Thesis and Research Journey

Question 1

Which result are you most proud of?

Answer 1

• development and application of the GDM-HMM for COVID-19 variants stands out
• working on this shortly after the pandemic and it was really interesting to see the effects of each variant all over the world shortly after reading / hearing the effects of the pandemic
• work itself captured real-world variant transitions while modelling uncertainty and overdispersion
• first piece of new novel work I produced

Question 2

What was the biggest challenge you faced during your PhD?

Answer 2

• effects of starting PhD during the pandemic started slowly
• changes within my supervisory team around a year in then there was
• lead supervisor illness

Question 3

How do you define “novelty” in your work?

Answer 3

• developing novel Bayesian hierarchal approaches for compositional data where a log-ratio transformation is unsuitable due to specific features (e.g. zeros, missing, count)
• developing and applying the GDM to compositional data, for both time series and spatial
• developing a HMM structure to compositional count time series

Question 4

What was the most surprising result from your research?

Answer 4

• how well the GDM performed over the GAM within the tree species work
• highlights the benefits of how the GDM accounts for the compositional structure
• need to account for the compositional structure to better model the data

Question 5

How does your work advance existing literature?

Answer 5

• extends and shows the strength of compositional data analysis beyond log-ratio methods
• provides Bayesian hierarchical frameworks that model compositional data with features such as zeros, missing values and counts
• generalised approached which are widely applicable for many compositional data applications

Question 6

How did your thinking about compositional data evolve over time?

Answer 6

• evolved from seeing compositional data as constrained proportions that sum to 1 to recognising the value in absolute values
• led to the importance of considering compositional data in its natural form by capturing relative and absolute information, whilst keeping zeros and missing values as they are

Question 7

How does your work contribute to the existing literature?

Answer 7

• provides generalised, Bayesian hierarchal approaches for compositional data with specialised features such as zeros, missing values or count structure
• enhances the existing literature by providing robust modelling techniques for advanced data challenges such as multilevel, non-smooth time series and spatial structures

Question 8

What’s the key message or contribution of your thesis?

Answer 8

• expanding the traditional definition of compositional data, allowing more innovative approaches to be applied
• proposed methods offer significant benefits across a wide range of compositional data applications, making them valuable tools for real-world analysis of compositional data
• allow for direct modelling of compositional data without the need for transformations or manipulations of values potentially discarding important information on the overall dynamics

Question 9

What was your contribution compared to your supervisors?

Answer 9

• leading on the literature review, understanding and presenting existing methods and showing the gaps in the literature
• leading on model development and producing the code for the models, including the MCMC modelling and posterior predictive checks
• supervisors guided the broader research framing and reviewed methodologies and results whilst supporting me to develop novel approaches

Question 10

What is the biggest criticism someone could make of your thesis?

Answer 10

• consider compositional data in the wider definiton than has been typically done and the group central to developing compositional data analysis methods could find fault with that

Question 11

How would you defend your work against a sceptical reviewer?

Answer 11

• highlight the challenges to trying to implement a log-ratio transformation in cases of zeros, missing values or counts
• show the benefits of modelling techniques for compositional data without a log-ratio transformation - widely applicable

Question 12

What are the ethical implications of your work?

Answer 12

• must be used with awareness of uncertainty and not treated as deterministic outputs - especially in forensic or policy settings

Question 13

What was the most difficult concept to grasp during your research?

Answer 13

• Understanding and implementing the GDM distribution
• never used and did not know the benefits it would have for compositional data
• never implemented the GD either so I had to grasp the ideas behind this family of distributions

Question 14

How did you stay motivated during the PhD?

Answer 14

• potential for real-world application and improvements for public health and forensic applications
• challenge of developing novel models
• support from supervisors, friends and family

Question 15

Which statistical concept did you master the most during this project?

Answer 15

• compositional data analysis
• Bayesian hierarchical methods

Question 16

What impact has this work had on your future career goals?

Answer 16

• developed a strong foundation for roles in Bayesian modelling and data science
• fuelled my interest in real-world applications and having an impact within statistical areas

Question 17

How did you manage reproducibility and data/code transparency?

Answer 17

• documented and wrote code in a clear and readable manner
• all code is uploaded to GitHub
• ensured that models and evaluations could be reproduced with shared scripts
• produced replicate datasets as data cannot be shared

Question 18

What would you tell a new PhD student starting a similar project?

Answer 18

• relax and enjoy the process - it will go by quickly but all come together
• keep models interpretable and generalisable.
• document all choices and tests, even failed ones
• believe in yourself