Imputation of Ordinal Outcomes: A Comparison of Approaches in Traumatic Brain Injury

Abstract

Loss to follow-up and missing outcomes data are important issues for longitudinal observational studies and clinical trials in traumatic brain injury. One popular solution to missing 6-month outcomes has been to use the last observation carry forward (LOCF). The purpose of the current study was to compare the performance of model-based single-imputation methods with that of the LOCF approach. We hypothesized that modelbased methods would perform better as they potentially make better use of available outcome data. The CENTER-TBI study (n = 4509) included longitudinal outcome collection at 2 weeks, 3 months, 6 months, and 12 months post injury; a total of 8185 GOSe observations were included in the database. We compared single imputation of 6-month outcomes using LOCF, a MI panel imputation, mixed effect model, a Gaussian process regression, and a multi-state model. Model performance was assessed via cross-validation on the subset of individuals with a valid GOSe value within 180 +/-14 days post-injury (n = 1083). All models were fit on the entire available data after removing the 180 +/-14 days post-injury observations from the respective test fold. The LOCF method showed lower accuracy (i.e. poorer agreement between imputed and observed values) than model-based methods of imputation, and showed a strong negative bias (i.e. it imputed lower than observed outcomes). Accuracy and bias for the model-based approaches were similar to one another, with the multi-state model having the best overall performance. All methods of imputation showed variation across different outcome categories, with better performance for more frequent outcomes. We conclude that model-based methods of single imputation have substantial performance advantages over LOCF in addition to providing more complete outcome data.

A genome-wide association study of outcome from traumatic brain injury

“…Subjects were predominantly men (53-75% across cohorts), with a mean age ranging from 34 to 61 years. The majority (59-91% across cohorts) of cases had mild TBI (GCS [13][14][15]. Differences in TBI severity were associated with differences in GOSE in each cohort: mortality was between 2% and 10%; combined death or severe disability between 9% and 25%; and complete recovery between 24% and 41% (Table 1, Supplementary Table 1).…”

“…For individuals in CENTER-TBI and TRACK-TBI in whom GOSE at 6 months was missing, but GOSE was measured at another time point within one year of the injury, missing 6-month GOSE values (499/2455 in the core CENTER-TBI and 274/1672 in the TRACK-TBI cohort) were imputed using a Markov multi-state model exploiting available longitudinal GOSE measurements. Used multi-stage approach relies on Markov assumptions and allows to model the probability of transitions between GOSE states and outperforms alternative panel imputation methods as discussed in detail by Kunzmann et al 14 Where no GOSE values were available at any time point, patients were excluded from the analysis. TBI severity was specified using the Glasgow Coma Score (GCS), with TBI classified as mild (GCS 13-15), moderate (GCS 9-12), or severe (GCS 3-8).…”

Psychometric Characteristics of the Patient-Reported Outcome Measures Applied in the CENTER-TBI Study

“…Since the GOSE-Q is not able to differentiate between vegetative state and lower severe disability, GOSE levels 2 and 3 were collapsed into one category. The missing values at six-months outcome assessments were imputed using a multi-state model; the imputation procedure is described elsewhere [ 47 ]. The GOSE was not subjected to reliability analyses, as it would require data from independent raters to provide interrater reliability, which was not available in the CENTER-TBI database.…”

Primary versus early secondary referral to a specialized neurotrauma center in patients with moderate/severe traumatic brain injury: a CENTER TBI study

“…For cases in which GOSE assessments had been performed outside the pre-specified window of 5-8 months, a multistate model was made by the CENTER-TBI statisticians to impute the 180-day GOSE [1,21]. This imputed GOSE variable was made by the CENTER-TBI statisticians and was directly extracted from the CENTER-TBI Neurobot dataset [22]. This enables all CENTER-TBI researchers to use the same outcome variable.…”