Whether it is okay to dichotomise a 4-point Likert-scale outcome item I'm using an already collected data-set within my research, meaning I was very limited with the outcomes that I could use. One of my two main outcome variables was an item using a 4-point Likert-scale (‘How often do you feel unhappy at school’ (1=all of the time to 4=never)). My supervisor previously asked me to dichotomise this variable, combining the response options ‘all of the time’ and ‘most of the time’, creating ‘high negative affect’, and combining ‘some of the time’ and ‘never’, creating ‘low negative affect ’. Though, I am having huge difficulties justifying in my paper why I dichotomised this variable. It may because only a small proportion of the participants had answered some of the response options e.g. 315/11,717 had answered one of the response options. Though, I'm not sure if this is enough reason to dichotomise it. I would really appreciate any support if anyone knows other reasons that she may have suggested to dichotomise it, or ways that I can justify this decision in my paper (my supervisor is on holiday and so I can't ask them).
Thank you so much all for your support in this, I really do appreciate it! I am just adding a bit of other information as some were asking for more details regarding why my supervisor may have asked me to do this. Judging from my notes, what she had said was that if you had a 3-level variable, you essentially calculate the odds of being at each level, as in that is what doing a linear regression would do (though mine is a 4-level variable so not sure if this also applies to this) and that it would not be ideal, and so making the variable binary would be a better option. Even at the time, she had said we would need to justify it conceptually and statistically and that perhaps I could try and find something about the dimensionality vs non-dimensionality of this construct, but she never really explained further than this.
 A: You're not violating your codes by saying 1 and 2 are not happy and 3 and 4 are happy. Maybe the decision to recode is based on prior evidence that suggest that anything more than a dichotomous response is meaningless, similar to what @BruceET said in his comment.
That said, recoding your outcome variable to be a dichotomous indicator for happiness at school would probably not be recommended in the general world of statistics. This site reflects that attitude (the types of answers you'll typically get on this site are about doing things as correctly as possible, which is very useful). But, an MA thesis is more about doing things "good enough". So, the answer might be more about the nature of academia than about statistics. I'm in my MA too, and my supervisor wants me to do things the simplest way possible while still being methodologically defensible, mostly because he knows that this approach sets up students well for their oral defense – they can give good answers based on a deep understanding of their simpler methodologies.
