Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2017 Jan 1.
Published in final edited form as: Can J Behav Sci. 2016 Jan 1;48(1):78–88. doi: 10.1037/cbs0000035

A Linguistic Inquiry and Word Count Analysis of the Adult Attachment Interview in Two Large Corpora

Theodore E A Waters 1, Ryan D Steele 2, Glenn I Roisman 3, Katherine C Haydon 4, Cathryn Booth-LaForce 5
PMCID: PMC4824682  NIHMSID: NIHMS739802  PMID: 27065477

Abstract

An emerging literature suggests that variation in Adult Attachment Interview (AAI; George, Kaplan, & Main, 1985) states of mind about childhood experiences with primary caregivers is reflected in specific linguistic features captured by the Linguistic Inquiry Word Count automated text analysis program (LIWC; Pennebaker, Booth, & Francis, 2007). The current report addressed limitations of prior studies in this literature by using two large AAI corpora (Ns = 826 and 857) and a broader range of linguistic variables, as well as examining associations of LIWC-derived AAI dimensions with key developmental antecedents. First, regression analyses revealed that dismissing states of mind were associated with transcripts that were more truncated and deemphasized discussion of the attachment relationship whereas preoccupied states of mind were associated with longer, more conflicted, and angry narratives. Second, in aggregate, LIWC variables accounted for over a third of the variation in AAI dismissing and preoccupied states of mind, with regression weights cross-validating across samples. Third, LIWC-derived dismissing and preoccupied state of mind dimensions were associated with direct observations of maternal and paternal sensitivity as well as infant attachment security in childhood, replicating the pattern of results reported in Haydon, Roisman, Owen, Booth-LaForce, and Cox (2014) using coder-derived dismissing and preoccupation scores in the same sample.

Keywords: Adult Attachment Interview, attachment, LIWC, linguistic structure

The most widely used and well-validated assessment in the field of developmental psychology for studying attachment in adults is the Adult Attachment Interview (AAI; Main & Goldwyn, 1984–1998). The development of the AAI (George, et al., 1985; Main, Goldwyn, & Hesse, 2003–2008) was based on an effort to understand how adults’ attachment representations are reflected in their discourse about their early childhood experiences (Hesse, 2008). Through careful analysis, eventually informed by the insights of the linguistic philosopher Grice (1975) about collaborative discourse, Main and her colleagues discovered which aspects of parents’ narratives about their childhood experiences predicted whether their own children would be classified as secure or insecure in the Strange Situation procedure (Van IJzendoorn, 1995).

The primary AAI scoring method developed by Main and Goldwyn (1984–1998) consists of a set of ratings on each transcript that inform assignment of individuals to one of three mutually exclusive primary attachment categories (secure-autonomous, dismissing, preoccupied) that reflect the coherence of the discourse produced in the interview. The majority of adults—described as secure-autonomous— freely and flexibly evaluate their childhood experiences, whether described as supportive or more challenging in nature. The next largest group of individuals—described as dismissing—defensively distance themselves from the emotional content of the interview either by normalizing harsh early memories or by idealizing their caregivers. The smallest group consists of preoccupied adults, who are unable to discuss their childhood experiences without becoming overwhelmed emotionally (see Hesse, 1999, for more details). In addition to these three categories, coders also classify individuals as unresolved if their discourse becomes psychologically confused while talking about loss or abuse experiences.

Increasing evidence suggests that the coherence of AAI narratives is not only predictive of the quality of adults’ relationships with their own children and romantic partners (e.g., Crowell & Feldman, 1988; Crowell, Treboux, Gao, Fyffe, Pan, & Waters, 2002; Holland & Roisman, 2010; Roisman, Madsen, Henninghausen, Sroufe, & Collins, 2001; Van IJzendoorn, 1995) but is also non-trivially associated with direct observations of the maternal and paternal sensitivity young adults experienced during childhood (e.g., Haydon, Roisman, Owen, Booth-LaForce, & Cox, 2014). That said, little is known about what specific features of AAI narrative discourse drive these associations, in part because the way in which discourse is summarized by AAI coders (i.e., via assignment of participants to the attachment categories noted above) obscures a great deal of information about the microstructure of the interview discourse.

In an effort to better understand the significant linguistic features of AAI discourse, researchers have recently turned to text analysis programs to deconstruct and evaluate linguistic elements associated with the AAI classifications. Most notably, four studies (two published articles; Borelli, David, Rifkin-Grabol, Sbarra, Mehl, & Mayes, 2012; Cassidy, Sherman, & Jones, 2012; and two unpublished dissertations O’Hara, 2007; Stone, 2003) have analyzed the linguistic structure of AAIs with the Linguistic Inquiry Word Count program (LIWC; Pennebaker, Booth, & Francis, 2007). LIWC was originally developed for the analysis of narrative discourse produced in expressive writing research and classifies words from a text file into more than seventy linguistic, psychological, and non-psychological categories (e.g., pronouns, affect, occupations; Pennebaker & Chung, 2011; see Frattaroli, 2006, for meta-analytic review of expressive writing findings).

Theory and prior research have suggested that individuals higher on dismissing states of mind appear motivated to minimize or deny the potential negative effects of their early caregiving experiences (e.g., Kobak, Cole, Ferenz-Gillies, Fleming, & Gamble, 1993). Main (2000), for example, argues that individuals who produce dismissing AAI discourse have developed an attentional strategy whereby they avoid discussing and re-experiencing the distress caused by difficult attachment experiences with caregivers from childhood. As a result of this attentional strategy, dismissing individuals are expected to produce narratives that are brief, incomplete, and attempt to minimize the impact their negative attachment experiences had on them. Consistent with these predictions, LIWC-based analyses of AAI transcripts have revealed that dismissing individuals use fewer negative emotion words (Borelli, et al, 2012), more negations (e.g., “no”, “not”; Cassidy, et al., 2012), more words in the present tense (e.g., Stone, 2003), and fewer words overall (Cassidy et al., 2012; O’Hara, 2007). A similar theory-consistent pattern has emerged for preoccupied individuals. Theory and prior research suggest that preoccupied individuals' attentional strategy results in confusion regarding their early caregiving experiences, which causes vague, long-winded, and at times angry and conflicted AAI narratives (e.g. Kobak et al., 1993). Research examining LIWC-derived linguistic features in the AAI has supported these claims. For example, results indicate that preoccupied adults use more anger-related words, such as "hate" (Cassidy et al., 2012; Borelli et al., 2012), more anxiety related words (Stone, 2003), more swear words (Cassidy et al., 2012) and more words overall in their AAI transcripts (e.g., Cassidy et al., 2012).

In addition to obtaining a number of theory-consistent findings regarding the linguistic correlates of AAI classifications, several of these prior studies imply that it might be possible to use LIWC, or similar text analysis programs, to actually automate coding of the AAI. In the first of two unpublished dissertations in this literature, Stone (2003) studied a sample of 164 individuals and a subset of 20 LIWC variables to examine the linguistic correlates of AAI classifications. This investigation used a three-way AAI classification system—secure, dismissing, and preoccupied—and found that AAI transcripts could be correctly classified 61% of the time based on LIWC scores (secure, 79.8%; dismissing, 52.9%; preoccupied, 8.3%). In a second dissertation, O’Hara (2007) studied a sample of 63 individuals and 25 LIWC variables to examine the fit between linguistic variables and AAI coder classification. O’Hara (2007) omitted preoccupied and unresolved transcripts from her analysis due to limited sample sizes, but was able to correctly classify 100% of the secure and dismissing participants based on LIWC data. Subsequently, Cassidy et al. (2012) used a sample of 136 individuals and 44 LIWC categories to examine the linguistic correlates of AAI classifications and found that 14 of those categories were associated with AAI classifications and accurately classified 71% of the AAI transcripts. However, one limitation of the Cassidy et al. (2012) study was that the authors were limited to the use of only six LIWC variables to classify individuals on the AAI within the context of the discriminant function analysis they conducted due to the small number of individuals (n = 7) in the preoccupied category (classification agreement between two human coders in Cassidy et al. [2012] was 79%).

