Recovery of Cognitive Function After Traumatic Brain Injury: A Multilevel Modeling Analysis of Canadian Outcomes

Cognitive Recovery After TBI: Domain-Specific Outcomes 

In addition to research targeting global outcomes, a number of studies have also employed more specific measures to examine the recovery of discrete cognitive functions after TBI.4, 51, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68 Results from these studies suggest that a wide range of cognitive impairments is apparent immediately after moderate to severe TBI.4, 53, 57, 59, 60 Importantly, however, there exists preliminary evidence that recovery takes place at a differential rate across cognitive domains.5, 53, 59, 61 For example, learning and memory, complex attention and speed of processing, and complex language functions (eg, inferential semantics) appear to recover more slowly and/or less than other functions.53, 55, 59, 68, 69 Conversely, some findings imply that visual perceptual abilities and verbal intelligence recover relatively more quickly.52, 63, 67 It should be noted, however, that the latter finding is confounded by the use of the Wechsler Adult Intelligence Scales, for which verbal and nonverbal intelligence vary in their composition of timed versus untimed tests. Nevertheless, these studies collectively serve to suggest that cognitive domains recover in a separable manner over time.

Recovery as a Function of Time Postinjury 

Past studies have also queried the temporal characteristics of cognitive recovery after TBI. For example, whether the greatest degree of improvement occurs in the early, middle, or late phases of recovery has important implications for the timing of service delivery. In a retrospective study of 876 persons with TBI, Mazaux et al70 reported that early (rather than later) intervention improved outcome. Moreover, the argument that recovery benefits from the proper timing of intervention is also indirectly supported by related research showing that exercise increased neuroplasticity, but only during discrete time windows.71 Similarly, investigators have suggested that different neurobiologic mechanisms may be active at different stages of recovery.72, 73, 74 For example, Kolb74 identified 2 neuroplastic responses after TBI, each with different time courses: an early-onset, “injury-induced” neuroplastic mechanism, which decreases dendritic spine density, and a later-onset, “experience-dependent” mechanism, which increases neuronal spine density and is longer-lasting.

Conversely, there is also evidence that the brain can maladaptively rewire as a function of intervention timing.75, 76 Taub and Morris77 demonstrated that forced use of an impaired limb (eg, paretic arm or leg) during the early stages of recovery can result in poorer recovery. In contrast, DeBow et al78 found that the use of constraint induced movement therapy immediately postinjury results in poorer outcomes in animals. As well, Humm et al79 demonstrated that intensive, very early behavioral stimulation can hinder spontaneous neurophysical and neurochemical repair of the brain after injury. Overall, these findings suggest that the time course of recovery is an important consideration when attempting to deliver therapy at the right time, so as to improve outcome and avert harm.

Methodologic and Statistical Limitations of Past Research 

Although several important findings have emerged in the literature reviewed here on post-TBI cognitive recovery, many central questions remain unanswered. Chief among these questions relates to the general temporal characteristics of cognitive recovery and whether performances in certain cognitive domains recover at differential rates. Contributing to the murky nature of this literature is the lack of comparability, from both design and measurement perspectives, across studies. For example, participant samples have varied significantly across demographic and injury characteristics. Discrepant test batteries have been delivered across disparate testing schedules, spanning from as little as 3 months to as long as 5 years between assessments. Variable sample sizes have resulted in unequal power to detect change across studies. Most previous studies have employed 2 assessments, allowing for the modeling of linear effects only; this is problematic in the face of existing evidence suggesting recovery is best characterized as curvilinear (namely asymptotic). Collectively, these limitations serve to recommend large-scale, prospective studies that employ comprehensive cognitive testing at more than 2 occasions. The primary advantage of such a study is the ability to ascertain the timing and nature of recovery across cognitive domains within a single sample, thereby overcoming problems inherent in cross-study comparisons.

