Using machine-assisted topic analysis to expedite thematic analysis of free-text data: Exemplar investigation of factors influencing health behaviours and wellbeing during the COVID-19 pandemic

Abstract

Objectives

Investigate the use of machine learning to expedite thematic analysis of qualitative data concerning factors that influenced health behaviours and wellbeing during the COVID-19 pandemic.

Design

Qualitative investigation using Machine-Assisted Topic Analysis (MATA) of free-text data collected from a prospective cohort.

Methods

Free-text survey data (2177 responses from 762 participants) of influences on health behaviours and wellbeing were collected among UK participants recruited online, using Qualtrics at 3, 6, 12 and 24 months after the COVID-19 pandemic started. MATA, which employs structural topic modelling (STM), was used (in R) to discern latent topics within the responses. Two researchers independently labelled topics and collaboratively organized them into themes, with ‘sense checking’ from two additional researchers. Plots and rankings were generated, showing change in topic prevalence by time. Total researcher time to complete analysis was collated.

Results

Fifteen STM-generated topics were labelled and integrated into six themes: the influences of and impacts on (1) health behaviours, (2) physical health (3) mood and (4) how these interacted, partly moderated by (5) external influences of control and (6) reflections on wellbeing and personal growth. Topic prevalence varied meaningfully over time, aligning with changes in the pandemic context. Themes were generated (excluding write-up) with 20 h combined researcher time.

Conclusions

MATA shows promise as a resource-saving method for thematic analysis of large qualitative datasets whilst maintaining researcher control and insight. Findings show the interconnection between health behaviours, physical health and wellbeing over the pandemic, and the influence of control and reflective processes.