Swapping Nationality and Language in Personas Fails for Multilingual Mental Health Datasets

Generating multilingual mental health dialogue datasets by modifying nationality and language parameters in English-centric clinical personas produces datasets where LLMs show inconsistent depression severity evaluations across Mandarin, Bengali, and Hindi versus the English baseline. Simple persona localization is insufficient for culturally valid multilingual clinical data.