Glasgow Outcome Scale Measures and Impact on Analysis and Results of a Randomized Clinical Trial of Severe Traumatic Brain Injury

Abstract

The original unstructured Glasgow Outcome Scale (uGOS) and the newer structured interviews GOS and the Extended GOS (GOS-E) have been used widely as outcomes in severe traumatic brain injury (TBI) trials. We compared outcome categories (ranging from dead [D] to good recovery [GR]) for each measure in a randomized trial of transfusion threshold and the implications of measure choice and analysis methods for the results of the trial.

We planned to explore patient symptomology possibly driving any discrepancies between the patient's uGOS and GOS scores. Category correspondence between uGOS and GOS scores occurred in 160 (88.4%) of the 181 analyzed cases. The GOS-E and GOS instruments incorporated more behavioral/cognitive/social and other components, leading to a worse outcome in some cases than for the uGOS. Choice of outcome measure and analysis led to incongruous conclusions. Dichotomizing uGOS into favorable outcome (GR and moderate disability [MD] categories) versus unfavorable (severe disability [SD], vegetative state [VS], and D categories), we observed a significant effect of transfusion threshold (odds ratio [OR] = 0.51, p = 0.03; adjusted OR = 0.40, p = 0.02). For the same dichotomization of GOS and GOS-E, the effect was not statistically significant but the ORs were similar (ORs between 0.57 and 0.68, p > 0.15 for all). An effect was not detected using ordinal logistic regression or sliding dichotomy method for all three measures. Differences in categorizations of subjects between moderate and severe disability among the scales impacted conclusions of the trial. In future studies, particular attention should be given to implementing GOS measures and describing the methodology for how outcomes were ascertained.

Introduction

Studies of traumatic brain injury (TBI) interventions during the acute hospital stay have failed to identify effective treatments.^1–3 As a result, choice of global outcome and its evaluation for clinical trials in TBI continues to be controversial^4,5 with many efforts aimed at improving measurement and data collection procedures, reducing heterogeneity in the outcome data,⁵ and evaluating the outcome data with various analytic techniques.^6,7 The original unstructured Glasgow Outcome Scale, hereafter referred to as the uGOS,⁸ had been accepted widely as the primary outcome measure in clinical trials and other TBI research for 40 years.⁹ It includes five categories: dead (D = 1), vegetative state (VS = 2), severe disability (SD = 3), moderate disability (MD = 4), and good recovery (GR = 5). However, the uGOS was found to have poor reliability between general practicioners versus psychologists.¹⁰

Wilson and colleagues developed two structured interview forms intended to be more objective and reliable than the uGOS.¹¹ The first, the structured interview Glasgow Outcome Scale (GOS), measures the same five categories as the uGOS but this modified version involves specific questions on consciousness, independence inside and outside of the home, work, social/leisure activities, and family/ friendship relationships. It was designed to be more objective and reliable than the uGOS.¹¹ The second, the Extended GOS (GOS-E), increased the number of categories by subdividing SD, MD, and GR by adding additional questions (e.g., return to normal life to distinguish between upper and lower GR) thus producing eight categories: D (1), VS (2), lower SD (3), upper SD (4), lower MD (5), upper MD (6), lower GR (7), and upper GR (8). The GOS-E quickly gained acceptance and essentially replaced the uGOS and GOS as a primary outcome in TBI trials. Indeed, GOS-E scores have been reported to have higher validity and greater sensitivity to change¹² and to produce a small consistent increase in efficiency.¹³ Of course, these measures have limitations in delineating various aspects of behavioral/cognitive/social and other outcomes.^14–16 Table 1 provides an overview of the three GOS measures. In a brief review of the literature, no studies were identified that have examined the differences in the categorizations of individuals using the uGOS, GOS, and GOS-E outcomes on a group of individuals with a serious TBI.

Table 1.

