Abstract
Two studies were conducted to develop and validate a six-item scale for measuring context-specific attributions regarding the extent to which people either blame or exonerate partners during couples’ conflicts. Context-specific attributions pertain to appraisals made during a single episode of relationship conflict, and the scale was expected to be distinct from existing attribution scales measuring people’s schemas regarding the types of attributions they typically make. Study 1 included 2,452 people in marriage or cohabitating relationships; Study 2 included 172 people in dating relationships, and participants in both studies completed Internet questionnaires. Item response theory was used to create an attribution scale using the fewest number of items to discriminate reliably across the full range of attribution levels. The resulting scale produced an expected pattern of convergent and divergent correlations with other context-specific measures, including two types of underlying concerns and three types of emotion. The context-specific attribution scale explained variance in these criterion variables that could not be explained by other existing scales that assess attributions at the schematic level.
Keywords
When people experience conflicts in marriage or in committed romantic relationships, they often make cognitive attributions regarding the extent to which they believe their partners are responsible and culpable for negative relationship events (Fincham & Bradbury, 1992). These attributions are important in both clinical work and research with couples. Research finds that attributions are associated with levels of relationship conflict (Bradbury & Fincham, 1990; Fincham, Bradbury, Arias, Byrne, & Karney, 1997; Marshall, Jones, & Feinberg, 2011), the use of negative communication (Bradbury, Beach, Fincham, & Nelson, 1996; Bradbury & Fincham, 1992; Miller & Bradbury, 1995), and levels of relationship satisfaction or distress (Bradbury & Fincham, 1990; Karney & Bradbury, 1995, 2000; Tashiro & Frazier, 2007). Accordingly, clinicians often seek to modify relationship attributions in cognitive–behavioral approaches to couples therapy (Hrapczynski, Epstein, Werlinich, & LaTaillade, 2012). Thus, in research, attribution assessment is necessary for identifying processes that influence conflict behavior and relationship functioning, and in clinical work with couples, it is necessary for selecting targets for intervention and for tracking treatment progress.
When assessing attributions, it is important to distinguish between schematic-level assessments and context-specific assessments. A schematic-level assessment pertains to a person’s general schema regarding a partner’s typical culpability and probable future culpability across a variety of different situations and contexts. In contrast, a context-specific assessment pertains to the cognitive appraisals a person makes about his or her partner in regard to a single, specific episode of relationship conflict. For example, a person would be making a schematic-level attribution if he or she thinks, “If we have a conflict, my partner is probably at fault,” whereas a person would be making a context-specific attribution if he or she thinks, “In our conflict last Thursday, my partner was at fault.”
There are several reasons why it might be important to assess attributions at the context-specific level. Studies have identified many aspects of conflict interaction that function at this level and that tend to change across different contexts (Nichols, Backer-Fulghum, Boska, & Sanford, 2015; Sanford, 2005, 2007a, 2012; Sanford & Grace, 2011). For example, Sanford (2012) found that people often experience changes in the types of emotion they experience across different episodes of conflict with a partner, and moreover, these changes in emotional experience predict corresponding changes in behavior. This suggests that attributions might also change across contexts, and if so, a context-specific measure would best capture how attributions actually function in a relationship. In addition, it is important for clinicians and researchers to use context-specific measures when assessing how people respond to interventions. This point is illustrated by results from a meta-analysis conducted by Blanchard, Hawkins, Baldwin, and Fawcett (2009) investigating the effectiveness of couples’ education programs. Their results suggest a pattern in which studies using context-specific measures of communication had more sensitivity for detecting change than studies using schematic measures. If a variable functions at a context-specific level, or if researchers want to assess immediate changes occurring in an interpersonal relationship, then the maximum precision will be gained by using a context-specific measure. As a first step in this line of work, a context-specific measure of attributions needs to be developed and validated.
One of the challenges in developing new instruments for measuring constructs pertaining to interpersonal relationships is selecting items that reliably discriminate between individuals across all levels of the target constructs. For instance, Funk and Rogge (2007) found that existing measures of relationship satisfaction worked well for differentiating between people at the low end of the continuum who were experiencing varying degrees of relationship distress, but failed to discriminate well at the upper end between people experiencing varying degrees of satisfaction. In order to create a scale that reliably discriminated across all levels of satisfaction, Funk and Rogge used item response theory (IRT) to select a set of items that best differentiated between people across all levels of relationship satisfaction. In a similar way, IRT would be useful for creating a measure that discriminates well across all levels of context-specific attribution.
