Predicting falls in people with Parkinson's disease: impact of methodological approaches on predictors identified

Background: Previous investigations of falls predictors in people with Parkinson's disease (PD) have used various statistical methods and categorization of falls outcomes. The impact of methodological differences on falls predictors has not been investigated.

Objectives: To describe similarities and differences in predictors modelled for fall rates [negative binomial (NB), Poisson Inverse Gaussian (PIG) and quantile regression] and previously-reported predictors of time to second fall (Cox regression), i.e. past falls, motor fluctuations, disability, levodopa dose and balance impairment. To investigate whether predictors from quantile regression vary across subsets of fallers based on fall frequency.

Methods: Participants with PD (n = 229) were followed-up for 12 months. NB and PIG regression were used to determine predictors of fall rates, with the best fitting model reported. Quantile regression was used to determine predictors at higher (62nd, 70th, 80th) percentiles of the falls distribution. Univariate and multivariate analyses were performed.

Results: Predictors of fall rates were the same in NB and PIG multivariate models, with the PIG model fitting our data better. Past falls, disability and levodopa dose were associated with fall rates from PIG and quantile regression. Freezing of gait was associated with fall rates from PIG regression. Disease severity predicted less (70th percentile, approximately 2-4) and more (80th percentile, approximately ≥ 5) frequent falls, and anteroposterior stability also predicted less frequent falls (p < 0.05), from quantile regression.

Conclusions: Not all predictors of time to second fall were predictors of fall rates. Quantile regression revealed some divergent predictors depending on the percentile of fall frequency examined.