# Best method to analyse survey results with multiple choice of answers

I have designed a survey with multiple choice answers. Each question contains same set of answers

• Strongly agree
• Agree
• Disagree
• Strongly disagree
• Don't know

There are 25 questions and questions are grouped into 5 areas. I need to analyze each area.

What is the best approach to analyze the survey results which I collected from a number of people? Is it a good method to calculate the mean response from each area as I need area-wise analysis of questions? Can any one suggest a best option? I am not a statistician

• Such a response format is typically called a Likert item (or a Likert scale but that's somewhat confusing because the whole point of Rensis Likert's work was to combine several such items to form a single scale). If you use that term to search for more info, you will see that there is a lot of material on the web and several questions on this site addressing various aspects of this type of data. – Gala May 4 '13 at 10:38
• Beyond that, you will need to give more details on your data and what you want to find out if you want specific links or advice. – Gala May 4 '13 at 10:39
• I just want to test the awareness of people about a topic. The people selected by me is a closed group. I need to analyze the results collected as part of my survey and wish to submit a report baesd on it. – Prasanth May 4 '13 at 16:41
• Obviously people are willing to help, you already got several tips that should have helped you a lot. If you want more, don't offer a bounty but ask a specific question! “Analyze the data” or “Test the awareness” is so broad as to be meaningless. What's the problem with generating some graphs or stating X% of people strongly agree with X or Y? What else do you want to learn? Without significant scale design effort, sophisticated study design (longitudinal or comparing several conditions) or perhaps some other data, such questionnaires don't allow for much more than that. – Gala May 6 '13 at 12:54

I upvoted simply because the downvote was uncalled-for.

1. Nobody can tell you how to analyze anything unless they have some idea of why you want to analyze it. That said:
2. If you just want to display the data, make a bunch of bar charts. People can see who put what, and how frequently. Done.
3. Do you want to say how two variables are related? Maybe that agreeing with one question predicts disagreeing with another? That is regression or one of its flavors. Do some basic reading. The simple stuff isn't really that hard, and excel can do a lot of really basic stuff.
4. Do you want to group people into clusters? People who think alike? That is a lot harder. PCA, kmeans, etc. Let us know what your goals are and we'll try to point you in the right direction.
• I have different categories of questions as mentioned. I want to analyse each question category separately. – Prasanth May 6 '13 at 6:09
• Then just make a bunch of bar graphs – generic_user May 6 '13 at 11:56
• @Prasanth: ACD wants you to expand on the word "analyze". What do you want to see? Give examples of the kinds of statements you'd like to be able to make. How do the categories enter into what you want to do? Techniques can be simple ("Twice as many people strongly agree with X as with Y") or complicated, and we can't really help you unless you specify what you mean by "analyze". And to be honest, you won't succeed if you don't understand what you mean by "analyse" clearly enough to explain it to us. – Wayne May 6 '13 at 12:56
• @Wayne: I want to measure the degree of understanding on a particular area based on the questions of that area. So I added questions based on each area. The response collected from a group of people should show their knowledge and understanding on that area. Likewise I wish to make an analysis on each area. I don't wish to plot graphs for each question. I want the results based on each area. – Prasanth May 6 '13 at 14:27
• Do you have some measure of the effectiveness of the "business continuity management system"? Can you call it $y$ and say that $y$ has some distribution of values? If not, you probably don't want statistics, but rather some other sort of qualitative logical argumentation, maybe bolstered by some bar graphs. – generic_user May 7 '13 at 11:01

Following up on points 3 and 4 by ACD: I would suggest that you take a look at Latent Class Analysis which handles properly your categorial data, assuming that your grouping of questions represents different latent constructs. In case you are interested in connections (e.g. correlated error terms) or even causal links between your "groups", there are also Latent Class Models which are kind of a mix between Structural Equation Models and Latent Class Analysis.

• This is not entirely unreasonable given the vagueness of the original question but it seems not to address the OP's concerns since his comments suggest that he is not interested in clustering or comparing individual participants at all. Also, the sample size is very small (40-50 people all coming from the organization that should be evaluated). How useful is LCA in this context? – Gala May 8 '13 at 9:46
• Indeed, the sample size could lead to problems with identifiability and robustness of the fit measures due to sparse data in the contingency tables. – non-numeric_argument May 8 '13 at 11:51
• I think that you would need something like an aggregation or clustering/comparing if you are actually interested in the outcome of an organiziational process but you have only collected data about the knowledge of individuals. – non-numeric_argument May 8 '13 at 12:00
• If so, what is the best method? – Prasanth May 9 '13 at 9:24
• Possibly, the best starting point to make sense of your data is an analysis using measures of tendency and dispersion (e.g. median and standard deviation) and graphs to describe the distribution of your results as ACD already recommended. Your assumption is that the aggregated answers of the individuals you questioned are representative for the level of (everyday) knowledge in your organization. In case you are not content with mere description and you are looking for certain underlying but hidden types or structures in your data, you should try cluster analysis or latent class analysis. – non-numeric_argument May 13 '13 at 8:24