Overall, the available data suggest that the use of LIWC with the AAI has produced a number of theory-consistent findings regarding the linguistic qualities of attachment narratives and provided some initial indication that text analysis software may be useful toward the goal of automating the coding of the AAI. However, there are a few important limitations associated with the extant literature. First, the sample size in each of these studies was modest. This is an important issue because, across the emerging literature examining the linguistic structure of the AAI, the sample size of the investigations have been negatively associated with the percentage of transcripts that were correctly classified. That is, as the sample size increased, the percentage of transcripts that were correctly classified decreased. Further, previous research in this area was limited by small numbers of preoccupied transcripts and therefore did not fully capture variability in the AAI.

Second, all prior studies in this area used the standard categorical coding system for the AAI (Main & Goldwyn, 1984–1998). Although research on the AAI using this system has clearly been productive (Bakermans-Kranenburg & Van IJzendoorn, 2009), factor analytic and taxometric research with large samples (Fraley & Roisman, 2014; Haltigan, Leerkes, et al., 2014; Haltigan, Roisman, & Haydon, 2014; Haydon, Roisman, & Burt, 2012; Roisman, Fraley, & Belsky, 2007) suggest that adults vary on two modestly correlated (rather than mutually exclusive) AAI state-of-mind dimensions (henceforth referred to as coder-derived AAI dimensions): dismissing states of mind and preoccupied states of mind. More specifically, dismissing states of mind reflect the degree to which adults freely evaluate or defensively discuss their attachment histories, whereas preoccupied states of mind reflect the degree to which adults become emotionally entangled in their narrative while discussing their early attachment experiences.

A third limitation of the four previous studies that used LIWC to classify individuals on the AAI pertains to how variables were selected for analysis. LIWC is capable of classifying words into more than 70 categories. However, none of the previous studies in this area has explored the full range of LIWC indicators. It is possible that a more comprehensive approach might account for more variation in individual differences in AAI narratives. Further, use of the full range of LIWC indicators in an exploratory and descriptive fashion in a large, discovery sample—along with confirmatory cross-validation in a second large sample—would allow attachment researchers to get a more complete picture of the linguistic features of dismissing and preoccupied AAI transcripts, including features that are not explicitly attended to during traditional coding.

In the current paper, we attempted to address each of these limitations. More specifically, we used two large corpora of AAIs (Ns = 826 and 857) to address three research questions. First, we attempted to replicate and extend prior findings concerning associations between AAI state-of-mind dimensions and LIWC categories using the full set of LIWC variables. Given previous theory and research arguing that dismissing individuals tend to provide brief, generalized descriptions of prior interactions with caretakers, we hypothesized that their transcripts would contain fewer words overall, more negations (e.g., no, not), more words describing tentativeness (e.g., maybe, perhaps) indicating less concrete and more generalized discourse, and fewer words per sentence. We also hypothesized that dismissing transcripts would contain fewer conjunctions (e.g., but, whereas) as Cassidy et al. (2012) noted that conjunctions are used to keep a conversation going. Finally, given dismissing individuals’ tendency to minimize and normalize emotional content within the interview, we expected that their transcripts would contain more non-relationship experience words (e.g., job, work). In contrast, given the tendency of preoccupied individuals to become angry, passive, or fearful when discussing early caregiving experiences, we hypothesized that their transcripts would contain more anger-related words (e.g., hate, kill), more swear words (e.g., damn, piss), and more words describing negative emotions (e.g., hurt, worthless). Furthermore, as preoccupation often occurs in AAIs of excessive length that contain grammatically entangled sentences, we hypothesized that such transcripts would contain more words overall and more words per sentence.

Second, we sought to provide some clarity concerning the potential of using text analysis software to automate coding of the AAI. In addition to regressing the AAI state-of-mind dimensions onto nearly the entire set of LIWC variables in order to examine what proportion of the total variation in the AAI state-of-mind dimensions could be accounted for the LIWC variables, we created LIWC-derived prediction equations to cross validate these results. Importantly, no prior studies that have used LIWC and the AAI have examined the convergence between the coder-derived AAI dimensions used in AAI research and LIWC-derived AAI dimensions. To redress this, in the current study LIWC-derived AAI dimensions were created in each sample by summing the values output by the LIWC software (see the Method section and Table 2 for more details) after they were multiplied by their unstandardized regression coefficients identified from multiple regression analyses. In an effort to evaluate the effectiveness of using LIWC-based output as a means of automating the coding of the AAI, we conducted cross-validation analyses in which we examined associations between LIWC-derived AAI dimensions and coder-derived AAI dimensions in the sample that had not been used to create the LIWC regression weights.

Table 2.

Unstandardized regression coefficients used to construct the LIWC-derived dismissing and preoccupied state-of-mind dimensions and the associated standardized regression coefficients

