Prompting strategies for coding in AI-assisted qualitative data analysis

Introduction: Coding is a central step in many qualitative data analyses (QDA). Artificial intelligence (AI) has made inroads into the field with performance dependent on the prompting strategy used. This work examines the effectiveness of different prompting strategies on AI performance. Methods: Coding and theme analyses were performed on abstracts from a 16-study meta-analysis examining the association between air pollution and preeclampsia incidence. These were also performed in a second related corpus containing fewer formal sources. Coding was performed with R, NVivo 14, and ChatGPT 4.1. Several prompting strategies in increasing degree of complexity (zero-shot, few-shots, and chain-of-thought – CoT) were evaluated regarding their accuracy against manual coding. Direct thematic coding with AI was also examined. Results: Zero-shot prompting (with or without CoT) provided the best results for structured text. Requesting a full word-by-abstract matrix increases sensitivity. Supplementing CoT with general topic information and role playing made it the best strategy for more colloquial text. Directly inquiring thematic coding within AI showed good performance. Some prompting strategies are more prone to hallucinations. The latter are associated with errors in the frequency counting process while coding. Discussion: This work offers performance assessment of an AI tool for coding and theme extraction through a variety of prompting strategies combined with head-to-head comparison with popular QDA tools using readily available data. During user interaction, AI chatbots provide suggestions regarding potential next steps which may enhance results during coding. AI tools may require replication to ensure consistent findings and mitigate their probabilistic nature. Integrating results from at least two tools (AI and traditional) provides more comprehensive and accessible results for QDA.