Professor Angie Wade

The construction of cross-sectional covariate-related centiles: repeat measures, joint and ordinal outcomes

11 February, 2019
UOW Innovation Campus, North Wollongong

Co-Speaker Prof Thomas Astell-Burt is a Founding Co-Director of the PowerLab and an NHMRC Boosting Dementia Research Leadership Fellow. Thomas will outline recently published findings by the PowerLab in the area of environment and health in Australia and overseas.

Angie Wade is Professor of Medical Statistics at University College London (UCL), UK, within the Great Ormond Street Institute of Child Health. She has over 170 peer-reviewed research publications, has been an expert reviewer in high profile legal cases, is one of 8 statistical editors of the British Medical Journal and has collaborated widely within the field of paediatrics. Her PhD investigated the use of likelihood-based methods for the construction of cross-sectional age-related standards and she has maintained an interest in this particular area of statistics for the last 20 years. Professor Wade founded and directs the UCL Centre for Applied Statistics Courses (CASC) which is the primary provider of short non-examined training courses in statistical methods for non-statisticians in the UK.

Seminar abstract: Reference ranges allow clinical identification of an individual who is an outlier with respect to the normal population. Methodologies for creating covariate-adjusted centiles when the outcome is numeric are well established. In this short talk, I will first give a brief overview of centile construction methods where covariate adjustment is necessary, and follow this by outlining 3 areas that I have been involved in where the basic method is inadequate and the methodology is extended: (1) Repeat measurements from individuals are often correlated. Where the number of measurements per individual is variable and/or related to the value of those measurements, a biased dataset may be obtained. Selection of a single measure per individual is one solution but wasteful of the data. The correlation structure can be incorporated into the maximum-likelihood approach and this methodology is contrasted with select one only, which is shown to lead to biased estimates. (2) Where there are joint outcomes, such as measurements from each eye per patient or several assessments of lung health, then bivariate centiles may be appropriate. (3) For ordinal outcomes, the assignment of centiles to individuals has received limited investigation. Proportional odds models with forms selected to mimic data features, such as asymmetric logistic models for developmental applications where a child progresses through several stages before attaining adult levels, can be incorporated into appropriate software to provide a useful clinical tool.