UIUC sample SECCYD sample
LIWC Variable Dismissing Preoccupied Dismissing Preoccupied
b β b β b β b β
Constant 1.067670 -- −0.942235 -- 1.145750 -- 0.070953 --
Word count −0.000014 −.12** 0.000032 .43** −0.000025 −.20** 0.000037 .52**
Words per Sentence −0.000915 −.02 0.000029 .00 −0.004028 −.09** −0.000528 −.02
Six letter words 0.012208 .05 0.008898 .05 0.006006 .04 0.002969 .03
1st person singular 0.002892 .01 0.009381 .05 −0.026204 −.09* −0.015000 −.09*
1st person plural −0.025436 −.04 −0.035904 −.08 −0.053018 −.07* −0.047601 −.12**
2nd person 0.034678 .03 0.029943 .05 −0.031816 −.03 0.050108 .07
3rd person singular −0.029341 −.06 0.032505 .11* 0.011832 .03 0.012880 .05
3rd person plural −0.052438 −.05 −0.010425 −.02 −0.050537 −.05 −0.037240 −.06*
Impersonal pronouns −0.023120 −.06 0.010088 .04 0.006306 .02 −0.003761 .02
Article −0.012392 −.02 −0.016976 −.05 −0.047211 −.08* −0.023386 −.07
Auxiliary verbs −0.004326 −.01 −0.006953 −.03 −0.044664 −.12** 0.000453 .00
Past tense −0.005612 −.02 0.008966 .04 0.006726 .02 −0.015905 −.08
Present tense −0.018862 −.06 0.012003 .06 0.041564 .13** 0.002352 .01
Future tense −0.033638 −.03 −0.001695 .00 −0.010731 −.01 −0.000138 .00
Adverbs −0.026746 −.07 −0.022495 −.09* 0.015623 .04 0.004972 .03
Prepositions −0.052283 −.14* 0.001733 .01 −0.058385 −.16** −0.015069 −.07
Conjunctions −0.036908 −.10* −0.013119 −.06 −0.072586 −.26** −0.014552 −.09
Negations 0.245092 .35** 0.077514 .18** 0.109274 .25** 0.051824 .22**
Quantifiers −0.078743 −.09 −0.006876 −.01 −0.011451 −.01 0.001374 .00
Numbers −0.058902 −.06 0.007431 .01 0.046296 .05 −0.005263 −.01
Swear words 0.667899 .07* 0.608359 .10** 0.448328 .07* 0.090477 .03
Family −0.043950 −.06 0.042212 .09* 0.018910 .03 0.020163 .05
Friends 0.010776 .00 −0.067269 −.03 −0.088530 −.03 0.062219 .03
Humans −0.005266 .00 0.050538 .04 0.034799 .02 −0.024668 −.02
Positive emotions 0.001575 .00 −0.024384 −.05 −0.000753 .00 −0.014034 −.05
Negative emotions −0.184698 −.17* 0.006254 .01 −0.056636 −.06 0.001082 .00
Anxiety −0.027793 −.01 0.011103 .01 −0.249811 −.10** 0.000043 .00
Anger 0.160657 .07 0.135056 .10 0.185751 .10* 0.137263 .14**
Sadness −0.062997 −.03 −0.067441 −.04 0.092727 .04 0.097404 .08
Insight −0.055519 −.09* −0.017312 −.05 −0.059575 −.10** −0.022435 −.07
Causation 0.052682 .05 0.066173 .09* 0.034281 .03 0.010662 .02
Discrepancy −0.024637 −.03 −0.019001 −.04 0.044915 .06 −0.002100 −.01
Tentative 0.073443 .18** 0.008209 .03 0.039392 .10** 0.004751 .02
Certainty −0.003417 .00 −0.018324 −.03 −0.063060 −.08** −0.039696 −.09**
Inhibition 0.310225 .09** 0.071853 .03 0.028774 .01 0.059974 .03
Inclusive −0.022482 −.06 0.003976 .02 0.022363 .08 0.013069 .08
Exclusive −0.006011 −.01 −0.008217 −.03 −0.014649 −.04 −0.027600 −.14**
See 0.147708 .06 0.043908 .03 −0.093297 −.04 0.014386 .01
Hear −0.045632 −.02 0.101858 .09** 0.049735 .03 0.014614 .02
Feel −0.088838 −.04 0.042611 .03 −0.125372 −.05 0.008236 .01
Body −0.163825 −.05 −0.204456 −.11** −0.061054 −.02 −0.025270 −.02
Health 0.123659 .05 0.053370 .04 0.089134 .04 0.028174 .02
Sexual −0.053548 −.01 −0.211051 −.08* −0.453077 −.12** −0.025652 −.01
Ingestion 0.038673 .01 0.044750 .02 0.099498 .03 0.013356 .01
Motion −0.023854 −.02 0.053398 .08* 0.027619 .03 0.039157 .09*
Space 0.025309 .04 −0.001617 .00 0.032224 .06 0.000152 .00
Time 0.042444 .10* 0.006926 .03 0.030725 .07 0.003898 .02
Work 0.061151 .07 0.005659 .01 0.040703 .04 −0.026774 −.05
Achievement 0.012721 .01 −0.049637 −.05 −0.053015 −.04 −0.028474 −.04
Leisure −0.028046 −.02 0.001909 .00 −0.007569 −.01 −0.047795 −.09**
Home −0.028745 −.02 −0.041542 −.04 0.031720 .02 −0.016717 −.02
Money 0.103720 .03 0.089006 .04 0.010813 .00 0.022890 .01
Religion −0.291912 −.07* −0.111774 −.04 0.025822 .01 0.065023 .03
Death 0.113045 .04 −0.018240 −.01 −0.032290 −.01 −0.080155 −.04
Assent 0.020335 .03 0.035162 .09* 0.009450 .02 −0.010046 −.04
Nonfluencies −0.000628 .00 0.012618 .08 −0.002806 −.01 −0.004298 −.03
Fillers −0.027684 −.16* 0.002106 .02 −0.016653 −.10* −0.002690 −.03
R 2 .38 .36 .54 .48
Open in a new tab

Note

**

p<.01,

*

p<.05

Third, we examined whether the LIWC-derived AAI dimensions were associated with a set of theory-driven developmental antecedent variables. Specifically, we examined associations between LIWC-derived dismissing and preoccupied AAI dimensions with direct observations of maternal sensitivity, paternal sensitivity, and security in infancy and early childhood. Further, we compared the magnitude of associations between the LIWC-derived AAI dimensions and the antecedent data with the associations between coder-derived AAI dimensions and the same antecedent data. Prior research by Haydon et al. (2014) using the NICHD Study of Early Childcare and Youth Development (SECCYD) sample, a normative-risk cohort follow from birth through age 18 years (see Booth-LaForce & Roisman, 2014, for more details), produced evidence that both dismissing and preoccupied states of mind were correlated with lower levels of maternal and paternal sensitivity throughout childhood and adolescence, as well as more attachment insecurity in infancy and early childhood. As such, we were able to examine the extent to which LIWC-derived AAI dimensions (created using a separate, large sample of AAIs) produced converging evidence with coder-derived AAI dimensions from the SECCYD with respect to associations with three key attachment-relevant antecedents.

Method

Participants and Procedure

Two large sets of AAI transcripts were analyzed for this report. The first corpus (University of Illinois at Urbana-Champaign [UIUC] sample) is based on the set of AAIs acquired from college and community participants primarily living in and around Champaign, IL that was the focus of Haydon et al. (2012; N=826).1 The second (SECCYD sample) consists of the AAIs acquired from the SECCYD cohort around age 18 years (N=857; Booth-LaForce & Roisman, 2014). See Table 1 for descriptive data.

Table 1.

Demographic data for participants included in the current study

Sample Size
(N)		 Gender
(% Male)		 Age
M (Range)		 Ethnicity
(% White)
UIUC Sample
College Samples
Roisman, Tsai, & Chang (2004) 60 47% 21.53 (18–30) 50%
Roisman (2006) 100 50% 18.85 (18–24) 66%
Roisman, Fortuna, & Holland (2006) 32 59% 19.15 (18–24) 56%
Dyadic Studies
Roisman (2007), Roisman, Holland et al. (2007), Roisman, Clausell, Holland, Fortuna, & Elieff (2008) 164 51% 37.77 (18–77) 91%
Roisman et al. (2008) 120 50% 34.56 (19–61) 87%
Holland & Roisman (2010); Roisman et al. (2008) 230 50% 20.44 (18–25) 76%
Fortuna, Roisman, Haydon, Groh, & Holland (2011) 120 50% 20.43 (18–25) 73%
UIUC Sample Total 826 50% 25.77 (18–77) 76%
SECCYD Sample Total 857 49% 17.8 (17–19) 78%
Open in a new tab

Measures

AAI (George, Kaplan, & Main, 1985)

The AAI is a semi-structured interview that requires approximately an hour to administer and consists of 20 questions with follow-up probes, during which participants are asked to describe their early experiences with their primary caregivers. The protocol is meant to provide a window into adults’ current state of mind with regard to attachment experiences during childhood. In addition to questions concerning the quality of the early relationship with caregivers, additional probes focus on any loss or abuse the participant might have experienced (see Hesse, 2008, for additional information).

