Novel statistical methods to help understand social and spatial factors associated with mental health outcomes in Cheshire and Merseyside

Description

Mental health service users in Liverpool City Region have 20 years lower life expectancy than the UK average. Liverpool also has some of the highest levels of socio-economic deprivation in the UK, resulting in a ‘syndemic’ of poor mental health and poverty. This PhD project seeks to explore how trajectories of severe mental illness vary time, place and socio-economic circumstances, in ways that may otherwise be over-aggregated in crude measures of deprivation.

The award-winning CIPHA project, combined with the new UK Mental Health Mission, offer unprecedented detail of mental health service user data within a general population of 2.7 million Cheshire & Merseyside residents. The NHS and social care data are linked to novel geospatial data for exploring determinants of mental health (care) trajectories.

Current statistical models tend to consider either spatial (change over space) or longitudinal (change over time) correlation. This ignores the complex interactions between these two factors which are difficult to tease out. Accounting for both types of correlation within a single model is likely to reveal additional patterns within mental health incidence data. But such models are computationally intensive, making their use potentially prohibitive in large datasets like CIPHA. This project will develop state-of-the-art Bayesian modelling algorithms to make inference scalable in the large CIPHA datasets. These models will explore associations between social and spatial factors with mental health outcome incidence, including how associations may be dynamic over space and time.

OBJECTIVES:

1. Create dashboards with interactive patterns of mental health and care outcomes, including comorbidities, and showing changes over time and place.
2. Utilise spatial models to look at impact of deprivation and associated geospatial measures on incidence rates in the region.
3. Develop novel Bayesian models for incidence accounting for both spatial and longitudinal data.
4. Use complex spatial-longitudinal models for mental health incidence rates and explore whether COVID-19 has impacted on these rates.
5. Identify hotspots and clusters of fast-changing (improving, worsening or stable) mental health indicators over time, and generate hypotheses as to the determinants.

NOVELTY:

This project will develop novel statistical algorithms to accurately account for spatial and longitudinal correlations in mental health incidence data. We will also harness the rich longitudinal linked data resource available through the CIPHA platform (www.cipha.nhs.uk) and the M-RIC data lake. This will allow unprecedented ability to combine data from GP records, hospital episodes, prescription data, social care contacts and small area socio-economic data, consumer and administrative (including welfare and education) data. The size, and richness of this data resource offers great potential to identify new trends in mental health incidence in the region.

EXPERIMENTAL APPROACH:

Spatio-temporal models are likely to be computationally intensive in the large datasets available to this project. The student will therefore develop novel methods for fast Bayesian inference. A potential route to this will be through the use of Variational Bayes methods. We will develop mean-field variational Bayes algorithms for extensions of the Besag-York-Mollie type Bayesian hierarchical Poisson models. We will compare accuracy, and computational cost of the developed algorithm to alternatives such as full Bayesian Markov Chain Monte-Carlo, or integrated nested Laplace algorithms.

POTENTIAL IMPACT:

This project will provide an exciting mix of outcomes. The student will develop state-of-the-art Bayesian algorithms. We anticipate this will lead to at least one methodology paper describing variational Bayes algorithms for spatio-temporal models (target: Journal of the American Statistical Association). The methods developed will be accompanied by software to fit the models in other datasets, which will be compiled into an R-package and made openly available via CRAN (cran.r-project.org). This will allow other researchers to use the models in other applications and increase the impact of the work done during this project. The project will provide additional knowledge related to mental health outcomes. We expect this to be published in relevant clinical journals and fed into various policy makers locally via M-RIC (via open policy briefs).