Longitudinal data subject to irregular observation: developing methods for variable selection, causal inference, and measurement error

点击翻译按钮获取中文摘要

We live in a data-rich society, and many datasets include information on the same individuals repeatedly over time (known as longitudinal data). However, often the times at which data is collected are related to the outcome of interest. For example, in a study of newborn growth, newborns who grow slowly are likely to have their weight measured more often. Taking the average of all the weight measurements over time will underestimate the growth rate. The measurement times are informative: the frequency of measurements gives us important information about the value of the measurements. We must take the measurement frequency into account to get an accurate picture of the trajectory over time. Methods for analysing this type of data are limited in what they can do. This work will focus on: (a) handling datasets for which a large number things are measured at each time point; (b) enabling researchers to detect whether one variable causes a change in another; (c) correctly accounting for measurement error; (d) describing the behaviour of a system according to whether it is in a high or low risk state. (a)Many datasets contain a large amount of information. For example, data on water quality might include the levels of thousands of contaminants; we need to filter out which contaminants are important. We will develop ways of doing this for longitudinal data with informative measurement times. (b)It is often straightforward to show whether one quantity tends to increase as another increases (association), but harder to show whether a change in one quantity causes a change in another (causality). Our ability to determine causality with longitudinal data with informative measurement times is limited. We will develop approaches to address this gap. (c)Data is typically measured subject to error. For example, height is never measured perfectly. This can lead to bias unless it is accounted for. We will find ways of doing this with longitudinal data with informative measurement times. (d)It is often helpful to describe expected outcomes conditional on risk status. For example, for a patient with a chronic disease subject to relapse and remission, it may be more helpful to describe health during periods of relapse, health during periods of remission, and the proportion of time spent in remission. We will develop ways of doing this using longitudinal data with informative observation times. In Canada we collect a large amount of data as part of usual societal operations. For example, many patients provide consent for their medical records to be used for research. This data can address questions that are specific to Canada, for example, to determine which groups of people are at risk of poorer health outcomes. This requires careful handling of informative observation in order to provide reliable results. The methods we develop will equip researchers to do this, and so generate high quality evidence on which to base societal decisions.