Development and practical application of landmarking in studies of time-varying exposures and survival

Context

It is of great importance to understand the effects of the features of an individual on their survival - that is, for example, on mortality or disease diagnosis rates. These features, which we refer to collectively as 'exposures', may be lifestyle factors, treatments received, or clinical measurements such as blood pressure. It is often of interest to study exposures which are changing in an individual over time ('time-varying exposures') and how they relate to survival. By understanding these relationships we can gain insight into biological mechanisms. This helps towards the development of new treatments and can inform public health policy. Understanding how time-varying exposures impact on health and survival can also guide clinicians in their decision making, enabling more individualised treatment and prognosis for patients.

To study relationships between time-varying exposures and survival requires detailed data on large populations collected over long time periods. Data of this type can be obtained from patient data registries and electronic health records, which are for example obtained from records of participating GP practices, and banks of biological material collected from volunteers (biobanks).

Aims

To analyse the types of data mentioned above requires complex statistical methods and this is still a developing area of research. This fellowship would enable me to develop statistical methods which enable researchers to perform analyses which answer a range of important questions regarding the relationships between time-varying exposures and survival.

There are a number of different questions which we may wish to try to investigate, including:

- For a person with a particular diagnosis and knowing that person's time-varying exposure measurements up to the present time, what is their predicted probability of surviving up to at least 5 years from now?

- Which exposures are having a causal impact on survival and therefore would be the most important targets for treatment?

- What are the impacts on survival of different patterns over time in time-varying exposures?

Addressing these questions becomes especially complicated when there are lots of variables to be considered simultaneously which are all influencing one another over time. It is a big challenge to try to disentangle the associations of different time-varying exposures with survival in order to try to establish causal effects of the measurements we are primarily interested in. To address the above questions currently requires the use of quite different complex statistical techniques. This means that researchers are discouraged from using them and we do not get the best information out of the available data. My aim in this research is to develop statistical methods based on an approach called 'landmarking' to provide a way of addressing the types of questions given above in a similar way and in a way which is intuitive and accessible to a wide range of researchers, including clinicians.

Part of my work will be to incorporate into the statistical methods the capacity to handle the difficulties which come from using data in practice. I will focus on two particular challenges, which are the occurrence of missing measurements in some individuals at some time points, and the fact that many exposures are measured with some degree of error.

It is important to me to provide methods which are widely accessible. I will provide software code which can be used to implement my methods.

Applications

The methods developed will have important applications in electronic health records data and will therefore be able to have an impact on the many health conditions which can be studied using these databases. As part of this fellowship I will apply the statistical methods developed to study questions concerning the survival of people with cystic fibrosis using data from a large patient registry.

It is vital to understand the effects of time-varying exposures on survival, for example to gain insight into biological mechanisms and to inform treatment decisions. Data such as electronic health records, which contain patient-level variables recorded longitudinally, enable us to study these effects.

My proposed research focuses on (1) making dynamic predictions of survival to a future time horizon based on an individual's measurements up to a given time; (2) estimating the causal effects on survival of patterns in time-varying exposures, including continuous and binary exposures and intermediate events, taking into account time-varying confounding.

These investigations can currently require quite different complex statistical methods, including joint modelling of exposure and survival data, multi-state modelling and causal inference methodology. A unified approach does not exist but is essential to enable applications from a wide range of researchers. I will develop statistical methods based on an approach called 'landmarking' to provide a unified and intuitive approach. A key advantage of landmarking is that it is based on the familiar method of Cox regression modelling. However, landmarking is currently limited to dynamic prediction and by lack of extensions to accommodate practical constraints.

In Part 1 I will devise how to use multiple imputation to handle missing data and measurement errors in longitudinal measurements used in the setting of dynamic prediction. In Part 2 I will develop landmarking to enable estimation of the causal effects of time-varying exposures on survival, allowing for time-varying confounding. Despite the importance of nested case-control and case-cohort studies, especially in biobanks, there have been few attempts at extensions to this setting. In Part 3, I will address this by extending the methods of Parts 1 and 2. I will place an emphasis on practical application and provide software code where needed.

The statistical methods developed in this project will enable statisticians, epidemiologists and other researchers to answer important questions concerning relationships between time-varying exposures and survival using the landmarking approach. The methods will be especially aimed at and important for researchers analysing data from electronic health records, patient data registries and biobanks. Specifically, researchers will be able to develop dynamic prediction models and to estimate the causal effects of time-varying exposure on survival, including in the complex situation in which there exists time-varying confounding.

The methods developed will enable researchers to make the best use of the data available to them and to incorporate the practical difficulties of missing data and measurement error, which are inherent in observational data, in order to obtain unbiased estimates. The landmarking approach is quite intuitive and based on methods already familiar to any researchers. This will enable researchers to have an in-depth understanding of the methods they are using and to be aware of the assumptions being made. My provision of example code and new software where needed will be essential to encourage uptake of the methods and will result in better and increased use of electronic health records data and related data sources.

Through my extensions of landmarking to the case-control setting, researchers faced with nested case-control and case-cohort data, which will arise in particular from biobanks, will be able to develop dynamic prediction models and estimate causal effects as if full-cohort data were available. This work will be novel and will enable questions to be addressed using case-control data which have not previously been possible. The availability of methods which can be applied to case-control-type data may also impact on how biobank resources are used, resulting in savings in cost and enabling more studies to be performed using the available biological material.

Once the methods developed as part of this fellowship become known and used by other researchers, the methods will in the future be applied to study a range of health conditions. The impact will be improved knowledge of the causal effects of exposures on ill health and improved models for predicting survival for individuals. Consequently, doctors will be able to make better-informed decisions about treatment and offer more individualised and prognosis for their patients. Patients will ultimately benefit by receiving more appropriate treatment and by having a better understanding of their condition and their prognosis. Improved understanding of the effects of time-varying exposures on survival may also ultimately impact on health policy makers.

A direct impact of the work of this fellowship will be for the cystic fibrosis community, including clinicians and patients, by application of my methods to data from a large patient registry. The impact of this work will be a dynamic prediction model for survival based on clinical characteristics and improved knowledge about the causal effects of lung function, bacterial airway infections and lung transplantation on survival.

My overall goal is for the work proposed in this application to encourage and enable more and better use of data from electronic records and related data sources. As a result, I hope that this work will ultimately have a positive impact on how the use of electronic records data is viewed by the general public in the UK, which in turn will encourage people to give consent for their data to be used in the future and have a positive impact on future research.

My collaborations will be an important route through which the impacts will be realised during the project and beyond. The skills and connections I am able to develop as part of this fellowship will set me in good stead to ensure that the impacts continue beyond the fellowship and have maximum benefit.