Relating Covariation Information to Causal Dimensions through Principles of Contrast and Invariance

Frank J. Van Overwalle and Francis P. Heylighen
Vrije Universiteit Brussel, Belgium

We are grateful to Vivianne De Schaepdrijver and Nils Burckhard for serving diligently as experimenters. We also sincerely thank Dennis Hilton for his very valuable suggestions and Daan Van Knippenberg for his statistical advice. For correspondence write to : Frank Van Overwalle, Vrije Universiteit Brussel, Department of Psychology, Pleinlaan 2, B-1050 Brussels, Belgium.


This paper examines the proposition that covariation information guides judgments about the dimensionality of attributions on the basis of causal principles of contrast and invariance, which are derived from Mill's methods of difference and agreement respectively. It is argued that the standard attribution categories specified in earlier research (e.g., person, occasion and stimulus) represent just one extreme of the attributional dimensions and require the principle of contrast, whereas additional attributional categories reflecting the opposite extreme of the dimensions (e.g., external, stable, general) require the principle of invariance. In three studies, subjects were given covariation information, and were asked to rate the properties of the likely cause along the dimensions of locus, stability, globality and control. In line with the predictions, consensus with others, consistency in time, distinctiveness between stimuli and contingency of one's actions showed the strongest effects on judgments of locus, stability, globality and control respectively. Similar results were obtained in a fourth study, where subjects had to judge the influence of eight causes with varying dimensional properties. Moreover, these judgments were rated somewhat higher given causes requiring the principle of invariance rather than the principle of contrast.


How do people infer a causal explanation[1 for some event ? Theory and research in the attribution literature that people make use of covariation information about the event to solve an attributional question. Drawing on Mill's (1872/ 1973) ]joint method of agreement and difference, Kelley (1967) postulated that an effect will be attributed to the condition with which it covaries : "the effect is attributed to that condition which is present when the effect is present and which is absent when the effect is absent" (Kelley, 1967, p. 194). Although Kelley's model received some support (McArthur, 1972, 1976; Ruble & Feldman, 1976), the original covariation notion has subsequently been refined and complemented by several alternatives such as the template-matching model (Orvis, Cunningham & Kelley, 1975), the logical model (Hewstone & Jaspars, 1987; Hilton & Jaspars, 1987; Jaspars, 1983), the abnormal conditions focus model (Hilton & Slugoski, 1986) and, most recently, the probabilistic contrast model (Cheng & Novick, 1990) and the dispositional model (Hilton, Smith & Kim, 1994).

Although the models that developed from Kelley's (1967) original model differ strongly with respect to how covariation information is used and processed, virtually all of them agree in one important respect : they all assume that people's causal attributions can be captured in three responses, including something about the person, something about the occasion, something about the stimulus, and a combination of these three factors. These three responses, which we have termed standard responses, have been widely used and considered sufficient in the literature to represent people's entire attributional universe.

This untested assumption stands in sharp contrast with findings from research on the dimensionality of attributions. Research on every-day causal attributions has revealed that they can be categorized along four fundamental dimensions of locus (internal vs. external), stability (stable vs. variable over time), globality (general vs. situation-specific), and control (changeable versus unchangeable by one's actions; see Abramson, Seligman & Teasdale, 1978; de Jong, Koomen & Mellenbergh, 1988; Lunt, 1988; Van Overwalle, 1989; Weiner, 1986). In the most simple case, these dimensions are viewed as dichotomies and thus reflect eight possible causal categories. Clearly, three of these categories are also represented by the standard responses specified in earlier attribution models : Attributions to the "person" reflect internal explanations, attributions to the "occasion" reflect variable explanations, and attributions to a "stimulus" represent specific causes.

However, as noted before, earlier theory and research that developed from Kelley's (1967) approach was limited to these three standard causal categories and did not specify responses to represent the opposite causal categories of external, stable and general attributions. Moreover, none of the responses reflected the dimension of causal control. We argue that these additional causal categories are an inherent part of a person's causal thinking and psychology, and that these responses should therefore be included in any general theory of the attribution process. For instance, failure to arrange a date with a girl can be explained not only by the fact that the actor did not call the girl sufficiently in advance (a joint attribution to the actor, occasion and the stimulus), but also by strict religious rules (an external, stable and general attribution). Most theories in applied attributional domains typically specify a larger variety of attributions, including the non-standard categories discussed above, for example in theories of motivation and emotion (e.g., Weiner, 1986), learned helplessness (e.g., Peterson & Seligman, 1984; consumer behavior (Folkes, 1984), marriage (e.g., Bradbury & Fincham, 1990) and others. Below, we will present a model of causal inference that specifies principles of contrast and invariance, through which people can make attributions to both standard and non-standard (opposite) causal factors. As these principles are derived from Mill's (1872/1973) joint method of agreement and difference, we have likewise termed our model a joint model of causal attribution.

Principles of Contrast and Invariance

The proposed joint model represents a more comprehensive integration of Mill's (1872/1973) methods of both difference and agreement than earlier attribution models. Kelley's (1967) covariation model and its later theoretical extensions mainly developed on the method of difference (e.g., Jaspars, 1973; Hilton & Slugoski, 1986, Cheng & Novick, 1990), and left relatively unspecified the complementary method of agreement (for a similar argument see Hilton, Smith & Kim, 1994). However, Mill (1872/1973) proposed both methods as "potent instruments" for causal inquiry. We will illustrate that these two methods are required to make attributions to standard and non-standard causes.

Mill's method of difference compares the target event where the effect is present to other events in which the same effect is absent, and specifies the cause as the one condition that differs between these two events.[2 Cheng and Novick's (1990) probabilistic contrast model is an instantiation of this method, and is based on their ]contrast principle which specifies that "a factor is designated a cause if the proportion of times the effect occurs when that factor is present is greater (by some criterion) than the proportion of times the effect occurs when the factor is absent" (Cheng & Novick, 1990, p. 561). Thus, if the contrast between those proportions is sufficiently large, causality is attributed to that factor. For example, the information that Mary always plays with children while most other persons do not, indicates that Mary is "different" and designates something special about Mary as the cause.

Borrowing from Kelley (1973), Cheng and Novick (1990) presented three information variables which reflect a high or low contrast between a target event and comparison events : Consensus reflects the extent to which a target person's outcomes resemble those of other persons, consistency denotes the frequency with which similar effects occurred in the past, and distinctiveness indicates the degree to which the effect of the target stimulus differs from that of other stimuli. For example, imagine that John failed a maths exam. The low consensus information that nobody else failed reveals that the cause is something about the person, John; the low consistency information that John never failed before indicates that the cause is something about the current occasion; and the high distinctiveness information that John passed the other exams reveals that the cause is something about the stimulus, the maths exam. In sum, the condition that reflects the "difference" is designated as the cause. Conversely, Cheng and Novick (1990, 1991) predicted that high consensus, high consistency and low distinctiveness refer to a lack of contrasts and therefore these factors are seen as causally irrelevant or as mere enabling conditions.

