Modelling pain in population-based cohort studies: Biases, causal inference, and latent class methodologies

Population-based cohort studies present unique opportunities to investigate research questions in areas such as population health and well-being. Addressing challenges in the statistical modelling of pain in older adults using cohort studies is the main goal of this thesis, motivated by two large cohort studies of ageing: the Irish Longitudinal Study on Ageing (TILDA) and the American Health and Retirement Study (HRS).

Firstly, the presence of pain-related attrition bias, mortality bias, and measurement bias due to differences in reporting styles (reporting heterogeneity) in TILDA is investigated. Evidence of mortality bias and reporting heterogeneity is found. Sex and socioeconomic disparities in pain previously reported in other countries are also observed in TILDA. Next, the causal effect of pain exposure on 20-year mortality in American older adults is estimated using HRS data. Propensity score methods are applied to adjust for measured confounding bias identified using a directed acyclic graph. Results suggest that pain likely causes a modest increase in mortality hazard, though the results are also compatible with no effect. Additionally, modifiable common causes of both pain and mortality are highlighted as potential targets for intervention.

Issues around the measurement of pain are also addressed. While pain is often modelled using a single measure such as pain intensity, it is desirable to identify a more holistic measure that incorporates multiple different aspects of pain experience. Latent class analysis (LCA) could be used for this task, however LCA model selection is particularly challenging for large datasets. The adaptation of fit indices from structural equation modelling for use with LCA is proposed to aid LCA model selection. The performance of the proposed indices is assessed using two simulation studies, and the indices show some potential when interpreted using an “elbow” rule. Finally, the proposed fit indices are applied to aid the development of an LCA model using various pain-related variables in the HRS dataset. Three distinct pain experience latent classes are identified, characterised by different patterns of pain impact and pain medication use. The latent classes are also found to differ across sociodemographic characteristics, with female sex and indicators of poorer socioeconomic background most common in the highest impact pain class.

In summary, this thesis addresses multiple challenges related to biases, causal inference, and latent class methodologies, developing approaches to strengthen the pain research evidence base. Findings are discussed within the context of current literature, and directions for future research are outlined.