Implicit Causality Bias in Adults with Traumatic Brain Injury

Abstract

Introduction

Individuals with moderate or severe traumatic brain injury often experience impairments in pragmatic language functions. Pragmatic language has been studied primarily in connected language genres such as narratives. It may be, however, that individuals with traumatic brain injury also miss microscopic cues, such as social cues embedded in single word meanings or sentence structure. The current study examined one type of sentence-level pragmatic language cue: implicit causality bias. Implicit causality bias is the attribution of an interpersonal transitive verb action to either the subject noun phrase or object noun phrase of a sentence, and is an inherent property of English-language verbs.

Method

In this study, 19 adults with traumatic brain injury and 18 typical adults were asked to provide sensible and spontaneous completions to 96 sentence fragments. Each fragment contained one interpersonal transitive verb and two noun phrases to which the cause of the verb could be attributed.

Results

Adults with traumatic brain injury showed significantly less implicit causality bias than typical adults, and also made more errors in assigning the causality of a clause.

Conclusions

These results challenge assumptions regarding intact implicit processing in adults with traumatic brain injury, and reveal mechanisms by which communication could fail in everyday social interactions.

1: Introduction

Individuals with moderate or severe traumatic brain injury (TBI) often show impairments in pragmatic aspects of communication (Coelho, Liles, & Duffy, 1994; Mentis & Prutting, 1987; Milton & Prutting, 1987), particularly in pragmatic inference (Dennis & Barnes, 1990; McDonald, 1993). Pragmatic inference is the process of integrating context cues and knowledge, to add meaning to information that has been explicitly provided. A common pragmatic inference problem for people with TBI is understanding speech acts such as sarcasm, which requires the listener to infer that a speaker means the opposite of what he or she says (McDonald, 1992; McDonald, 1999; McDonald, 2000). Individuals with TBI also may miss the implied gist of narratives (e.g., missing the moral of a story), or fail to bridge semantic gaps (e.g., hearing “The children were cooking dinner” and then “The family had to order pizza,” and inferring that something must have happened to the dinner; Chapman et al., 2006; Dennis & Barnes, 1990). Comprehension of sarcasm and other implied information requires elaborative inference, i.e., inferring meaning via controlled and strategic use of information such as world knowledge and social norms (Johnson & Turkstra, 2012). Elaborative inference is described as having a high cognitive load (Swinney & Osterhout, 1990), because the person must access information outside of current working memory contents and hold and manipulate this information in working memory (Moran & Gillon, 2005; Swinney & Osterhout, 1990). Given the prevalence of working memory impairments among adults with TBI, it is not surprising that elaborative inference is often impaired in this group, both on laboratory tasks (Bibby & McDonald, 2005; Channon & Watts, 2003; Dennis & Barnes, 2000; Dennis, Purvis, Barnes, Wilkinson, & Winner, 2001; Ferstl, Guthke, & von Cramon, 2002; Martin & McDonald, 2005; Moran & Gillon, 2005) and in conversations with everyday partners (Johnson & Turkstra, 2012), and that inference accuracy has been linked explicitly to working memory scores (e.g., Dennis & Barnes, 2000; Moran & Gillon, 2005; Turkstra, 2008).

By contrast, automatic inferences are low-cognitive-demand inferences that require little world knowledge or strategic use of information, and are made rapidly and accurately by typical adults (McKoon, Greene, & Ratcliff, 1993; Swinney & Osterhout, 1990). Automatic inferences include presupposition (e.g., understanding that the phrase “Jane knows that lunch was delivered” entails that lunch was delivered; Dennis & Barnes, 2000), and also implicit semantic associations, such as that the phrase “Jane knows that lunch was delivered” implies that it is lunch-time (McKoon et al., 1993). There have been very few studies of automatic inference in TBI, and results are mixed. Bergemalm and Lyxell (2005) tested automatic inference by asking adults with or without TBI to complete sentences with missing words (e.g., May…order…later?), and found no significant difference between groups. Sentences were common phrases from restaurants or shops, however, and might have been known to participants, so results could have reflected cued recall of social scripts rather than inference. Johnson and Turkstra (2012) measured inference in extemporaneous conversations between adults with or without TBI and their self-nominated familiar partners, and again found no difference between groups and very low error rates overall. As in the Bergemalm and Lyxell (2005) study, however, familiarity of topics could have provided cues to support inference. Interactions were in person, so non-verbal cues also could have played a role, although elaborative inference errors were still significantly higher in the TBI group.

In a third study, Dennis and Barnes (2000) asked children with mild or severe TBI to judge the truth of sentences that contained factive predicates such as know and realize, as in the example in the previous paragraph; non-factive predicates (e.g., “Jane believes that lunch was delivered” does not entail that lunch was delivered); implicative predicates (e.g., “Jane remembered that lunch was delivered” entails that lunch was delivered); and non-implicative predicates (e.g., “Jane wants to remember that lunch was delivered” does not entail that lunch was delivered). Inferences based on these predicates are automatic for school-aged children, as evidenced by near-ceiling scores in the typically developing children included in the study (M = 22.9/24). Accuracy was significantly lower in children with severe TBI (M = 18.6), but not in those with mild TBI (M = 20.4), and pragmatic inference scores were significant predictors of scores on a standardized test of speech acts. That is, word-level automatic inference processes were impaired, and predicted performance on a connected language test.

