On Tuesday 4 December, the RSS Medical Section hosted an afternoon of presentations and discussion on the topic of ‘Multiple Imputation: 40 years on, where are we now?’ The event was held at 15 Hatfields on Chadwick Court in London and organised by Section chairman Robin Mitra, who is a Lecturer in Statistics at Lancaster University.
Presentations were given by highly regarded named in the field of Multiple Imputation. Ian White, professor of statistical methods for medicine at the MRC Clinical Trials Unit at University College London, began proceedings with a fantastic overview entitled ‘Multiple imputation: the universal panacea, and its limitations’. Multiple imputation is a complex technique that has become particularly popular in the last 10 years and is the most flexible solution to missing data. Robust software has developed over time to broaden its scope for various data scenarios as research groups have found new and different questions about missing data.
Tra my Pham, a research associate in medical statistics at University College London, gave the second presentation on ‘Population-calibrated multiple imputation for a binary/categorical covariate in categorical regression models’. The Calibrated-δ adjustment MI utilises population-level information to generate an intercept adjustment in the imputation model. This can improve on standard MI under missing not at random (MNAR) and is a useful tool for sensitivity analysis to assess impact of potential departures from missing at random (MAR).
The final presentation was given by James Carpenter, professor of medical statistics at the London School of Hygiene and Tropical Medicine and the MRC Clinical Trials Unit at University College London, and Suzie Cro, research fellow at Imperial Clinical Trials Unit at Imperial College London, titled ‘Sensitivity analysis for missing trial outcomes: what can it for you?’ Having discussed reference-based MI, there are clear advantages, in that explicit specification of numerical sensitivity parameters is not required, assumptions are intuitive and easily accessible and different missing data assumptions are readily incorporated. On involving experts in elicitation methods, the positives are that trial reports are read and interpreted by experts, who use them for licensing, funding, and treatment decisions, and, when interpreting trials with missing outcome data, they make an implicit assumption about the missing values, which we better try to capture and quantify this.
A Q&A was held following the presentations. Multiple topics were covered: for instance, how do you test for MNAR? James Carpenter suggested that, given that The Health Improvement Network (THIN) dataset has been proven to be close to survey data, comparisons can be drawn between THIN and the relevant dataset to test those assumptions. On model with multiple covariates, Ian White suggested imputing consistently across all the multiple models.
The event was preceded by the AGM of the RSS Medical Section, which saw six committee members stand down and the three nominated new members confirmed. A special mention, and gift of flowers, was given to Gill Price, senior lecturer at the University of East Anglia, to celebrate her long-term effort as meetings secretary.
Presentations from this event are available below:
- Multiple imputation: a universal panacea and its limitations (PPTX download) Ian White
- Population-calibrated multiple imputation for a binary/categorical covariate in categorical regression models (PDF) Tra My Pham, James R Carpenter, Tim P Morris, Angela M Wood, Irene Petersen
- Sensitivity analysis for missing trial outcomes - what can it do for you? (PDF) James R Carpenter and Suzie Cro