We argue, however, that these latter information variables can result in causal attributions by applying Mills' method of agreement. The method of agreement compares the target event to other events in which the same effect is invariably present, and specifies the cause as the one condition that is also invariably present among a number of otherwise varying conditions : "If two or more instances of the phenomenon under investigation have only one circumstance in common, the circumstance in which alone all the instances agree, is the cause of the given phenomenon" (Mill, 1872/1973, p. 390). For example, if Mary always plays a particular game, causality is attributed to this game because that factor is invariably present together with her playing behavior. The method of agreement thus enables attributions to factors that contain no contrasts, in contradiction to the predictions of Cheng and Novick (1990, 1991). In particular, we propose that when contrasts are absent, people will infer that the cause does not fall on the standard extreme of the dimension, and that it must therefore necessarily fall at the opposite (i.e., non-standard) extreme. For example, high consensus information that all students failed the exam reveals that the cause is external to the target person (as opposed to something internal); the high consistency information that John always failed in the past reveals that the cause is stable (as opposed to some temporary occasion); and the low distinctiveness information that John failed on all courses reveals that the cause is something general (as opposed to something specific about the stimulus). We define this instantiation of the method of agreement as the invariance principle, because we predict that non-standard attributions to the opposite extremes of the causal dimensions require invariant, as opposed to contrastive information variables.

Table 1 presents the predictions of our joint model which incorporates both principles of contrast and invariance. The information variables are listed in the first column in their conventional labels, while the causal categories are depicted in the last column. The information patterns of contrast and invariance that establish the inferential linkages between information variables and causal categories are shown in the middle column.

In the bottom panel of Table 1, a less familiar information variable, contingency, is introduced. It is tentatively proposed that this variable is diagnostic for perceived control over causes. The notion of contingency stems from the learned helplessness literature where it is often used synonymously with the notion of objective control (Abramson, Seligman & Teasdale, 1978). In the present context, we use the term contingency only to reflect the stimulus characteristics of an event, and we use the term control to refer to perceived properties of a cause. Following Abramson et al. (1978), an outcome is said to be noncontingent if the occurrence of the outcome is not related to a person's behavior, that is, the outcome remains invariant whether or not a particular action is made by the actor. For example, a child may not recuperate from an illness regardless of what the doctor prescribes; and our student, John, may continue to do poorly no matter how hard he tries. As these examples illustrate, low contingency reveals that the agent has little control over the cause of the event. Conversely, an outcome is contingent when after some action it differs from the outcome when the action was not made. In such instances, people may infer that the cause was the action under the control of the agent. Thus, the joint model predicts that contingency information determines inferences of causal (un)controllability.

Table 1

Predictions of the Joint Model

Information Variables   Factor      Causal                                             
                        Pattern     Categories                                         
Consensus               low         Persons       contrast       Internal (Person)     
                        high                      invariant      External              
Consistency             low         Time          contrast       Variable (Occasion)   
                        high                      invariant      Stable                
Distinctiveness         high        Stimuli       contrast       Specific (Stimulus)   
                        low                       invariant      Global                
Contingency             high        Actions       contrast       Controllable          
                        low                       invariant      Uncontrollable        

Note. Between parentheses on the right are the standard attribution responses predicted by earlier theory and research (e.g., Cheng & Novick, 1990)

The linkages between information variables and dimensional categories shown in Table 1 have been pointed out in earlier attribution research (e.g., Kammer, 1984; Meyer, 1980; Read, 1987; Read & Stephan, 1979), although an explicit theoretical rationale of the underlying attribution process was never provided. As mentioned earlier, many earlier models based on the method of difference specify only standard attributions (shown between parentheses) and their combinations (e.g., Cheng & Novick, 1990; Försterling, 1989; Hewstone & Jaspars, 1987; Jaspars, 1983; Hilton & Slugoski, 1986). Most of these models predict that no attributions are possible following the other information variables listed in Table 1, although Cheng and Novick's (1990) contrast model allows for (relatively weak) attributions under invariant conditions depending on the default assumptions perceivers make about unspecified information (e.g., the behavior of other people in other situations; see also footnote 5). In contrast, the joint model predicts that when there is invariance over one or more information variables, the perceiver may still infer strong causality in terms of external, stable, global and/or uncontrollable attributions. For example, the fact that all living creatures (high consensus) in the whole universe (low distinctiveness) always die sooner or later (high consistency) whatever their attempts to avoid this (low contingency), can be sufficiently explained by an external, global, stable and uncontrollable factor such as God or Laws of Nature.

In the following sections, four empirical studies will be presented to illustrate that subjects infer causal attributions in a manner consistent with the proposed information-dimension linkages through the application of both principles of contrast and invariance of the joint model. Before turning to our empirical data, however, we will first review prior work that bears upon the joint model.

Evidence from Previous Research

Several studies in the attribution literature provide support for our hypothesis that people make attributions in accordance with the proposed information-dimension linkages. Most empirical work comes from prior studies by Weiner and others, relating Kelley's information patterns with single attributions of ability, effort, task difficulty, and luck (see Table 2). In two exemplary experiments by Frieze and Weiner (1971), subjects were presented with information patterns involving achievement outcomes and then had to indicate how much of an actor's outcome was due to skill, effort, task difficulty and luck. Consensus was manipulated by giving the percentage of others who succeeded or failed at the given task, consistency was indicated by the percentage of the actor's past successes or failures at the same task, and distinctiveness was provided by the percentage of outcomes for the actor in other tasks. In other studies, information about the event was manipulated experimentally. For example, Feather and Simon (1971a, 1971b) varied consistency information by letting the subjects first fail or succeed on "practice" items before beginning the "test" items. The subjects who failed or succeeded on both practice and test items were thus given high consistency experiences, whereas subjects who attained different outcomes on practice and test items were exposed to low consistency experiences. Unfortunately, none of these studies manipulated the contingency variable.

Table 2 summarizes the findings of single-attribution investigations that used similar informative or experimental procedures and reported relevant significance tests.[3 All studies involved an achievement task. Subjects ranged from fourth grade elementary school children to college students, and attributions were given from the perspective of the self or others. To analyze these findings, however, it is first necessary to make some reasonable assumptions with regard to the dimensional location of ability, effort, task and luck. A possible categorization, based on Weiner (1986), is shown in Table 2 for each relevant information variable. The entries in Table 2 further denote whether consensus, consistency and distinctiveness is ]high or low. Attributions following this information are shown on top of the table. For example, the first row shows that low consensus is followed by ability, effort and luck attributions, whereas high consensus is associated with attributions to task difficulty.