In summary, there is strong evidence that elaborative inference processes can be impaired in individuals with TBI, but the story for automatic inference is less clear. As automatic inferences are a key component of everyday communication, and errors could lead to communication breakdowns, it was important to address this gap in knowledge.

In this study, we examined automatic inference in response to cues embedded in single words, specifically in verbs. Many verbs naturally embody underlying causes, accompaniments, and results that can appear in any context without being formally introduced or defined (Chafe, 1972). Understanding these features of verbs is critical in everyday communication. Many simple, active sentences are composed of two noun phrases (NPs), which are either nouns or constructions that function syntactically as nouns. Simple transitive sentences follow the form subject-verb-object, or NP1-verb-NP2 (e.g., “Susan helped David.”). In this grammatical context, certain verbs imply attributes to either the subject (NP1) or object (NP2) of the sentence. For example, the cause of the sentence “The mother scolded her son” is not explicitly stated. However, the verb “scold” naturally implies that the second noun phrase (NP2) is the cause of the statement, specifically that the son’s behavior is the cause of his mother’s scolding (Ferstl, Garnham, & Manouilidou, 2011). Sentence (1) below is an example in which the cause of the verb is attributed to the second noun phrase (NP2), which is the expected attribution; whereas sentence (2) contradicts expectations by attributing the cause of the verb to the mother, thereby treating the verb as a NP1 type verb (Garvey & Caramazza, 1974).

1. The mother scolded her son because he (NP2) admitted his guilt.

2. The mother scolded her son because she (NP1) discovered his guilt.

The phenomenon shown in this example is referred to as implicit causality (IC), which reflects intuitions about the causal agent of an event that are embedded within a verb’s semantics (Hartshorne, 2014). These intuitions have been interpreted as linguistic structures (e.g., semantic features of verbs), reflections of general world knowledge, and high-level cognitive functions (Ferstl, Walther, Guthke, & von Cramon, 2005; Hartshorne, 2014). Work by Rohde and colleague’s shows that IC verbs yield expectations about discourse direction, particularly regarding explanations and elaborations (Rohde, Kehler, & Elman, 2006). They have demonstrated that surface-level accounts of pronoun resolution and interpretation do not account for the range of findings in the literature (e.g., NP1 preference: Crawley, Stevenson, & Kleinman, 1990; grammatical parallelism: Sheldon, 1974; thematic role preference: Stevenson, Crawley, & Kleinman, 1994). However, pronoun biases unexplained by grammatical, thematic, or event-related accounts, can be modeled by the unfolding discourse (Rohde, Kehler, & Elman, 2007). The term “bias” is often paired with IC because it is necessary to make judgments about a statement’s cause in order to measure IC (Ferstl et al., 2011). IC bias is a unique construct because it links a conceptual network of causal relationships to the concrete linguistic process of pronoun resolution, or determining the pronoun to which a verb refers (Hartshorne & Snedeker, 2013).

Accurate IC bias is an essential skill for comprehending text and spoken language (Caramazza, Grober, Garvey, & Yates, 1977; McKoon et al., 1993). To understand language in everyday contexts, we must have the skills to understand causal relations between events and states described within that language (Ferstl et al., 2011; Stuss & Alexander, 2000). Garvey and Caramazza (1974), who first introduced IC bias, found they could predict comprehension times of sentences based on their causal attribution. For example, people make faster judgments for sentence (1) than for sentence (2) above, because the first sentence follows natural expectations whereas the second sentence violates those expectations. These same skills are employed in inference making and in general comprehension tasks required for routine social interactions, online comprehension tasks, and judgment tasks (Caramazza et al., 1977; Ferstl et al., 2011; McKoon et al., 1993). As the aforementioned studies suggest, IC bias is a crucial aspect of language in interpersonal conversation, written narration, and comprehension and interpretation of text.

Given the mixed findings about automatic inference in adults with TBI, it is possible that IC bias is impaired in this group. If so, this impairment could contribute to pragmatic communication problems. Specifically, an individual with TBI might misunderstand a sentence because he or she incorrectly attributes the agency of a verb, and this misattribution could lead to social errors. Consider the following example: “The boss criticized the employee.” This is a NP2-biased sentence, because the verb “criticized” implies that the NP2 (i.e., the employee) was the most likely agent to cause the verb (or the criticism) to occur. A person with TBI might misinterpret this statement as being the boss’s problem and think that no action is required on the part of the employee. Such a misunderstanding in a workplace could appear disrespectful toward an authority figure and show lack of initiative to improve (if the person with TBI is the employee). Instead, it would be more appropriate to interpret the statement as motivation for the employee to improve his or her work performance. This example highlights the degree to which a misinterpretation of IC bias can affect social communication.