Additionally, past studies have been hampered by their use of customary statistical models of change. That is, the behavioral and clinical sciences have traditionally analyzed change over time as a characteristic of groups at the expense of considering change as a characteristic of persons.80 Methods typically applied to longitudinal data for the purpose of studying mean group differences in change over time include mixed-model ANOVA, multivariate/repeated measures ANOVA, and analysis of covariance. These means-as-outcome approaches regrettably sacrifice the heterogeneity of individual change as a meaningful parameter. One could imagine, for example, the unlikely scenario in which change for half of a sample of subjects, whose performance was above the mean at time 1, was exemplified by a decline, such that their time 2 performance was now equivalent to the scores of the lower half of the sample at time 1. Imagine that, in parallel, the lower scoring half of the sample experienced a similarly robust increase, such that their time 2 performance was now at a level similar to that at which the top half of the sample performed at time 1. In this illustrative example, one can visualize that mean group performance scores would not change over time, yet at the individual level very robust and important change has occurred. A second problem with traditional analytic techniques is that they are optimally designed to describe linear change. This is especially problematic for characterizing recovery from TBI given consistent indications that it is in fact nonlinear and strongly asymptotic.66 An extension of ANOVA models that allows for analysis of curvilinear functions is polynomial trend analysis. However, design restrictions associated with these techniques still limit their overall flexibility. In particular, it is generally necessary that all participants have data at all time points and that participants with missing data be excluded from the analysis. In longitudinal TBI studies, in which missed appointments and dropout rates are high, the exclusion of patients with missing data would result in much smaller and potentially biased samples. Second, these techniques allow the incorporation of discrete, but not continuous, predictors of change.81

Multilevel Modeling: A Solution to Limitations of Previous Recovery Research 

In recent decades, a statistical approach has been developed that overcomes the obstacles described here: multilevel modeling (also known as mixed-effects modeling, hierarchical linear modeling, covariance component modeling, or random-coefficient modeling). This analytic technique was developed to evaluate complex patterns of variability, with a focus on nested sources of variability. A prototypical example of a nested data structure is children nested within families. Multilevel modeling can also be readily applied to longitudinal analyses because repeated observations across time can be conceptualized as nested within a single participant. This approach has several advantages. First, change can be modeled at the individual level by parameters (namely intercepts and slopes plus random error) from individual growth curves (ie, a curve composed of time point data graphed for each individual subject). Employing intercepts and slopes as outcomes circumvents the difficulties inherent in modeling change at the group level only. In addition, however, multilevel modeling provides other notable general advantages, including improved estimation of individual effects, modeling cross-level effects, and partitioning of variance-covariance components. Detailed discussion of these topics is beyond the scope of this article; however, the interested reader is directed to one of the many excellent texts on the subject.81, 82 Such general advantages confer several specific rewards in the case of longitudinal design—namely, greater flexibility in terms of data requirements for variable number and spacing of time points across subjects. Because observations are viewed as nested within persons rather than as the same fixed set for all persons, the number and timing of observations can vary randomly across participants. Moreover, all effects and their interactions are tested simultaneously, thereby allowing for the ascertainment of moderator effects from any given level of the model (eg, individual, group, community). This is generally attractive to the longitudinal researcher because growth trajectories can be conditionalized on third variables emanating from any level of the model. For example, it is feasible to test how individual change varies as a function of level 2 variables (eg, sex, experimental treatment), level 3 variables (eg, community, country of origin), and so on.

Previous Studies of Recovery From Brain Injury Using Multilevel Methods 

A handful of studies have used multilevel modeling for examining the nature of recovery after TBI. Spikman et al66 studied 60 patients with a closed head injury over the course of 1 year postinjury with repeated assessments of attention (ie, Stroop Color-Word Test, Paced Auditory Serial Addition Task, Reaction Time Discrimination Task, Reaction Time Dual Task, Wisconsin Card Sort Perseverative Error Score, and Trail-Making Test). Results indicated that, across most tasks, performance improved for both patient and control groups. (Exceptions to this rule were noted on the perseverative error score and reaction time distraction task, on which control participants failed to show performance change over time.) Overall, patients demonstrated greater recovery than control subjects, which was nonlinear in nature and occurred more quickly than controls within the first 6 months postinjury. In addition, advanced age negatively impacted recovery, and more severely injured patients showed greater improvement than less severely injured patients (although this was at least partially attributable to the lower starting point, allowing more room for improvement).