Studies relating Information Variables to Single Attributions


Study            Ability          Effort           Task             Luck             

Consensus information
                 intern[a]        intern           extern           extern           
Frieze &         Low              Low              High             Low              
Weiner (1971,                                                                        
study 1)                                                                             
Frieze &         ns               Low              High             ns               
Weiner (1971,                                                                        
study 2)                                                                             
Weiner & Kukla   Low              Low              ----             ----             
(1970, study 6)                                                                      
Fontaine         Low              Low              High             Low              
(1975, study 1)                                                                      
Fontaine         ns               ns               ns               ns               
(1975, study 2)                                                                      
Read & Stephan   Low              Low              High             Low              
Frieze &         ns               ns               High             Low              
Bar-Tal (1980)                                                                       

Consistency information
                 stable[a]        unstable         stable           unstable         
Frieze &         High             ns               High             Low              
Weiner (1971,                                                                        
study 1)                                                                             
Frieze &         High             Low              High             Low              
Weiner (1971,                                                                        
study 2)                                                                             
Feather &        High             ----             ----             Low              
Simon (1971a)                                                                        
Feather &        High             ns               ns               Low              
Simon (1971b)                                                                        
Nicholls (1975)  High             ns               ns               Low              
Read & Stephan   High             ns               ns               Low              
Frieze &         High             Low              High             Low              
Bar-Tal (1980)                                                                       

Distinctiveness information

                 global[a]        global           global           specific         
Frieze &         Low              ns               ns               High             
Weiner (1971,                                                                        
study 1)                                                                             

Note. Entries denote whether consensus, consistency and distinctiveness is high or low. The attributions given following this information are shown on top of the table. Horizontal lines (---) indicate that the attribution was not measured. Insignificant results are denoted by ns.

[a] Presumed causal position.

As can be seen from Table 2, the findings are generally in agreement with the assumed linkages between information and dimensions. Low consensus between people was followed by internal attributions to ability and effort, whereas high consensus increased external attributions to task difficulty. High consistency with past performance increased stable attributions to ability and task difficulty, whereas low consistency led to unstable attributions of effort and luck. Low distinctiveness increased attributions to global causes such as ability, whereas high distinctiveness led to specific attributions of luck. Some attributions, however, provided rather weak support for the hypotheses, particularly in the case of effort and task difficulty following consistency and distinctiveness information. Presumably, this is due to the fact that the properties of stability and globality are less explicit for these causes (Weiner, 1983). Indeed, variable effort is sometimes perceived as stable, whereas difficulty is at times seen as moderately variable. Similarly, general effort may be confounded with specific preparation for a particular exam.

However, the predictions were strongly contradicted for luck attributions, as this external cause followed after low instead of high consensus. Read and Stephan (1979, p. 199) suggested several reasons for this unexpected finding. First, people may view luck as belonging to the individual and in that sense perceive it as internal. If people view luck as internal, then the fact that low consensus leads to greater attributions to luck becomes consistent with predictions. A second explanation provided by Read and Stephan is that because of its random (i.e., unstable and specific) nature, luck cannot reasonably be invoked to explain the independent outcomes of a large number of people at one time. Attribution to an extremely unstable and specific factor precludes its use to explain outcomes for which there is high consensus. According to this explanation, properties on one dimension can sometimes override properties on other dimensions. This may follow from the fact that causal dimensions correlate with one another (rs from .19 to .68, Anderson, 1983).

An important limitation of single-attribution studies is that dimensions have to be inferred by the researcher from the given causes. Therefore, we tested the information-dimension linkages predicted by the joint model more directly by analyzing the effect of the information variables on perceived causal dimensionality; that is, the subjects themselves rated the properties of the most likely cause. To our knowledge, only two studies have been published which measured perceived dimensionality, although only indirectly. Meyer (1980) measured dimensionality through a factor-analysis of nine attributions and found the predicted linkages between consensus and locus, and between consistency and stability. Kammer (1984) asked her subjects to rate a number of causes varying in globality, and showed an effect of distinctiveness on specific versus global causes.

The effects of covariation information on dimensional judgments were tested in three studies. The first study involved a single event, set in an achievement context (i.e., losing a computer game), while the second study took place in a social context (i.e., being complemented for a dress at a party). These events were deliberately set in an unfamiliar context (i.e., a newly developed computer game, a party in a foreign culture) so as to avoid preconceived attributions that might interfere with the information manipulated. The third study involved sixteen different achievement and social events, taking place in more familiar contexts. Our prediction is that information on consensus, consistency, distinctiveness and contingency determines causal properties of respectively locus, stability, globality and controllability.

Study 1



Subjects in the first study were 42 freshmen from the communications department at the Vrije Universiteit Brussel (Belgium). During a regular class hour, they were invited to participate in a study on "how people make attributions". After receiving a booklet with written instructions on the first three pages, the subjects completed the questions individually. The whole procedure took about 30 minutes, and subjects were thanked for their participation.


A scenario was introduced in a small paragraph on the first page of the booklet. It described "a boy, Peter, who lost the XB03 game during a test session when trying out a Gambit tactic". The subjects had to imagine themselves as researchers who had to infer the reasons for winning or losing in order to improve the initial test version of this newly developed computer game. After this introduction, a "situation" was depicted on each subsequent page which involved a set of four sentences providing additional information on consensus, consistency, distinctiveness and contingency. The wording of these information variables is illustrated below (the sentence parts that alter given invariant versus contrastive information variables are shown in italic) :

- Everyone / Nobody except Peter lost the XB03 game (high/low consensus);

- On prior test sessions, Peter always lost / sometimes won on the XB03 game (high/low consistency);

- Peter lost / won on every other computer game (low/high distinctiveness);

- Peter lost / won on the XB03 game no matter what or how much he did / when he tried some other tactics (low/high contingency).

In total, there were sixteen situations (pages) involving all possible combinations of information variables. Following each situation, the subjects were asked to analyze carefully the information and to rate the properties of the cause on four 7-point scales (the dimensions are indicated between square brackets) : resides within Peter -- resides outside of Peter [locus]; influences only this game (specific) -- influences many different games (global) [globality]; is not controllable by Peter -- is controllable by Peter [control]; is temporary (variable) -- is permanent (stable) [stability]. Each dimension was fully defined with the aid of an example in the introductory pages of the booklet to preclude any lack of understanding. The situations (pages) in the booklet were presented in one random order and its reverse. Within each situation, the information variables were counterbalanced between subjects in four different (Latin square) orders.

Results and Discussion

The data were analyzed with a 2 (high vs. low consensus) x 2 (high vs. low consistency) x 2 (high vs. low distinctiveness) x 2 (high vs. low contingency) within-subjects ANOVA with the four attribution measures as dependent variables. As noted before, the attributional dimensions may not be entirely independent and, to control for such interactions, the ANOVAs were conducted for each attribution dimension with the other three attribution measures as covariates. The means and F-values are shown in the top panel of Table 3. The findings were generally in line with the joint model as the causal dimensions were most strongly determined by the predicted information variables. These predicted effects are shown by the significant F-values on the main diagonal of the top panel in Table 3. As can be seen, the first three predicted effects relating the information variables of consensus, consistency and distinctiveness with the dimensions of locus, stability and globality are highly significant, p<.0001, whereas the predicted effect of contingency on control only reaches the .05 level.

Table 3

Mean Causal Judgments and F-values following the Information Variables : Studies 1 - 3.

