Generally speaking, what would be the idea behind dichotomizing ordinal, categorical variables for binary logistic regression. To be clear, I'm not talking about the dependent variable but only explanatory variables or predictors. As an example, let's say someone were interested in how interest in politics or just any kind of politica attitudes affects the likelihood of casting a ballot (1;0). Let's further assume that the data basis for such an analysis would be survey data and that the explanatory variables would be ordinal variables that are coded at least on a 5-point scale (so that you could theoretically also treat them as 'metric').
Though my first instinct would be to use the whole scale when adding this variable to the equation, I have seen many studies doing it differently and dichotomizing variables like that: for instance, "strongly agree" and "agree" are lumped together and coded
1 and the rest goes
Does this have to do with getting a more robust measure given that if you lump categories together you avoid dealing with very low cell frequencies (though I cannot imagine that this would be the reason) or does it have to do with the link function and the fact that the outcome variable is also dichotomous? I'm not sure if that is really worth accepting that you lose information if you dichotomize. And furtheremore, I'm not sure if by linearizing (is that a word?) the effect, which you implicitly do in this process, would do justice to the link function and the way the model is estimated.
So ... what would be the reasoning? And is there a gold standard?
EDIT: I'm aware of @Tom's question from 2013 (What is the benefit of breaking up a continuous predictor variable?), but in contrast to his post, I'm exclusively interested in binary logistic models and the practice of dichotomizing ordinal variables (because it seems to be related to logistic regression models and the nature of the outcome variable).