It might also be possible to maximize a scale’s ability to discriminate between people, and also maximize content validity, by including items that are worded to assess both poles of the target construct. In the case of attributions, this would include items assessing both: (a) negative attributions regarding the extent to which a person blames the partner and believes the partner is responsible for causing a negative event and (b) positive attributions regarding the extent to which a person exonerates the partner from blame and believes the partner acted reasonably. However, research with couples finds that items measuring positive and negative types of relationship sentiment sometimes form two distinct factors rather than two poles of a single dimension (Mattson, Rogge, Johnson, Davidson, & Fincham, 2013). Thus, if positive and negative items are included on a single scale, it is important to demonstrate that they actually measure two poles of the same construct, and not two different constructs.
In developing a context-specific measure of attributions, it will also be important to test the new instrument’s construct validity by examining correlations with other context-specific variables. For example, emotion is a context-specific variable that often changes within people across different episodes of relationship conflict (Sanford, 2012) and previous research has identified three types of negative emotion that are particularly salient during conflict interactions: sometimes termed hard, soft, and flat emotion (Sanford, 2007a, 2007b; Sanford & Rowatt, 2004). Hard emotion includes feeling angry and irritated and is associated with asserting power and control (Jacobson & Christensen, 1998). Soft emotion includes feeling sad and hurt and is associated with expressing vulnerability (Jacobson & Christensen, 1998). Flat emotion includes feeling apathetic and indifferent, and drawing from the two-dimensional theory of affect (Watson & Tellegen, 1985), it can be described as an emotion that is low in both positive and negative affect. Attributions should be correlated with both hard and soft emotions and significantly less correlated with flat emotions. Lazarus’s (2001) appraisal theory suggests that people are likely to display anger when they blame another for a negative event, and accordingly, several studies find correlations between attributions and hard emotion (Fincham, Beach, & Nelson, 1987; Fincham & Bradbury, 1992; Sanford, 2005; Senchak & Leonard, 1993). In addition, when people make negative attributions, they may perceive they have been devalued by their partners or that they have lost a desired level of intimacy and closeness in their relationships. Because soft emotion is associated with the desire for closeness, intimacy, and commitment (Clark, Pataki, & Carver, 1996; Leary & Springer, 2001; Sanford, 2007b), and with perceptions of being devalued (Leary & Springer, 2001), soft emotion should also be correlated with making negative attributions. In contrast, people are observed to express flat emotion when they withdraw from conflict (Coan & Gottman, 2007), and flat emotion may be most associated with appraisals that would influence a person’s decision to withdraw (such as issue importance, efficacy to resolve conflict, and expected outcomes) and less associated with appraisals regarding a partner’s responsibility and blame. In sum, this means that attributions should not merely demonstrate convergent correlations with both hard and soft emotion, but attributions should also demonstrate divergent validity by having a weaker association with flat emotion.
Attributions should also be associated with couples’ underlying concerns during conflict interactions. Underlying concerns are peoples’ primary reasons for feeling distressed during conflicts, and previous research finds that concerns often fall along two basic dimensions: perceived threat and perceived neglect (Sanford, 2010b, Sanford & Grace, 2011; Sanford & Wolfe, 2013). A perceived threat occurs when a person believes that his or her partner is being assertive, controlling or aversive, and threatening one’s status or power in the relationship. In contrast, perceived neglect occurs when a person believes that a partner is failing to show investment or commitment or failing to make a desired contribution to the relationship. As with emotion, research finds that underlying concerns are context-specific and people show meaningful variation in their concerns across different episodes of conflict (Sanford & Grace, 2011). Importantly, a measure of attributions should have different levels of association with each type of concern. Attributions involve judgments about a partner’s culpability and motivation, and these types of judgments may be especially relevant for evaluating a partner’s level of commitment and investment in a relationship, but may be less salient for evaluating the extent to which a partner’s behavior is assertive, aversive, or potentially harmful to one’s own self interests. In other words, people might sometimes view a partner as being innocently aversive, but rarely as being innocently uncommitted. Accordingly, Sanford (2010b) finds that attributions are more strongly associated with perceived neglect than with perceived threat.
In developing a new measure of context-specific attributions, it is important to examine associations between the new scale and existing measures of both schematic attributions and relationship satisfaction. For example, the Relationship Attribution Measure (RAM; Fincham & Bradbury, 1992) is a widely used and well-validated measure of schematic attributions. It asks respondents to rate the types of attributions they would typically make in response to a set of hypothetical scenarios, and it includes two scales, one pertaining to appraisals involving cause and one pertaining to appraisals involving responsibility and blame. A context-specific measure of negative attributions would be expected to be correlated with the RAM, because they are both measuring types of attributions, however, these measures should not be redundant. When predicting context-specific criterion variables, such as emotion and underlying concerns, a context-specific measure of attributions should explain unique variance beyond what can be explained by the RAM. In addition to being distinct from the RAM, a new measure of context-specific attributions should also be distinct from measures of relationship satisfaction. It is sometimes difficult to develop new questionnaire scales measuring different aspects of relationship functioning because, due to “sentiment override” (Weiss, 1980), people sometimes respond to items on the basis of their global feelings of relationship satisfaction. Thus, it is important that a new measure of context-specific attributions explains unique variance in criterion variables, not only after controlling for scores on the RAM but also after controlling for relationship satisfaction.