Similarly, Wong et al67 used multilevel modeling to establish whether PIQ recovers at a different rate than VIQ in a sample of 319 patients with TBI. Results indicated that PIQ recovered at a rate that was almost 4 times slower than VIQ. Moreover, recovery for both domains was asymptotic. Recovery in both domains was moderated by length of coma (with greater impact observed on PIQ versus VIQ), while sex and age of subjects did not add significantly to recovery prediction. More recently, Chu et al54 performed a comparable analysis on patterns of recovery within the verbal memory domain (ie, on the RAVLT). They found considerable variability in initial performance at 1-year postinjury with only modest improvements over the subsequent course of 5 years (ie, an average growth rate of .13 words, of a total possible 80 words a year). Between-subject variation was also large; it was estimated that 84.9% of variation in outcome was between persons while the remaining 15.1% was within persons. Like the results of Spikman et al,66 results from this study demonstrated that both age and length of PTA significantly moderated verbal memory recovery at 1 year postinjury. Sex and performance during acute rehabilitation were not significant predictors of verbal memory performance at the 1-year point.

Collectively, the studies reviewed here converge to reveal that recovery from TBI is largely asymptotic over the first year postinjury and significantly moderated by age and injury severity. They also demonstrate the value of multilevel modeling for investigating these phenomena. An important drawback, however, is that each study investigates a relatively narrow domain of cognitive recovery (ie, attention, IQ, or memory). Yet, data from several of these experiments imply that recovery across domains is uneven. Therefore, the main purpose of the current study was to use multilevel methods to examine differential recovery curves as a function of cognitive domain within the first year postinjury of a TBI.

Participants 

The 75 patients with TBI in this study were recruited to a larger study investigating the natural history of cognitive and motor recovery in the inpatient, Neurorehabilitation Program of the Toronto Rehabilitation Institute, a large, urban, Canadian rehabilitation hospital. Inclusion criteria for the study were as follows: (1) acute care diagnosis of TBI, (2) PTA of 1 hour or more and/or GCS score of 12 or less either at their emergency admission or the scene of accident and/or positive computed tomography or magnetic resonance imaging findings, (3) age between 18 and 80 years, (4) able to follow simple commands in English based on the speech language pathologist intake assessment, and (5) competency to provide informed consent for study or availability of legal decision-maker. Exclusion criteria included (1) orthopedic injuries affecting both upper extremities and/or both lower extremities (relevant to the larger study, which was investigating motor recovery as well); (2) diseases primarily or frequently affecting the central nervous system, including dementia of Alzheimer type, Parkinson disease, multiple sclerosis, Huntington disease, lupus, and stroke, ascertained via medical records and/or screening of family members for patients over 60 years regarding any definite or possible prior diagnosis of dementia; (3) history of psychotic disorder; (4) failure to emerge from PTA by 6 weeks postinjury, as measured by the Galveston Orientation Amnesia Test; (5)TBI secondary to other brain injury (eg, a fall caused by stroke); and (6) failure on a test of symptom validity (Test of Memory Malingering)83 at any of the assessments.

The present sample was predominantly composed of young to middle-aged (mean age ± SD, 37.37±15.49y) men (80%) with a high school education (mean level of education ± SD, 12.71±2.78y). Most participants were injured as a result of a motor vehicle collision (55.7%), followed by falls (32.9%), assaults (8.69%), and sports injuries (2.9%). On average, participants had experienced a moderate to severe brain injury (mean lowest/GCS score ± SD, 6.97±3.59), which resulted in an approximately 40-day acute care hospitalization (mean acute care length of stay ± SD, 38.03±17.17d). Premorbidly, most participants were employed as minor professionals or technical workers (35.7%), followed by machine operators or semiskilled workers (31.4%), skilled craftsmen or clerical staff (20%), professionals (10%), and unskilled laborers (1.4%). The average ± SD estimated premorbid IQ (indexed via the North American Adult Reading Test84 or Wechsler Test of Adult Reading85) for the sample was 100.43±12.51.

Materials 

Cognitive tests were selected to measure the following cognitive domains: premorbid IQ,85 language skills,86, 87 visuospatial skills,86, 87 verbal and visuospatial attention/concentration,88 speed of processing,1, 89, 90 learning and memory,1, 88 executive functioning,1, 86, 87 and general intellectual function.86, 87 Table 1 lists the cognitive tests used in the current study. All tests have demonstrated adequate validity and reliability for brain-injured populations.1, 91 In some cases, comprehensive assessment within a domain (eg, executive function) could not be achieved with standardized, clinical tests; in such cases, experimental tests were used (eg, Sustained Attention to Response Test,92 Modified Hayling Sentence Completion Task—computerized administration). Tests were selected to have minimal timed, manual motor demands; where unavoidable (eg, choice reaction time and Trail-Making Test part B), subtraction tests (eg, simple reaction time; Trail-Making Test part A) were used to control for motor contributions, and alternate forms were employed to minimize practice effects. The battery required approximately 4.5 hours to administer to severely brain-injured patients.