Consistent with convention, the AAIs in the current report were audio-recorded and transcribed verbatim by trained interviewers and transcribers. The AAI Q-Sort (Kobak, 1993) was used by raters—all of whom were trained using the Main and Goldwyn (1984–1998) AAI categorical coding system—to code the AAI narratives (see Haydon et al., 2012; Booth-LaForce & Roisman, 2014). The AAI Q-sort contains 100 descriptors related to discourse, attachment-related states of mind, and inferred parent experiences as assessed by the AAI. The cards are sorted into a forced normal distribution ranging from least characteristic to most characteristic of the transcript and these distributions are then correlated with prototype sorts. As already noted, Haydon et al. (2012) demonstrated with the UIUC sample used in the current report that adults’ AAI narratives vary with respect to two weakly correlated state-of-mind dimensions: dismissing states of mind (e.g., “adjectives supported by vague or shallow memories versus adjectives supported by detailed episodic memories”) and preoccupied states of mind (“is confused or overwhelmed with information about parents versus information about parents is adequate and well organized”). As such, this analysis used the same operationalization of the state-of-mind dimensions as reported in Haltigan et al. (2014) with respect to the SECCYD sample and in Haydon et al. (2012) with respect to the UIUC sample, both of which leveraged the dismissing and preoccupied prototype scores of the Q-sort (see Spangler & Zimmermann, 1999). For the sample of AAIs included in this paper, 37% (309/826) of the cases from the UIUC sample and 21% (178/857) of the cases from the SECCYD sample were double-sorted. Of the subset of double-coded cases in each sample, coders were reliable (≥0.6 using the Pearson–Brown prophecy formula) on 82% of cases from the UIUC sample and 90% of cases from the SECCYD sample. For both samples, in cases in which two coders were not in agreement, a third coder sorted the case; the final sort was computed by averaging the two sorts that were most highly correlated. Overall, the SECCYD and UIUC samples had mean dismissing state-of-mind scores of −.02 (.40) and −.21 (.41), and mean preoccupied state-of-mind scores of −.23 (.22) and −.18 (.26), respectively.

LIWC (Pennebaker et al., 2007)

The LIWC software program uses an internal dictionary of 4,500 words to classify words found in text files into over 70 categories. The categories include general descriptors (e.g., word count, words per sentence), linguistic components (e.g., adverbs, conjunctions), psychological processes (e.g., social and cognitive processes), and non-psychological processes (e.g., work). LIWC variables are organized hierarchically such that summary-level LIWC variables contain sets of individual-level LIWC variables. For example, the “Social processes” summary-level variable contains the “Family”, “Friends”, and “Humans” individual-level variables. As such, in the current research only the individual-level LIWC variables were used in analyses, with only three word-based LIWC variables omitted from analyses. More specifically, percentage of words in the transcript captured by LIWC dictionary was omitted because it was not relevant and two other LIWC variables (total pronouns and personal pronouns) were fully captured by other summary- and individual-level variables and thus redundant. Additionally, all punctuation-related variables were also omitted from analyses, as we were interested in the extent to which LIWC could aid in classifying individuals on the AAI as a function of their verbal behavior.

After processing the text files, the LIWC program creates an output file containing variables that reflect each LIWC category as a percentage of total word count (the only exceptions are total word count and words per sentence, which are count derived variables). For example, if a transcript received a score of 10.2 on the adverbs LIWC category, this would indicate that 10.2% of the total number of words classified in the transcript were adverbs. See Cassidy et al. (2012) for more information regarding the reliability and validity of LIWC. The most recent version of the LIWC software program (Pennebaker et al., 2007) was used to prepare the data for analysis in this report.

AAI transcript preparation

Transcripts were prepared following the directions for editing oral transcripts by Pennebaker et al. (2007). More specifically, transcribers’ comments were eliminated, fillers used in everyday speech were converted (e.g., “you know” was changed to “youknow” so that “know” was not counted toward cognitive state variables), and nonfluencies were changed to facilitate word identification by the LIWC software program. Additionally, slang words within the transcript were translated (e.g., “gotta” was changed to “got to”). Finally, AAI-specific changes to the transcripts were made to facilitate processing by the LIWC software program. More specifically, interviewers’ text was eliminated from the transcripts, and ellipses and double-dashes (denoting pauses within the AAI) were changed to single periods as suggested by J. W. Pennebaker (personal communication, September 12, 2012).

Developmental Antecedents (SECCYD sample)

Proportion of times secure in early childhood

Security in early childhood with the mother was assessed using the Strange Situation Procedure (SSP; Ainsworth, Blehar, Waters, & Wall, 1978; 2014) at 15 months, the Attachment Q-Set (AQS; Waters & Deane, 1985) at 24 months, and the Modified Strange Situation Procedure (MSSP; Cassidy, Marvin, & the MacArthur Working Group on Attachment, 1992) at 36 months. For all individuals with at least two of the three early attachment assessments, a proportion score was calculated by dividing the number of times the child was classified as secure by the total number of assessments (see Groh et al. (2014) for more details regarding this composite).

Parental sensitivity

Direct observations of maternal sensitivity were acquired at 6, 15, 24, 36, and 54 months; Grades 1, 3, and 5; and age 15 years. Assessments of paternal sensitivity were collected at 54 months; Grades 1, 3, and 5; and age 15 years. Sensitivity was assessed while children and their mother/father were videotaped as the target participants completed age-appropriate tasks (e.g., Owen, Vaughn, Barfoot, & Ware, 1996). More information regarding the tasks and scoring system can be found in Booth-LaForce, Groh, Burchinal, Roisman, Owen, and Cox (2014) and Fraley, Roisman, Booth-LaForce, Owen, and Holland (2013). As reported in Booth-LaForce et al. (2014), internal consistencies for the composite measures of sensitivity collected from the full sample ranged from .70–.85 for mothers and .71–.82 for fathers across assessments. Maternal (α = .79) and paternal (α = .78) ratings were averaged over time to create total summary measures of antecedent maternal and paternal sensitivity.

Results

Analytic Plan

This analysis addressed three research questions. First, we examined which aspects of language use were influential in AAI state-of-mind coding. To answer this question, we conducted regression analyses to identify which LIWC variables were non-trivially (i.e., β ≥ .10) associated with the AAI dismissing and preoccupied state-of-mind dimensions. Second, we examined whether LIWC could be used to automate coding of the AAI, using trained and reliable raters’ AAI coding as a gold standard. To address this second question, we used the unstandardized regression coefficients obtained from initial regression analyses to create a set of LIWC-based dismissing and preoccupied states of mind equations for each sample. These prediction equations were then cross-validated in the other sample in relation to coder-derived AAI dimensions. Third, after obtaining initial evidence that the LIWC-derived AAI dimensions showed a good deal of convergence with coder-derived AAI dimensions, we examined whether LIWC-derived AAI dimensions shared associations with caregiving and attachment security in childhood that were comparable in magnitude to associations previously documented in the SECCYD dataset with coder-derived AAI state-of-mind scores (i.e., Haydon et al., 2014).

What aspects of language use are especially influential in AAI state-of-mind coding?

Dismissing states of mind

In the UIUC sample we regressed the (coder-derived) dismissing state-of-mind dimension scores on the full set of individual-level LIWC variables. Several LIWC variables were non-trivially associated with dismissing states of mind. More specifically, individuals with higher dismissing scores used fewer words overall, prepositions, and conjunctions, more negations (e.g., no), fewer words indicative of negative emotions (e.g., hurt), more words describing tentativeness (e.g., maybe), more time-related words (e.g., end), and fewer filler-related words (e.g., you know) (see Table 2).

A parallel analysis of dismissing states of mind was undertaken in the SECCYD sample. Limiting our scope to the LIWC variables producing non-trivial associations, results revealed that individuals with higher dismissing state-of-mind scores used fewer words overall, fewer auxiliary verbs (e.g., will), more words in the present tense, fewer prepositions, fewer conjunctions, more negations, fewer anxiety-related words (e.g., nervous), more words indicative of anger, fewer words relating to insight (e.g., think), more words suggesting tentativeness, fewer words relating to sex (e.g., love), and fewer filler-related words (see Table 2).