In sum, an instrument for assessing context-specific attributions would ideally have three characteristics. First, the measure would use the fewest number of items to discriminate reliably across the full range of attribution levels. Second, it would demonstrate construct validity by showing an expected pattern of correlations with criterion variables such as measures of emotion and underlying concerns. Finally, although measures of context-specific attributions should be moderately correlated with a measure of schematic attributions, it should explain variance in context-specific criterion variables that cannot be explained by a schematic-level measure.
Study 1
In Study 1, a new context-specific attribution measure was developed and tested with a large, Internet sample of married and cohabiting participants that included people in both distressed and nondistressed relationships and included people reporting on both recent and distant episodes of relationship conflict. All participants completed a pool of items measuring both blaming and exonerating attributions. In order to decrease the number of items and still be able to discriminate reliably across the full range of attribution levels, results from a graded response model (GRM; Samejima, 1969) were used to select a set of items. In addition, this study addressed three questions. First, to what extent do attributions correlate with criterion variables regarding emotion and underlying concerns? Second, to what extent do attributions predict criterion variables after controlling for relationship satisfaction? Third, are there differences between criterion variables in the magnitude of association with attributions?
Method
Participants
Study 1 used a sample of 2,452 people who were either married (94.5%) or cohabiting with a partner (5.5%). The age of the participants ranged from 18 to 87 years (M = 40.33, SD = 12.80) and, for the married participants, length of marriage ranged from less than 1 year to 51 years (median = 10 years, M = 13.4 years, SD = 12.0 years). The sample was 66.7% female, 68.3% White, 7.9% African American, 10.9% Hispanic or Latino, 8.8% Asian, and 4.1% reporting other racial or cultural backgrounds. Annual family income ranged from less than US$10,000 to more than US$500,000 (median = US$92,500, M = US$112,700, SD = US$98,591). The mean score on the Couples Satisfaction Index was 50.21 (SD = 21.41), indicating that, in comparison to the norming sample for the Couples Satisfaction Index (M = 61, SD = 17, as reported by Funk & Rogge, 2007), the present sample included a wider range of distress scores and greater representation of highly distressed people.
Procedure
Participants visited an interactive website which allowed them to create an individual account, complete an anonymous questionnaire, submit responses, receive personalized feedback, and view a resource bank of information for couples. Approximately 60% of the participants were recruited by psychology students over the course of eight semesters. Students invited their married parents, relatives, and friends to complete the questionnaire, and they received credit when invitees sent e-mails to the instructor stating that they either completed a questionnaire or declined the invitation. The remaining participants discovered the questionnaire via an online search or from unsolicited links displayed on news websites, blogs, and other webpages (e.g., ABC News and ScienceDaily). If both members of a couple wanted to complete an assessment, the first member to use the program was instructed to send a link to the second member, thereby identifying their profiles as belonging to a single couple. There were 581 cases in which both members of a single couple completed assessments, and in these cases, one partner was selected at random for inclusion and the other dropped; hence, all 2,452 participants were independent. At the beginning of the questionnaire, participants were asked to “Think about a single, specific episode of conflict in your relationship.” A text box was given and participants were asked to write a brief description of the conflict. Participants then completed several measures pertaining to the identified conflict. To be included in the data analysis, participants needed to demonstrate a sufficient level of persistence by completing all sections of the questionnaire and also to respond affirmatively to the question “Will you be submitting serious answers to this questionnaire?”
Measures
Context-specific attributions
Participants responded to a pool of 14 attribution items. Items in this pool specifically focused on assessing the responsibility–blame dimension of attributions regarding judgments that a person is at fault for violating a moral imperative (Fincham & Bradbury, 1992). Blaming attributions were conceptualized as involving appraisals that a partner acted on culpable motives, and exonerating attributions were conceptualized as involving appraisals that a person’s motives were valid and understandable. The pool included seven attributional blame items that comprise a scale previously developed by Sanford (2010a), and also seven new attributional exoneration items designed to capture different ways people might view a partner’s motives and behaviors as valid and understandable. The entire pool of items can be found in Table 1. All items were rated in regard to the specific conflict interaction participants identified at the beginning of the questionnaire, and they were rated on a 5-point scale (1 = disagree strongly, 2 = disagree, 3 = agree somewhat, 4 = agree, 5 = agree strongly). Exonerating items were reverse scored, thereby making it possible to create a total score, indicating the extent of blame or negativity in attributions, by averaging across items.
Thresholds and Discrimination Parameters for the Context-Specific Attribution Scale.
