Hello everyone,
First question:
I have 4 instruments for example reading, writing, spelling, and phonology. In each instrument, there are 5 to 6 sub-questions related to the instruments. If I want to prove that all sub-questions are the predictor to determine the factor influence the result of instruments, can I use multiple regression or perhaps hierarchical regression?
Second question:
Let say I used multiple regression and when I entered all sub-questions, the result of Adjusted R-squared give me 1.000 with the sig in ANOVA give me a 'nothing empty blank' number. I do notice that this is the unacceptable result.
However, when I used hierarchical regression, I can see clearly that the Adjusted R-squared are increasing towards the incremental of the sub-question entered which in the end is 1.000. All the sig in Anova is 0.000. Therefore, with all the respect I need to ask for an expert opinion on what should I do? Should I delete the sub-questions?
Thanks in advanced.
Dear Siti,
I'm not quite sure what your dependent variable is. It sounds like you might need to run a reliability analysis on each of your groups of questions in turn. We recommend using Principal Component Analysis for this. Once you have validated your scales you can then decide what kind of analysis to do with them.
Dear Peter,
My dependent variables are:
1. Reading performance for reading instrument with 5 sub-questions
2. Phonology performance for phonology instrument with 5 sub-questions
3. Spelling performance for spelling instrument with 6 sub-questions
4. Writing performance for writing instrument with 4 sub-questions
If I understand correctly, you have 20 questions and you believe that these measure four constructs, and you would like to see if the pattern of the associations for the 20 questions is consistent with your hypothesized four constructs. Latent variable and principle component methods seem applicable. Both have variants depending on measurement level of the 20 questions and the 4 constructs (and taxometric methods may help if you are not sure about the metric of the 4 constructs).
Have I understood this correctly?
Hi Siti, what did you do?
Did you add the individual items per concept to an overal sumscore per concept and did you perform a regression on the sumscore with the original items? In that case you will most of the time get as a result an r-square of 1 or near 1.
It sounds to me you should need to perform an PCA, IRT or factoranalysis. PCA
if you think the items 'cause' your concepts, but that would be rather strange in most reseach fields.
If you think that the concepts (like "reading performance") 'cause' the answers on the 5 items that belong with this concept, you might want to use IRT or factoranalysis.
A quick, but not completely appropriate would be to test each concept and its items using a ML estimation with a one factor model. The setup in SPSS would look like:
FACTOR
/VARIABLES VAR01 to VAR05
/MISSING LISTWISE
/ANALYSIS VAR01 to VAR05
/PRINT INITIAL EXTRACTION
/CRITERIA FACTORS(1) ITERATE(25)
/EXTRACTION ML
/ROTATION NOROTATE.
With only one factor rotation is no option. This setup tests if a one factor model fits the data. Most of the time this does not fit very well. In this case it might be because you use an interval type analyses for what I think to be ordinal data. You could perform confirmatory ordinal one factor analysis, but that can't be done in SPSS.
Siti,
Your first question refers to the assessment of the validity of your instruments. You have two options here to do so: 1- exploratory factor analysis and 2) confirmatory factor analysis (CFA). Since you know (based on theories or whatever) that which sub-question are assumed to measure a given instrument, therefore you must use a CFA method, but not EFA. In CFA, you need to fit a model in which the sub-questions (i.e., observed variables) are assumed to measure a latent construct.
To my knowledge, SPSS does not run CFA, but LISREL, AMOS, STATA, R, ... are able to do that. On the basis of your results, you need to assess the factor loadings as well as fitness statistics (e.g. Chi-square statistics).
Hope this helps.
Nader
Your problem with using multiple regression may be that you are including too many variables. If the overall score is determined exactly from the subscores, then you need to leave one of the subscores out. For example, if you have a dummy variable for gender, you include only one variable that takes a value of 1 or 0, not two variables with one for male and one for female. However, it sounds as if your dependent variable may be defined in terms of the subscores; in this case, they are related because that is how you have defined them; your measure of the dependent variable should be separate from the measures of the independent variables. Otherwise, all you are doing is recomputing the formula you created in the first place.
Thank you for your suggestions, however, currently, my sample is quite limited to only 30 samples. Therefore I am a bit worried to use PCA cause as far as I know the minimum number to used PCA must be 100 over.Please, kindly advice.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS