Using the concept that consensus, consistency and distinctiveness are evaluated using the ANOVA model, Kelley (1967; 1971; 1972 as cited in Kelley 1973) predicted the patterns of information that would lead to specific attributions. He suggested that an information pattern of high consensus, high consistency and high distinctiveness (as seen in the scenario above) would prompt an attribution of causality to the particular entity (in this case the girl). Kelley (1973) also suggested that a pattern of low consensus, high consistency and low distinctiveness (e.g. Jack, alone smiles at the girl, he has always smiled at the girl yet he also smiles at other girls) suggests an attribution in terms of the person (i.e. Jack). Finally, Kelley (1973) suggested that an attribution is made in terms of the particular circumstances at that time if consensus is low, consistency is low and distinctiveness is high (e.g. Jack, alone smiles at the girl, he doesn’t usually smile at the girl, and he doesn’t smile at other girls).
Although Kelley predicted the attributions that may arise from certain combinations of consensus, consistency and distinctiveness information, he never verified them through experimentation. Consequently, the first investigation into Kelley’s predictions by McArthur (1972) is notable as it grounded much subsequent research into the covariation-based theory of attribution. McArthur’s (1972) study allocated either high or low values to each of the three types of information, thus producing eight possible patterns of attribution data. The attributions made by participants about these configurations were consistent with Kelley’s (1967; 1971; 1972 as cited in Kelley 1973) predictions.
More recent, naturalistic investigations into Kelley’s attribution model only partially support covariation-based theory. Peterson’s (1980) archival investigation into attributions for victory and defeat in a sporting newspaper found that football players’ and coaches’ ignored obvious covariation in some instances. Subsequently, although successful teams attributed a win late in the season to themselves and unsuccessful teams attributed a late-season defeat to themselves, these covariation-congruent attributions did not extend to other circumstances. Notably, successful teams attributed a late defeat to factors within the teams, and unsuccessful teams ascribed a late win to themselves although the obvious covariation throughout the season refuted these attributions. However, it is debatable whether public statements of attribution made in a newspaper accurately reflect processes in the more personal types of attribution Kelley (1973) theorized about.
Peterson’s (1980) experiment is not the only investigation that does not wholly verify a covariation-based attribution theory. Much evidence has been cited concerning the limitations of Kelley’s (1967; 1971; 1972 as cited in Kelley 1973) theory and the methodology used by McArthur in the verification of predictions made by the ANOVA model. Prototypical instances. Unnaturalistic?
McArthur’s (1972) study provides a large empirical basis for support of Kelley’s covariation model (Hewstone, 1989). However, methodological flaws and limitations in McArthur’s investigations are evident. As noted by the author herself, this investigation of Kelley’s predictions only assesses attributions made about the behavior of other people, despite the fact that Kelley intended his model to apply to ones own behavior as well as that of another person (McArthur, 1972). Equally, McArthur observed that her study would only asses judgements made by participants when they provided with prepackaged information about a certain event. However, in real life attribution processes participants do not receive such neat informational inputs (Fischhoff, 1976, as cited in Major, 1980) and may not actually search for or employ information on consensus, consistency or distinctiveness (McArthur, 1972). Cordray and Shaw (1978) suggest that this is a fundamental flaw in the covariation theory and a drawback for McArthur’s methodology. They suggest that the use of a within-subjects design may have prompted participants to believe that they should respond differently to each information configuration. Furthermore, the very structured stimuli used would encourage participants to give a response congruent with standard logic (Cordray & Shaw, 1978). It has been suggested that “evidence that subjects utilize covariation information to infer causality under conditions that encourage subjects to behave logically merely shows that they can be logical” (Cordray & Shaw, 1978, p.281). Therefore, a test of whether participants actually use covariation information, rather than whether they are capable of using, it is needed (Cordray & Shaw, 1978).
Several studies have investigated participants’ actual utilization of consensus, consistency and distinctiveness data when the information is not given in a prepackaged form. The results prove problematic for the covariation-based model of causal attribution (Hewstone, Strobe & Stephenson, 1996). Garland, Hardy and Stephenson (1975, as cited in Major, 1980) asked participants what types of information they would like if they had to make an attribution about a particular circumstance. Only 23% of the requests made by participants could be allocated to one of the informational categories described by Kelley. Furthermore, Major (1980) found that when provided with a total of 36 pieces of consensus, consistency and distinctiveness information (12 pieces from each category) to base their attributions, participants sampled, on average, only 9 pieces. In addition to this, 23% of participants failed to sample from at least one category, with the highest total of 17% of participants failing to utilize consensus information in their attributions.
These results suggest that participants are not assessing covariation in the way that Kelley (1973) suggests a naïve scientist would. Their collection of data is by no means exhaustive and particular categories of information that Kelley (1973) deemed necessary to identify covariation using an ANOVA model are often overlooked. However, the notable underutilisation of consensus information is not entirely surprising due to investigations in other areas of social psychology. Consensus information concerns the extent to which other actors show the same behavior towards the target entity, namely the base-rate for the occurrence of this phenomena (Kassin, 1979). However, Kahneman & Tversky, (1972) found that participants tended not to apply non-intuitive base rates in their estimates of the probability of certain occurrences and instead used preconceived heuristics to interpret the data. These findings are important for covariation-based attribution theory as they suggest that participants are unlikely to found attributions on sample base-rates when they have normative expectancies derived from previous causal observations. In line with this, Nisbett and Borgida (1975, as cited in Kassin, 1979) found that there were no differences between the attributions made by participants who were given consensus information and participants who were not provided with base-rate data. However, Kassin (1979) suggests that consensus information is not entirely redundant, as participants will use it when prior expectancies have been neutralized or if the information is particularly salient.
As well as the conceptual and methodological problems concerning the validity of an attribution model based on the covariation of consensus, consistency and distinctiveness information, questions about the use of an ANOVA model have also been raised. Firstly, a theory of attribution based on covariation is problematic as statistical principles affirm that correlation does not denote causation. Nor does causation imply correlation (Hewstone, 1989). An equally valid problem concerning this covariation model is the fact that it is hard to assess the cognitive processes actually used by participants in ascribing causality. Although the attribution arrived at may conform to those predicted by Kelley, it does not necessarily mean that the participants have computed a complex statistical technique such as ANOVA in their heads (Hewstone, 1989).
Two, more recent models of the covariation of consensus, consistency and distinctiveness, attempt to overcome the difficulties encountered when using an ANOVA as the basis for everyday assessment of causality. The Abnormal Conditions Focus model proposed by Hilton and Slugoski (1986) suggested that participants select the condition seen to be abnormal when contrasted to consensus, consistency and distinctiveness information, as the cause of the occurrence; if consensus is low (i.e. no one else performs the action) then the particular person is abnormal and causality is attributed to them (Hewstone, 1989)
Tie back in the garland study to introduce a combined model.