Hello Agathe,

Yes, it can be frustrating when an intended model does not seem to be corroborated by your data set.

What could the causes be?

1. Poor indicator variables that should be modified or replaced;

2. An atypical or unmotivated sample;

3. A sample that is too small relative to the total number of variables (and the number of unique relationships among them);

4. Some combination of the above.

In order to make recommendations, it would be helpful to know:

1. What specific research question/s you are trying to address via your thesis research;

2. What variables you are using, how each is quantified, and how data were collected;

3. Whether the same or similar variables had previously been shown to be well-represented by some specific factor structure;

4. If no prior evidence (as in #3) exists, then are your indicator variables and/or measures ones that have been researched for relationships in the past, or whether your work is the first time these have ever been used;

5. What you've tried thus far, and what results you've obtained.

Good luck with your work.

David Morse ,thank you for answering me!

I am writing a thesis about CSR in consumers decision making when buying plant based alternatives. I have done a survey for my research and the variables in the factor analysis are Likert scale/continuos. Alot the questions that passed the factor analysis does fit my research purpose and highlights responsible consumers etc. However, the opposite variables measuring the respondents who dont value CSR in decision making all got excluded. They showed negative value in the FA and measured opposites. This ofc, makes sense. But since it wont correlate with the other questions cronbachs is negative etc. Therefore I cant keep them for some of my hypotheses regarding this dimension of the research.

Not having an alpha of a certain value is neither good nor bad. It just describes one aspect of your data. Trying to alter things to make things appear not as they are is questionable (at best).

Cronbach's alpha is appropriate for a very specific model (all the loadings being the same with only a single dimension). Do you reason to believe this? Also, since this sometimes happens, do any of your items require reverse coding?