Elizabeth Armstrong and Kris Rogers, ANZHFR and UNSW School of Public Health
E Armstrong: ANZ Hip Fracture Registry: data visualisations for different audiences
K Rogers: Strategies for dealing with missing data in clinical trials
The July 2022 meeting featured two presentations delivered via Zoom by Elizabeth Armstrong (former Manager of the Australia and New Zealand Hip Fracture Registry) and Dr Kris Rogers (Senior Biostatistician at The George Institute for Global Health).
Elizabeth presented visualizations from the ANZHFR 2021 Annual Report and Kris’ presentation discussed advantages and disadvantages of various methods to obtain parameter estimates in the presence of missing data and methodologies for causal inference from non-trial data.
ANZHFR is a registry that collects data on the care of older people in Australia and New Zealand, admitted to hospital with a fracture of the proximal femur.
In 2021, 86 hospitals who contributed to the patient level report (64 in Australia and 22 in New Zealand) and the 117 hospitals who contributed to the facility level results.
Elizabeth presented a number of visualizationsfor key KPIs in the surgical care of patients who received hip replacement. The outlier report for the 16 quality indicators in the clinical care standard is an effective graphic display for the measured in standard deviations from the average performance of all hospitals. Performance outliers are flagged when the care metric lies 2 to 3 standard deviations from the overall hospital average performance.
For the determination of KPI attainment across various aspects of care, hospitals must have contributed at least 10 patient records during the relevant calendar year to be included in the patient level report.
Following on Elizabeth’s presentation, Kris discussed a range of methodologies for dealing with missing data, drawing comparisons between simpler approaches (Last Observation Carried Forward, mean imputation, multiple imputation) and other more advanced methods that use full information maximum likelihood, weighted GEEs (General Estimating Equations), Bayesian methods and causal inference methods.
Missing data in randomised studies is often an underappreciated issue. The goal of a randomised experiment is to draw a causal inference on the effect of a treatment on an outcome. In the context of randomised controlled trials (RCT), random assignment of treatment ensures that the average effect difference between the two groups can only be attributable to treatment. However, in the presence of missing data, the assumption of ignorability (e.g., the method of the data collection does not depend on the missing data), is very rarely met. Technically, the presence of missingness negates the benefits of randomisation, leaving the researcher with an observational dataset.
An interesting final point Kris raised, and which is often overlooked, is that data in observational studies needs to contain common support for the exposure in order to correctly apply causal inference methods for observational data. This can be done using propensity score models and visualising the areas of overlap of the propensity score in the exposure groups.
By Gabriella Lincoln