Less overall detection of IC bias, or less accuracy in detecting IC bias, would not only help explain pragmatic communication problems in adults with TBI, but might also support the use of IC tasks in clinical assessment. In addition, everyday communication partners could be made aware of potential differences between their interpretation of causality and those of their partners with TBI, which could support more effective communication. Thus, we asked if adults with TBI would show IC bias and if not, what their responses would be. The hypotheses were that:

1. adults with TBI would show less IC bias than a comparison group of uninjured adults; that is, adults with TBI would be less likely than their uninjured peers to assign bias to either NP category; and

2. for items in which adults with TBI did assign bias, they would be significantly less likely than their uninjured peers to align with predicted NP1 and NP2 biases.

2: Material and methods

2.1: Participants

Participants were 19 adults with TBI and a comparison group of 18 uninjured adults. All participants were recruited from the Midwestern United States. Demographic data are listed in Table 1. Inclusion criteria were: age 18–65 years, hearing and vision adequate to complete study tasks, and no history of neurological disease affecting the brain, or language or learning disability (premorbidly for participants in the TBI Group). Participants were all native English speakers who spoke English as their primary language and had oral language skills sufficient to complete study tasks. In addition, all TBI group participants had moderate-to-severe injuries (see Table 2), based on nationally accepted criteria (Malec et al., 2007), were more than 6 months post injury, and were out of post-traumatic amnesia. Two participants in the TBI group had more than one injury, and time post-injury was calculated from the first injury. Participants were excluded if they had severe dysarthria or aphasia, defined as a score of less than 93.8 on The Bedside Western Aphasia Battery – Revised (Kertesz, 2006). Participants were matched group-wise for age, race, sex, and education (Table 2). There was a marginally significant between-groups difference in age, t(35)=1.66, p=0.06. All participants self-identified as Caucasian.

Table 1.

Participant characteristics.

	TBI Group n=19	Comparison Group n=18
Average Age (Years; Months)	42;10	36;3
Females:Males	9:10	10:8
Highest Level of Education Completed
High School	0	2
Some College	7	2
Associates Degree	4	3
Bachelor’s Degree	7	10
Graduate Degree	1	1
WAIS-IV PSI Scaled Score	84.37 (15.04)	103.06 (19.63)
M(SD)
CVLT II Delayed Recall	−1.19 (1.12)	0.58 (0.81)
M(SD)
Trail-Making Subtest B	−0.68 (1.85)	0.30 (2.00)
M(SD)
LCQ-Self	65.32(13.75)	52.13(9.42)
M(SD)
LCQ-Other	66.21(17.21)	47.88(13.68)
M(SD)

Notes: TBI = traumatic brain injury; WAIS-IV PSI = Wechsler Adult Intelligence Scales – Fourth Edition Processing Speed Index; CVLT-II = California Verbal Learning Test – Second Edition. LCQ = La Trobe Communication Questionnaire.

Table 2.

Injury characteristics of participants with TBI.

Demographic and Injury Characteristics					Cognitive Test Measures			Implicit Causality Measures
Participant ID	Age (Yrs;Mos)	Time Post TBI (Yrs;Mos)	Injury Mechanism	Injury Severity	WAIS-IV PSI Scaled Score	CVLT II Delayed Recall	Trail-Making Subtest B	NP1 Bias	NP2 Bias	No Bias
TBI 1	50;11	32;2	MVA	Severe*	59	−0.50	N/A	36	40	20
TBI 2	25;3	9;9	MVA	GCS = 3	111	1.50	2.19	22	68	6
TBI 3	36;4	14;7	Assault	N/A	114	1.00	0.13	23	68	5
TBI 4	37;2	19;6	MVA	N/A	102	0.50	0.04	37	53	6
TBI 5	26;11	8;8	MVA	GCS = 6T	83	0.00	−0.29	34	49	13
TBI 6	42;11	8;5	MVA	GCS < 8	84	−0.50	−0.23	21	60	15
TBI 7	59;2	1;3	Fall	GCS = 14	94	1.00	0.74	53	32	11
TBI 8	27;0	10;8	MVA	GCS = 6	120	0.00	0.09	29	60	7
TBI 9	44;5	12;11	Fall	GCS = 9	87	−1.00	1.53	32	55	9
TBI 10	58;6	5;2	MVA	N/A	102	−4.00	1.74	24	62	10
TBI 11	52;9	35;8	Sports	N/A	97	1.00	1.75	46	36	15
TBI 12	28;8	4;3	MVA	GCS = 3T	76	−1.50	−3.86	36	20	40
TBI 13	45;6	4;6	MVA	GCS = 3T	79	−0.50	−0.83	16	28	52
TBI 14	54;11	43;2	Assault	N/A	92	−1.00	−2.02	32	40	24
TBI 15	48;3	15;7	MVA	GCS = 4	94	1.00	1.43	44	26	26
TBI 16	34;6	15;11	MVA	Severe*	76	−1.50	−4.85	41	51	4
TBI 17	58;7	4;4	Fall/Hit	N/A	102	1.00	−4.91	29	46	21
TBI 18	33;3	12;8	MVA	GCS = 5	79	0.00	−1.85	28	36	32
TBI 19	50;5	5;0	MVA	N/A	89	1.00	0.01	45	48	3

Notes: MVA= motor vehicle accident. GCS = Glasgow Coma Scale score. N/A = No information pertaining to severity of TBI in medical records; Subject TBI 1 did not have Trail-Making Subtest B score.

