The analysis of falls counts from falls prevention trials in people with Parkinson’s

Abstract

Falls are a common recurrent event for People with Parkinson’s (PwP) and may result in injuries and loss of independence in daily activities. Falls prevention trials evaluate whether an intervention is effective in reducing falls. The traditional analysis is the logistic regression, but Negative Binomial (NB) models have become widely used recently. The distribution of the falls count is usually heavily skewed, with a relatively small mean and a few outlying large numbers. These large counts are a challenge in modelling falls count because they may have great influence in model estimation, especially when there is imbalance between groups.

This thesis focuses on examining the statistical methods used in analysing falls counts, especially the NB model. Diagnostic plots specifically designed to assessing the influence of outliers on NB modelling are developed in this context, so that the outliers can be easily identified.

The falls counts during a pre-randomisation baseline period is usually strongly correlated with the falls counts during an outcome period. Approaches to incorporating the baseline count in modelling outcome falls counts are examined in three motivating datasets and simulations carried out generating data resembling the characteristics of real data with respect to the methods used to collect the falls count. Data from trials with prospectively collected outcome counts and retrospectively collected baseline counts are examined using an actual dataset and simulations to check whether this design impacts on model estimation. Overall, including the logged baseline count as a covariate in NB regression was shown to have satisfying power and to be robust when the underlying assumption does not hold.

Some alternative count response models to the standard NB model are also considered: Poisson Inverse Gaussian models for heavily skewed data; zero-inflated NB to check for potential zero-inflation in falls counts; right-censored/right-truncated NB to reduce the influence of large falls counts; finite mixture Poisson model to accommodate the frequent fallers as a subpopulation; and random-effects NB models to explore the possibility of modelling longitudinal falls counts. They all show potential in dealing with specific issues in analysing falls data.

More information

Identifiers

ORCID for Han Zheng: orcid.org/0000-0003-3134-6947

Catalogue record

Export record