Underlying concerns
The 16-item Underlying Concerns Inventory (Sanford, 2010b) was used to measure levels of perceived threat and perceived neglect during the conflict interaction participants identified at the beginning of the study. The perceived neglect scale includes items such as “I felt overlooked” or “My partner seemed uncommitted,” and the perceived threat scale includes items such as “I felt accused” or “My partner seemed demanding.” Previous research finds that the instrument fits an expected two-dimensional factor structure (Sanford, 2010b) and that each scale correlates with specific types of desires for conflict resolution (Sanford & Wolfe, 2013). In the present study, Cronbach’s alpha was .93 for both threat and neglect.
Emotions
The Couples Emotion Rating Form (CERF, Sanford, 2007a) was used to assess hard, soft, and flat emotion during the conflict interaction participants identified at the beginning of the study. The hard emotion scale includes four items measuring feelings of anger, annoyance, and irritation; the soft emotion scale includes four items measuring feelings of sadness, hurt, and concern, and the flat emotion scale includes four items measuring feelings of indifference, disengagement, and boredom. The CERF was developed and validated in a series of studies (Sanford, 2007a, 2007b; Sanford & Rowatt, 2004) finding that the CERF fits an expected factor structure, corresponds with observer ratings of expressed emotion, and that changes in emotion predict corresponding changes in communication behavior, cognition, and conflict resolution (Sanford & Grace, 2011). In the present study, Cronbach’s alphas were .84, .83, and .83 for hard, soft, and flat emotion, respectively.
Relationship satisfaction
The 16-item version of the Couples Satisfaction Index (Funk & Rogge, 2007) was used to assess overall relationship satisfaction. The measure was developed using IRT analysis to select discriminating items from a pool of items drawn from several existing measures. The Couples Satisfaction Inventory (CSI) demonstrates strong convergent validity with other measures of satisfaction (Funk & Rogge, 2007). Sample items include the following: “My relationship with my partner makes me happy” and “I have a warm and comfortable relationship with my partner.” In the present study Cronbach’s alpha was .98.
Analytic Approach
The first step in analysis was to test the factor structure of the item pool. An assumption of IRT is that the item pool is essentially unidimensional, and although parameter estimates can be robust to violations of this assumption (Harrison, 1986; Kirisci, Hsu, & Yu, 2001), it is important that the pool have a single dominant factor. In the present study, the minimum criteria were that the first component should explain at least 20% of the variance (as recommended by Reckase, 1979), the acceleration factor for a scree plot (indicating the point where the slope of the curve changes abruptly; Raîche, Walls, Magis, Riopel, & Blais, 2013) should indicate one factor, and if a single principal component is extracted, all the items should load greater than .55 on that component (which Comrey & Lee, 1992, define as “very good” loadings). After testing dimensionality, data were analyzed using a GRM (Samejima, 1969), which is an item response model appropriate for ordered categorical responses. Each item in the GRM is characterized by a discrimination parameter and Ki−1 category threshold parameters, where Ki is the number of categories in the response scale on item “i.” The discrimination parameters specify how well each item discriminates between people (Gummelt, Anestis, & Carbonell, 2012). The item thresholds indicate the levels for the latent attribution construct (in z score units) at which each response option for each item best discriminates. As such, they provide information on the extent to which each item is sensitive across the entire range of attribution levels (Gummelt et al., 2012). Results from the GRM were used to create a final scale composed of a small number of items that discriminated reliably across the full range of attribution levels. After selecting items for the final scale, a test information function (TIF) was computed for this scale by summing the amount of item information provided by all items at each level (see Baker, 2001).
Correlations and regression equations were used to test associations between the final scale and criterion validity variables. To provide information on how the new scale might function across different groups or different contexts in which it might be used, analyses were also run after splitting the sample into three groups. The first group included 492 people in highly distressed relationships. People were included in this group if their score on the CSI was more than 1 standard deviation below the mean (<29), and notably, this produced a substantially more distressed group than would have resulted from using the originally suggested distress cutoff for the CSI (which was 51; Funk & Rogge, 2007). The remaining nondistressed people were further divided into two groups: 907 people who completed assessments in regard to recent conflicts occurring within the past week and 1,053 people rating nonrecent conflicts occurring more than a week ago.
Results
Preliminary tests regarding factor structure suggested that the entire pool of 14 items included a single dominant factor. In a principal component analysis, the first component explained 50% of the variance, the scree plot had an acceleration factor of one, and all items loaded greater than .63 on the first component. Although this provided sufficient evidence for a single dominant factor, it is notable that two components produced eigenvalues greater than one (7.05 and 1.69, respectively).