^*Per physician notes.

2.2: Characterizing the Sample

As recommended by the Interagency Traumatic Brain Injury (TBI) Outcomes Workgroup (Wilde et al., 2010), participants completed several cognitive tests to characterize the sample for comparison to other TBI studies. Results are shown in Table 1. Groups differed significantly on the Wechsler Adult Intelligence Scales—Fourth Edition (Wechsler, 2008) Processing Speed Index (WAIS-IV PSI; t(35)=3.26, p<.001; and delayed recall on the California Verbal Learning Test II (CVLT II; Delis, Kramer, Kaplan, & Ober, 2000), t(35)=5.49, p<0.001; and a marginally significant difference on the Trail Making Test Subtest B (Tombaugh, 2004), t(35)=1.52, p=0.07.

Participants and their nominated close others completed the La Trobe Communication Questionnaire (LCQ; Douglas, O'Flaherty, & Snow, 2000), a measure of perceived communication ability. While there is no standard criterion for impairment on the LCQ, scores for 16 of 19 individuals in the TBI group were more than 1 SD higher than the control group average, either by self-report (n=11) or other-report (n=5). The remaining three had no other-report and their self-reports were within 1 SD of the comparison group average.

2.3: Experimental Stimuli

Stimuli were 96 English interpersonal transitive verbs (i.e., action verbs that express a ‘doable’ activity and have a direct object or person who receives the action of the verb) that were categorized by word length, frequency, and emotional valence in a previous study conducted by Ferstl and colleagues (2011). Average age of acquisition (AoA) was age 8.17 years (SD = 2.72 years), according to a megastudy of AoA ratings for English content words (Kuperman, Stadthagen-Gonzalez, & Brysbaert, 2012). Verbs represented four categories from the revised action-state taxonomy of Rudolph and Fösterling (Rudolph, 1997; Rudolph & Főrsterling, 1997), who grouped verbs by criteria such as whether they were observable, i.e., feel or experience vs. do (e.g., disappoint vs. slander); and attribution to the subject vs. the object of the NP, i.e., NP1 vs. NP2. Rudolph and colleagues divided interpersonal verbs into four categories: agent-patient (e.g., kissed), agent-evocator (e.g., criticized), stimulus-experiencer (e.g., charmed), and experiencer-stimulus (e.g., adored). The current experiment consisted of 24 verbs in each of these four categories. We used a MANOVA to compare these four verb categories on verb length, frequency (using CELEX database; Baayen et al., 1995), or valence (positive vs. negative), and there was no significant difference according to category, F(3, 92)=1.06, p=0.39.

To create the NP stimuli, we chose 96 female and 96 male most-common baby names of 2012 in the United States, as ranked by the Official Social Security Administration. Seven adults reviewed this name list to identify gender-ambiguous, extremely unusual, and old-fashioned names, which were then deleted, resulting in a list of common names that could clearly be assigned to binary gender pronouns. We then randomly assigned one female (F) and one male (M) name to each of the 96 verbs, to construct sentences in the format: (1) “M verbed F because…” and (2)“F verbed M because…” (see Table 3). Two stimulus lists were created such that each verb appeared once per list and each list had an equal number of female NP1s and male NP1s. Since half of the sentences had a female NP1 and half had a male NP1, stimuli were counterbalanced to account for any potential sex-based differences in IC bias. By using mixed-gender pairs, we could clearly identify the pronoun signaling the entity selected as the causal factor. That is, a response using “she” signaled that the causal factor was the female (F) NP, and “he” signaled that the causal factor was the male (M) NP. Each stimulus was in the format NP1 + Verb + NP2 + Prompt (“because…”).

Table 3.

Sample stimulus characteristics.

	NP1	Verb	NP2	Semantic Category	Anticipated NP Bias Score
1.	Amelia	criticized	Jason	Do	NP2
2.	Michael	frightened	Lily	Feel/Experience	NP1
3.	Jennifer	loved	Andrew	Feel/Experience	NP2
4.	Christopher	scorned	Michelle	Do	NP2
5.	Olivia	interrupted	Thomas	Do	NP1
6.	William	married	Allison	Do	NP1
7.	Sophia	appalled	Samuel	Feel/Experience	NP1
8.	Patrick	helped	Emily	Do	NP2
9.	Victoria	comforted	David	Feel/Experience	NP2
10.	Brandon	appreciated	Anna	Feel/Experience	NP2
11.	Sarah	resented	John	Feel/Experience	NP1
12.	Jacob	blamed	Samantha	Do	NP2

Notes: NP = noun phrase.

