SPSS Statistics

Hello Everyone and thank you for your help in advance.  I am helping a friend with her graduate degree research.  As part of her research project she administered a Likert scale questionnaire of 15 questions to 74 subjects.  The data are self reported.  The Likert scale results are categorical and ordinal and consequently require non-parametric stats (?), unless the data are transformed?  She is hoping to find a positive effect on subject learning due to a training program she developed.  The main data are the results of the 15 question survey on the subject's perceptions of their training experience, eg.  "the training program enhanced my learning".  

My friend cites a paper in her thesis draft that has a similar experimental design.  The authors of that paper analyzed their data with a series of chi-square tests.  They provide little detail on how they conducted their analyses and did not have any regard for an experiment-wide alpha level.  This study had about the same # of subjects and questions in a Likert scale format. 

 So my main question is, given this brief summary, is a chi-square or G-test analysis is the best way to proceed with analyzing this study? 

If I proceed with a chi-square analysis, I plan to do a chi-sq test for each of the 15 questions where the null hypothesis is 14.8 (15) observations (subjects) choosing each of the 5 categorical answers.  The total # of subjects is 74, there are 5 answer options, 74/5=14.8.  The editors of the above journal didn't worry about the # of tests overall in the study, so I probably shouldn't either?  In general, if I do 15 separate chi-sq tests my experiment-wise alpha level would be 0.05/15=0.0033?

Also would a Spearman correlation analysis be a viable option to be done instead of or in addition to the chi-sq tests?  I'm not sure about this but, would a correlation analysis compare the subject's answers to question 1 to their answers in questions 2-15.  The analysis would be based on 74 answer values to q1 compared to 74 answer values to  q2, 74 answers to q3, etc. through the 15 questions of the survey?  The results would be a correlation matrix, with q1-q15 across the x axis and q1-q15 down the Y, with statistic values between -1 and +1, indicating whether there was any relationship between how the aggregate subject's answered q1 and q2, for example.  Any significant correlations could then be interpreted, for example subjects that tended to strongly agree with question 12 also strongly agreed with question 14 (if there is a significant, positive r value).

Lastly we would like to compare this survey data to a 5 question, quiz given to subjects following the training. One of these questions is a true/false question.  The other 4 are multiple choice questions.  So there are correct and incorrect responses for 5 questions for the subjects and an overall score for the # of correct answers for a subject.  Unfortunately there is no way to identify individual subject data from the survey analysis and the multiple choice post training quiz. Also there were 74 subjects in the survey analysis and only 40 subjects that took the post test.  So far, we are thinking that the survey data and the post-test data will have to be analyzed separately.  If anyone has any ideas of how to compare the survey data results and the post test results in a statistical analysis I would like to hear them.  One, speculative idea I had was to treat the results of the 5 question multiple choice quiz (the post test answers) as 5 more questions to add to the correlation matrix that I am proposing with the survey data. I could also add a variable that includes the overall # correct on the post test.  Now my correlation matrix would be 21 variables by 21 variables?, with the last 6 not planned to be included from the start.  This last idea may be over the line and unnecessary.  I suppose the best result it could show is that the subject's responses to one of survey questions, a question with a strongly agree 5 response, was positively correlated with the overall post test score.  Subjects that strongly agreed that the training program was helpful tended to have a higher score on the post test (?) 

Thank you and please let me know where I have made any mistakes in my thinking.  I am happy to answer any questions.    

Hello, mike

There is much debate about the proper qualities of the Likert scale. Non the less there is a significant body of literature that specifically emphasizes that if a Likert type variable distribution has normal distribution properties (e.g. central tendency, dispersion, skewness, and kurtosis), it can be analyzed with parametric tests such as T-tests and ANOVAs.

in my experience, most parametric tests are even robust enough to overcome normality issues if they are modest.

Hello Everyone and thank you for your help in advance.  I am helping a friend with her graduate degree research.  As part of her research project she administered a Likert scale questionnaire of 15 questions to 74 subjects.  The data are self reported.  The Likert scale results are categorical and ordinal and consequently require non-parametric stats (?), unless the data are transformed?  She is hoping to find a positive effect on subject learning due to a training program she developed.  The main data are the results of the 15 question survey on the subject's perceptions of their training experience, eg.  "the training program enhanced my learning".  

My friend cites a paper in her thesis draft that has a similar experimental design.  The authors of that paper analyzed their data with a series of chi-square tests.  They provide little detail on how they conducted their analyses and did not have any regard for an experiment-wide alpha level.  This study had about the same # of subjects and questions in a Likert scale format. 

 So my main question is, given this brief summary, is a chi-square or G-test analysis is the best way to proceed with analyzing this study? 

If I proceed with a chi-square analysis, I plan to do a chi-sq test for each of the 15 questions where the null hypothesis is 14.8 (15) observations (subjects) choosing each of the 5 categorical answers.  The total # of subjects is 74, there are 5 answer options, 74/5=14.8.  The editors of the above journal didn't worry about the # of tests overall in the study, so I probably shouldn't either?  In general, if I do 15 separate chi-sq tests my experiment-wise alpha level would be 0.05/15=0.0033?

Also would a Spearman correlation analysis be a viable option to be done instead of or in addition to the chi-sq tests?  I'm not sure about this but, would a correlation analysis compare the subject's answers to question 1 to their answers in questions 2-15.  The analysis would be based on 74 answer values to q1 compared to 74 answer values to  q2, 74 answers to q3, etc. through the 15 questions of the survey?  The results would be a correlation matrix, with q1-q15 across the x axis and q1-q15 down the Y, with statistic values between -1 and +1, indicating whether there was any relationship between how the aggregate subject's answered q1 and q2, for example.  Any significant correlations could then be interpreted, for example subjects that tended to strongly agree with question 12 also strongly agreed with question 14 (if there is a significant, positive r value).

Lastly we would like to compare this survey data to a 5 question, quiz given to subjects following the training. One of these questions is a true/false question.  The other 4 are multiple choice questions.  So there are correct and incorrect responses for 5 questions for the subjects and an overall score for the # of correct answers for a subject.  Unfortunately there is no way to identify individual subject data from the survey analysis and the multiple choice post training quiz. Also there were 74 subjects in the survey analysis and only 40 subjects that took the post test.  So far, we are thinking that the survey data and the post-test data will have to be analyzed separately.  If anyone has any ideas of how to compare the survey data results and the post test results in a statistical analysis I would like to hear them.  One, speculative idea I had was to treat the results of the 5 question multiple choice quiz (the post test answers) as 5 more questions to add to the correlation matrix that I am proposing with the survey data. I could also add a variable that includes the overall # correct on the post test.  Now my correlation matrix would be 21 variables by 21 variables?, with the last 6 not planned to be included from the start.  This last idea may be over the line and unnecessary.  I suppose the best result it could show is that the subject's responses to one of survey questions, a question with a strongly agree 5 response, was positively correlated with the overall post test score.  Subjects that strongly agreed that the training program was helpful tended to have a higher score on the post test (?) 

Thank you and please let me know where I have made any mistakes in my thinking.  I am happy to answer any questions.