A GRM was estimated using the entire pool of 14 items, and the threshold and discrimination values are listed in Table 1. These values, along with item information curves, were used to identify the best six items for inclusion on a context-specific attribution scale. First, to maximize sensitivity, four items with the highest overall discrimination were selected. These were also the four items with the highest peaks on the item information curves. Then, to maximize the range of discrimination, one item with the highest threshold and one item with the lowest threshold were selected. On the item information curves, these were the items providing the most information at 3 standard deviations above and 3 standard deviations below the mean. The selected items are shown in Table 1, and the final scale included three attributional blame items and three reverse-scored, attributional exoneration items.
The final set of six items was clearly unidimensional. In a principal component analysis, the first component explained 56% of the variance, the scree plot had an acceleration factor of one, and all items loaded greater than .64 on the first component. In addition, only one component produced an eigenvalue greater than one (eigenvalues were 3.36 and 0.95 for the first two components). A GRM was estimated using only the final set of six items. Discrimination parameters for each item are displayed in Table 1, and the values are quite similar to the values from the first model using all 14 items. The TIF for the final scale was calculated and it is depicted in Figure 1. When interpreting these results, it is helpful to note that an information value of 5 is analogous to a reliability of .8 in classical test theory (see Thissen, 2000). As can be seen in the figure, the TIF was above 5 and was relatively flat across attribution levels ranging from 2 standard deviations below the mean to 2 standard deviations above. This suggests that the context-specific attribution scale discriminated both adequately and equally across a wide range of attribution levels.

Test information function for the six-item context-specific attribution scale.
Scores for the new six-item context-specific attribution scale were calculated by reverse scoring the exoneration items and averaging across all six items. Cronbach’s alpha was .84. Means and standard deviations were calculated separately for the three groups. Within the nondistressed participants, there was not a significant difference between people with recent and nonrecent conflicts (M = 2.97, SD = 0.74, and M = 2.91, SD = 0.84, respectively). As might be expected, the distressed group scored significantly higher than the pair of nondistressed groups, distressed group M = 3.65, SD = 0.76, d = .90, t(2450) = 17.90, p < .001. There was also a significant gender difference in attribution scores, d = .48; t(2450) = 11.11, p < .001, with women scoring higher (M = 3.21, SD = 0.82) than men (M = 2.82, SD = 0.81).
Correlations were computed between the context-specific attribution scale and each of the criterion variables (including three types of emotion and two types of underlying concerns) and also relationship satisfaction. Correlations were computed separately for each of the three groups, and t tests were used to test for differences in magnitude between correlation pairs. Results are listed in Table 2, and all hypothesized correlations were significant for all groups. Consistent with hypotheses, attributions were moderately correlated with both hard emotion and soft emotion, and across all groups, these correlations were significantly larger in magnitude than the correlation between attributions and flat emotion. In addition, attributions were strongly correlated with perceived neglect, and across all groups, this correlation was larger in magnitude than the correlation between attributions and perceived threat.
Correlations Between Attribution Scales and Criterion Variables.
Note. CSA = context-specific attribution scale; RAM = Relationship Attribution Measure.
p < .05. **p < .01.
A series of regression equations were used to estimate the extent to which the attribution scale predicted each criterion variable after controlling for relationship satisfaction, and the standardized beta weights from these results are listed in Table 3. After controlling for ratings of relationship satisfaction, negative attributions still explained unique variance in hard emotion, soft emotion, perceived threat, and perceived neglect (but not flat emotion). These equations were also tested including relationship length as a covariate, and results were nearly identical. In sum, the attribution scale demonstrated expected differences in magnitude of correlations with different criterion variables, these results were similar across three different groups (distressed participants, non-distressed participants reporting recent conflicts, and non-distressed participants reporting non-recent conflicts), and all hypothesized convergent relationships remained significant after controlling for relationship satisfaction.
Study 1 Regression Results: Attributions and Satisfaction Predicting Criterion Variables.
Note. CSA = context-specific attribution scale. CSA ΔR2 values indicate the unique variance explained by CSA controlling for satisfaction.
p < .05. **p < .01. ***p < .001.
Study 2
Study 2 was designed to address three goals. The first goal was to test the extent to which results from Study 1, regarding convergent and divergent correlations with measures of emotion and underlying concern, could be replicated using other populations of people in relationships. Whereas Study 1 used a sample of married and cohabiting people, Study 2 used a younger sample of college students in romantic relationships. The second goal was to compare the new context-specific attribution measure with the RAM (Fincham & Bradbury, 1992), a schematic attribution measure. The RAM contains two scales, one for causal attributions and one for responsibility–blame attributions, and both scales were expected to correlate with the new context-specific measure. However, the new context-specific measure was also expected to be distinct from the RAM scales, as well as from relationship satisfaction. Specifically, it was expected to explain unique variance in the criterion variables (regarding emotion and underlying concern) after controlling for scores on both RAM scales and relationship satisfaction. Moreover, the unique variance explained by the context-specific measure was expected to be significantly greater than the unique variance explained by each of the RAM scales.