2.4: Procedure

The two counterbalanced stimulus lists of 96 items were each presented in two different randomized orders, creating four versions of the experimental material. Stimuli were presented via a web-based questionnaire powered by Qualtrics research software. Participants were tested individually in a laboratory setting, controlled to have minimal distractions. Participants saw each stimulus on a computer screen (e.g., “Michael valued Amy because…”), and were required to type a response before proceeding to the next item. Participants were instructed to provide “sensible and spontaneous completions” to each item. Participants were instructed not to revise their answers and were told there was no time limit to the task. Task completion time was 30–60 minutes, depending on participant response speed. If participants had motor dexterity impairments, examiners typed their responses. The University of Wisconsin-Madison institutional review board approved all procedures and participants were paid for their time.

2.5: Data Analysis

Participant responses beginning with a pronoun “he” or “she” were scored as NP1-biased or NP2-biased based on the corresponding verb in the fragment. In addition, no-bias continuations were defined as sentence continuations that were neither NP1- nor NP2-biased. Examples of no-bias responses included continuations with the pronoun “they” or general continuations with no agent (e.g., “…because it was fun.”). Percent accuracy of NP continuations was calculated for each participant, with accuracy defined as percent agreement with data collected by Ferstl and colleagues (2011). To compute inter-rater reliability, two raters scored all responses. Reliability was 98% and all disagreements were resolved by discussion.

2.6: Statistical Analysis

The first hypothesis – that adults with TBI would assign less IC bias than the comparison group – was tested using a two-sample t-test comparing TBI and uninjured groups on number of no-bias items. The second hypothesis – that adults with TBI would assign IC bias less accurately than the comparison group – was tested using a 2x2 analysis of variance on all continuations that included a bias (i.e., NP1 and NP2 continuations), with a between-subjects factor of group (TBI, comparison) and a within-subjects factor bias type (NP1, NP2 continuations) as predictors and percent correct as the dependent variable. The criterion p value was .05 and effect sizes were calculated using Cohen’s d. Data were analyzed using Stata™ statistical analysis software. We did not expect age effects, as stimuli were words mastered in early childhood, but given the marginally significant group difference in age we calculated the correlation of age with all dependent variables prior to testing study hypotheses.

3: Results

Age was not significantly correlated with the either the production of any continuation type (NP1: r=−0.11, p=0.51; NP2: r=−0.02, p=0.88; No Bias: r=0.11, p=0.50), or overall accuracy of continuations produced (r=−0.05, p=0.77). As the first hypothesis predicted, there were significantly more no-bias responses in the TBI group than the comparison group, t(35)=2.75, p=0.007, d=0.83. Accuracy scores supported the second hypothesis, showing a significant main effect of group, with higher scores in the comparison group the TBI group, F(1, 188)=28.71, p<0.001, d=0.73; and a marginal main effect of bias type, with higher scores on NP2 items than NP1 items, F(1, 188)=4.07, p=0.045, d=0.29. There was no significant interaction of group by bias type, F(1, 188)=0.70, p=0.40, and there was no a priori expectation that there would be. Results for each verb are presented in the Appendix. Across the 96 verbs, there was a slight bias for both groups to produce a NP2/object continuation (e.g., Johnny harmed [NP1 bias] Susie because she [NP2 completion] stole his ball). This tendency was also found in a larger corpus of 720 verbs examined by Hartshorne and Snedeker (2013). Descriptive results of continuation types produced by participant groups are shown in Table 4. Analysis of variance results are shown in Table 5.

Table 4.

Continuations produced by participants in the TBI and comparison groups. Data are mean percentages with SDs in parentheses.

Continuation Type		TBI Group	Comparison Group
Analysis 1	NP1 Bias	0.33 (0.08)	0.37 (0.11)
	NP2 Bias	0.48 (0.12)	0.54 (0.13)
	No Bias	0.19 (0.14)	0.09 (0.07)
Analysis 2	NP1 Bias	0.41(0.24)	0.40(0.30)
	NP2 Bias	0.59(0.25)	0.60(0.30)
Percent Correct	NP1 Bias	0.63(0.19)	0.72(0.18)
	NP2 Bias	0.72(0.17)	0.77(0.17)

Notes: The denominator for Analysis 1 was total number of continuations (96). The denominator for Analysis 2 was total number of biased continuations (i.e., excluding No Bias responses).

Table 5.

Analysis of variance

Source	Partial SS	df	MS	F	p	d
Bias	1136.83	1	1136.83	4.07	0.045*	0.29
Group	8016.60	1	8016.60	28.71	<0.001***	0.73
Bias* Group	195.85	1	195.85	0.70	0.403
Residual	52489.92	188	52489.92

Notes:

^*. Correlation is significant at the 0.05 level (2-tailed).

^**. Correlation is significant at the 0.01 level (2-tailed).

^***. Correlation is significant at the 0.001 level (2-tailed).

To explore potential links between cognitive functions and bias performance, we correlated IC measures with cognitive test scores. The criterion p value was set to 0.005 (.05/9) to correct for multiple comparisons. There were no significant correlations between IC measures and cognitive test scores (Table 6).

Table 6.

Pearson correlations between implicit causality (IC) performance and cognitive tests.