Table 1. Neuropsychological Tests Administered and Corresponding Cognitive Domains

	Cognitive Domain	Neuropsychological Test	Reference
	All Memory	WMS-III Logical Memory I	Wechsler, 1997⁎88
		WMS-III Logical Memory II	Wechsler, 1997⁎88
		RAVLT–Long Delay	Lezak, 1995‡1
		RAVLT–Short Delay	Lezak, 1995‡1
		RAVLT–Total	Lezak, 1995‡1
		Rey Visual Design Learning Test–Total	Lezak, 1995‡1
	Attention Span	WMS-III Digit Span Forward	Wechsler, 1997⁎88
		WMS-III Visual Span Forward	Wechsler, 1997⁎88
	Executive General	Stroop Color-Word Test–Interference	Lezak, 1995‡1
		WAIS-III Matrix Reasoning	Wechsler, 1997†87
	Executive Timed	Symbol Digit Modalities Test (oral)	Smith, 198289
		Verbal Fluency	Lezak, 1995‡1
	Executive Working Memory	WMS-III Digit Span Backward	Wechsler, 1997⁎88
		WMS-III Visual Span Backward	Wechsler, 1997⁎88
	Language	WAIS-III Vocabulary	Wechsler, 1997†87
	Manual Motor	Grip Strength	Lezak, 1995‡1
	Logical Memory	WMS-III Logical Memory I	Wechsler, 1997⁎88
		WMS-III Logical Memory II	Wechsler, 1997⁎88
	Memory–RAVLT	RAVLT–Long Delay	Lezak, 1995‡1
		RAVLT–Short Delay	Lezak, 1995‡1
		RAVLT–Total	Lezak, 1995‡1
	Visual Memory	Rey Visual Design Learning Test–Total	Lezak, 1995‡1
	Motor Speed	Grooved Pegboard	Lezak, 1995‡1
		Trail Making Test part A	Lezak, 1995‡1
	Speed of Processing Motor	Symbol Digit Modalities Test (written)	Smith, 198289
		Trail-Making Test part B	Lezak, 1995‡1
	Speed of Processing Simple	Stroop Color-Word Test–Color Naming	Lezak, 1995‡1
		Stroop Color-Word Test–Word Naming	Lezak, 1995‡1
	Verbal Abstraction	WAIS-III Similarities	Wechsler, 1997†87
	Visuospatial	WAIS-III Block Design	Wechsler, 1997†87
		WAIS-III Matrix Reasoning	Wechsler, 1997†87

Abbreviation: WMS-III, Wechsler Memory Scale–3rd Edition.

⁎ Refers to reference: Wechsler D. Wechsler Memory Scale–3rd Edition (WMS–III). San Antonio: Psychological Corp; 1997. † Refers to reference: Wechsler D. Wechsler Adult Intelligence Scale–3rd Edition (WAIS-III). San Antonio: Psychological Corp; 1997. ‡ Refers to reference: Lezak MD. Neuropsychological Assessment–3rd edition. New York: Oxford Univ Pr; 1995.

Procedure 

The current study employed a prospective, repeated-measures design. Patients were tested at 2, 5, and 12 months postinjury. The cognitive battery was divided into 5 blocks of tests, with a fixed order of tests within each block designed to minimize interference between tests (eg, verbal memory test contained nonverbal tests between learning and delayed recall phases). Test blocks were matched as much as possible for the number of timed tests and effortful tests. Each block contained a maximum of 1 memory test. Block order was counterbalanced, but each participant received the same block order across testing sessions. Cognitive tests with known practice effects contained 2 or more alternate forms. In some cases, the same form was administered a second time (ie, where only an original plus 1 alternate form was available); however, this occurred only between the 2-month and 12-month assessments, and thus a gap of approximately 10 months separated the 2 occasions of testing. Order of alternate forms was counterbalanced across subjects. The 2-month testing window ranged from 1.5 to 2.5 months postinjury and took place during the inpatient stay. Neuropsychological assessment was administered over a maximum 72-hour period, with individual testing sessions ranging from 0.5 hour to 3 hours, as tolerated by the patient. The 5-month window ranged from 3.5 to 5.5 months postinjury, and all testing took place over a 2-day period. The 12-month window ranged from 11 to 13 months postinjury, and again, all testing took place during the same 2-day period.