Method
Participants
The participants in this study included 172 (136 female, 36 male) undergraduate students who were in a romantic relationship. The length of relationship was less than 1 year in 44.8% of the cases, and more than 1 year for the remaining 55.2%. The sample was 64.5% White, 9.3% Asian, 8.7% African American, 13.4% Hispanic and Latino, and 4.1% other. Most participants (96.5%) were dating and not living with their partner, while 3.5% were cohabiting with their partners.
Procedure
Participants were recruited via a university department subject pool, and were included in the study if they reported that they were currently in a romantic relationship. As in Study 1, participants completed an online questionnaire. At the beginning of the questionnaire, participants were instructed to “Think about a single, specific episode of conflict in your relationship,” and to write a brief description of their conflict in a text box. Participants were then asked to complete several questionnaires regarding the identified conflict.
Measures
Context-specific attributions, emotions, underlying concerns, and relationship satisfaction
Participants completed the new six-item, context-specific attributions scale that was developed in Study 1. As in Study 1, they also completed the CERF (Sanford, 2007a) and the Couples Underlying Concern Inventory (Sanford, 2010b). Cronbach’s alphas were .84 for the context-specific attribution scale, .88, .79, and .80 for hard, soft, and flat emotion, respectively, and .91 and .92 for threat and neglect, respectively. Participants also completed the 32-item version of the CSI (Funk & Rogge, 2007), which is a longer version of the same instrument used in Study 1 and in Study 2, Cronbach’s alpha was. 94.
Relationship Attribution Measure
The 24-item version of Fincham and Bradbury’s (1992) RAM was used to measure schematic attributions. Items were modified to be appropriate for dating partners by changing the words “wife” and “husband” to “partner.” Respondents were given a series of four hypothetical scenarios involving common types of negative interaction (e.g., “Your partner criticizes something you say”), and for each scenario, they were asked to rate three types of negative causal attributions (e.g., “My partner’s behavior was due to something about him or her”) and three types of negative responsibility–blame attributions (e.g., “My partner’s behavior was motivated by selfish rather than unselfish concerns”). Cronbach’s alpha was .80 for the causal attribution scale and .90 for the responsibility–blame attribution scale.
Results
As an initial step in data analysis, means and standard deviations were calculated. The average score on the new, context-specific attribution scale was 2.66 (SD = 0.85; range = 1-5). Similar to Study 1, there was a significant gender difference in attribution scores. d = .37; t(170) = 2.13, p = .04, with women scoring higher (M = 2.72, SD = 0.88) than men (M = 2.43, SD = 0.68). Compared with the sample of married people used in Study 1, the sample of dating people used in Study 2 reported lower levels of blaming attribution, d = .50, t(2624) = 6.33, p < .001, and higher levels of relationship satisfaction, using satisfaction scores based only on items common to both studies, d = .63, t(2624) = 7.25, p < .001.
Correlations were computed between the context-specific attribution scale and the other variables, and these are listed in Table 2. As in Study 1, the context-specific attribution scale produced a significantly stronger correlation with perceived neglect than it did with perceived threat (which, in Study 2, was not significantly correlated with the context-specific attribution scale). Also consistent with Study 1, the context-specific attribution scale was significantly correlated with both hard emotion and soft emotion, and the scale’s correlation with hard emotion was significantly stronger than its correlation with flat emotion. However, in contrast to Study 1, the context-specific attribution scale’s correlation with hard emotion increased to a magnitude that was significantly larger than its correlation with soft emotion, and the scale’s correlation with flat emotion increased to a magnitude that was no longer significantly different from its correlation with soft emotion.
Correlations were also computed between the context-specific attribution scale and the two attribution scales from the RAM. The context-specific attribution scale was strongly correlated with the RAM Responsibility Attributions scale (r = .55, p < .001) and moderately correlated with the RAM Causal Attributions scale (r = .39, p < .001). These results provide further support for the convergent validity of the context-specific attribution scale.
A series of regression equations were used to estimate the extent to which the context-specific attribution scale explained unique variance in the context-specific criterion variables after controlling for satisfaction and schematic attributions. A separate equation was estimated for each of the five different context-specific criterion variables (hard emotion, soft emotion, flat emotion, perceived neglect, and perceived threat), and each criterion was regressed on the context-specific attribution scale, relationship satisfaction scale, and both of the RAM scales. The results are listed in Table 4. This analysis differs from Study 1 in that the equations controlled not only for satisfaction but also for the two schematic attribution scales, yet the results were largely consistent across both studies. Specifically, context-specific attributions explained unique variance in both types of underlying concern and in all the emotion variables except for flat emotion.
Study 2 Regression Results: Attributions and Satisfaction Predicting Criterion Variables.