Glasgow Outcome Scale (GOS) Measures

	Scoring categories	How data were acquired	Advantages	Limitations	References
Unstructured GOS	D = 1; VS = 2; SD = 3; MD = 4; GR = 5	Patient in-person visits (hospital or home) were preferred. If not possible, then phone interviews and/or interviewing of nurses, physicians, or others with knowledge of the patient's status.The same trained individual collected data for all GOS measures on a particular participant. The same ordering was obtained for all subjects: uGOS, then GOS, then GOS-E.	Interrater reliability is 95% agreement. Considered clinical trials “gold standard” for outcome in adults with TBI until GOS/GOS-E.	Lack of standard interview, modest interrater reliability; disproportionate weighting of physical disability relative to cognitive impairment and behavioral disturbance.	References41–45
GOS	D = 1; VS = 2; SD = 3; MD = 4; GR = 5		Interrater reliability is 92–95% agreement. Interrater weighted kappa = 0.85. Emphasis on change from pre-injury functioning and subject's capacity to perform adaptive activities, occupational, and social roles.	Some investigators have reported misclassification rates of 17–40% for GOS outcomes in clinical trials. Compared with GOS-E, not as sensitive to 3-month change. Considered as blunt outcome.	References11,12,43,46–50
GOS-E	D = 1; VS = 2; Lower SD = 3; Upper SD = 4; Lower MD = 5; Upper MD = 6; Lower GR = 7; Upper GR = 8		Interrater reliability is 78% agreement. Interrater weighted kappa = 0.84. Interrater kappa = 0.85. More associated with measures of functional outcome, neuropsychological status, and affective functioning, compared with GOS.	Low ceiling that does not adequately represent the range of impairment within GR categories. Considered as blunt outcome. May fail to capture the multi-faceted effects of TBI and its sensitivity to subtle changes, especially in the cognitive dimension.	References11,12,43,47,48,50–53

D, Death; GR, Good Recovery; GOS, Glasgow Outcome Scale; GOS-E, Extended Glasgow Outcome Scale; MD, Moderately Disabled; SD, Severely Disabled; TBI, traumatic brain injury; uGOS, unstructured Glasgow Outcome Scale; VS, Vegetative State.

The choices of a GOS primary outcome measure and whether any of the outcome categories are collapsed determines the statistical analysis method used. The uGOS, GOS, and GOS-E measures are ordinal by nature yet often dichotomized for the analysis, but not always dichotomized at the same threshold. The most common dichotomy is between the SD and MD categories (poor outcome [1–3; D, VS, and SD] versus good outcome [4–5; MD and GR] for uGOS and GOS¹⁷ and poor outcome [1–4] versus good outcome [5–8] for GOSE^{1–3,18–24}). Other dichotomies^25–28 and categorizations²⁹ of the GOS-E have been used as well.^30–34 However, dichotomizations of the GOS-E can be problematic due to possible misclassifications of individuals, particularly close to the threshold, and the categories close to the cutpoint possibly being closer to the individuals across the cutpoint versus on the same side of the cutpoint. For example, if dichotomizing between SD and MD, individuals assigned an upper SD might be more like those with a lower MD than those with a lower SD (eg. [1–3] vs. [4–8]).

A common analytic choice is whether or not to take the baseline neurological status into account when analyzing the GOS outcome. Many trials utilize and power for a simple comparision of proportions, which makes underlying assumptions that there are no differences between intervention groups in baseline neurological status among patients with TBI. Some analyses by nature consider baseline neurological status. For example, the sliding dichotomy analysis takes into account the baseline injury by splitting the observations into tertiles of severity and identifies a binary split on the outcome for each tertile.³⁵ Another approach to account for baseline neurological status is a regression model with adjustments for baseline injury severity. The choice of binary versus ordinal analyses, adjustments for baseline injury severity, and which statistical test to use should be specified a priori in randomized trials. However, the choice is not obvious and the impact of these types of choices could potentially lead to incongruous results. Therefore, understanding some of these differences is important when designing a clinical trial.