                             Locus          Stability      Globality      Control        
    Study 1                                                                              
   Consensus      High       4.47           3.87           3.70           3.57           
                  Low        2.64           3.94           3.84           4.17           
                  F          25.06****      ----           ----           ----           
  Consistency     High       3.63           4.52           3.74           3.68           
                  Low        3.48           3.28           3.80           4.07           
                  F          ----           26.81****      ----           ----           
Distinctiveness   Low        3.46           4.11           5.23           3.90           
                  High       3.64           3.69           2.31           3.85           
                  F          ----           ----           63.73****      ----           
  Contingency     Low        4.02           4.41           3.75           3.09           
                  High       3.09           3.39           3.80           4.65           
                  F          ----           ----           ----           5.13*          
    Study 2                                                                              
   Consensus      High       4.36           3.74           3.92           3.57           
                  Low        3.28           3.76           4.05           4.20           
                  F          23.03****      ----           ----            6.04*         
  Consistency     High       3.89           4.32           3.84           3.95           
                  Low        3.92           3.17           4.12           3.82           
                  F          ----           26.58****      ----           ----           
Distinctiveness   Low        3.81           4.09           5.39           4.03           
                  High       4.01           3.40           2.58           3.74           
                  F          ----           ----           53.81****      ----           
  Contingency     Low        4.11           4.05           4.08           3.12           
                  High       3.70           3.44           3.88           4.66           
                  F          ----           ----           ----           25.40****      
    Study 3                                                                              
   Consensus      High       3.40           3.48           3.56           4.23           
                  Low        2.48           3.97           4.00           4.54           
                  F          14.18***       6.74*          ----           ----           
  Consistency     High       2.94           4.42           3.76           4.31           
                  Low        2.93           3.03           3.80           4.45           
                  F          ----           67.92****      ----           ----           
Distinctiveness   Low        2.81           4.15           4.76           4.35           
                  High       3.06           3.30           2.80           4.41           
                  F          ----           ----           27.62****      ----           
  Contingency     Low        2.87           3.85           3.98           4.21           
                  High       3.01           3.60           3.57           4.55           
                  F          ----           ----           4.90*          4.38*          

Note. The entries reflect attributions to external, stable, global and controllable causes respectively. Degrees of freedom for F = 1 and 38 for Study 1; 1 and 36 for Study 2; 1 and 60 for Study 3. F-values that are not significant are denoted by a horizontal bar. Expected tests are in italic.
* p < .05. ** p < .01. *** p < .001. **** p < .0001.

The ANOVA also revealed a number of interactions. However, given that there were 44 possible interactions over the three studies, we limit the discussion in this and the next two studies only to those interactions that are significant beyond the .01 level to avoid undue type II errors. The ANOVA showed a significant interaction between consensus and consistency on locus attributions, F(1,38) = 16.26, p<.001. The means indicated that the predicted discrepancy between internal and external attributions given consensus information is increased when consistency is high (M = 2.45 vs. 4.80) as opposed to low (M = 2.83 vs. 4.14). A possible explanation is that repeated failure on the computer game increased the subjects' confidence in their locus ratings. The ANOVA also revealed an interaction between consistency and contingency on perceived stability, F(1,38) = 12.18, p<.01. The means showed that the predicted stable attributions following high consistency were increased after low contingency (M = 5.30) as opposed to high contingency (M = 3.74). This seems to point to a confound between contingency and consistency. Indeed, as low contingency indicates that outcomes remain identical even after multiple actions by the actor, this necessarily implies that the outcome was consistent over some time as well, because different behaviors by one person are typically performed one after another. While developing the scenarios for these studies, extreme care was taken to avoid such confounds by taking larger time frames for the consistency information (e.g., several test sessions) and shorter ones for the contingency variable (e.g., several tactics within one test session). Apparently, though, it was not possible to exclude these confounds entirely.

Study 2


Subjects in the second study were 40 freshmen from the psychology department at the same university. The procedure and materials were similar in format to those of Study 1. The present study described "a girl, Nawal, who was complimented on her outfit at a party when she wore her pink dress". Subjects had to imagine themselves as anthropologists who studied the habits and rites of a non-Western culture. The information variables provided were phrased as follows :

- Everyone / Nobody except Nawal was complimented on her outfit (high/low consensus);

- At prior parties, Nawal was always / never complimented on her outfit (high/low consistency);

- Nawal was complimented on her outfit everywhere she went / nowhere else (low/high distinctiveness);

- Nawal was / was not complimented on her outfit no matter what or how much she did / when she wore one of her other dresses (low/high contingency).

Next, subjects rated the properties of the cause on the following 7-point scales : resides within Nawal -- resides outside of Nawal; influences only this party (specific) -- influences many parties (global); is not controllable by Nawal -- is controllable by Nawal; is temporary (variable) -- is permanent (stable).

Results and Discussion

The data were analyzed as before, that is, with a 2 (high vs. low consensus) x 2 (high vs. low consistency) x 2 (high vs. low distinctiveness) x 2 (high vs. low contingency) within-subjects ANOVA and with the four attribution measures as dependent variables. The results are shown in the middle panel of Table 3. Consistent with our predictions, the F-values for the predicted information-dimension linkages (see main diagonal) were all very strong, p<.0001. There was, however, a small unexpected main effect of low consensus on perceived causal control, F(1,36) = 6.04, p<.05. A possible explanation is that subjects may have brought with them some preconceived notion that social events caused by an internal agent (as indicated by low consensus) also imply that the cause was under the actor's control to some degree. The ANOVA further revealed that none of the interactions reached the .01 level.

Study 3


Subjects in the third study were 64 freshmen from the economics department at the same university. The procedure and materials were similar in format to those of Study 1 and 2, although this study involved 16 scenarios rather than just one. Eight scenarios described an achievement-related event (e.g., "John attained a low mark this week on a math test when he studied the material from book A") and eight scenarios involved a social event (e.g., "Mark was able to convince his friends to go to a Woody Allen movie when he showed them a good review"). Each scenario was provided in a single sentence at the top of a page of the booklet, followed by four information variables and attribution scales phrased in a similar manner as in the first two studies. The scenarios were combined with the information variables in four random orders.

Results and Discussion

The data were analyzed as before. The means and F-values of the third study are depicted in the bottom panel of Table 3. As can be seen, the predicted main effects of the first three information variables on the related causal dimensions are highly significant, p<.0001, whereas the effect of contingency on controllable attributions is quite weak, p<.05. Moreover, there are two small unexpected main effects. The ANOVA revealed an effect of consensus on stability attributions, F(1,60) = 6.74, p<.05. The means show that the cause is judged to be more stable when consensus is low (M = 3.97) as opposed to high (M = 3.48). A possible explanation is that low consensus tends to be attributed to dispositional traits, which also contain stable properties and therefore result in stable judgments as well (see Hilton et al., 1994). There was also a small unexpected effect of contingency on globality attributions, F(1,60) = 4.90, p<.05, showing that low contingency leads to more global attributions (M = 3.98) than high contingency (M = 3.57). Perhaps, this is again due a confound between two information variables, in this case between contingency and distinctiveness. As low contingency denotes that different behaviors by an actor did not change the outcome, it may imply that the outcome under different material circumstances or stimuli (brought about by the action) would not change either, resulting in more global attributions.

