Hi,
I have a survey dataset (Likert scale) which contains some No Opinion answers. How should I deal with them in my dataset? Estimation, omission, converting to midpoint (0) or ...?
What is the acceptable threshold for them in relation with the whole number of valid data in the dataset?
Mentioning scientific references would be appreciated.
This looks like a homework assignment.
Have you looked at your textbook first?
It's not a homework and I did search the available resources, including searching in the internet, and could not find a comprehensive, reliable answer.
Because... there is no simple answer for this question. It depends on many factors, like: what is percentage of NA/DK answers, what are you going to do with data (e.g. how are you going to analyze them - what technique), what was original question, etc. There are a lot of methodological papers about dealing with this answers (though I don't remember any particular ones right now, sorry... Just google for them). Summarizing, without any further details it's difficult to give more adequate answer.
Thanks Marcin,
I am going to do CFA and then SEM with the data. I have 37 different measures with 410 respondents. In total, around 10% are No Opinion answers. Out of 37 measures, around 10 are the ones with the most No Opinion answers.
In this case, I suggest you to be... empirical. I mean try do two models - one with nonresponses taken as missing values (e.g. exclude them from analyze) and second with nonresponses recoded as a middle position on scales. Then just compare results. Based on my experience I assume no big diffrences in what you will get, but empirical test will be the best proof.
However, reading you previous answer about proportion of nonresponses (e.g. 10% of your sample) I will rather include them as a middle position, Think about like this: exluding them will be loss of 10% of your study sample.
@Marcin
Good idea. Should I in the first model exclude the cases with missing data or just run the model with missing values included but not replaced? This will affect the sample number of each model!
Dont know is a response. Usually it is valuable to describe those individuals who answer dont know. What are their characteristics? It will also depend on their proportion to impact your results. Usually dont knows and also missings are not at ramdom. If you are excluding them in the maim model, you have to describe who are those. I just wrote a paper where we have missings and dont know because mother could not remember. So I described who were those mothers. The first table I presented was a comparison of characteristics of dont´know and those who answer the question. No statistical significance was observed, but some tendence of more don´t knows among poorest moms. Therefore I discussed all possible iinfluence of that in my final results. Just describe and be honest about results. More sofisticated work you could impute data, not only for missings but for dont knows and see the impact your final results. Giving all information to whom will read will anable the reader to make their own judjment.
The first step in any useful data analysis, particularly one that aspires to use fairly sophisticated analytical techniques like SEM which are fairly dependent on assumptions about the distributions of the data, is to really understand your variables and what they mean - not just take somebody else's word for it. A careful examination on a variable by variable basis is likely to reveal that there are at least two different kinds of "no opinion" variables: (a) those in which the respondent might actually not know anything about the issue in question, and (b) those in which the respondent might know the relevant information but simply have no preference with regards to the measurement scale. There is actually a third category in which the respondent might know the information and have a preference but simply choose not to share it with you, but this is probably less common than the first two. If you can distinguish between the variables that might be giving rise to these different kinds of responses, you understand that you have to treat them differently.
I'm attaching
a link to an article that might help disentangle these different kinds of non-responses. There is no one-size-fits-all model for dealing with nonresponse.
As a side note, you might want to consider using a method like partial least squares rather than CFA/SEM. It provides much the same kind of modeling approach, but requires far fewer assumptions about your data and offers some more precise modeling techniques. If you haven't considered it, it's worth looking into. I'm also attaching a link on this question.
http://orm.sagepub.com/content/10/2/195.short
Article The Partial Least Squares Approach to Structural Equation Modeling
- Badges
- Science topic