Large Language Models (LLMs) are crucial for qualitative analysis because they offer automation and interpretive information. Also, the computation time for LLM is much shorter than that of software-assisted manual qualitative analysis. However, LLM hallucinations can lead to misleading or incorrect outputs that pose a significant challenge to reliability and accuracy. This study identified and examined the root causes of 12 types of hallucinations in LLM-based qualitative analysis. To mitigate these hallucinations, a systematic system prompts refinement, spurious noise filtering, and controlled batch processing of transcripts were adopted to optimize and enhance the reliability and precision of LLM-based qualitative research results.