The ANOVA also revealed three significant interactions. There was an interaction between consistency and contingency on stability judgments, F(1,60) = 20.75, p<.0001. This is a very robust interaction because it also appeared in Study 1 (p<.01) and Study 2 (p<.05). The means indicated that stable attributions following high consistency are further increased when contingency is low (M = 4.78) as opposed to high (M = 4.06). As intimated before, this result is probably due to an inherent confound of the contingency variable with time-related consistency information, because behaviors of one person cannot be performed in a time-less vacuum and thus necessarily involve some lapse of time. That is, low contingency implies that the actor's outcomes (following several behavioral attempts) remained identical over some period of time, and thus may increase the predicted effect of high consistency on stable attributions. There was also a significant interaction between distinctiveness and contingency on globality judgments, F(1,60) = 16.82, p<.001. The means indicated that global judgments following low distinctiveness are increased when contingency is low (M = 5.18) as opposed to high (M = 4.33). This interaction extends the unexpected main effect of contingency on globality attributions discussed above, and is probably due to a confound between low contingency and low distinctiveness. Finally, there was also a three-way interaction between consistency, consensus and distinctiveness on stability attributions, F(1,60) = 11.05, p<.01. The means indicate that low consensus and low distinctiveness combine to increase stability attributions after high consistency information is given. This interaction is difficult to interpret, and is perhaps a result of the unexpected main and interaction effects discussed above.

Discussion of Studies 1 - 3

In all three studies, the predicted main effects of information variables on causal dimensions were found to be significant. However, there were also a number of unexpected main and interaction effects. Most of these unexpected effects were found in Study 3, which used scenarios involving familiar events, as opposed to the scenarios in Studies 1 and 2, which took place in contexts that were less familiar to the subjects. Hence, a possible explanation is that preconceived beliefs about the events (i.e., scripts; Read, 1987) and their likely causes distorted the effect of the information variables. An alternative explanation may lie in the source of the effects. Indeed, the majority of the unexpected effects (i.e., 5 out of 7) involve contingency information. Moreover, the predicted main effect of this variable was the weakest overall.

This gives some cause for concern. Is contingency not the predecessor of causal control as hypothesized in the joint model ? Or is contingency to some extent confounded with the other information variables so that its predicted main effect stands out less ? In an attempt to answer this question, we conducted an additional study in which we manipulated only contingency information over the same 16 scenarios, and we asked another 44 subjects to rate the perceived control of the cause. (We also provided another phrasing for the contingency variable, but this manipulation had no effect.) The results revealed a strong effect of contingency on controllability attributions, F(4,40) = 21.63, p<.0001. Consistent with the joint model, the cause was seen to be substantially more controllable given high contingency (M = 5.65) as opposed to low contingency (M = 3.66). This finding indicates that contingency determines attributions of control to a large extent when manipulated in isolation, and suggests that the weak main effect of this variable in Studies 1 and 3 is perhaps due to confounds with the other information variables.

Study 4

The previous studies demonstrate that inferences about causal dimensionality are strongly determined by the predicted information variables. However, this research is far from conclusive. Showing that subjects appear to be sensitive to both extremes of the causal dimensions does not necessarily imply that they applied principles of both contrast and invariance. One might argue, for example, that subjects based their judgments only on contrasts in the information, and that the apparent use of the invariance principle is an artifact of the bipolar dimension scales used. For example, indicating a decreased influence of an internal (person) cause necessarily results in greater external judgments, because internal and external lie at the opposite extremes on the same judgment scale. Although earlier research on single attributions reviewed above indicated that subjects do make attributions to single causes in accordance with the proposed joint model, the crucial properties of these causes were assigned post-hoc by experimenters rather than by the subjects themselves.

Hence, this study was designed to provide a more stringent test of our hypotheses. Subjects directly indicated the causal likelihood of a number of single causal factors, the dimensional properties of which were assigned on the basis of prior pilot testing on the same population. We predicted that attributions to single causes would be determined by the information variables as proposed by the joint model. Moreover, we also predicted that these attributions would be based on both principles of contrast and invariance. This can be tested directly by comparing attributional judgments following contrastive as opposed to invariant information patterns. Some earlier models, such as the covariation models of Kelley (1973) and Försterling (1989), the natural logic model of Hewstone and Jaspars (1987), and the default abnormal conditions focus model of Hilton and Slugoski (1986) predict that subjects will infer no attributions when the stimulus information involving that factor is invariant.[4 The probabilistic contrast model of Cheng and Novick (1990) predicts that causal inferences are possible depending on the default assumptions people make about other information that is not given by the experimenter (the ]non-configurational cells, e.g., what other people do in other situations). However, the contrast model also predicts that regardless of these assumptions, attributions should be less strong given invariant rather than contrastive information.[5 In contrast, we predicted that subjects would infer an attribution given invariant information and we expected these attributions to be equally as strong as inferences following contrastive information.




Subjects were 40 freshmen taking an introductory psychology course, who participated for a partial course requirement. During the last part of a regular class hour, they were invited to participate in a study on "how people make attributions". The subjects completed the questions individually. The whole procedure took about 30 minutes, after which subjects were thanked for their participation.


The subjects received two booklets with brief scenarios depicting an actor who attained a negative outcome at a job interview or in an exam. The application scenario ran as follows : "Annie is unemployed. One day she applies for a job as a secretary in a firm. After an interview with the personnel manager, she has been told that she will not be hired". The exam scenario read as follows : "John is a student. During the first exam session, he has been told that he failed the math oral exam".

After this introduction, one information variable was provided on each subsequent page. This involved a sentence describing one of eight information variables (i.e., high or low consensus, consistency, distinctiveness or contingency) phrased in a manner similar to that used in Studies 1 - 3. Subjects were instructed to analyze carefully this information. Next, eight possible causes were listed on the same page. Subjects had to indicate the degree to which each cause had influenced the outcome on a 7-point scale ranging from no influence (1) to strong influence (7). Within each scenario, the causes were presented in one random order or its reverse. The information variables (pages in the booklet) were also presented in one random order and its reverse. The order in which the two scenarios were presented was also randomized.


The causes for each scenario were selected on the basis of a pilot study with 175 students from the same university who participated voluntarily. In this pilot study, subjects received a description of four scenarios and then indicated for each scenario the property of about 60 causes for one of the four dimensions on a 6-point scale. The cause "luck" was not included since this factor does not conform to the predictions of joint model as shown in earlier research (see earlier section). Two target causes were then selected for each dimension on the criteria that they received very high (above 4.5) or very low (below 2.5) means on that dimension, while the means for the remaining dimensions were relatively moderate (between 2.5 and 4.5). It should be noted that it was not always possible to satisfy this latter criterion as most causes received high or low ratings on more than one dimension. One scenario was not selected for the main study because too many causes failed to satisfy the latter criterion (as they showed loadings on more than one dimension). Another scenario was eliminated because the causes failed to show extreme ratings on the globality dimension.

