comparing two items of a questionnaire I have developed a questionnaire that assesses the reasons why adolescents do not adhere to their asthma treatment. 
I have an item that asks 'How often are you supposed to take your asthma medicine?' and one that asks 'How often do you actually take yuor asthma medicine?' 
Both are assessed on a Likert scale (1= daily; 2= 5x a week; 3= 3x a week; 4= less frequent)
I would like to compare the answers in order to see if adolescents actually take their medicine as they are supposed to. Should I just look at the frequencies and compare them or is their a statistical test that I could conduct?
 A: I think what I would do is create a variable for each respondent which would be dichotomous: yes this person takes their medication on the correct schedule, or no they don't.  From there, you calculate the proportion of respondents that take their medication on schedule (e.g. "0.75, or 75% of respondents take their medication on schedule").  
Then, you can create a confidence interval around this proportion. Be sure to use a technique that is appropriate for determining a confidence interval for a proportion, and appropriate for your sample size.  The interpretation depends upon if this interval includes certain proportions of interest. I don't know what these would be.  It depends on your experience and opinion, but as examples:  If the interval includes 1.0, then that is good news, as the proportion for your sample is not statistically different than everyone taking their medication on schedule.  If the interval includes 0.50, then your sample is no better than 50/50 at taking their medication on schedule.  If the interval is below 0.50, then, I imagine, that is bad news.
The same type of result could be obtained with a binomial test, testing against different proportions of interest (e.g. 0.9, 0.5, 0.25), but I think the confidence interval approach is more elegant in this case.
A: I am not sure if that can help you. By default what you are asking can provide a residual between the supposed medication intake and the actual medication intake by subtracting the items pairwise. And then subject the result to a residual analysis with which I am not very familiar at the moment. Residual analyses are usually performed on correlation matrices such as root mean square error etc. You may get some ideas by studying residual analysis formulas.
A much simpler solution could be to treat your data as experimental data and subject them to a within subject design t test. If the difference between condition 1 supposed medication intake and condition 2 actual medication intake is significant. Then you have a difference from an inferential test.
within subject design t test is a very simple and effective solution.
I would be happy to see some ideas on residual analysis here as well.
