Are scores significantly different to a neutral response? Please help...

#1
Hello

I am working with a likert scale with response ranges from 1-5. A score of '3' indicates a neutral response.

I want to calculate whether, for each of 20 items, subjects have responded in a way that is significantly different from a neutral response. I.e. do they indicate an opinion/preference rather than just a neutral response.

I think that to do this I want to be looking at whether the responses show a significant deviation from the neutral midpoint?

I don't know what test to use to do this. I thought about adding together the times people scored a 1,2,4,5 for each item and comparing it to the times people scored a 3, but I'm not sure if using the combined score of 1,2,4 and 5 would work. I want to run this in SPSS, and am thinking a Chi Square test may be appropriate?

In all honesty I am rather confused about how to do this and would be so grateful if someone could offer some advice?

Thanks a lot in advance to anyone kind enough to help :)
 
#3
Thanks so much!

Apologies for the further question: If I were to do it that way, would I just put the original variables in (i.e. insert each of the 20 items scored from 1-5) as the 'test variables'?

I ask this because I have also recoded them into 3 categories (strongly agree/agree; neutral; strongly disagree/disagree) and I wonder if this is what I should put in rather than the 5 category options: I've tried both and get slightly different results with each one.

I'm familiar with independent t-tests, ANOVAs and chi square tests, but haven't used one-way t-tests before so if you could possibly clarify it for me that's be a massive help?

Many many thanks for this...
 
Last edited:
#4
If you want to do separate analyses for each question and if you use SPSS, just create a column for each question, the rows are subjects, and use the entire scale (1-5), if you can assume that the scale is a continuous one. If it´s not a continuous scale you should not use t-tests, ANOVA, etc, but something more suitable for categorical data. If it´s ordinal then there are non-parametric versions of t-tests you might use. Just google it. Then locate one-sample t-test in SPSS, and go (put test value as 3 and throw in the variables you want to look at, each question is a variable). You would have to do a separate analysis for each variable but you can just put them all in at once.

If you want to have a summary over all questions just take the mean for each person, and test those values against 3 as above, so that you have one variable that is the mean for each person.

I see no point in recoding the data so that you lose information if it is not necessary to do so.
 
Last edited:
#6
Hello

I am working with a likert scale with response ranges from 1-5. A score of '3' indicates a neutral response.

I want to calculate whether, for each of 20 items, subjects have responded in a way that is significantly different from a neutral response. I.e. do they indicate an opinion/preference rather than just a neutral response.

I think that to do this I want to be looking at whether the responses show a significant deviation from the neutral midpoint?
You need to clarify what hypothesis you really want to test. It's not entirely clear what your research question is. The way I'm reading what you've said above, you want to test whether people respond "neutral" vs. "non-neutral." A one-sample t-test will test whether the average response differers from 3, which is neutral.

If, on a particular item half of your subjects responded "5" and the other half responded "1", you would have a mean of 3, even though none of your subjects actually responded "3".

A one-sample t-test would not reject the null hypothesis, even though not a single subject responded neutrally.

If you really want to test whether people are responding neutral or not, you need to create a new categorical variable in which anyone responding 3 gets a "neutral" and anyone responding 1,2,4, or 5 gets a "non-neutral."

I'm not sure you want a chi-square test, because what is your other variable? Do you want to compare the 20 questions? This might be what is confusing you, because it sounds like your original instinct was right.

You probably want a one-sample test of proportions, which would test if the proportion of people who choose a non-neutral response is different from a certain value, presumably chance. You'd have to decide how to define chance--it could be 4/5 if you think each value is equally likely, or it could be 1/2, if you think neutral and non-neutral are equally likely.

Karen