Next, we examined more systematically the degree to which the two samples produced convergent evidence regarding the strength of the association between LIWC variables and dismissing state-of-mind scores. Importantly, the mean of the absolute values of the differences in the beta coefficients produced by the two regression analyses reported above was relatively close to zero (M = .06, SD = .05) suggesting that individuals from both samples who had higher dismissing state-of-mind scores used similar types and frequencies of words within the AAI. Indeed, results revealed that 82% of the LIWC predictors (47 of 57 LIWC variables) produced beta coefficients within .10 standardized units (i.e., small or trivial differences) across the two regression analyses. In addition, of 14 LIWC variables non-trivially associated with dismissing state-of-mind scores in either sample, 6 (43%) were observed in both samples. These LIWC variables were word count, prepositions, conjunctions, negations, tentativeness, and fillers. In other words, individuals higher on dismissing state of mind used notably fewer words overall, fewer prepositions, fewer conjunctions, more negations, more words indicative of tentativeness, and fewer fillers in both samples (see Table 2).

Preoccupied states of mind

In a parallel fashion we regressed the (coder-derived) preoccupied state-of-mind dimension scores on the full set of individual-level LIWC variables in the UIUC sample. Results revealed that preoccupied individuals used more words overall, more third-person singular pronouns (e.g., she), more negations, more swear words (e.g., damn), more words indicative of anger (e.g., hate), and fewer words relating to the body and bodily processes (e.g., hands) (see Table 2). Similarly, individuals with higher preoccupied states-of-mind scores used more words overall, fewer first-person plural pronouns, more negations, more anger-related words, and fewer words indicative of exclusivity (e.g., without) in the SECCYD sample (see Table 2).

The mean of the absolute values of the differences in the beta coefficients generated by the two regression analyses reported above were relatively close to zero (M = .05, SD = .04) suggesting that individuals who had higher preoccupied state-of-mind scores used similar types and frequencies of words within the AAI in both samples. Indeed, results revealed that 88% of the LIWC predictors (50 of 57 LIWC variables) produced beta coefficients within .10 standardized units (i.e., small or trivial differences) across the two regression analyses. In addition, of 8 LIWC variables non-trivially associated with preoccupied states of mind in either sample, 3 (38%) were observed in both samples. More specifically, individuals with higher preoccupied states-of-mind scores used notably more words overall, more negations, and more words indicative of anger in both samples.

Convergence between LIWC-derived and coder-derived state-of-mind dimensions

Although not yet emphasized, the multiple regression analyses presented thus far revealed that, in aggregate, LIWC variables explained remarkably large proportions of the total variation in both coder-derived AAI dimensions in both samples (see Table 2), with LIWC ratings accounting for 38 and 54% of the variation in dismissing states of mind and 36 and 48% of the variation in preoccupied states of mind in the UIUC and SECCYD datasets, respectively. That said, it is important to note that these R2s had the potential to overestimate the degree to which LIWC-related variation accounts for AAI coders’ assessments of dismissing and preoccupied states of mind as regressions by default maximize the convergence between predictors and criteria within a given sample. As such, before it might be reasonable to conclude that LIWC could be used to automate AAI coding, it was crucial as a next step to cross-validate the regression weights generated in each sample.

To that end, we used the unstandardized regression coefficients produced from the initial set of multiple regression analyses reported in Table 2 to create prediction equations to code AAI dimensions via an algorithm based on LIWC analysis of the other dataset. We refer to the dismissing and preoccupied state-of-mind dimensions created by using the opposite-sample prediction equations as LIWC-derived AAI dimensions. More specifically, we multiplied the unstandardized regression coefficients produced from the UIUC AAIs with the LIWC output based on the SECCYD AAIs to create a UIUC weights-based dismissing dimension and a UIUC weights-based preoccupied dimension in the SECCYD dataset. Using a similar procedure, a SECCYD weights-based dismissing dimension and a SECCYD weights-based preoccupied dimension were created in the UIUC dataset. We then computed bivariate correlations between the coder-derived AAI dimensions and the opposite-sample LIWC-derived weights AAI dimensions in each dataset.

Results in the SECCYD sample demonstrated that the UIUC weights-based dismissing dimension explained 40% of the variation in the coder-derived dismissing dimension (r = .63) and the UIUC weights-based preoccupied dimension explained 32% of the variation in the coder-derived preoccupied dimensions (r = .57). In short, LIWC-based coding of AAI data from the SECCYD accounted for as much or more of the variation in coder-derived dimensions in the UIUC data as in the sample in which the weights were originally estimated (i.e., the UIUC sample). We also used the LIWC weights derived from the SECCYD sample in a similar manner in the UIUC dataset. In contrast to the first set of analyses, the SECCYD weights-based AAI dimensions explained somewhat less (though still substantial) variation in each coder-derived AAI dimension in the UIUC sample. More specifically, the SECCYD weights-based dismissing dimension explained 26% of the variation in the coder-derived dismissing dimensions (r = .51) and the SECCYD weights-based preoccupied dimension explained 27% of the variation in the coder-derived preoccupied dimensions (r = .52) in the UIUC sample.

On the basis of the results of the correlations presented above, we concluded that UIUC weights-based AAI dimensions in particular might have the potential to automate AAI coding given their high degree of convergence with coder-rated AAI dimensions in both samples. In order to more rigorously test this hypothesis, we next examined to what degree these LIWC-derived AAI dimensions based on the UIUC beta weights could reproduce associations previously documented in the literature between early caregiving and (coder-derived) AAI states of mind in the SECCYD.

Comparing developmental antecedents of LIWC-derived and coder-derived AAI state of dimensions

We used the SECCYD dataset to attempt to replicate findings obtained by Haydon et al. (2014) regarding the developmental antecedents of the coder-derived AAI dimensions using UIUC weights-based AAI dimensions. As first reported in Haydon et al. (2014) and presented in Table 3, the (coder-derived) dismissing and preoccupied state-of-mind dimensions were both negatively associated with maternal sensitivity, paternal sensitivity, and the proportion of times secure in infancy and early childhood. The UIUC weights-based dismissing and preoccupied dimensions generated the same patterns of results (see Table 3).

Table 3.

Comparing correlations between the coder-derived and the UIUC weights-based AAI dimensions with maternal sensitivity, paternal sensitivity, and the proportion of times secure in early infancy and childhood in the SECCYD sample

DVs Maternal sensitivity Paternal sensitivity Prop. times secure
Coder-derived
Dismissing −.33** −.16** −.11**
Preoccupied −.20** −.11** −.13**
UIUC weights-based
Dismissing −.22** −.09* −.07*
Preoccupied −.35** −.20** −.17**
Open in a new tab

Note

**

p<.01,

*

p<.05.

Prop. times secure = proportion of times secure in early infancy,

In order to determine whether the magnitude of these associations varied for coder-derived versus UIUC weights-based AAI dimensions, we computed Steiger’s (1980) Z-tests. The magnitude of associations between both early maternal sensitivity (Z = 3.84, p < .01) and paternal sensitivity (Z = 2.48, p < .05) with dismissing state-of-mind scores were significantly larger using the coder-derived dismissing scale than the LIWC-derived dismissing dimension. In contrast, the size of the associations between both early maternal sensitivity (Z = −4.91, p < .01) and paternal sensitivity (Z = −2.78, p < .01) with preoccupied state-of-mind scores were significantly larger using the LIWC-derived preoccupied dimension than the coder-derived preoccupied scale. The magnitude of the associations between infant security and the two dismissing dimensions, however, did not differ significantly (Z = 1.21, p = .23), nor did the correlations between infant security and the two preoccupied dimensions (Z = −1.41, p = .18).

