Statistical Models of Selection and Causation
Valid causal inference remains the key goal motivating the analysis of data from observational studies and randomised clinical trials (RCTs). Recent advances in the development of models of causation have been made in statistics, econometrics and related disciplines, alongside the capacity to cope with different selection effects, principally confounding, missing data and non-adherence to random allocation. These models allow us to view causal inference as a highly specific form of missing data analysis.
Marginal structural models (MSMs) were originally introduced to adjust for time-dependent confounders in observational studies. We hypothesise that MSMs can be innovatively applied for the analysis of non-adherence in RCTs, since they provide estimates of unbiased as-treated effects, conditional on subject’s previous exposure history. Estimating the causal parameters from a MSM can be achieved by g-computation, inverse probability of treatment weighting or double robust (DR) estimators. DR procedures allow for unbiased estimators under the misspecification of either the data model or the model for the coarsening mechanism, and can offer increased efficiency. In all studies where missing data is ubiquitous, methods include multiple imputation and inverse probability weighting; DR procedures have also recently been proposed.
The methods and their implementation are demonstrated using four statistically designed simulation studies, analysed by a new approach to compare efficiency on paired data sets. The CUtLASS trials, investigating the efficacy of neuroleptic medication in patients suffering from severe schizophrenia, are also analysed using these methods to verify whether the conclusions of the ITT analysis hold after allowing for selection effects.
MSMs are shown to be applicable for analysing non-adherence in RCTs, and the overall results show the DR estimator to be more efficient than alternative estimators, retain the DR property, and be easily implemented in standard software, for all situations considered. The analysis of the CUtLASS trials reveals that the class of atypical drugs perform no better than conventionals, and that Clozapine is better than the alternative atypicals, with respect to quality of life and symptom severity.
Duration of the project
September 2003 - September 2006
Medical Research Council PhD Studentship
Members of the project
|Professor Richard Emsley||PhD student|
|Professor Andrew Pickles||Supervisor|
|Professor Graham Dunn||Supervisor|
- Emsley, R., Lunt, M., Pickles, A., & Dunn, G. (2008). Implementing double-robust estimators of causal effects. Stata Journal, 8(3), 334-353. . Publication link: b3645507-d1d4-42f8-bff1-a398d5e71b40