Data Analysis 

Multilevel modeling 

The longitudinal multilevel model used in the analysis of the first 3 waves of data has the following form:

where Yit is an outcome of a test administered to the ith subject on the tth occasion at time Tit measured in months postinjury. The expression (Tit–5) is equal to 0 if Tit is less than 5 and to Tit–5 otherwise. Consequently, the expression

represents a linear spline in T with a knot at 5 months: β0i is the expected level of Y for the ith subject at month 0, βli is the expected rate of recovery a month before the fifth month, and βli is the change in the expected recovery rate at the fifth month. Thus, (βli+βli) is the expected recovery rate after the fifth month. The error term, εit, represents random variability in test results from one occasion to the other for a given subjects. Its expected magnitude is related to the test-retest reliability of a particular test.

Longitudinal multilevel models allow the intercept β0i and the slopes βli and βli to vary randomly from subject to subject. As discussed, domains whose components are externally standardized can be treated as if they are externally standardized themselves, thus allowing comparisons between domains. Various aspects of the trajectories of different domains can be compared, including initial impairment, the rate of recovery, and impairment at 1 year postinjury. In addition, pertinent baseline differences between participants were statistically controlled for by modeling these variables as covariates. Covariates were chosen based on 2 criteria: (1) they demonstrated a significant, unique (ie, jointly significant partial correlations) association with the outcome variable of interest (ie, cognitive domain) and (2) were rationally deemed to be time-invariant. Consequently, covariates diverge as a function of the cognitive domain being tested. Broadly speaking, covariates constellations were composed of a selection of the following 3 variables: age, estimated premorbid IQ, and length of acute rehabilitation hospitalization. This procedure was an attempt to minimize baseline differences rather than assess the impact of these variables on recovery curves (for data regarding the impact of moderator variables on recovery curves, see Green93). For purposes of statistical inference, multiple domains form a multivariate multilevel response vector whose analysis using likelihood-based methods is much more complex than that of a univariate response variable. A relatively simple and robust method for between-domain comparisons is based on an analysis that uses bootstrapping. The subject-to-subject variability of interdomain comparisons was estimated by using 4000 bootstrap resamples of the subjects. The P values for these comparisons are then adjusted for multiple comparisons using the Bonferroni-Holms method. This approach allows valid comparisons between domains that generalize to the population from which the study subjects are deemed to have been drawn.

Study Limitations 

Limitations of the current study include its relatively modest sample size. As a consequence, power to detect statistical differences and the ability to produce stable parameter estimates may have been compromised. This possibility is bolstered by the fact that several domains with similarly large recovery slopes did not survive correction for multiple comparisons. This raises the possibility that, with increased participants, recovery slopes from several more domains would be revealed as differentially impacted. Taken together, these observations serve to caution the reader against interpreting the current data as perfectly valid in the absence of replication from independent samples. In addition, a potential confound of the current analyses is variability across domains of inherent measurement error. Unreliability serves to bias the measurement of change such that unreliable measures frequently underestimate the magnitude of change when raw scores are converted to standard scores. Consider a given expected change in a raw score measure. If the measure has high reliability, then its variance will come mainly from interperson variability. If it has low reliability, the interperson variability will be inflated by measurement error. Because the standard score in each case is obtained by dividing the raw score by its SD, the expected difference for the raw score with low reliability will transform into a smaller difference in standard scores. Therefore, in order to estimate the effect of reliability on slope parameters in the current analysis, test-retest reliability coefficients were computed for each domain. Next, reliabilities were correlated with slope parameters from the 2-month to 5-month recovery epoch. Reliabilities were only modestly and nonsignificantly correlated with slopes from this section of the recovery curve (r=.124; P=.752), undermining the likelihood that the variability in magnitude of slopes could be entirely accounted for by differences in domain reliabilities.