Discussion

The results from this study revealed that a variety of LIWC variables accounted for significant variation in coder-derived dismissing and preoccupied states of mind in two large samples. Indeed, in aggregate, LIWC variables explained more than a third of the variation in coder-derived dismissing and preoccupied state-of-mind dimensions in both the UIUC and SECCYD samples. Moreover, when we used the regression weights estimated in the UIUC sample to automate coding of dismissing and preoccupied states of mind dimensions in the SECCYD sample, LIWC-derived AAI dimensions were as strongly associated with coder-derived AAI dimensions in the SECCYD as they had been in the UIUC dataset.

Building on this result, we found that LIWC-derived state-of-mind dimensions were also significantly associated with the quality of early caregiving in theory-consistent ways. Specifically, LIWC-derived dismissing and preoccupied state-of-mind scores (based on regression weights estimated in the UIUC sample) were negatively related to maternal and paternal sensitivity and proportion of time secure in infancy. These associations paralleled those reported in the SECCYD sample using coder-derived AAI state-of-mind dimensions (Haydon et al., 2014), and were actually larger in magnitude than those previously reported findings in the case of relations between LIWC-derived preoccupied states of mind and maternal and paternal sensitivity. That said, the LIWC-derived dismissing scores showed significantly smaller, although still theory-consistent associations with observed parental sensitivity when compared to the coder-derived dismissing dimension.

In this study we used a total of 57 variables from the LIWC dictionary in an effort to get a more complete descriptive picture of linguistic variation in the AAI associated with coder-derived states of mind. Notably, many of the significant variables that emerged from these analyses were consistent with prior theory and research (e.g. Borelli et al., 2012; Cassidy et al., 2012; Kobak et al., 1993; Hesse, 1998). For example, coder-derived dismissing states-of-mind scores were associated with the production of notably (i.e., βs ≥ .10) fewer words overall during the AAI and, in terms of percentages, more negations, fewer prepositions, fewer conjunctions, fewer fillers, and more tentativeness across the two samples. These findings suggest that dismissing state-of-mind scores are reflected in expected ways in the microstructure of AAI narratives and that these narratives are indeed more truncated and de-emphasize/avoid discussion of attachment relationships, as would be expected.

Coder-derived preoccupied state-of-mind scores were in contrast (but as hypothesized) associated with notably more words per transcript, and had higher percentages of negations and anger words across both samples. Interestingly, preoccupation was not associated with other emotion word variables (e.g., negative emotion, sadness). This suggests that the primary emotional state that characterizes preoccupation in AAI transcripts is anger and not a more diffuse negative emotional tone.

Beyond gaining a deeper understanding of how dismissing and preoccupied states of mind manifest in narrative microstructure, as aforementioned we also examined the extent to which LIWC-derived state-of-mind scores were associated with earlier assessments of observed parental sensitivity and infant attachment security. As LIWC-derived AAI dimensions replicated the original findings reported by Haydon et al. (2014), our results suggest that automation of AAI coding with text analysis software is indeed feasible. AAI training and coding can be an expensive and time-consuming endeavor (e.g. Van IJzendoorn, 1995) and the ability to acquire comparable results with LIWC-derived dimensional coding has the potential to increase the accessibility of the AAI for researchers with limited resources.

Although the LIWC-derived scores were strongly associated with the coder-derived scores and performed well in terms of their associations with key developmental antecedents, it is important to emphasize that LIWC-derived AAI dimensions were by no means a perfect match with coder-derived scores. Further, the LIWC-derived scores based on the UIUC sample cross-validated somewhat better than those based on the SECCYD sample, a result we speculate might be due to the fact that the UIUC sample included more age-related diversity than the SECCYD sample. In addition to sample-based differences, LIWC-derived dismissing scores somewhat underperformed (though the preoccupied dimension actually outperformed) coder-derived scales in terms of the magnitude of the associations they produced with maternal and paternal sensitivity. Nonetheless, at least in the context of research focused on directional predictions, automated coding of the AAI has clear potential.

For other applications in which precision is more critical—for example using AAIs to inform clinical practice—there is still much to learn about the linguistic profiles of dismissing and preoccupied state-of-mind dimensions before automated coding of the AAI can be implemented with confidence. That said, LIWC is only one of many programs available to conduct textual analyses of transcripts. Several other programs use different dictionaries, allow for more context-specific analyses, and are able to disambiguate semantic context fairly reliably (see Alexa & Zuell, 2000, for a review of fifteen additional text analysis programs). Software capable of a more nuanced analysis of AAI transcripts may offer additional insights into the linguistic features of adult’s narratives about attachment relationships and how those features relate to early experience with caregivers.

One aspect of AAI narrative that we were unable to examine with the approach taken in this manuscript was the links between unresolved discourse and linguistic microstructure. Although the AAI Q-Sort system employed here captures a wide range of attachment-related experiences, there are relatively few items focused on unresolved discourse with respect to loss, abuse, and trauma. Given the clinical and theoretical significance of unresolved attachment (e.g. Moran, Bailey, Gleason, DeOliveira, & Pederson, 2008), analysis of unresolved discourse with LIWC, or other software, may provide useful descriptive information as well as insights into the underlying representations associated with unresolved attachment in the AAI. That said, Roisman et al. (2007; see also Haltigan, et al., 2014) found that AAI subscales reflecting unresolved status and preoccupation loaded on a single factor (rather than separately, as is anticipated by the Main and Goldwyn [1998] classification system). At this stage it is unclear whether an investigation into unresolved discourse would yield unique results or results that converge with those reported for preoccupation in this manuscript and this remains an issue for future research.

Research going forward should seek to expand the scope of narrative assessments of attachment subjected to linguistic analysis in an effort to better understand the development and generalizability of the types of linguistic variation associated with early experience reported here. For example, predictions regarding the linguistic microstructure of attachment narratives are not confined to adulthood or to interviews focused on early experiences with caregivers. Extension of text analysis into earlier age ranges (e.g., Child Attachment Interview; Shmueli-Goetz, Target, Fonagy, & Datta, 2008) or to interviews focused on different attachment relationships (e.g., Romantic Partners: Current Relationship Interview, Crowell & Owens, 1996; Parenting: Parent Development Interview; Slade, 1999) would inform our understanding of if and when linguistic features indicative of dismissing and preoccupied states of mind emerge and generalize across relationship domains.

Initial work in these areas has produced promising results. For example, Borelli, West, Decoste, and Suchman (2012) found associations between adults’ positive emotion language while being interviewed about negative experiences, and the quality of care and parenting they provided for their children. In a childhood sample, Borelli, Sbarra, Mehl, and David (2011) found that dismissing children’s narratives about their caregivers demonstrated lower levels of experiential connectedness (e.g., first person pronoun use) thought to indicate an avoidant/distancing emotion regulation strategy. However, research has yet to examine the broader set of LIWC variables or the utility of those variables in capturing coder-derived attachment categories or dimensions across ages or relationship domains. In short, although many questions remain, a growing body of research demonstrates that automated text analysis of attachment narratives has the potential to be both informative and practical.

Acknowledgments

The authors would like to acknowledge the effort of Cory Buenting, Lauren Wruble, and Laura Baer, who helped prepare the data for analysis.

