Dear Kenan,

You can calculated the mean scores of all latent constructs by adding the responses on each individual items and dividing by the number of items. Please do not forget to reverse code the items if you have any item that is reverse coded. Then you can run correlation analysis in SPSS. Common practice is to use mean score so that you can report mean and standard deviation of your variables. But before proceeding for correlation analysis, you need to assess the normality of data. If your data is normally distributed, then only you can do correlation analysis (Pearson's correlation), otherwise you have to use non-parametric tests.

If you're just looking for association, factor by factor, you can go with a nonparametric correlation measure (Spearman's rho) & p-values.

If you're looking for prediction, one factor by another, you should use gamma proportional reduction in error (http://web.csulb.edu/~msaintg/ppa696/696bivar.htm).

A multivariate analysis is possible whenever you will have enough observations per factor (limit varies with method/approach).

Your can use Principal Component Factoring. This is a data reduction method and it is the best method to reduce data especially on likert scaling.

Use median and mode, and for curve fitting u shouls use anova or ordinal models.

Dear All, thank you very much for your valuable answers.

I will put here some dummy data, just to clear to my self what exactly I think I should do, so that you can point me regarding your answers, whats exactly should I do. It is much easier for me to understand answers if we use some example (even with dummy data).

OK, so I will make it simple:

Lets say I have just  2 factors:

F1: e-marketing orientation of company

F2: marketing performance of company

Factor1 (measured by likert scale with 4 questions)

Respondent;Q1;Q2;Q3;Q4

1;3;4;4;5

2;2;4;4;5

3;3;4;2;5

...

x;4;4;3;5

Factor2 (measured by likert scale with 4 questions)

Respondent;Q1;Q2;Q3;Q4

1;4;4;4;4

2;5;4;4;5

3;5;4;3;5

...

x;4;4;3;5

So, how my final data for testing correlation between those 2 factors (does higher Factor 1 leads to higher Factor 2) should looks like?

Should I calculate mean score like:

Factor1 (measured by likert scale with 4 questions)

Respondent;Q1;Q2;Q3;Q4;MeanScoreF1

1;3;4;4;5;(3+4+4+5)/4

2;2;4;4;5;(2+4+4+5)/4

3;3;4;2;5;(3+4+2+5)/4

...

x;4;4;3;5;(4+4+3+5)/4

same for Factor2 and than test correlation between MeanScoreF1 and MeanScoreF2?

OR

to use MedianScoreF1 vs MedianScoreF2 (calculate Median for every row and not mean, as Likert scale are ordinal)

OR

to use ModScoreF1 vs ModScoreF2 (calculate Mod for every row and not mean)

OR

to use SumOfItemsF1 vs SumOfItemsF2

where SumOfItems is:

Factor1 (measured by likert scale with 4 questions)

Respondent;Q1;Q2;Q3;Q4;SumOfItemsF1

1;3;4;4;5;(3+4+4+5)

2;2;4;4;5;(2+4+4+5)

3;3;4;2;5;(3+4+2+5)

...

x;4;4;3;5;(4+4+3+5)

Thank you very much.

Gamma statistics will fit to your case scenario. 

http://www.harding.edu/sbreezeel/460%20files/statbook/chapter14.pdf

With four questions in a factor, it is not accurate to determine median or mod ...  If all your factors have 4 questions, the mean and the sum will lead to the same conclusion, if not, the sum privilege the factors with more questions ...

On the other hand, I totaly agree with Frank B. K. Twenefour for the using of PCA ... every question can be considered as a "sub-factor" and eventually the questions of the same factor wil finished in the same region of PCA plan ... moreover, the position of the factors one toward other will give you a clear idea about their correlation ... 

Let's look at this PCA presentaion as an example ... points represents the questions, optimally, the questions of the same factors are represented by close points ... correlation coefficient between two factors F1 and F2 is approximately equals to the cosine of the angle F1OF2 where O is the origin of axes .. so in this example, F3 and F4 are highly and positively correlated (cos 0 = 1), F2 is highly but negatively correlated to F3 and F4 (cos 180° = -1) and finally F1 is not correlated to the other factors (cos 90° = 0)

Dear colleagues, thank you very much for your valuable inputs.

Samer: PCA is something that I will do before any inferential statistics, for scale purification (even if most scales I used are already tested, I will test to validate everything). My problem is not "what should I put in which scale" but how to test relationship between 2 constructs that are measured with 2 different likert scales.

After a lot of additional reading and viewing some SPSS tutorials on youtube, I think Arjun and Sandro have gave me best directions how to go.

Sandro, thank you for providing a link to that textbook, it seems a very good reading.

Sandro and Arjun: I have 2 more questions: Gamma seems to be a solution "when working with ordinal level data that is ranked in a small number of response categories (ex. High, Medium, Low). Can you tell me is this means that I should transform my responses to fit into one of 3 categories, or I can use 5 categories (as my liker scale is 5-point).

And my second question, if I have 2 constructs, C1 is measured with likert scale (lets assume I will use mean score of scale and that we have normal distribution), and other construct C2 is budget expressed as percentage of total budget (so it is ratio data).

So, if I need to test relationship between C2 (ratio data), and C1 (mean score, normal distribution), what kind of statistics should I use? I want to know does higher C2 (ratio data) leads to higher C1 (mean score of likert scale).

Thank you very much.

 Transform C2 into categorical variable 1 for 0 to 0.2, 2 for 0.2 to 0.4, 3 for 0.4 to 0.6, 4 for 0.6 to 0.8 and 5 for 0.8 to 1 and you will have the same type of variable as C1, than use chi-2 test