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

Statistical models for outcome prediction are central to traumatic brain injury research and critical baseline risk adjustment. Glasgow coma score (GCS) pupil reactivity crucial covariates in all such but may be measured at multiple time points between the of hospital subject a variable degree unreliability and/or missingness. Imputation missing data undertaken using full imputation or by simple substitution measurements from other points. However, it is unknown which strategy best more predictive. We evaluated pseudo- R 2 logistic regression (dichotomous survival) proportional odds (Glasgow Outcome Score—extended) different strategies on The Collaborative European NeuroTrauma Effectiveness Research Traumatic Brain Injury (CENTER-TBI) study dataset. Substitution were easy implement, achieved low levels missingness (<< 10%) could outperform without need computationally costly calculations pooling final models. While model performance was sensitive strategy, this effect small absolute terms clinical relevance. A emergency department discharge assessments working back when these generally performed well. Full had advantage preserving time-dependence models: pre-hospital found relatively unreliable predictors survival outcome. predictive later model-dependent. In conclusion, imputing GCS response can perform well alternative many cases.