| Summary: | Observational health data are a rich resource that present modelling challenges due to data complexity. If inappropriate analytical methods are used to make comparisons amongst either patients or healthcare providers, inaccurate results may generate misleading interpretations that may affect patient care. Traditional approaches cannot fully accommodate the complexity of the data; untenable assumptions may be made, bias may be introduced, or modelling techniques may be crude and lack generality. Latent variable methodologies are proposed to address the data challenges, while answering a range of research questions within a single, overarching framework. Precise model configurations and parameterisations are constructed for each question, and features are utilised that may minimise bias and ensure that covariate relationships are appropriately modelled for correct inference. Fundamental to the approach is the ability to exploit the heterogeneity of the data by partitioning modelling approaches across a hierarchy, thus separating modelling for causal inference and for prediction. In research question (1), data are modelled to determine the association between a health exposure and outcome at the patient level. The latent variable approach provides a better interpretation of the data, while appropriately modelling complex covariate relationships at the patient level. In research questions (2) and (3), data are modelled in order to permit performance comparison at the provider level. Differences in patient characteristics are constrained to be balanced across provider-level latent classes, thus accommodating the ‘casemix’ of patients and ensuring that any differences in patient outcome are instead due to organisational factors that may influence provider performance. Latent variable techniques are thus successfully applied, and can be extended to incorporate patient pathways through the healthcare system, although observational health datasets may not be the most appropriate context within which to develop these methods.
|
|---|