Robertson and colleagues reported on the results of a double-blind randomized clinical trial for the acute treatment of serious closed head injury (CHI), referred to as the EPO Severe TBI trial.³⁶ The trial design was factorial with randomization to erythropoietin versus placebo and to a transfusion threshold of 7 g/dL versus 10 g/dL. There was no significant interaction between the two factors. Acute treatment with erythropoietin did not improve the rate of favorable outcome on GOS at 6 months post-injury nor did the maintenance of hemoglobin concentration at ≥10 g/dL versus ≥7 g/dL with transfusions. Data for the uGOS, the GOS, and the GOS-E were collected at 6 months post-injury, but only the findings for the GOS were reported as this was the pre-specified primary outcome. These trial data provide a unique opportunity to compare all three outcomes. In the current study, we analyzed the randomization to the transfusion threshold because there was a non-significant trend for a worse outcome with the 10 g/dL transfusion threshold group.

The purpose of the current post hoc study was twofold. First, we wanted to compare the findings for the uGOS versus the structured interview GOS and GOS-E in terms of agreement of assigned outcome categories. Some degree of disagreement in category assignment was expected because of the additional detail provided by questions in the GOS and the GOS-E including addition of a section on return to normal life that focuses on behavioral/cognitive/social aspects of recovery in a way that the uGOS does not. Second, we wanted to assess the impact of the choice of pre-specified and alternative analysis methods on the trial results for the hemoglobin transfusion trigger (TT) randomization groups. The EPO Severe TBI trial serves both as a case study as well as an opportunity to conduct a sensitivity analysis of the robustness of the trial results of the TT groups.

Methods

Study design

The background, hypotheses, method, and primary analytic procedures of the EPO Severe TBI trial along with results and their interpretation can be found in the article by Robertson and colleagues.³⁶ In summary, individuals with a CHI admitted to two Level 1 trauma centers were enrolled if they did not follow commands post-resuscitation and also could be enrolled within 6 h of injury. Those with a Glasgow Coma Scale (GCS) of 3 and fixed dilated pupils, penetrating head injury, life-threatening systemic injuries defined as an Abbreviated Injury Score >4 in any organ system other than head, and pregnancy were excluded. Enrolled patients were randomized to one of two hemoglobin levels for triggering a transfusion (7 g/dL or 10 g/dL). Detailed demographic and injury characteristics of the 200 patients in the study are given in Table 1 in the article by Robertson and colleagues.³⁶ Individuals with severe polytrauma and with spinal cord injury were excluded from the study. Of the 181individuals with acquired 6-month outcomes, 142 (78%) were assessed within the window of 184 ± 14 days.

Materials

A form for the uGOS⁸ ranked patients on a scale of 1 (D) to 5 (GR). The structured interview forms for the GOS and the GOS-E were taken from Wilson and associates¹¹ and translated into Spanish. All descriptions and questions in the GOS and GOS-E were translated by several Spanish-speaking individuals (including a physician and technicians) from several Hispanic cultures. The translations were then reviewed by the team until team members were satisfied. No formal validation was conducted for the Spanish translations, although the assessments were done by people who were fluent in both Spanish and English. The effects of systemic injuries on disability were not considered separately from the effects of the brain injury in assigning the uGOS, GOS, and GOS-E scores.

Training and quality control

A clinical neuropsychologist (HJH) and trained staff conducted the outcome assessments blinded to study randomization groups. All the neuropsychology personnel with the exception of the clinical neuropsychologist (HJH) were fully bilingual in English and Spanish in terms of reading, writing, and speaking because of the high proportion of Spanish speakers in the TBI population in Houston, Texas, where the trial was conducted (50% in this trial). If Spanish was used for the interactions with any person involved in the research, answers were recorded as they were given in Spanish so that any regional or national variations in the meaning of a word or phrase could be established afterward and appropriate translations made. All examiners had BA or BS degrees from a university and were trained over several months by HJH and other personnel experienced in interviewing individuals with TBI, family members, and significant others. Training included going over the interview materials and related articles, observation of interviews by other examiners and HJH, and then supervision of their interviews initially.