The research presented in this paper was supported by a series of grants (including a Beckman Award) from the Research Board at the University of Illinois at Urbana–Champaign, an NIMH National Research Service Award (MH19893-04), and a Wayne F. Placek Award from the American Psychological Foundation to G. I. Roisman, as well support by the Eunice Kennedy Shriver National Institute of Child Health & Human Development of the National Institutes of Health under Award Numbers R01 HD054822 to C. Booth-LaForce and F32 HD078250 to Theodore E. A. Waters. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Footnotes

1

Sixteen participants from a study of engaged and married couples (Roisman, 2007; Roisman, Holland, et al., 2007) were not included in the current analysis but were included in Haydon et al. (2012) as these participants were part of a subsample in which the AAI was not originally transcribed but instead coded directly from the audiotapes (see Roisman, 2007). As the AAIs have to be in text format to be processed by the LIWC software program, we made every attempt to transcribe the audiotapes containing the AAIs that had not been previously transcribed. However, degraded audio quality led to the exclusion of a small number (n = 16) of these audiotapes.

Contributor Information

Theodore E. A. Waters, New York University – Abu Dhabi

Ryan D. Steele, University of Minnesota

Glenn I. Roisman, University of Minnesota

Katherine C. Haydon, Mount Holyoke College

Cathryn Booth-LaForce, University of Washington.