	WAIS-IV PSI Scaled Score		CVLT II Delayed Recall		Trail-Making Subtest B
	r	p	r	p	r	p
No Bias	−0.242	0.143	−0.103	0.537	−0.278	0.095
NP1 Bias	−0.137	0.411	0.251	0.129	−0.093	0.584
NP2 Bias	0.332	0.589	−0.087	0.240	0.331	0.592

4: Discussion

Pragmatic communication deficits in individuals with TBI have been well documented in narratives and conversations, but little is known about automatic inference and word-level pragmatic processes. The current study focused on response to implicit causality (IC) cues, a word-level process that involves automatic inferences and could play a role in pragmatic communication impairments in adults with TBI. IC is a well-established feature of interpersonal transitive verbs that create a bias to reference causally implicated subjects. Adults with TBI were significantly less likely than their uninjured peers to respond to IC bias, and when they did they were more likely to make causal attribution errors. Results were consistent with previous findings of Dennis and Barnes (2000), who found that children with severe TBI were less accurate than their uninjured peers at making automatic inferences about mental states from written text, and suggest that not all automatic inference processes are intact. Results differed, however, from previous findings by Bergemalm and Lyxell (2005), and Johnson and Turkstra (2012), perhaps because both studies included cues beyond those embedded in the individual words.

Our findings also differed from those of McDonald, Saad, and James (2011), who studied bias in adults with TBI using the Implicit Association Test (IAT). The IAT evaluates gender bias by examining participants’ response times to gender-associated words. Participants with TBI saw the words male, female, strong, and weak, in either stereotypically compatible pairs (female/weak, male/strong) or incompatible pairs (male/weak, female/strong). IAT research predicted that participants would have slower reaction times to items that contradicted social stereotypes than to items congruent with social stereotypes (e.g., a slower response to male when it follows the word weak than when it follows strong). Overall, participants with TBI had slower and more variable response times than comparison peers, but the TBI group IAT effect (mean reaction time difference between compatible and incompatible pairs) was virtually identical to that of the comparison group (24.5% vs. 24.7%). McDonald and colleagues interpreted this finding as evidence that implicit social processes are intact in adults with TBI. It may be that gender stereotypes have stronger associations than verb causality, or that, as noted below, participants with TBI had subtle word retrieval or word association impairments that were not detected on standardized tests but affected results. McDonald (1999) proposed that inference impairments in TBI are hierarchical, with more errors on inference types that require multiple cognitive steps or are based on ambiguous utterances. Our stimuli may be more ambiguous than those associated with gender stereotypes, and thus more vulnerable to error.

This study asked if adults with TBI showed IC bias, as a first step in examining text-level cues. The next step will be to identify mechanisms underlying the lack of response to IC cues. One candidate mechanism is underlying impairments in cognitive functions such as working memory or executive functions. Although the TBI group had lower scores than the comparison group on cognitive tests, there were no significant correlations between task scores and test scores. These tests were used to characterize the sample, however, rather than test hypotheses about mechanisms, and stimuli were constructed to minimize cognitive demands, so further work is needed. A second candidate mechanism is word-finding impairments. While TBI is not typically associated with impairments in formal aspects of language, and all participants scored in the normal range on an aphasia test, word-finding deficits have been reported in adults with TBI (e.g., Barrow et al., 2003; King, Hough, Walker, Rastatter, & Holbert, 2006; Hough, 2008). It is possible that the WAB Bedside Test was not sufficiently sensitive to rule out formal language deficits such as word-finding impairments. There is some evidence of sub-clinical lexical impairments in adults with TBI, such as higher error rates than uninjured peers on a lexical priming task (Russell, Arenth, Scanlon, Kessler, & Ricker, 2012; although opposite findings were reported by Chobor & Schweiger, 1998). Much of the evidence, however, derives primarily from confrontation naming tasks, which require precise lexical access and thus are influenced by factors such as inefficient search strategies, which have been shown in adults with TBI (McWilliams & Schmitter-Edgecombe, 2008). We attempted to minimize demands on word retrieval in several ways: participants were not required to give unique or sensible completions, sentence meanings did not figure in scoring (other than the pronouns), average age of acquisition was about age 8 years, and, as noted earlier, participants had unlimited time. Nevertheless, it is possible that subtle changes in word finding contributed to the higher number of no-bias items in the TBI group. The group difference in bias errors is unlikely to be explained by word-retrieval problems, as there is no a priori reason that word finding would affect NP1- vs. NP2-type verbs.

Whatever the mechanism underlying the finding of reduced IC bias in adults with TBI, this finding may have implications for everyday social communication. IC bias cues are part of a family of semantic and syntactic priming cues that speed sentence comprehension by constraining the number of possible meanings of words following the cue (see reviews in Friederici, Steinhauer, & Frisch, 1999; Gernsbacher, 1991). For example, a listener will verify that the sentence The boy went fishing and caught a trout is true faster than the sentence The boy went fishing and caught a bird, because the word fishing has primed trout not bird. If a person with TBI fails to attribute causality to a verb, or attributes it incorrectly, subsequent sentences may be either misinterpreted or understood more slowly because the original attribution must be re-considered. Using the example in the introduction, to typical listeners the sentence, The boss criticized the employee, implies that the employee is at fault (an NP2 bias). If the listener erroneously attributes the cause to the boss (e.g., being a grouchy person), then a subsequent action of firing the employee would seem particularly unfair. Although these cues are subtle and sometimes missed by typical adults, they may add to other communication impairments in adults with TBI (e.g., slow information processing, reduced working memory capacity, impairments in understanding abstract language; Novack, Alderson, Bush, Meythaler, & Canupp, 2000), and thus have additive effects on interpersonal communication.