Information for these measures was determined either in-person in a variety of settings (e.g., intensive care, step-down, and other floor units of various facilities as well as neuropsychology offices and places of residence) or over the telephone. Two individuals usually made visits to other facilities and any residences. Information was obtained from the patient, next of kin, significant others and/or caretakers, and medical personnel as needed or possible. Also, information was occasionally obtained from records of facilities and practices once appropriate signed release of confidential information had been obtained. Neuropsychology personnel collected additional information based on discussion and also regular quality control meetings that always included HJH. The same trained individual collected data for all GOS measures on a particular participant. The same ordering was used for all subjects: uGOS, then GOS, then GOS-E. This ordering initially focused on allocation to a uGOS and then GOS category, and then involved limited GOS-E-specific questioning to further determine the GOS-E category.

Data analysis

Relationships among category assignments on the uGOS, GOS, and the GOS-E

Cross-tabulations between the uGOS versus GOS, the uGOS versus GOSE, and the GOS versus GOS-E assignments 6 months post-injury were calculated to determine the percentage of discrepancies between assignments. Pairwise cross-tabulations also allowed us to discover whether there was a pattern of one outcome measure systematically giving a better or worse outcome score compared with another outcome measure. Possible symptoms related to the modifications in the GOS and GOS-E that might have led to some of these discrepancies are described. Cohen's kappa statistic³⁷ and weighted kappa³⁸ were used to assess agreement between outcomes. The weighted kappa gives less penalty to categories that were close to each other.

Statistical analysis

To assess any differences in the transfusion threshold effect on outcomes, we used an overall statistical approach similar to that reported in the primary article,³⁶ which included an intent-to-treat principle with multiple imputation of missing values and a complete case analysis. The following specific analysis methods were used for each outcome: unadjusted and adjusted binary and ordinal logistic regression, and a sliding dichotomy. Multiple imputation of missing 6-month uGOS, GOS, and GOS-E data was completed as outlined in the article by Robertson and colleagues.³⁶

Binary logistic regression was conducted to analyze efficacy of transfusion threshold on 6-month recovery status with pre-specified outcome dichotomization into poor (1, 2, 3) and good outcome (4, 5) for the uGOS and GOS scores, whereas GOS-E scores were dichotomized between upper SD (1–4) and lower MD (5–8). Ordinal logistic regression was also conducted with no dichotomization of the uGOS, GOS, or GOS-E scores. Adjustments were made for pre-specified covariates of injury severity including the Injury Severity Score (ISS) and the International Mission for Prognosis and Analysis of Clinical Trials (IMPACT) score³⁹ given for the core + CT + laboratory probability model. An additional covariate, presence of epidural hematoma, was also included in the adjusted model because this variable was found to be imbalanced between the transfusion threshold groups at baseline.

The sliding dichotomy method was used to take the baseline prognostic score into account to increase the statistical power as to whether a participant improved or not. We chose to use the IMPACT score as the baseline prognostic score.³⁹ The data were split into baseline prognostic risk using tertile thresholds of the IMPACT score for all subjects in the study. The distribution of uGOS, GOS,or GOS-E scores within each tertile was tabulated for the two TT groups in separate analyses. For each prognostic band, the dichotomy was chosen so that, for the pooled data, the split was as close as possible to having half of the subjects below and half above the split. Pearson's χ² test with Yates' continuity correction was conducted for the difference in the proportion with good recovery between the groups, where each subject's good recovery category was dependent on the sliding dichotomy.

Statistical significance was determined by a two-sided p-value of <0.05. Due to the exploratory nature of these analyses, no adjustment was done for multiplicity. Significant p-values should be interpreted with caution. All analyses were conducted using R version 2.13.1 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Baseline demographics between the two TT groups are summarized in Table 2.

Table 2.

