Propensity Score Diagnostics: Providing evidence that a propensity based estimate is unbiased

Whilst clinical trials are the gold standard for evaluating the effect of a treatment, there are many situations in which a trial is not practical for reasons of cost, size, time or ethics. In such cases obtaining accurate estimates of the effect of a treatment from observational data can be vitally important. Propensity scores are widely used for the analysis of observational data, since it has been shown that a correctly specified propensity score can produce an unbiased estimate of the effect of an exposure by ensuring that the distributions of all confounders are the same in unexposed and exposed subjects conditional on the propensity score. However, there are no guidelines to help researchers determine whether or not their propensity score is correctly specified and achieves this aim. The fact that researchers tend to assume that their model is adequate with no justification has led to understandable scepticism about the value of propensity score methods.

Researchers need further guidance if they decide that their propensity score model is not adequate. Conventional statistical model selection criteria may produce sub-optimal propensity scores by including variables that are strongly associated with the exposure, but not with the outcome, and hence are not confounders, leading to inefficient estimates of the effect of exposure. Strategies specific to propensity score models are required.

This PhD project will aim to compare existing methods of assessing propensity score models based purely on distributions of confounders, with newer methods incorporating the strength of the association between the confounder and the outcome. The outcomes of the project will be a measure of the extent to which the propensity score can be expected to have removed confounding, and a procedure to follow to improve the propensity score model if it is found to be inadequate.

This project will align with the current MRC highlight notice concerning the use of observational data in clinical decision making. Clearly, the primary focus will be skills in biostatistics, but also in communicating complex statistical ideas to non-statisticians clearly and convincingly.

Additional training opportunities beyond those already outlined include:
Attending the UK Research in Musculoskeletal Epidemiology (UK-RiME) partnership summer school in advanced epidemiological methods
Interaction with Centre for Musculoskeletal Research's Research User Group of patient and carer representatives