Covariation model
Also known as: Kelley's covariation model
We decide a cause by checking what the effect reliably varies with.
What it means
Kelley's account of how people make causal attributions when they have information across multiple observations, by attributing an effect to the factor with which it covaries. Observers weigh three kinds of information: consensus (do other people respond the same way to this stimulus?), distinctiveness (does this person respond this way only to this stimulus?), and consistency (does this person respond this way across time and settings?). High consensus, high distinctiveness, and high consistency point to the stimulus; low consensus, low distinctiveness, and high consistency point to the person. The model treats lay causal reasoning as a quasi-statistical analysis of variance. While normatively elegant, people often deviate from it, especially by underusing consensus information — a finding that helped reveal the fundamental attribution error.
Reading the cube
The model is usually drawn as a two-by-two-by-two cube: consensus, distinctiveness, and consistency each high or low, giving eight cells. Beyond the person and stimulus patterns, a third canonical outcome is a circumstance attribution, produced when consensus is low, distinctiveness is high, and consistency is low: the behavior happened just this once, so we credit the moment rather than the actor or the thing. Consistency does special work here. It behaves less like a third vote and more like a gate. When consistency is low, neither a person nor a stimulus attribution is licensed, and observers fall back on transient circumstances or withhold judgment. The three cues are not read independently; the whole configuration is interpreted at once, which is why fragmentary information leaves an attribution unsettled.
What the evidence shows
McArthur (1972) ran the first full test, pairing sixteen sentence-events across emotions, accomplishments, opinions, and actions with all eight information patterns. The broad predictions held: distinctiveness and consistency moved judgments strongly in the directions the model specifies. But consensus was weak, explaining only a small share of the variance in person attributions relative to distinctiveness. Hewstone and Jaspars (1987) later showed that a formal logical model, closer to Kelley's own analysis-of-variance analogy than the loose verbal account, fit people's choices better. The recurring lesson is that observers can use covariation information roughly as prescribed when it is handed to them cleanly and completely, but they weight the three cues unequally rather than combining them as a statistician would.
The consensus puzzle
Consensus underuse hardened into a textbook fact and a plank in the case for the fundamental attribution error: if people discount how others behave, they over-attribute to disposition. But the finding is partly an artifact of method. McArthur's respondents could pick single causes or their combinations, but a genuine person-situation interaction had little explicit room in that format, so consensus effects had nowhere obvious to land. When later studies supplied richer response formats and made the base-rate meaning of consensus concrete rather than abstract, its influence grew. The honest reading is that consensus is genuinely underweighted against a statistical ideal, but not ignored, and its apparent neglect was inflated by how the early experiments framed the question. The direction of the bias is real; its size was overstated.
What it does not explain
Kelley's model chooses among candidate causes; it does not generate them. It assumes the observer already holds the relevant comparisons, multiple people, occasions, and stimuli, which a single encounter rarely provides. For those one-shot cases Kelley proposed separate causal schemata, prior templates about how causes combine. Cheng and Novick (1990) recast the machinery as a probabilistic contrast computed over a focal set of causes the reasoner has already selected, which improved the fit to data and exposed the gap plainly: the model is silent on where that candidate set comes from. The practical upshot is to treat covariation as a disciplined way to test causes you have already named, not a method for discovering which causes were worth naming in the first place.
Examples
If everyone laughs at a comedian (high consensus), you laugh only at this comedian (high distinctiveness), and you do so every time (high consistency), you credit the comedian, not yourself.
Your laptop freezes only in one app, every time you open it, and other users report the same — high distinctiveness, high consistency, high consensus, so you blame the app, not the machine.
A manager notices one team member complains about every project, nobody else complains about this one, and she does it week after week — low distinctiveness and low consensus point to the person.
Nearly every diner pans one dish (high consensus), while praising the kitchen's other plates (high distinctiveness), and the complaint recurs across visits (high consistency): the fault sits with that dish, not fussy customers.
A striker misses in every match (high consistency), against keepers his teammates routinely beat (low consensus), and in training too (low distinctiveness): the shortfall reads as his finishing, not bad luck.
First described in Harold Kelley (1967).
Key references
- Cheng, P. W., & Novick, L. R. (1990). A probabilistic contrast model of causal induction. Journal of Personality and Social Psychology, 58(4), 545-567. doi.org/10.1037/0022-3514.58.4.545
- 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(4), 663-672. doi.org/10.1037/0022-3514.53.4.663
- Kelley, H. H. (1973). The processes of causal attribution. American Psychologist, 28(2), 107-128. doi.org/10.1037/h0034225
- McArthur, L. A. (1972). The how and what of why: Some determinants and consequences of causal attribution. Journal of Personality and Social Psychology, 22(2), 171-193. doi.org/10.1037/h0032602