Demographic and Injury Characteristics of Patients

	Transfusion Threshold 7 g/dL (n = 99)	Transfusion Threshold 10 g/dL (n = 101)
Age, median (25th–75th)	28 (21–48)	31 (24–45)
Gender, n (%)
Female	14 (14.1)	12 (11.9)
Male	85 (85.9)	88 (87.1)
Male living as female		1 (1.0)
Race/Ethnicity, n (%)
Asian	3 (3.0)	3 (3.0)
Hispanic	50 (50.5)	53 (52.5)
Black	20 (20.2)	23 (22.8)
White, non-Hispanic	26 (26.3)	22 (21.8)
Mechanism of Injury, n (%)
Assault	7 (7.1)	15 (14.9)
Fall/Jump	18 (18.2)	9 (8.9)
Automobile accident	58 (58.6)	58 (57.4)
Motorcycle accident	14 (14.1)	17 (16.8)
Other	2 (2.0)	2 (2.0)
Enrollment motor GCSa, n (%)
4–5	63 (63.6)	66 (65.4)
1–3	36 (36.4)	35 (34.7)
ED pupil reactivity, n (%)
Both reactive	63 (63.6)	58 (57.4)
One reactive	14 (14.1)	9 (8.9)
Neither reactive	22 (22.2)	34 (33.7)
ED Marshall CT scan category, n (%)
Diffuse injury 1 or 2	49 (49.5)	40 (39.6)
Diffuse injury 3 or 4	23 (23.2)	23 (22.8)
Mass lesion	27 (27.3)	38 (37.6)
Injury Severity Score, median (25th–75th)	29 (25–38)	29 (25–35)

^aMotor GCS = motor component of the Glasgow Coma Scale, which ranges from 1 (no motor response) to 6 (follows commands).

CT, computed tomography; ED, emergency department.

Relationships among category assignments for the uGOS, GOS, and the GOS-E

Category assignment for the 6-month post-injury uGOS versus the GOS appears in Table 3. Category correspondence for the uGOS and GOS occurred in 160 (88.4%) of the 181 cases and Cohen's kappa was 0.84 (0.96 for weighted kappa). Specifically, correspondence was 100% for D and VS categories. The 21 (11.6%) discrepancies arose in the assignment of SD, MD, and GR scores. Table 4 describes the relationship between the three outcomes. The uGOS outcome was better (higher category) than the GOS-E outcome in 18 of 21 discrepancies (p = 0.001, one-sample test of proportions). An MD uGOS category was given rather than a SD GOS to 5 subjects (of which 3 were lower SD and 2 upper SD on the GOS-E) and a GR uGOS category was given rather than an MD GOS (upper MD on the GOS-E) to 13 subjects. In the remaining three discrepancies, outcome was poorer on the uGOS, SD as opposed to MD on the uGOS versus GOS (lower MD on the GOS-E), respectively.

Table 3.

Correspondence of 6-month uGOS and GOS Assignments (n = 181)

	uGOS
GOS	D	VS	SD	MD	GR
D (1)	31	0	0	0	0
VS (2)	0	8	0	0	0
SD (3)	0	0	69	5	0
MD (4)	0	0	3	32	13
GR (5)	0	0	0		20

Bold italics denote discrepancies in score assignments.

D, Death; GR, Good Recovery; GOS, Glasgow Outcome Scale; MD, Moderately Disabled; SD, Severely Disabled; uGOS, unstructured Glasgow Outcome Scale; VS, Vegetative State.

Table 4.

Correspondence of 6-month GOS-E and GOS Assignments (n = 180)

	GOS					uGOS
GOS-E	D (1)	VS (2)	SD (3)	MD (4)	GR (5)	D (1)	VS (2)	SD (3)	MD (4)	GR (5)
D (1)	31	0	0	0	0	31	0	0	0	0
VS (2)	0	8	0	0	0	0	8	0	0	0
Lower SD (3)	0	0	66	0	0	0	0	63	3	0
Upper SD (4)	0	0	7	0	0	0	0	5	2	0
Lower MD (5)	0	0	0	20	0	0	0	3	17	0
Upper MD (6)	0	0	0	28	0	0	0	0	15	13
Lower GR (7)	0	0	0	0	6	0	0	0	0	6
Upper GR (8)	0	0	0	0	14	0	0	0	0	14

One patient with an incomplete GOS-E could not be included in this table. Bold italics denote discrepancies in score assignments.

D, Death; GR, Good Recovery; GOS, Glasgow Outcome Scale; GOS-E, Extended Glasgow Outcome Scale; MD, Moderately Disabled; SD, Severely Disabled; uGOS, unstructured Glasgow Outcome Scale; VS, Vegetative State.

