Correlation with very different sample sizes I am new to statistics (self-taught since I need it for my thesis) and I have a question that I haven't been able to find an answer to yet. For my thesis I want to find out whether there is any relationship between how students and their teachers answer a likert scale questionnaire (several items forming a scale). I want to see whether there are any similarities/if teachers influence their students thinking in certain areas. I have questionnaires from 3 teachers and their 160 students. The questionnaire I used is based on an existing one but it has never been analysed in this way. What calculation or correlational analysis would best answer my research question? 
I am very thankful for any tips or answers! 
 A: What I would do:


*

*Group the students according to the teachers and check for
correlations to each answer.

*Check if this correlates with the teachers answer. 
EDIT: 


*

*You could take the average value of the student's answers (or their sample median, or 25% quantile, ...) and compare those to their teacher's answers. Doing so, you get two datasets per class: one for the teacher and one (averaged) for the students. These you can easily correlated.

*If you want to take it one step further, you could weight the calculated correlations (from above) by the students variance. 


*Check if this correlates with the other student's answer -- a certain topic might be sexy, indep. of the teacher.


With your dataset it will be impossible to insure that your conclusions are valid for the teachers population, because you can't assume that the three teachers are representative for the whole population. However, if you have more than 20 questions per sheet, the evaluation whether or not there exists a correlation between the answers of these particular teachers and students might be possible.
EDIT: 
In my experience, one should not take all student's and teacher's answers on evaluation sheets too seriously. Questions are interpreted and the answers do strongly depend on the current mood of the person. Therefore, it might be wise to evaluate only those questions which invoked extrem answers (bad and good) in a second "interpretation" of the data.
A: In my opinion, the vast differential between the two sets of variables (3, 160) means that you really can't statistically answer comparative questions between the two samples. Full exhaustive correlation cannot be performed, due to the small sample size of the teachers. The second problem you have in re the teacher sample is that a Likert-style question, usually means that there are either seven levels (desirable) or 5 levels (acceptable), each of which should ideally have a discernable midpoint (unlike a 4-point scale). Thus,in all likelihood, not every answer that COULD be given by a teacher in the Likert item WAS in fact encountered. You therefore have no ability to answer whether the responses differed according to the introduction of an external variable. It's like asking 3 people what their favorite number is from one to ten, having results 1, 4, and 4 and declaring 4 everyone's favorite number.
