Very nice, The initial analysis (AM_1.out) tests the strict linear hypothesis: 1-->1, 2-->2, etc. The second analysis (FA_AM.docx) looks for non-linear effects and finds local regression along the domain of the attributes.

Very nice, The initial analysis (AM_1.out) tests the strict linear hypothesis: 1-->1, 2-->2, etc. The second analysis (FA_AM.docx) looks for non-linear effects and finds local regression along the domain of the attributes.

I just looked at the output again. The TABLE command in UniODA was used. This is a utility function that makes analysis of square contingency tables close to effortless. However, the data must be categorical.

In this application the data are ordered: the scale of concordance has four values ranging from 1 (strongly disagree) to 4 (strongly agree).

Because the data are ordered, the TABLE command can't be used.

Don't use the TABLE functionality. Instead, input the data into a UniODA program, and omit the CAT command (the data are not categorical). In the first analysis the same DIRECTIONAL hypothesis tested in the TABLE command analysis is used to assess perfect consensus. In the second analysis the correctly-classified observations from the first analysis are eliminated, and a non-directional hypothesis will be evaluated. Everything will likely turn out the same (in analysis and write-up) as occurred for the categorical analysis, with the possible exception of p-values (which would be assessed for ordered data, not categorical data).

The ODA book presents discussions and examples of multicategorical designs involving ordered attributes, and the ODA Journal (http://ODAJournal.com) also has articles discussing new exact maximum-accuracy range tests for decomposing statistically significant multicategorical bivariate effects.

1) The variable FA was obtined by multiple responses  (1 to 4) from  FA1 to FA5. The table has the number of responses to both variables

2) So in free format the FA lines goes from 4 to 20 (count responses ). And if i treat them as continuous I think I can only make non-directional hypotheses!.

3) But if I analyse each of the 5 variables FA and 6 variables MA, i can make directional hypothesis. 

4) what do you think is the best way to analyse data by ODA?

5) The relationship between responses in both composite variables  it is not guaranteed that   is  ordinal.

For example, when analyzing the relationship between the number of car accidents and age, assuming that both variables are classified as ordinal, it is not guaranteed that the relationship between both obey an order, because  we know that  the younger and the older ones have the greatest number of accidents. Thus, the relationship between those two variables would be analyzed as if it were  nominal variables (= categorical according ODA)

6) if  input data in free format, how can I know what are the correct classifications? if data was in table format it is easy to see that, but in free format i really do not know!

Sorry, I don't understand clearly what you are asking, and I have many project deadlines to meet. 

Regarding (2), in the ODA paradigm variables are only categorical or ordered. The concept of "continuous" is irrelevant.

Regarding (3) when a researcher is first learning to use ODA methods and software I recommend running each analysis individually, and continuing each separate analysis until it is completed correctly, or until the reason underlying the failure is clear. If the analysis does not run, but the UniODA program is correct, then the problem resides in the data. Usually the latter case is the result of having too few or too many class categories given the metric of the attribute. If you come across a specific analysis that fails to run, and there is nothing wrong with the data or the program, then posting a question (and including the program and data) is a good idea.

Regarding (4), there are many articles on this topic, and many relevant discussions in the ODA book. The search box on the ODA webpage may be useful for finding specific topics. Of course, a solid understanding of UniODA facilitates clarity in understanding of more complex ODA methods, so time spent mastering UniODA is an excellent investment.

Regarding (5), scores of any unidimensional construct measured using a scale ranging from opposite poles (e.g., worst to best) are, by definition, ordered. Here is a brief, open access paper on incorrect treatment of ordered data:

http://optimalprediction.com/files/pdf/V1A12.pdf

Regarding (6), if I understand what you are asking, I use SAS to manipulate data sets. So, if I wish to explore sample strata that were identified by ODA models, then I use SAS. There are plenty of software systems available for manipulating and examining subpartitions of samples. ODA (is the only) "mission-specific" software that is specifically designed to identify maximum-accuracy models.

PS:

Regarding (2), directional hypotheses are stated in terms of class categories, not in terms of attribute metric, if that was your meaning.

In terms of attribute metric the default is ordered. If some attributes are (multi)categorical then they are specified using the CAT command. If all of the attributes are (multi)categorical that can also be specified using the CAT ON command. Examples of these and other analyses are presented in the ODA book and eJournal.

If I uderstand your question very well, then I think Principal Component Factoring  can be used on your data to extract latent factors and then you can go ahead and perform regression analysis on them.

Dear Frank, 

Because of violation of normality and because we have only a scale from 1 to 4, we should not aplly PCF...that is why I should aplly optimal data analysis (ODA) .

But i apreciate your interest in this question. Thank you very much in the same way.

Helena