Note. CSA = context-specific attribution scale; RAM = Relationship Attribution Measure. CSA ΔR2 values indicate the unique variance explained by CSA controlling for the other three predictors.
p < .05. **p < .01. ***p < .001.
As expected, several of the standardized beta weights listed in Table 4 appeared to be larger for context-specific attributions than for the two schematic attributions. To provide a direct test of this possibility, all of the regression equations were reestimated using LISREL 9.1, and in each equation, the standardized beta for the context-specific attribution scale was constrained to equal the standardized beta for one of the two RAM scales. This produced a model with one degree of freedom testing the null hypothesis that the two standardized beta weights were equal. The standardized beta weight for the context-specific attribution scale was significantly stronger than the standardized beta weights for both RAM scales in predicting hard emotion, but differences did not reach significance for soft emotion and perceived neglect. Notably, in testing the associations with perceived threat (which were expected to be small), the beta weight for the context-specific attribution scale became negative, and significantly different from both of the RAM scales.
Discussion
A new measure of context-specific attributions was developed and validated in a series of two studies, with the first using a large sample of married and cohabiting couples and the second using a smaller sample of dating couples. IRT was used to maximize scale discrimination using the fewest number of items. The final scale included six items: three blaming attribution items regarding the extent to which a person believed the partner was at fault for causing a conflict and three reverse-scored exonerating attribution items regarding the extent to which a person believed his or her partner had viewpoints and feelings that were valid and reasonable. The scale was unidimensional and it demonstrated an expected pattern of correlations with criterion variables such as measures of emotion and underlying concern. Moreover, the context-specific attribution scale explained unique variance in these context-specific criterion variables even after controlling for two schematic-level attribution scores from the RAM and a measure of relationship satisfaction. In sum, the new measure is short, yet discriminates well across a wide range of attribution levels; it is distinct from other scales measuring schematic-level attributions and relationship satisfaction, and potentially useful for investigating components of conflict interaction that occur at a context-specific level.
An important feature of the new scale is that it produced different levels of association with two types of underlying concern. In samples of both married people and dating people, context-specific attributions were strongly correlated with concerns involving perceived neglect, and significantly less correlated with concerns involving perceived threat. The difference was especially pronounced in the regression results from the sample of dating people. When predicting neglect in this sample, the beta was positive, but when predicting threat, there was a suppressor effect with a negative beta. The observed differences between correlations are consistent with previous research on attributions and underlying concerns (Sanford, 2010b), which finds that negative attributions are more strongly associated with perceived neglect than perceived threat. This makes theoretical sense. Attributions involve appraisals about whether a partner violated standards for trust and cooperative interaction in a relationship, and these appraisals may be especially important for evaluating a partner’s level of commitment and investment (components of perceived neglect), but may be less important for evaluating a partner’s level of power and for determining whether a partner is voicing complaints (components of perceived threat). The results of the present study are consistent with the possibility that each type of concern is associated with different types of appraisal (Sanford, 2010b), and this builds on other research indicating that each type of concern is also associated with a different set of desires (Sanford & Wolfe, 2013). Notably, tests for differences between correlations involving types of concern were only significant for the new context-specific measure and not for the RAM scales measuring schematic-level attributions. This is consistent with the fact that underlying concerns are presumed to function at a context-specific level, and it suggests the new measure is particularly useful for investigating interpersonal process at this level.
As expected, in both samples, the new scale was strongly correlated with feelings of anger and it was significantly less correlated with a measure of flat negative emotion (feeling bored and apathetic). Moreover, in a regression equation predicting anger, the standardized beta weights for the new scale were significantly larger than the beta weights for the schematic-level attribution scales from the RAM. The effects regarding anger are noteworthy because anger is an especially important variable in attribution theory. Appraisal theory (Lazarus, 2001) suggests that blaming attributions lead to feelings of anger, and accordingly, several previous studies found correlations between attributions and anger (Fincham & Bradbury, 1992; Fincham et al., 1987; Sanford, 2005; Senchak & Leonard, 1993). To the extent that anger is best understood as a context-specific variable that can change within people across different episodes of conflict (Sanford, 2012), it makes sense that the context-specific measure of attributions would be especially well suited for capturing the association with anger.
The present study provides some preliminary results regarding similarities and differences across different groups and different assessment contexts. For example, women consistently scored higher than men on the new context-specific attribution measure. While it is not clear if women actually make more negative attributions, or if they are simply more likely than men to seek help from relationship websites when making negative attributions, it is notable that similar gender differences have been found in studies using schematic-level attribution measures (Marshall et al., 2011). The present study also found that people in distressed relationships report higher levels of negative attributions; however, the pattern of correlations with criterion variables was essentially the same for distressed and nondistressed people. It was also the same for people reporting on recent conflict episodes and people reporting on conflicts that occurred more than a week ago. Although the present study did not directly assess validity in contexts where participants were specifically instructed to recall recent (or nonrecent) conflicts, the results are at least consistent with a study by Backer-Fulghum and Sanford (2015) which found that the validity of a self-report measure of conflict communication was not noticeably moderated by the recentness of the conflicts being assessed.
