers) who are interested in measuring people's perception, attitude, behaviour as it fundemantally affects the way how a particular research is designed, structured, and undertaken. This then results in a seriels of questions, e.g. what analysis objective(s) should be set up, what analysis techniques should be applied in order to gain the objectives (e.g. in your case, whether PCA or FA is more appropriate to use) ...
Have fun with SPSS![/quote]