Good afternoon,
During the data quality stage, if your dataset includes only categorical variables and as a result, you cannot use imputation methods, to which degree do you judge that the quality of your data is bad 5%, 10%, or more? I think it depends on your research project, research question, and models performed but if you can add the rate of missing values of your study, it can help me to have an idea of this percentage.
Thanks for your help,
Ines Francois
Hello Inès François ,
I presume that you're talking about instances other than ignorable missing values. (An example of this would be a skip logic survey, such as: Q1: Are you married? If not, skip to question #4. Anyone answering "No" to Q1 shouldn't have responded to Q2 or Q3.)
If the sheer amount of missing data doesn't exceed 5%, you likely don't have to worry much about working with only complete cases. Some sources say that up to 10% can be tolerated.
However, it is important to check whether missingness on a variable is in any way related to other characteristics of a case, which could include demographic attributes, and/or the way a case responded to some other item. If it is, then you must be more cautious about how to proceed.
You can impute responses for categorical variables; it's just more easily done for continuous variables.
Finally, do note that a lot of times, cases with "no response" or "refusal to respond" are treated as if they had elected an otherwise viable option (hence, a three-category item could functionally be treated as a four-category item, with missing responses assigned to the fourth category). This is because numerous studies have shown that sometimes, these folks _are_ different in one or more important ways from those that did respond.
Good luck with your work.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis