spatial modeling | JH

Bayesian Spatio-temporal modeling with Department of Health data

Mon, 01 Aug 2022 00:00:00 +0000

The Bayesian modeling project is a collaborative project involving seven epidemiologists from the Department of Health, Western Australia (DoHWA), three researchers from the Queensland University of Technology (QUT) and the project director from FrontierSI. The project manager was Dr Alex Xiao, a senior epidemiologist from DoHWA.

From August to November 2021, I worked on this project as an intern, supervised by Dr Susanna Cramb and Dr Alex Xiao. I learned a behemoth amount in relation to Bayesian modeling and computation, goal setting and, of course, the core skill of effective communication. During the internship I was responsible for the following exciting tasks

Exploring the wealth of health data available at DOHWA for mapping
Assessing, deciding and running a variety of Bayesian spatio-temporal models
Optimizing and simplifing Bayesian R packages (nimble and CARBayesST) by crafting wrapper functions in R
Prepare two written reports (one approx. 10 pages and the other approx. 100 pages)
Prepare slides and R code for three days of intensive training
Help conduct said training by travelling to Perth!

Fortunately, I continue to work on this project in a consultation capacity as a QUT research assistant.

The modifiable areal unit problem (MAUP) in disease mapping

Fri, 01 Oct 2021 00:00:00 +0000

Between October 2021 and March 2022 I worked as a casual research assistant for a small research project named ‘Analysis of biases in aggregated spatial data’. The project was funded by the Centre for Data Science at the Queensland University of Technology. I was fortunate to be part of a large research team which included two chief investigators, two co-investigators, two research assistants, and one external collaborator. My responsibilities in the project included

Developing a simulation experiment to generate large spatial datasets
Creating R, bash and AZTools code to carry out the intensive simulation experiment
Fitting Bayesian spatial models
Visualizing and interpreting the simulation results
Helping to create a shiny app to dynamically explore the results
Preparing a manuscript describing the work for publication in a peer-reviewed journal

Australian Cancer Atlas

Thu, 01 Apr 2021 00:00:00 +0000

Details on the internationally-renowned Australian Cancer Atlas (ACA) can be found here. I implore you to check it out now… because it’s soon getting a major makeover!

The PhD project

Thu, 01 Apr 2021 00:00:00 +0000

The page to rule all pages. Well not really, but definitely true for my stats work now!

I am incredibly fortunate to be working with a very large collection of researchers on the second-phase of the Australian Cancer Atlas (ACA). The current version provides spatial maps of cancer incidence and excessive deaths at the small area level. My PhD work relates to cancer risk factors, which the 2019 Australian Burden of Disease Study estimated could be attributable to 44% of the health burden of cancer in Australia. We want to include spatial maps of cancer risk factors in the second-phase of the ACA.

To enable policymakers a more nuanced and targeted approach to reduce the cancer burden in Australia, I will generate small area level prevalence estimates for several cancer risk factors using the 2017-18 National Health Survey data. To achieve this (hefty, but exciting) goal, the first stage of my PhD involved the development of novel small area estimation methods — particularly those for very sparse survey data.

Once I’ve generated the risk factor estimates at the small area level. We hope to explore the spatial (and non-spatial) dependence of risk factors at the small area level. This is exciting and novel work for several reasons. Firstly, we must adjust for the uncertainty of the risk factor estimates which haven’t been considered for proportions yet. Either that or I haven’t found any papers that do. Secondly, we will want to apply both spatial and non-spatial models to these data as they will help to answer unique questions about the state of cancer risk in Australia. Finally, I’m excited to be able to explore multivariate spatial priors and discover ways to use these efficiently in probabilistic programming languages such as Stan.

The final step – time permitting of course - is to generate a single map that captures cancer risk. A single map that can give a rough overview for policymakers and simplify the dissemination of the current state-of-health-play, as it were. A tool such as this, if done correctly, would be super helpful and hopefully not too upsetting or concerning. Methods to generate a single map has ties to dimension reduction, clustering or health index creation. My current solution (without having tried any!) is that Bayesian structural equation models or factor models could be employed. The benefit to these is their ability to capture the estimated uncertainty and propagate it to the resulting map. By assessing or estimating whether all the cancer risk factors come from a common underlying latent spatial field, we could infer the latent risk posed by a high prevalence of cancer risk factors.

I have about 1.5 years remaining, which is insufficient given all the interesting stats I wish to apply. Not only am I excited about exploring the stats behind the ideas I’ve very briefly discussed above, but I am enthusiastic about the potential public health impacts my PhD work may have