The discrepancies between the GOS and uGOS were mainly driven by the additional detailed questions included in the GOS that provided additional information in determing the oucome that was not present in the uGOS. For example, serious behavioral and cognitive problems, reduced ability to work, changes in social and leisure activities in the community, and psychological problems disrupting relationships with family and friends led to worse GOS categories compared with uGOS categories.

Analyses of hemoglobin transfusion trigger (TT)

uGOS analyses

The 7 g/dL TT resulted in significantly better outcome compared with the 10 g/dL TT on the uGOS. The percentage of favorable outcomes using the uGOS (defined as MD or GR) was 47% and 31% in the 7 g/dL and 10 g/dL TT groups, respectively. TT was statistically significant for both the unadjusted (p = 0.03) and adjusted (p = 0.02) binary logistic regression with dichotomization between MD and SD (Table 5) but not the unadjusted or adjusted ordinal logistic regression. Multiple imputation pooled logistic regression estimates resulted in identical inference in both adjusted (p = 0.03) and unadjusted (p = 0.02) models demonstrating significantly favorable outcome for the7 g/dL versus 10 g/dL TT groups (Table 6). The sliding dichotomy unadjusted analysis of the uGOS data using the tertiles in Table 7 was not significant (p = 0.72).

Table 5.

Binary and Ordinal Logistic Regression Results for Transfusion Threshold for uGOS, GOS, and GOS-E Outcomes

	uGOS		GOS		GOS-E
	OR (95% CI)	P-value	OR (95% CI)	P-value	OR (95% CI)	P-value
Unadjusted logistic regressiona
10 vs. 7 g/dL transfusion threshold	0.51 (0.28–0.93)	0.03	0.66 (0.36–1.22)	0.19	0.68 (0.37–1.24)	0.20
Adjustedb logistic regressiona
10 vs. 7 g/dL transfusion threshold	0.40 (0.19–0.86)	0.02	0.57 (0.27–1.21)	0.15	0.58 (0.27–1.23)	0.16
Unadjusted ordinal logistic regression
10 vs. 7 g/dL transfusion threshold	0.66 (0.39–1.13)	0.13	0.82 (0.48–1.39)	0.46	0.71 (0.42–1.19)	0.19
Adjustedb ordinal logistic regression
10 vs. 7 g/dL transfusion threshold	0.72 (0.41–1.27)	0.26	0.91 (0.52–1.59)	0.73	0.79 (0.45–1.36)	0.39

^auGOS, GOS-E, and GOS were dichotomized into {Good Recovery, Moderately Disabled} vs. {Severely Disabled, Vegetative State, Death}.

^bAdjusted for Injury Severity Score, the IMPACT laboratory model score, and presence of epidural hematoma.

CI, confidence interval; GOS, structured interview Glasgow Outcome Scale; GOS-E, Extended Glasgow Outcome Scale; uGOS, unstructured Glasgow Outcome Scale; OR, odds ratio.

Table 6.

Unadjusted and Adjusted Multiple Imputation Pooled Logistic Regression Estimates for uGOS, GOS, and GOS-E Outcomes Using All Cases

	uGOSa		GOSa		GOS-Ea
	OR (95% CI)	P-value	OR (95% CI)	P-value	OR (95% CI)	P-value
Unadjusted logistic regression
10 vs. 7 g/dL transfusion threshold	0.51 (0.28–0.94)	0.03	0.69 (0.37–1.26)	0.23	0.68 (0.37–1.24)	0.21
Adjustedb logistic regression
10 vs. 7 g/dL transfusion threshold	0.39 (0.18–0.83)	0.02	0.58 (0.27–1.23)	0.15	0.58 (0.27–1.24)	0.16

^auGOS, GOS, and GOS-E were dichotomized into {Good Recovery, Moderately Disabled} vs. {Severely Disabled, Vegetative State, Death}.

^bAdjusted for Injury Severity Score, the IMPACT laboratory model score, and presence of epidural hematoma.

CI, confidence interval; GOS, structured interview Glasgow Outcome Scale; GOS-E, Extended Glasgow Outcome Scale; uGOS, unstructured Glasgow Outcome Scale; OR, odds ratio.

Table 7.

