HOD: Handling missing data and time-varying confounding in causal inference for observational event history data

In medicine it is often important to obtain valid estimates of the effects (both beneficial and detrimental) of a new treatment. To do this, we typically compare outcomes in a group of patients who received the new treatment (treatment group) with those who did not receive the new treatment (control group). The randomised controlled trial (RCT) is the gold standard for obtaining these estimates of treatment effects because it fairly allocates patients to the two groups, which makes them likely to be comparable prior to the start of treatment, e.g. one group will not be older or younger, sicker or healthier and so on.

However, RCTs are very expensive and complicated to run, and are not necessarily appropriate for answering all questions about the effects of treatment. For example, a drug may cause cancer as a side-effect, but the cancers may only appear after several years of treatment. It is then unlikely that an RCT would be maintained for long enough to detect this effect. It would therefore be very useful to measure the effects of treatments by looking only at data about patients who received the treatments as part of their normal care (through "observational studies").

However unlike in RCTs, the investigators have no control over the assignment of patients to different treatment regimens in observational studies and therefore groups of patients given different drugs may differ in other ways as well. For example, patients with more severe disease may be more likely to be given drugs which are good at improving the disease but have unpleasant side-effects. If there is a difference in outcome found between the groups, it is not clear whether the difference is due to the fact that the groups are different beyond just the drugs received, or whether the difference was really caused by the treatment (i.e. it was a "causal effect"). One widely used method to make groups more comparable when estimating the causal effect is to calculate propensity scores. For each patient, his/her propensity score is the predicted probability of receiving a particular treatment based on that patient's characteristics at the time the treatment decision is made. Groups of patients with the same propensity score but different treatments should, on average, be comparable for all of their characteristics, and any differences in outcome between the groups should therefore be attributable to treatment.


The aim of this project is to extend standard methods for obtaining causal treatment effects so that they can be used when important information about patient characteristics is missing and when patient's treatment changes over time. Both of these situations are common in observational studies, thus it is important to have reliable and robust ways to deal with them. We propose a programme of methodological research to address the above situations in observational studies, with a particular focus on the effect of treatments on the time to clinical events (e.g. how long does a patient survive after a surgery, or how soon after the start of a new treatment do unpleasant side-effects start appearing). This project will provide a general framework and guidelines for practitioners who use observational data in medical research.

Observational studies play an important role in the evaluation of treatment effects on long-term outcomes, when randomised controlled trials are not feasible because of size, time, budget and ethical constraints. Because of the absence of randomisation in observational studies, it is crucial to adequately control potential confounding from various factors (time-invariant and time-varying) in order to obtain causal effects of treatments. There has been rich literature on how to control potential confounding in observational studies such as using standard techniques-propensity score (PS) methods. However, there are various important methodological issues that have not been addressed adequately in the existing literature, including 1) partially missing confounder data in PS estimation; 2) sensitivity analysis for unmeasured confounding; 3) time-varying confounding; 4) multi-state treatment and outcome processes.


In the present application we aim to propose a programme of methodological research to address the above issues for the analysis and interpretation of data from observational studies, with a particular focus on event history data. We will develop and validate diagnostic tools in measuring the balance between treatment groups in terms of both observed values of confounders and their missing data patterns. We will provide a detailed evaluation of different missing data methods and PS methods, using balance diagnostic tools developed. We will develop general Monte Carlo sensitivity analysis methods for unmeasured confounding and non-ignorable missing data in measured confounders for common models for event history data analysis. We will develop robust time-varying PS methods for obtaining causal treatment effects when there are missing data in important time-varying confounders and explore a multistate framework to handle time-varying confounding in more general treatment and outcome processes for observational event history data.

The immediate beneficiaries of this project will be the academic community involved in using data from observational studies to obtain causal effects of treatments, interventions or exposures.

In addition downstream beneficiaries will be clinicians and public health policy makers who wish to make healthcare decisions based on data from observational studies. Currently, there is no consensus in the causal inference community on how best to deal with missing data in propensity score estimation, complex time-varying confounding and unmeasured confounding in the analysis of observational event history data. This proposed project would provide a methodological framework for addressing these common problems and thus better inform healthcare decision making.