The target causes in the application scenario are listed below (the dimensional assignments and their mean ratings from the pilot study are given in parentheses) : Annie has no interest in a secretarial job (internal : 5.42), the personnel manager set high standards (external : 1.52), Annie lacks necessary aptitude for this job (stable : 5.04), Annie was ill during the interview (unstable : 1.42), Annie has low general intelligence (global : 5.10), the personnel manager is unfriendly (specific : 2.14), Annie was not well prepared for the interview (controllable : 5.12), Annie is physically handicapped (uncontrollable : 1.80). The means of the target causes on the remaining dimensions were on average 3.53.

For the exam scenario, the target causes were : John has no interest in maths (internal : 5.36), the maths examinator set high standards (external : 1.60), John has low general intelligence (stable : 5.26), John was tired (unstable : 1.62), John has a bad memory (global : 5.10), the maths exam took place at an inconvenient moment (specific : 2.10), John did not study regularly (controllable : 5.56), John was ill for a long time (uncontrollable : 1.80). On the other dimensions, the means of the target causes were on average 2.90.

Results and Discussion

To compute an index for each dimension, the ratings of the target cause with low loadings on that dimension were reversed (1 was changed to 7, 2 to 6, and so on) and then averaged with the ratings of the target cause with high loadings. For instance, for the locus dimension, the internal target cause and the (reversed) external target cause were averaged. Thus, each dimension index represents the aggregated ratings of two target causes. These dimension indices were then averaged over the two scenarios (see Table 4). The indices of each contrastive information condition were then subtracted from those of the corresponding invariant conditions, resulting in a difference score (see also Table 4). These difference scores indicate to what extent subjects differentiate causes with opposite properties (e.g., internal vs. external causes) given opposite conditions of the information variables (e.g., low vs. high consensus). In line with the joint model, the predicted information-dimension linkages showed the highest (absolute) difference scores (see main diagonal on Table 4).

Table 4:

Causal Attributions Weighted along their Dimensional Properties in function of the Information Variables : Study 4

Consensus         High           3.05           4.03           3.28           4.13       
                  Low            4.88           3.93           4.65           4.02       
                  Difference     -1.83          .11            -1.37          .11        
Consistency       High           4.62           4.62           4.71           4.19       
                  Low            3.01           2.67           2.70           3.73       
                  Difference     1.61           1.94           2.01           .46        
Distinctiveness   Low            4.41           4.84           5.06           3.86       
                  High           3.67           3.06           2.94           4.21       
                  Difference     .74            1.79           2.12           -.35       
Contingency       Low            3.30           4.08           4.07           3.58       
                  High           3.79           3.49           3.67           4.53       
                  Difference     -.49           .59            .40            -.96       
                  Contrast F     64.31*         39.14*         76.78*         31.05*     
                  Residual F     16.83*         21.37*         44.19*         9.39*      

Note. The entries reflect judgments to attributions with internal, stable, global and controllable properties respectively. Degrees of freedom for Contrast F = 1 and 39; and for Residual F = 2 and 38. Highest difference scores in each column are in italic.
* p < .0001.

Causal Explanations and Dimensional Properties : Experiment 4

Causal Explanation                         Property                                   

Application Scenario
Annie has no interest in a secretarial     i - - -                                    
The personnel manager set high standards   e - - -                                    
Annie was ill during the interview         - v - -                                    
Annie lacks necessary aptitude for this    - f - -                                    
The personnel manager is unfriendly        - - s -                                    
Annie has low general intelligence         - - g -                                    
Annie was not prepared for the interview   - - - c                                    
Annie is physically handicapped            - - - u                                    

Exam scenario
John has no interest in math               i - - -                                    
The math examinator set high standards     e - - -                                    
John was tired                             - v - -                                    
John has low general intelligence          - f - -                                    
The math exam was held at an               - - s -                                    
inconvenient moment                                                                   
John has a bad memory                      - - g -                                    
John did not study regularly               - - - c                                    
John was ill for a long time               - - - u                                    

Note. Dimensional Properties: i=intern, e=extern, v=variable, f=fixed, s=specific, g=general, c=controllable, u=uncontrollable.

The difference scores were analyzed with a 4 (dimension) by 4 (information variable) within-subjects ANOVA. The results showed significant mean effects for dimension, F (3,37) = 31.35, p<.0001, for information variable, F (3,37) = 25.58, p<.0001, and as expected, a significant interaction between dimension and information variable, F (3,37) = 31.35, p<.0001. To test our specific hypotheses, we conducted a series of planned comparisons. For example, we compared the difference scores for locus given consensus information and compared them with the difference scores of the other information variables. If these comparisons were significant then our hypotheses would be supported. Moreover, if no residual variance remained to be explained then the hypothesized variables were unique determinants of the attribution judgments. The F-values resulting from these comparisons are depicted in the bottom panel of Table 4. As can be seen, all contrast Fs testing our hypotheses are highly significant, p<.0001. As in the previous studies, the effect of contingency on controllable attributions was weakest, as shown by the smaller difference scores and the lower F-values. Unexpectedly, the residual Fs were also significant for all dimensions, p<.0001, although they attained lower values than the hypothesized contrasts.

There can be many reasons why the predicted effects were not unique. First, as the subjects who indicated the dimensional properties in the pilot study were not the same ones that provided attributional judgments, some noise may have entered the data due to individual differences in the interpretation of causal properties. As Weiner (1985, p. 555) cautioned, "the interpretation of specific causal inferences might vary over time and between people and situations". Thus, what might have been a clear instance of an internal cause for most subjects, may nonetheless have appeared as a somewhat external cause to other subjects. This noise can only be avoided if causes are individually tailored for each subject, which is very difficult to accomplish.

Second, as some causes have extreme loadings on more than one dimension, the derived dimension indices were not completely independent, resulting in noise in the dependent measures. For example, by providing a high rating on an internal cause which also has extreme loadings on some other dimension (say, stability), it may appear as if stability influenced that causal rating as well. As noted before, one scenario was excluded from the main study for this reason, although fully independent measures could not be obtained in the two remaining scenarios either.

Third, the scenarios may involve scripted events so that subjects may have relied on routine knowledge from which they extracted some stereotyped explanations which then received stronger ratings for that reason only (cf. Read, 1987). In sum, the information variables most strongly determined attributions to causes with the predicted dimensional properties. However, these effects were not unique because other information variables also had some influence, although it may be that these additional influences were due to methodological artifacts.

A second prediction of this study was that attributional ratings would be equally strong given invariant as well as contrastive information conditions. To test this prediction, we calculated the mean of all original causal judgments in both conditions. The mean judgment given invariant conditions was 4.55, which was slightly higher than the mean judgment under contrastive conditions, 4.43. This difference tended towards statistical significance, t(39) = 1.98, p=.055. This result contradicts many earlier attribution models which predict that causal inferences should be weaker or absent following invariant rather than contrastive information.