The current study had several limitations. First, the study sample size was small, which limits the ability to generalize these findings to the TBI population. Second, all participants in this study were Caucasian and resided in the Midwestern U.S., although IC bias is considered a universal cue, and therefore should be consistent across racial and ethnic groups. Third, as discussed, there may be limitations in generalizing results of the current task design to more naturalistic social communication, which typically contains a rich variety of inference cues. Fourth, the sentence completion task requires some sentence production, verb comprehension, and executive planning skills (e.g., central executive, verbal working memory). The nature of this task makes it difficult to isolate the mechanism underlying reduced IC bias in adults with TBI.

In the current sample, the TBI group was older than the comparison group, albeit not significantly, but stimulus verbs were chosen because they are typically acquired in early childhood, so age should not affect the results. Also, if there were experience effects related to age this would be expected to benefit the TBI group rather than the comparison group. The TBI group also had fewer years of education, primarily because the injury affected educational trajectory. Because of the early age of acquisition of the verbs, however, educational differences between groups also were unlikely to have affected results.

5: Conclusion

This study aimed to advance knowledge about pragmatic inference impairments in individuals with TBI by examining response to cues embedded within words. The findings suggest that while some types of word-level inference may be intact in adults with TBI, such as automatic bias from learned social stereotypes, response to other important word-level cues might be impaired. The results may reflect primary deficits in implicit cognitive processes such as inference, or secondary effects of lexical network disruption. In either case, reduced response to IC bias cues may have negative effects on sentence- and paragraph-level comprehension, perhaps extending to interpersonal communication. If replicated and extended to conversations, the finding of diminished response to implicit word-level cues can advance our understanding of communication breakdowns between individuals with TBI and their partners, and help us develop assessment and treatment strategies to support effective communication in everyday life.

Highlights.

• Adults with TBI have social communication impairments.

• It is not clear if these communication impairments extend to implicit social cues.

• Adults with TBI missed some implicit social cues associated with written verbs.

• When adults with TBI did respond, they were less accurate than uninjured peers.

• Results challenge assumptions of intact implicit processing in adults with TBI.

