Merge categories or drop one? On seeing the following table, I concluded there is a lot of sparseness in the 'very unsafe' category. My first reaction was to merge it with the 'unsafe' category, leaving me with 3 categories. But another option is to just delete the category. What are the considerations to make before choosing one option? Also, I am wondering if it is ok to keep 2 degrees of safe, and only 1 degree of unsafe. It would, to me, not make much sense when discussing the odds. 

edit in response to Nick Cox:
I assume I should remove the category since the conclusions seem out of line with theoretical expectations. For example, it is assumed that those living in more urban areas experience greater levels of fear. And indeed, the odds of feeling very safe (vs safe) are:
countryside > small city > big city. The odds of feeling safe (vs unsafe) have the same order. But then for unsafe (vs very unsafe) it becomes countryside > big city > small city, which is out of line with theory and the former trend.
I could say that, although urbanisation brings with greater feelings of unsafety, these feeling take on less severe degrees when initiated urbanisation proceeds to higher levels. But of course, I don't want to report such conclusions when these are due to data sparseness rather than social realities. 
 A: This is as much social science or psychology (insert your own field if different) methodology as statistics, but 


*

*The fact that "very unsafe" values are rare is part of the information in your data. As it was presented to the people taking the questionnaire as a possible category, keeping it as a separate category is ideal. It's as interesting as "Yes, the idea of going to Mars knowing that I will die there does appeal" might be as an answer, regardless of the presumably very small number of people who will agree. 

*Merging "very unsafe" and "unsafe" might be forced upon you if whatever ordinal response models you are fitting don't converge or turn out to have very unstable parameters. If your analysis is simpler or more descriptive having some percents small is not fatal, just part of what you have found. Here there is only an allusion to odds, but calculating odds (and related measures) from both fine and coarser versions of the data might make sense. 

*Dropping the "very unsafe" responses -- which you seem to be implying is a possibility -- strikes me as poor social science and poor statistics too. There are few  good reasons for ignoring data, but they do include deciding certain responses are all invalid or all irrelevant to your research problem. 