General Discussion

The basic goal of this article was to bring theory and research on the attribution process developed from Kelley's (1967) covariation model more in line with other theories on the dimensionality and social consequences of attributions (e.g., Abramson, Seligman & Teasdale, 1978; Weiner, 1986). In particular, we argued that the standard attribution responses should be extended to the opposite extremes of the fundamental dimensions of causality, and that the variables of contingency and the related dimension of control should be included. We further reasoned that Mill's (1872/1973) method of difference, as realized in the principle of contrast (i.e., Cheng & Novick, 1990), is applied to infer standard attributions and the attribution of control, whereas Mill's method of agreement, as realized in the principle of invariance, is used to infer the opposite non-standard attributions.

The data presented in this article were largely in agreement with the proposed joint model. A review of earlier relevant work demonstrated that subjects choose causal attributions in line with the proposed information-dimension linkages. Similarly, the evidence from four studies suggested that subjects not only judge the properties of causes on the basis of the predicted information variables (Studies 1 - 3), but that they also select single attributions on the same basis (Study 4). Moreover, Study 4 confirmed that these attributional judgments are made through the joint application of both principles of contrast and invariance. The finding that attributions were somewhat stronger following invariant rather than contrastive information is inconsistent with most earlier models which predict that when data are invariant, attributions should be weaker (Cheng & Novick, 1990) or even absent (Kelley, 1973; Försterling, 1989; Hewstone & Jaspars, 1987; Hilton & Slugoski, 1986). Admittedly, some authors (Hilton & Slugoski, 1986; Cheng & Novick, 1991) have suggested that invariant information patterns may be attributed to enabling conditions rather than causes. For instance, Cheng and Novick (1990) proposed that enabling conditions do not covary with the effect in a given set of observations (i.e., the focal set), but rather covary with the effect outside the focal set. Although this distinction was not made in our research, explanations in terms of enabling conditions require at least weaker causality judgments within a particular focal set, but stronger judgments were found in Study 4.

We proposed four covariation variables that may carry contrastive/invariant information about potential causes. Three of these have been suggested earlier by Kelley (1967) and were related to the causal dimensions of locus, stability and globality. A fourth information variable, contingency, was proposed to determine causal judgments of control. The results, however, suggest that this latter linkage is relatively weak. Perhaps the effect of contingency was undermined by confounds with the other information variables, because an additional study revealed that the impact of contingency was stronger if other covariation variables were omitted. Although the effect of contingency is rather weak, it is important to note that contingency was the sole determinant of perceived causal controllability, because the evidence from all four studies revealed that none of the other covariation variables influenced perceived control on its own or in interaction with other variables (except for a small main effect of consensus in Study 2).

Another possible explanation for the weak effect of contingency may lie in a fundamental flaw in the logical relation between contingency and control. Although low contingency necessarily indicates that control through one's actions was impossible because the outcome did not change, high contingency does not necessarily imply that the reverse implication of high control is true. It may be that the outcome changed through some trial and error, lucky guesses or other random behaviors which involve no knowledge on the part of the actor on how the outcome was changed. However, such knowledge is needed actually to control the outcome in the present as well as in the future. The idea that controllability involves a minimal level of knowledge has also been captured by Heider (1958, p. 113) as he stated that a person "is considered responsible, directly or indirectly, for any aftereffect he may have foreseen". Thus, it may be that to attribute full control, information is needed not only on the contingency of the actor's behavior, but also on the actor's state of knowledge concerning the consequences of his or her behavior.

The implications drawn from this research need some qualification because of the limited scope of the present studies : They involved only four dimensions of covariation and causality, were limited to achievement and social domains, were obtained from verbal summary information and did not address process mechanisms. However, not all of these limitations are equally problematic for the basic predictions of the joint model. For instance, it seems plausible that the principles of contrast and invariance apply to other covariation variables and causal dimensions than those we have investigated. Thus, for example, covariation information may be given in terms of countries rather than persons, so that locus is evaluated in terms of whether the cause is internal or external to the target country. It is also conceivable that perceivers have an even more specific causal hypothesis in mind that they want to test. Although the joint principles of contrast and invariance can be extended to accommodate such individual cases, our data do not reveal which rules or procedures are used to test specific causal questions. In particular, it is unclear how the perceiver goes from general covariation information to particular causes (see Study 4). Is this done with specific dimensions in mind, or is a causal hypothesis formed first and are its dimensional properties then tested against the information given ?

A recent study from our laboratory may be relevant here (Van Overwalle, Heylighen, Casaer & Daniëls, 1992). The data suggested that causal dimensions are automatically available upon reading the relevant information variables. Indeed, a unique prediction of the joint model is that there are unique linkages between eight information conditions and eight attribution categories. Given these unique one-to-one mappings, it seems likely that the repeated use of these linkages has become automated so that the mere presence of covariation information may activate the related attributional dimensions (cf., Shiffrin & Schneider, 1977; Kornblum, Hasbrouck & Osman, 1990).

Another concern is that our data were limited to success and failure in achievement and social domains, whereas other studies investigated various other events such as emotions, opinions and actions (e.g., McArthur, 1972). Insofar as these events are also desirable or undesirable from the actor's perspective, it seems logical that the same covariation and causal dimensions apply to them, although perhaps to a lesser extent. For example, some reservations should be made for contingency and controllability which may be of less relevance in the context of an action (e.g., Jack contributed a large sum of money) because contingency takes as implicit the assumption that performing such a social action is controllable in and of itself (although the successful completion or outcome of the action might be uncontrollable).

Another limitation of the present studies is that covariation information was presented in a verbal summary format. It is unclear to what extent such summary information made the relevant dimensions more salient than if raw data were presented about a number of events. Moreover, the underlying causality principles given pre-packed summary as opposed to raw information may not be the same. This may seriously restrict not only the present findings, but also those of other researchers using the same paradigm (e.g., Cheng & Novick, 1990; Hilton & Jaspars, 1987; Hilton & Slugoski, 1986; Hilton, Smith & Kim, 1994; Hewstone & Jaspars, 1987; Jaspars, 1983; McArthur, 1972, 1976; Orvis, Cunningham & Kelley, 1975; Ruble & Feldman, 1976). Although summary information may tell us a lot about the processes governing attributions based on social (pre-digested verbal) communication, it does reveal very little about spontaneous attribution search from actual (raw) social behavior. Perhaps, a more fundamental question is whether attributional inferences -- derived from summary or raw data --proceed on the basis of a set of rules as implied by the joint model and all previous covariation models, or whether another, evolutionary more primitive form of associative learning mechanism governs our causal thinking (see Shanks, 1991, 1993). Associative models assume that causal knowledge is represented in the form of mental associations between causes and effects, and that these associations are gradually learned. Schanks (1993) argues that these models can deal with the acquisition of causal knowledge over time and a number of other phenomena such as blocking and inhibition, which are not well captured by rule-based covariation models.

Although the present findings confirm that causality can be inferred by applying the principle of invariance, it should be realized that this principle may sometimes lead to spurious inferences, because one cannot guarantee that the cause inferred by applying the invariance principle is the only factor that co-occurs with the effect (e.g., attributing the sunset to the cock's crow). Mill (1872/1973) also warned that the method of agreement is "inferior" to the method of difference. It would be interesting, therefore, to delineate in future research the conditions under which people prefer one principle over another in causal reasoning.


