When drawing conclusions from data (causal inference), 'confounding' variables - variables other than the independent variable being assessed - can introduce spurious connections and bias. This project aims to apply recently developed tests for identifying certain types of potentially unmeasured confounding variables, which will in turn allow for assessing measured variables for biases. These tests are based on 'mixed models', statistical models that contain both fixed and random variables.
Explaining the science
There is a relative scarcity of methods for testing for unmeasured confounding. Most current methods use 'instrumental' variables, and whilst this can be a very powerful technique, it's not always possible to identify an 'instrument' in all situations. This project aims to develop a more general method that can be applied in a wider range of settings by exploiting the properties of random variables.
The project aims to investigate the use of mixed models in order to identify potentially unmeasured confounding when conducting causal inference.
Current causal inference methodologies make the assumption that there are no unmeasured confounders, which is in many cases not an accurate representation of a problem. Identifying if confounders are present is important in longitudinal observational studies as it enables for more reliable assessment of conclusion drawn from the inference process.
This project is part of the Data-centric engineering programme's Grand Challenge of 'Monitoring Complex Systems'.
Whilst the methodologies developed in this project will be applicable in a wide range of applications, the work is particularly focused on the transportation setting.
Other projects under the Grand Challenge of 'Monitoring Complex Systems' are investigating inference methodologies to answer questions about traffic interventions. The tests developed in this project will enable researchers and practitioners in those other projects to understanding whether there is likely to be bias when measuring the effects of such traffic interventions. This will allow for more reliable conclusions to be drawn about how well such interventions are actually functioning.