Unadjusted Analyses of uGOS, GOS, and GOS-E Using the Sliding Dichotomy Method

IMPACT probability of poor outcome	0–0.23	0.23–0.51	0.51–1	Total	P-value
uGOS favorable dichotomization threshold	> MD	> SD	> VS
TT7, n (%)	13 (46)	15 (41)	9 (39)	37 (42)	0.72
TT10, n (%)	11 (34)	6 (26)	26 (67)	43 (46)
GOS favorable dichotomization threshold	> MD	> MD	> VS
TT7, n (%)	23 (82)	12 (33)	9 (39)	44 (51)	0.73
TT10, n (%)	19 (59)	6 (26)	26 (67)	51 (54)
GOS-E favorable dichotomization threshold	> upper MD	> lower SD	> VS
TT7, n (%)	7 (25)	16 (44)	9 (39)	32 (37)	0.48
TT10, n (%)	8 (25)	6 (27)	26 (67)	40 (47)

GOS, structured interview Glasgow Outcome Scale; GOS-E, Extended Glasgow Outcome Scale; MD, Moderately Disabled; SD, Severely Disabled; TT7, 7 g/dL transfusion threshold; TT10, 10 g/dL transfusion threshold; uGOS, unstructured Glasgow Outcome Scale; VS, Vegetative State.

GOS analyses

The percentage of favorable GOS/GOS-E (MD or GR) was 43% and 33% in the 7 g/dL and 10 g/dL TT groups, respectively. TT levels (7 vs.10 g/dL) were not significantly related to 6-month outcome in all of the performed GOS analyses (see Tables 5, 6, and 7), including the multiple imputation pooled logistic regression estimates. Finally, the analysis incorporating the sliding dichotomy did not produce a significant hemoglobin TT group difference (Table 7, p = 0.73).

GOS-E analyses

TT levels (7 vs. 10 g/dL) were not associated with the GOS-E outcome for either unadjusted or adjusted binary logistic regression with dichotomization of the GOS-E between lower MD and upper SD (Table 5) or for the multiple imputation pooled logistic regression estimates (Table 6). The GOS-E sliding dichotomy unadjusted analysis, using the tertiles given in Table 7,was not significant (p = 0.48).

We examined other possible dichotomizations of our GOS-E outcome data because our data suggested that the increased emphasis on behavioral/cognitive/social and other problems appeared to be affecting the assignment of participants who previously might have an MD rating on the uGOS but a lower SD rating on the GOS-E. We explored the possibility that dichotomizing the GOS-E data between lower SD and upper SD might be most likely to produce a significant hemoglobin TT effect. A dichotomy between lower SD and upper SD produced a significant hemoglobin transfusion trigger effect for both an unadjusted logistic regression with an OR = 0.49 (p = 0.02), and an adjusted logistic regression with an OR = 0.40 (p = 0.01). No other dichotomizations were associated with significant effects. These findings should be viewed with caution given their exploratory nature. Adjustments for multiple comparisons would have resulted in all p-values being statistically non-significant.

Discussion

Our findings add to the discussion of the use of current global outcome measures in research in general and in clinical trials in several ways. Our data set uniquely collected uGOS, GOS, and GOS-E in the same patient population prospectively. The observed discrepancies between the outcome measures were mainly due to the GOS and GOS-E categories being lower than the uGOS, driven by the added behavioral/cognitive/social questions in the GOS and GOS-E structured questionnaires (see Table 3 and its discussion). Consideration of discrepancies in score assignment between the uGOS and the GOS-E appears to support the intent of Wilson and associates¹¹ to develop items that were more sensitive to behavioral/cognitive/social changes affecting outcome.

