For multinomial modelling problems, should one focus more on overall prediction accuracy or conditional accuracy?

Question

Right now I am on a social science project to establish a (sort of) multinomial logit regression model on a dataset. In my model, unobserved heterogeneity of the examples is