Appendix. Percent of completion types per group for each verb

Verb	Correct NP Assignment	Control Group Response Data			TBI Group Response Data
		NP1	NP2	No Bias	NP1	NP2	No Bias
acclaimed	2	0.28	0.56	0.17	0.26	0.58	0.16
accused	1	0.78	0.22	0.00	0.21	0.58	0.21
admired	2	0.00	0.94	0.06	0.00	0.74	0.26
adored	2	0.06	0.78	0.17	0.26	0.47	0.26
advised	2	0.56	0.44	0.00	0.32	0.53	0.16
alienated	1	0.39	0.44	0.17	0.42	0.53	0.05
angered	1	0.78	0.17	0.06	0.74	0.21	0.05
annoyed	1	0.89	0.06	0.06	0.58	0.16	0.26
appalled	1	0.72	0.22	0.06	0.32	0.53	0.16
applauded	2	0.11	0.72	0.17	0.26	0.47	0.26
appreciated	2	0.00	0.94	0.06	0.05	0.79	0.16
believed	2	0.17	0.67	0.17	0.26	0.63	0.11
blamed	2	0.50	0.39	0.11	0.26	0.32	0.42
blessed	2	0.22	0.61	0.17	0.42	0.32	0.26
bored	1	0.72	0.11	0.17	0.68	0.21	0.11
calmed	2	0.22	0.72	0.06	0.21	0.68	0.11
celebrated	2	0.06	0.56	0.39	0.05	0.32	0.63
charmed	1	0.78	0.11	0.11	0.58	0.32	0.11
cheated	1	0.89	0.11	0.00	0.63	0.21	0.16
cheered	2	0.22	0.72	0.06	0.11	0.63	0.26
cherished	2	0.22	0.44	0.33	0.26	0.47	0.26
comforted	2	0.11	0.83	0.06	0.21	0.53	0.26
condemned	2	0.06	0.83	0.11	0.21	0.53	0.26
congratulated	2	0.06	0.89	0.06	0.05	0.68	0.26
consoled	2	0.17	0.67	0.17	0.05	0.47	0.47
corrupted	1	0.67	0.33	0.00	0.53	0.16	0.32
criticized	2	0.44	0.50	0.06	0.63	0.26	0.05
delighted	1	0.83	0.06	0.11	0.68	0.21	0.11
deserted	1	0.61	0.39	0.00	0.47	0.21	0.32
despised	2	0.11	0.78	0.11	0.26	0.68	0.05
disdained	2	0.22	0.67	0.11	0.32	0.58	0.11
disliked	2	0.06	0.83	0.11	0.05	0.95	0.00
distrusted	2	0.22	0.72	0.06	0.21	0.74	0.05
divorced	2	0.33	0.61	0.06	0.42	0.21	0.37
dominated	1	0.78	0.22	0.00	0.53	0.21	0.26
dreaded	2	0.06	0.83	0.11	0.32	0.63	0.05
dreamed about	1	0.78	0.22	0.00	0.79	0.11	0.11
employed	2	0.00	0.89	0.11	0.26	0.58	0.16
enraged	1	0.78	0.17	0.06	0.84	0.05	0.11
envied	2	0.17	0.72	0.11	0.16	0.68	0.16
exhilarated	1	0.78	0.17	0.06	0.63	0.11	0.26
feared	2	0.11	0.78	0.11	0.21	0.63	0.16
fed	2	0.11	0.89	0.00	0.05	0.84	0.11
fought	1	0.67	0.22	0.11	0.37	0.32	0.32
freed	2	0.33	0.61	0.06	0.37	0.42	0.21
frightened	1	0.78	0.00	0.22	0.58	0.11	0.32
gladdened	1	0.67	0.28	0.06	0.42	0.37	0.21
grabbed	2	0.28	0.72	0.00	0.32	0.53	0.16
greeted	2	0.39	0.39	0.22	0.37	0.32	0.32
guided	2	0.44	0.56	0.00	0.42	0.53	0.05
harassed	1	0.67	0.28	0.06	0.79	0.16	0.05
harmed	1	0.44	0.50	0.06	0.47	0.37	0.16
hated	2	0.11	0.83	0.06	0.26	0.68	0.05
haunted	1	0.67	0.17	0.17	0.53	0.37	0.11
helped	2	0.39	0.61	0.00	0.32	0.58	0.11
honored	2	0.00	0.78	0.22	0.05	0.58	0.37
insulted	1	0.67	0.28	0.06	0.47	0.47	0.05
interrupted	1	0.61	0.39	0.00	0.58	0.42	0.00
invigorated	1	0.61	0.28	0.11	0.32	0.47	0.21
killed	1	0.50	0.50	0.00	0.26	0.37	0.37
loved	2	0.11	0.72	0.17	0.05	0.74	0.21
married	1	0.56	0.28	0.17	0.42	0.16	0.42
missed	2	0.11	0.78	0.11	0.11	0.63	0.26
nuzzled	1	0.72	0.28	0.00	0.53	0.26	0.21
passed	1	0.50	0.50	0.00	0.21	0.47	0.32
petted	2	0.33	0.61	0.06	0.21	0.63	0.16
pitied	2	0.06	0.61	0.33	0.05	0.74	0.21
praised	2	0.06	0.94	0.00	0.00	0.68	0.32
prized	2	0.06	0.89	0.06	0.16	0.74	0.11
protected	2	0.28	0.61	0.11	0.42	0.53	0.05
punished	2	0.17	0.83	0.00	0.32	0.58	0.11
reassured	2	0.17	0.83	0.00	0.32	0.47	0.21
rebuked	2	0.50	0.50	0.00	0.16	0.63	0.21
relaxed	1	0.50	0.39	0.11	0.21	0.53	0.26
repulsed	2	0.67	0.17	0.17	0.42	0.21	0.37
resented	1	0.06	0.83	0.11	0.16	0.63	0.21
respected	2	0.00	0.89	0.11	0.11	0.74	0.16
reviled	2	0.28	0.50	0.28	0.42	0.37	0.21
rewarded	2	0.06	0.94	0.00	0.11	0.63	0.26
ridiculed	2	0.56	0.44	0.00	0.47	0.47	0.05
saluted	2	0.17	0.61	0.22	0.11	0.68	0.21
scorned	2	0.17	0.83	0.00	0.21	0.79	0.00
sickened	1	0.72	0.17	0.11	0.47	0.26	0.26
slandered	1	0.61	0.28	0.11	0.53	0.32	0.16
sued	2	0.11	0.78	0.11	0.05	0.84	0.11
supported	2	0.39	0.44	0.17	0.32	0.47	0.21
thanked	2	0.06	0.94	0.00	0.11	0.74	0.16
treasured	2	0.28	0.61	0.11	0.32	0.53	0.16
trusted	2	0.06	0.56	0.39	0.26	0.42	0.32
valued	2	0.06	0.89	0.06	0.16	0.58	0.26
victimized	1	0.78	0.17	0.06	0.58	0.37	0.05
welcomed	2	0.33	0.61	0.06	0.21	0.63	0.16
worried about	2	0.17	0.78	0.06	0.11	0.79	0.11
wounded	1	0.50	0.33	0.17	0.58	0.16	0.26
wowed	1	0.83	0.17	0.00	0.68	0.16	0.16
yelled at	2	0.44	0.50	0.06	0.26	0.53	0.21

Notes: NP = noun phrase. Refer to Hartshorne and Snedeker (2013) for a more comprehensive list of verbs and their respective biases collected from healthy adults.