Imputation strategies for missing baseline neurological assessment covariates after traumatic brain injury: A CENTER-TBI study.

Ari Ercole,David W Nelson,Omar Bouamra,Fiona Lecky,Frederick A Zeiler,Lindsay Wilson,Daan Nieboer,Shubhayu Bhattacharyay,Andrew I R Maas,Ewout W Steyerberg,Hester F Lingsma,David K Menon,Abhishek Dixit,Simone A Dijkland,Corstiaan Den Uil

doi:10.1371/journal.pone.0253425

Abstract

Statistical models for outcome prediction are central to traumatic brain injury research and critical to baseline risk adjustment. Glasgow coma score (GCS) and pupil reactivity are crucial covariates in all such models but may be measured at multiple time points between the time of injury and hospital and are subject to a variable degree of unreliability and/or missingness. Imputation of missing data may be undertaken using full multiple imputation or by simple substitution of measurements from other time points. However, it is unknown which strategy is best or which time points are more predictive. We evaluated the pseudo-R2 of logistic regression models (dichotomous survival) and proportional odds models (Glasgow Outcome Score-extended) using different imputation strategies on the The Collaborative European NeuroTrauma Effectiveness Research in Traumatic Brain Injury (CENTER-TBI) study dataset. Substitution strategies were easy to implement, achieved low levels of missingness (<< 10%) and could outperform multiple imputation without the need for computationally costly calculations and pooling multiple final models. While model performance was sensitive to imputation strategy, this effect was small in absolute terms and clinical relevance. A strategy of using the emergency department discharge assessments and working back in time when these were missing generally performed well. Full multiple imputation had the advantage of preserving time-dependence in the models: the pre-hospital assessments were found to be relatively unreliable predictors of survival or outcome. The predictive performance of later assessments was model-dependent. In conclusion, simple substitution strategies for imputing baseline GCS and pupil response can perform well and may be a simple alternative to full multiple imputation in many cases.

Highlights

As a major worldwide cause of death, disability and socioeconomic burden [1], traumatic brain injury (TBI) continues to be an important area of research into novel or better stratified interventions or systems of care
There is some potential for ambiguity in defining the baseline Glasgow coma score (GCS) and pupil reactivity as both may be confounded by other contributors to unconsciousness such as hypoxia or hypotension before resuscitation
The GCS and GCSm and pupils were highly correlated between imputation strategies (Fig 2) illustrating their degree of equivalence

Summary

Introduction

As a major worldwide cause of death, disability and socioeconomic burden [1], traumatic brain injury (TBI) continues to be an important area of research into novel or better stratified interventions or systems of care. In common with other such models the presenting Glasgow coma score (GCS- or its motor component GCSm) and number of unreactive pupils are both important clinical parameters describing injury severity at presentation. Coupled with a general lack of documentation standardisation between pre-hospital and hospital care, there may be a high proportion of such early data that will be missing, unreliable or confounded by under-resuscitation [4]. Dealing with these missing covariates is a critical data curation task since baseline adjustment underpins almost all analyses and deserves particular methodological attention

Objectives

Methods

Results

Discussion

Conclusion