References

  1. Ainsworth MDS, Blehar MC, Waters E, Wall S. Patterns of attachment. Hillsdale, NJ: Erlbaum; 1978; 2014. [Google Scholar]
  2. Alexa M, Zuell C. Text analysis software: Commonalities, differences and limitations: The results of a review. Quality and Quantity. 2000;34(3):299–321. [Google Scholar]
  3. Bakermans-Kranenburg MJ, Van IJzendoorn MH. The first 10,000 Adult Attachment Interviews: distributions of adult attachment representations in clinical and non-clinical groups. Attachment & Human Development. 2009;11:223–263. doi: 10.1080/14616730902814762. [DOI] [PubMed] [Google Scholar]
  4. Booth-LaForce C, Groh AM, Burchinal MR, Roisman GI, Owen MT, Cox MJ. Caregiving and contextual sources of continuity and change in attachment security from infancy to late adolescence. In: Booth-LaForce C, Roisman GI, editors. The Adult Attachment Interview: Psychometrics, stability and change from infancy, and developmental origins. Monographs of the Society for Research in Child Development. Vol. 79. 2014. pp. 67–84. [DOI] [PubMed] [Google Scholar]
  5. Booth-LaForce C, Roisman G, editors. The Adult Attachment Interview: Psychometrics, stability and change from infancy, and developmental origins. Monographs of the Society for Research in Child Development. 2014;79:1–14. doi: 10.1111/mono.12110. [DOI] [PubMed] [Google Scholar]
  6. Borelli JL, Sbarra DA, Mehl M, David DH. Experiential connectedness in children’s attachment interviews: An examination of natural word use. Personal Relationships. 2011;18:341–351. [Google Scholar]
  7. Borelli JL, David DH, Rifkin-Graboi A, Sbarra DA, Mehl MR, Mayes LC. Language use in the Adult Attachment Interview: Evidence for attachment-specific emotion regulation. Personal Relationships. 2012;19:1–15. [Google Scholar]
  8. Borelli JL, West JL, Descoste C, Suchman NE. Emotionally avoidant language in the parenting interviews of substance-dependent mothers: Associations with reflective functioning, recent substance use, parenting behavior. Infant Mental Health Journal. 2012;33:506–519. doi: 10.1002/imhj.21340. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Cassidy J, Marvin RS the MacArthur Working Group on Attachment. A system for classifying individual differences in the attachment-behavior of 2 1/2 to 4 1/2 year old children. University of Virginia; 1992. Unpublished coding manual. [Google Scholar]
  10. Cassidy J, Sherman LJ, Jones JD. What’s in a word? Linguistic characteristics of Adult Attachment Interviews. Attachment & Human Development. 2012;14:11–32. doi: 10.1080/14616734.2012.636649. [DOI] [PubMed] [Google Scholar]
  11. Crowell JA, Feldman SS. Mothers' internal models of relationships and children's behavioral and developmental status: A study of mother-child interaction. Child Development. 1988:1273–1285. doi: 10.1111/j.1467-8624.1988.tb01496.x. [DOI] [PubMed] [Google Scholar]
  12. Crowell JA, Owens G. Current Relationship Interview and scoring system. State University of New York at Stony Brook; 1996. Unpublished manuscript. [Google Scholar]
  13. Crowell JA, Treboux D, Gao Y, Fyffe C, Pan H, Waters E. Secure base behavior in adulthood: Measurement, links to adult attachment representations, and relations to couples’ communication and self reports. Developmental Psychology. 2002;38:679–693. [PubMed] [Google Scholar]
  14. Fortuna K, Roisman GI, Haydon KC, Groh AM, Holland AS. Attachment states of mind and the quality of young adults’ sibling relationships. Developmental Psychology. 2011;47:1366–1373. doi: 10.1037/a0024393. [DOI] [PubMed] [Google Scholar]
  15. Fraley RC, Roisman GI. Categories or dimensions? A taxometric analysis of the Adult Attachment Interview. In: Booth-LaForce C, Roisman GI, editors. The Adult Attachment Interview: Psychometrics, stability and change from infancy, and developmental origins. Monographs of the Society for Research in Child Development. Vol. 79. 2014. pp. 36–50. [DOI] [PubMed] [Google Scholar]
  16. Fraley RC, Roisman GI, Booth-LaForce C, Owen MT, Holland AS. Interpersonal and genetic origins of adult attachment styles: A longitudinal study from infancy to early adulthood. Journal of Personality and Social Psychology. 2013;104:817–838. doi: 10.1037/a0031435. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Frattaroli J. Experimental disclosure and its moderators: A meta-analysis. Psychological Bulletin. 2006;132:823–865. doi: 10.1037/0033-2909.132.6.823. [DOI] [PubMed] [Google Scholar]
  18. George C, Kaplan N, Main M. The Adult Attachment Interview. University of California at Berkeley; 1985. Unpublished manuscript. [Google Scholar]
  19. Grice HP. Logic and conversation. In: Cole P, Moran JL, editors. Syntax and semantics III: Speech acts. New York: Academic Press; 1975. pp. 41–58. [Google Scholar]
  20. Groh AM, Roisman GI, Booth-LaForce C, Fraley RC, Owen MT, Cox MJ, Burchinal MR. Stability of attachment security from infancy to late adolescence. In: Booth-LaForce C, Roisman GI, editors. The Adult Attachment Interview: Psychometrics, stability and change from infancy, and developmental origins. Monographs of the Society for Research in Child Development. Vol. 79. 2014. pp. 51–66. [DOI] [PubMed] [Google Scholar]
  21. Haltigan JD, Leerkes EM, Wong MS, Fortuna K, Roisman GI, Supple AJ, O’Brien M, Calkins SD, Plamondon A. Adult attachment states of mind: Measurement invariance across ethnicity and associations with maternal sensitivity. Child Development. 2014;85(3):1019–1035. doi: 10.1111/cdev.12180. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Haltigan JD, Roisman GI, Haydon KC. The latent structure of the Adult Attachment Interview: Exploratory and confirmatory evidence. In: Booth-LaForce C, Roisman GI, editors. The Adult Attachment Interview: Psychometrics, stability and change from infancy, and developmental origins. Monographs of the Society for Research in Child Development. Vol. 79. 2014. pp. 15–35. [DOI] [PubMed] [Google Scholar]
  23. Haydon KC, Roisman GI, Burt KB. In search of security: The latent structure of the Adult Attachment Interview revisited. Development and Psychopathology. 2012;24:589–606. doi: 10.1017/S0954579412000181. [DOI] [PubMed] [Google Scholar]
  24. Haydon KC, Roisman GI, Owen MT, Booth-LaForce C, Cox MJ. Shared and distinctive antecedents of AAI state of mind and inferred experience dimensions. In: Booth-LaForce C, Roisman GI, editors. The Adult Attachment Interview: Psychometrics, stability and change from infancy, and developmental origins. Monographs of the Society for Research in Child Development. Vol. 79. 2014. pp. 108–125. [DOI] [PubMed] [Google Scholar]
  25. Hesse E. The Adult Attachment Interview: Historical and current perspectives. In: Cassidy J, Shaver P, editors. Handbook of attachment: Theory, research, and clinical applications. New York: Guilford; 1999. pp. 395–433. [Google Scholar]
  26. Hesse E. The Adult Attachment Interview: Protocol, method of analysis, and empirical studies. In: Cassidy J, Shaver PR, editors. Handbook of attachment: Theory, research, and clinical applications. 2nd ed. New York: Guilford Press; 2008. pp. 552–598. [Google Scholar]
  27. Holland AS, Roisman GI. Adult attachment security and young adults’ dating relationships over time: Self-reported, observational, and physiological evidence. Developmental Psychology. 2010;46:552–557. doi: 10.1037/a0018542. [DOI] [PubMed] [Google Scholar]
  28. Kobak RR. The Adult Attachment Interview Q-set. University of Delaware; Newark: 1993. Unpublished manuscript. [Google Scholar]
  29. Kobak RR, Cole HE, Ferenz-Gillies R, Fleming WS, Gamble W. Attachment and emotion regulation during mother-teen problem solving: A control theory analysis. Child Development. 1993;64:231–245. [PubMed] [Google Scholar]
  30. Main M. The organized categories of infant, child and adult attachment: Flexible vs. inflexible attention under attachment-related stress. Journal of the American Psychoanalytic Association. 2000;48:1055–1096. doi: 10.1177/00030651000480041801. [DOI] [PubMed] [Google Scholar]
  31. Main M, Goldwyn R. Adult attachment scoring and classification system. University of California at Berkeley; 1984–1998. Unpublished manuscript. [Google Scholar]
  32. Main M, Goldwyn R, Hesse E. Adult attachment scoring and classification system. University of California at Berkeley; 2003–2008. Unpublished manuscript. [Google Scholar]
  33. Moran G, Bailey HN, Gleason K, DeOliveira CA, Pederson DR. Exploring the mind behind unresolved attachment: Lessons from and for attachment-based interventions with infants and their traumatized mothers. 2008 [Google Scholar]
  34. O’Hara M. Psycholinguistic textual correlates and classifying the Adult Attachment Interview: A tool for foster care and adoption communities. New York City: The New School; 2007. (Unpublished doctoral dissertation). [Google Scholar]
  35. Owen MT, Vaughn A, Barfoot B, Ware A. The NICHD Study of Early Child Care Parent-Child Interaction Scales: Early childhood. University of Texas at Dallas; 1996. Unpublished manuscript. [Google Scholar]
  36. Pennebaker JW, Booth RJ, Francis ME. Linguistic Inquiry and Word Count: LIWC. Austin, TX: LIWC.net; 2007. [Computer software]. [Google Scholar]
  37. Pennebaker JW, Chung CK. Expressive writing: Connections to physical and mental health. In: Friedman HS, editor. Oxford handbook of health psychology. New York, NY: Oxford University Press; 2011. pp. 417–437. [Google Scholar]
  38. Roisman GI. The role of adult attachment security in non-romantic, non-attachment-related first interactions between same-sex strangers. Attachment & Human Development. 2006;8:341–352. doi: 10.1080/14616730601048217. [DOI] [PubMed] [Google Scholar]
  39. Roisman GI. The psychophysiology of adult attachment relationships: Autonomic reactivity in marital and premarital interactions. Developmental Psychology. 2007;43:39–53. doi: 10.1037/0012-1649.43.1.39. [DOI] [PubMed] [Google Scholar]
  40. Roisman GI, Clausell E, Holland A, Fortuna K, Elieff C. Adult romantic relationships as contexts of human development: A multi-method comparison of same-sex couples with opposite-sex dating, engaged, and married dyads. Developmental Psychology. 2008;44:91–101. doi: 10.1037/0012-1649.44.1.91. [DOI] [PubMed] [Google Scholar]
  41. Roisman GI, Fortuna K, Holland A. An experimental manipulation of retrospectively defined earned and continuous attachment security. Child Development. 2006;77:59–71. doi: 10.1111/j.1467-8624.2006.00856.x. [DOI] [PubMed] [Google Scholar]
  42. Roisman GI, Fraley RC, Belsky J. A taxometric study of the Adult Attachment Interview. Developmental Psychology. 2007;43:675–686. doi: 10.1037/0012-1649.43.3.675. [DOI] [PubMed] [Google Scholar]
  43. Roisman GI, Holland A, Fortuna K, Fraley RC, Clausell E, Clarke A. The Adult Attachment Interview and self-reports of attachment style: An empirical rapprochement. Journal of Personality and Social Psychology. 2007;92:678–697. doi: 10.1037/0022-3514.92.4.678. [DOI] [PubMed] [Google Scholar]
  44. Roisman GI, Madsen SD, Hennighausen KH, Sroufe LA, Collins WA. The coherence of dyadic behavior across parent–child and romantic relationships as mediated by the internalized representation of experience. Attachment & Human Development. 2001;3(2):156–172. doi: 10.1080/14616730126483. [DOI] [PubMed] [Google Scholar]
  45. Roisman GI, Tsai JL, Chiang KHS. The emotional integration of childhood experience: Physiological, facial expressive, and self-reported emotional response during the Adult Attachment Interview. Developmental Psychology. 2004;40:776–789. doi: 10.1037/0012-1649.40.5.776. [DOI] [PubMed] [Google Scholar]
  46. Slade A. Attachment theory and research: Implications for the theory and practice of individual psychotherapy with adults. In: Cassidy J, Shaver PR, editors. Handbook of attachment: Theory, research, and clinical applications. New York, NY: Guilford; 1999. pp. 575–594. [Google Scholar]
  47. Shmueli-Goetz Y, Target M, Fonagy P, Datta A. The Child Attachment Interview: a psychometric study of reliability and discriminant validity. Developmental Psychology. 2008;44(4):939–956. doi: 10.1037/0012-1649.44.4.939. [DOI] [PubMed] [Google Scholar]
  48. Spangler G, Zimmermann P. Attachment representation and emotion regulation in adolescents: A psychobiological perspective on internal working models. Attachment and Human Development. 1999;1:270–290. doi: 10.1080/14616739900134151. [DOI] [PubMed] [Google Scholar]
  49. Steiger JH. Tests for comparing elements of a correlation matrix. Psychological Bulletin. 1980;82:245–251. [Google Scholar]
  50. Stone LDA. Using language to detect and change attachment style. Austin, Texas: The University of Texas at Austin; 2003. (Unpublished doctoral dissertation). [Google Scholar]
  51. Van IJzendoorn MH. Adult attachment representations, parental responsiveness, and infant attachment: A meta-analysis on the predictive validity of the Adult Attachment Interview. Psychological Bulletin. 1995;117:387–403. doi: 10.1037/0033-2909.117.3.387. [DOI] [PubMed] [Google Scholar]
  52. Waters E, Deane K. Defining and assessing individual differences in attachment relationships: Q-methodology and the organization of behavior in infancy and early childhood. In: Bretherton I, Waters E, editors. Growing points of attachment theory and research. Monographs of the Society for Research in Child Development. 1–2. Vol. 50. 1985. pp. 41–65. [Google Scholar]

RESOURCES