Several statistical methods reported in the literature were used to analyze data from a single clinical trial data set using the three GOS outcome measures. To some degree, this permitted an examination of the relative usefulness of the various statistical techniques for a particular outcome measure and also a comparison of the findings across measures with a single statistical method. We found a disparity in the significance of the TT randomized group comparison depending on which outcome and statistical method were used. Binary logistic regression of the uGOS data unadjusted for and adjusted for ISS, the IMPACT laboratory model score, and presence of epidural hematoma with a dichotomy between MD and SD indicated a significantly better outcome for the 7 g/dL as compared with the 10 g/dL TT; this was the case also for multiple imputation pooled logistic regression estimates. No reliable effects appeared for the uGOS data with the unadjusted and adjusted ordinal logistic regression or the sliding dichotomy method. TT group effects were not found for the same statistical procedures when the GOS-E and GOS data were analyzed, although the ORs were generally consistent among all outcome assessments.

Strengths

This study has several strengths. (1) Outcomes with uGOS, GOS, and GOS-E measures were collected and findings were compared using data at 6 months post-injury from a randomized clinical trial. (2) Several analytic methods were utilized including unadjusted and adjusted binary and ordinal logistic regression, and a sliding dichotomy with the same data set. (3) Because racial/ethnic/cultural issues can affect the quality and type of information collected,⁴⁰ fully bilingual evaluators (reading, writing, and speaking) were hired given the large Hispanic population in the study.

Limitations

This study has several limitations that should be noted.

• (1)We had many individuals involved in the outcome assessment over the grant period, which could contribute some variation in the information collected, although a standard training procedure was followed. Additionally, evaluatiors themselves may differ in their knowledge of acute stages of TBI, the care of patients in the acute and later stages of recovery and its course, and their telephone as well as in-person interviewing and interpersonal skills. The Spanish translation of the evaluations may have differed because they were not formally validated. However, all staff were trained by the same clinical neuropsychologist over the course of the entire trial. Additionally, all outcome assessments were reviewed and discussed in periodic meetings with the same clinical neuropsychologist throughout the study. It cannot be ruled out that knowledge of one of the scores could have biased the assessment of the other scores during the discussions, but this seems unlikely because each instrument had definitions on administering and scoring that were instrument-specific.
• (2)Our study found differences in the statistical significance of the TT randomization group. TT was only statistically significant for the uGOS analyses as well as one of the dichotomizations of GOS-E, but it is not clear what drove this difference nor what the true association is. It is possible that the GOS-E did not lead to statistically significant results for the TT randomization groups because it focuses on issues that are directly related to the TBI rather than issues that might be more influenced by systemic injuries. However, because the exploratory analyses of the dichotomization cutoff showed a significant difference for GOS-E when the dichotomy was between upper and lower SD, it suggests that the level of disability was simply judged to be more severe with the GOS-E because of the emphasis on behavioral/emotional symptoms. It is also possible that the uGOS underestimates the functional status, compared with the GOS and GOS-E. Usually, the dichotomization of variables throws away some information that results in decreased power and thus less significant results. However, in the current study, the ordinal analysis resulted in non-significant results, whereas the binary analysis was significant for uGOS (Table 5). The proportional odds assumption was not violated, but the power may have been limited for detecting deviations from proportional odds. This could be investigated further in a larger sample. Importantly, many statistical tests have been conducted and therefore false-positive results are possible.
• (3)uGOS, GOS, and GOS-E do not distinguish between functional status as a result of the TBI from the effects of polytrauma. For example, in cases of severe injuries such as having accompanying spinal injury with the TBI, it is hard to assign a GOS that reflects the effect of only the TBI.¹¹ Here, the three GOS measures were collected in a similar manner and patients with severe polytrauma and patients with spinal cord injury were excluded from the trial. Therefore, we believe that this would have minimal effect on the results from this study.
• (4)The same individual collected all three outcomes on an individual subject. However, the staff members who collected outcomes discussed the scores based on each measure with HJH, and therefore there was a level of oversight for outcomes collection.

Conclusion

The choice of outcomes and analysis can lead to incongruous conclusions in a clinical trial. The GOS-E and GOS questions more specifically address behavioral/cognitive/social changes in the assessment of outcome, and resulted in a worse outcome in some cases than the uGOS. The differences in categorizations of subjects between MD and SD among the scales could have impacted the conclusions of the trial. In future studies of severe TBI, particular attention should be drawn to implementing GOS measures and describing the methodology for how outcomes were ascertained.

Author Disclosure Statement

No competing financial interests exist.