Abramson, L. V., Seligman, M. E. P., & Teasdale, J. D. (1978). Learned helplessness in humans : Critique and reformulation. Journal of Abnormal Psychology, 87, 49-74.

Anderson, C. A. (1983). The causal structure of situations : The generation of plausible causal attributions as a function of type of event situation. Journal of Experimental Social Psychology, 19, 185-203.

Bradbury, T. N. & Fincham, F. D. (1990). Attributions in marriage : Review and critique. Psychological Bulletin, 107, 3-33

Cheng, P. W., & Novick, L. R. (1990). A probabilistic contrast model of causal induction. Journal of Personality and Social Psychology, 58, 545-567.

Cheng, P. W., & Novick, L. R. (1991). Causes versus enabling conditions. Cognition, 40, 83-120.

de Jong, P. F., Koomen, W. & Mellenbergh, G. J. (1988). Structure of causes for success and failure : A multidimensional scaling analysis of preference judgments. Journal of Personality and Social Psychology, 55, 718-725.

Feather, N. T., & Simon, J. G. (1971a). Attribution of responsibility and valence of outcome in relation to initial confidence and success and failure of self and other. Journal of Personality and Social Psychology, 18, 173-188.

Feather, N. T., & Simon, J. G. (1971b). Causal attributions for success and failure in relation to expectations of success based upon selective or manipulative control. Journal of Personality, 39, 527-541.

Folkes, V. S. (1984). Consumer reactions to product failure : An attributional approach. Journal of Consumer Research, 11, 398-409.

Fontaine, G. (1975). Causal attribution in simulated versus real situations : When are people logical, when are they not ? Journal of Personality and Social Psychology, 32, 1021-1029.

Försterling, F. (1989). Models of covariation and attribution : How do they relate to the analogy of analysis of variance ? Journal of Personality and Social Psychology, 57, 615-625.

Frieze, I. H., & Bar-tal, D. (1980). Developmental trends in cue utilization for attributional judgments. Journal of Applied Developmental Psychology, 1, 83-93.

Frieze, I. H., & Weiner, B. (1971). Cue utilization and attribution judgments for success and failure. Journal of Personality, 39, 4, 591-606.

Heider, F. (1958). The psychology of interpersonal relations. New York : Wiley.

Hewstone, M. & Jaspars, J. (1987). Covariation and causal attribution : A logical model of the intuitive analysis of variance. Journal of Personality and Social Psychology, 53, 663-673.

Hilton, D. J., & Slugoski, B. R. (1986). Knowledge-based causal attribution : The abnormal conditions focus model. Psychological review, 93, 75-88.

Hilton, D. J., Smith, R. H., & Kim S. H. (1994). The process of causal explanation and dispositional attribution. Journal of Personality and Social Psychology, in press.

Hilton, D. J. & Jaspars, J. M. F. (1987). The explanation of occurrences and non-occurrences : A test of the inductive logic model of causal attribution. British Journal of Social Psychology, 26, 189-201.

Jaspars, J. (1983). The process of causal attribution in common sense. In M. Hewstone (Ed.) Attribution theory : Social and functional extensions (pp. 28-44). Oxford, England : Basil Blackwell.

Kammer, D. (1984). Die Kausaldimension Generalität : Vorauslaufende Bedingungen für das Zustandekommen globaler und spezifisher Attributionen. Zeitschrift für experimentelle und angewandte Psychologie, 1, 48-62.

Kelley, H. H. (1967). Attribution in social psychology. Nebraska Symposium on Motivation, 15, 192-238.

Kornblum, S., Hasbrouc, T., & Osman, A. (1990). Dimensional overlap : Cognitive basis for stimulus-response compatibility -- A model and taxonomy. Psychological Review, 97, 253-270.

Lunt, P. (1988). The perceived causal structure of examination failure. British Journal of Social Psychology, 27 171-179.

McArthur, L. (1972). The how and what of why : Some determinants and consequences of causal attribution. Journal of Personality and Social Psychology, 22, 171-193.

McArthur, L. (1976). The lesser influence of consensus than distinctiveness information : The person-thing hypothesis. Journal of Personality and Social Psychology, 33, 733-742.

Meyer, J. P. (1980). Causal attributions for success and failure : a multivariate investigation of dimensionality, formation and consequences. Journal of Personality and Social Psychology, 38, 704-718.

Mill, J. S. (1872/1973). System of logic (8th ed). In J. M. Robson (Ed.) Collected works of John Stuart Mill (Vols. 7, & 8). Toronto, Canada : University of Toronto Press.

Nicholls, J. G. (1975). Causal attributions and other achievement-related cognitions : Effects of task outcome, attainment value and sex. Journal of Personality and Social Psychology, 31, 379-389.

Orvis, B. R., Cunningham, J. D., & Kelley, H. H. (1975). A closer examination of causal inference : The role of consensus, distinction and consistency information. Journal of Personality and Social Psychology, 32, 605-616.

Peterson, C., & Seligman, M. E. P. (1984). Causal explanations as a risk factor for depression : Theory and evidence. Psychological Review, 91, 347-374.

Read, S. J. & Stephan, W. G. (1979). An integration of Kelley's attribution cube and Weiner's achievement attribution model. Personality and Social Psychology Bulletin, 5, 196-200.

Read, S. J. (1987). Constructing causal scenarios : A knowledge structure approach to causal reasoning. Journal of Personality and Social Psychology, 52, 288-302.

Ruble, D. N., & Feldman, N. S. (1976). Order of consensus, distinctiveness, and consistency information and causal attribution. Journal of Personality and Social Psychology, 34, 930-937.

Shanks, D. R. (1991). Categorization by a connectionistic network. Journal of Experimental Psychology : Learning, Memory and Cognition, 17, 433-443.

Shanks, D. R. (1993). Human instrumental learning : A critical review of data and theory. British Journal of Psychology, 84, 319-354.

Shiffrin, R. M. & Schneider, W. (1977). Controlled and automatic human information processing : II Perceptual learning, automatic attending and a general theory. Psychological Review, 84, 127-190.

Steven, L., & Jones, E. E. (1976). Defensive attribution and the Kelley cube. Journal of Personality and Social Psychology, 34, 809-820.

Van Overwalle, F., Heylighen, F., Casaer, S., & Daniëls, M. (1992). Preattributional and attributional determinants of emotions and expectations. European Journal of Social Psychology, 22, 313-329.

Van Overwalle, F. (1989). Structure of freshmen's causal attributions for exam performance. Journal of Educational Psychology, 81, 400-407.

Weiner, B., & Kukla, A. (1970). An attributional analysis of achievement motivation. Journal of Personality and Social Psychology, 15, 1-7.

Weiner, B. (1983). Some methodological pitfalls in attributional research. Journal of Educational Psychology, 4, 530-543.

Weiner, B. (1985). An attributional theory of achievement motivation and emotion. Psychological Review, 92, 548-573.

Weiner, B. (1986). An attributional theory of motivation and emotion. New York : Springer-Verlag.