Some particularly notable group differences were evident in the contrast between results from the sample of married people and the sample of college students in dating relationships. The sample of dating people reported more satisfaction and lower levels of attributional blame. This may reflect the fact that it is easier for dating people than for married people to dissolve a relationship when they are dissatisfied or when they believe a partner is highly culpable for causing relationship problems (Cupach & Metts, 1986). Also, in the sample of married people, context-specific attributions correlated more strongly with soft emotion than with flat emotion, as predicted. In contrast, in the sample of dating people, the difference between these correlations was not significant. While this may be due to the fact that the second study used a smaller sample and thereby had more error variance, it also may be due to the fact that the correlation between attributions and flat emotion was somewhat larger in the sample of dating people than the sample of married people. Previous research finds withdrawal from relationship is associated with flat emotion (Nichols et al., 2015), and thus, there may be reason to speculate that dating people are less committed and thereby more likely than married people to withdrawa from a relationship and experience flat emotion when they make blaming attributions. Taken together, the results suggest that there may be differences in attribution processes between marriage relationships and dating relationships; however, these ideas need to be tested directly before any firm conclusions can be made.
Overall, the results revealed several ways in which the context-specific measure of attributions was distinct from the schematic-level attribution scales. The context-specific measure was more strongly correlated with anger, and more sensitive to the distinction between threat and neglect. At the same time, it is important to note that the context-specific measure was correlated with the schematic-level scales, and this correlation was especially large for the RAM scale measuring attributions regarding responsibility and blame. This large correlation highlights the difficulty in teasing apart the difference between effects occurring at a context-specific level and effects occurring at a schematic level. Specifically, if a study finds a correlation between a schematic-level measure of attributions and an outcome, it is not clear whether this correlation reflects a schematic-level process, or a spurious correlation driven by processes occurring at the context-specific level in which context-specific attributions are a third variable that correlate with both the schematic-level measure and the outcome. The new context-specific measure of attributions could be used in future studies to tease these effects apart. Specifically, if context-specific attributions are assessed across multiple episodes of conflict, multilevel modeling can be used to distinguish context-specific effects (which involve within-person variance) from schematic effects (which involve between-person variance). Other research on conflict in couples has used this approach and found that associations between variables involving the experience of emotions, communication behavior, conflict disengagement, the perception of partner emotions, and the experience of underlying concerns all occur at a context-specific level (Nichols et al., 2015; Sanford 2007a, 2012; Sanford & Grace, 2011). In sum, attributions may have important effects at the context-specific level, and the new context-specific attribution scale may be useful in future research testing these effects.
Although the new measure shows promise as a measure of context-specific attributions, it is important to note that participants in the current studies were assessed only on a single occasion, and consequently, it was not possible to test the extent to which the new measure was sensitive to changes occurring within people over time and across different contexts. A key benefit to using a context-specific measure lies in the measure’s ability to assess within-person change, and it will be important to test this in future research. It is also important to note that the current project used a sample of married and cohabiting people with relatively high-income levels, and a sample of dating people who were all college students. While it was beneficial to test the new measure using two different populations, it is not clear how results will generalize to other populations that may be of interest to clinicians and researchers, such as people currently receiving couples therapy, people completing assessments in a laboratory setting, people experiencing particular types of relationship problems such as aggression, and low-income populations. In addition, both studies relied on data collected over the Internet, so the environments in which these assessments were taken could not be controlled, thereby increasing the chances of distractibility (Clifford & Jerit, 2014) and multitasking (Chandler, Mueller, & Paolacci, 2014).
The present studies build other work using IRT to create scales for assessing interpersonal relationships, such as Funk and Rogge’s (2007) study developing a measure of relationship satisfaction (the CSI). Like the present study, Funk and Rogge also found that it was possible to create highly discriminating scales using only a small number of items. In addition, studies using IRT provide results regarding test information curves which can be informative, especially when compared across studies. For example, Funk and Rogge (2007) found that test information curves for satisfaction scales tend to drop precipitously around 1 standard deviation above the mean, whereas the present study found the information curve for attributions to remain relatively flat across a wide range of attribution levels. In other words, satisfaction may have a low ceiling, and there may be an absence of meaningful differences between varying extremes of high satisfaction; and in contrast, there may be meaningful differences between people with varying extremes of high (and low) partner exoneration. Taken together, these studies show how IRT can be useful in constructing efficient scales for assessing interpersonal relationships and informative in clarifying the nature of the constructs assessed by these scales.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported in part by a grant from the Baylor University Research Committee.
