The Multicue Integration Model
Our impressions of other people integrate numerous pieces of information – the social groups they belong to, their facial expressions, their behavior, and more. For decades, one of the central questions in person perception has been how those pieces are combined into a single judgment.
That question turns out to be remarkably hard to study, for a reason that has less to do with theory than with measurement.
One judgment, many sources
Suppose you ask people how threatening they find Bob. Bob is Black, his behavior comes across as somewhat threatening, and he's smiling. A participant judges him as threatening.
What produced that judgment? It could reflect heavy reliance on the stereotype attached to his social group. It could reflect the threatening behavior being weighted heavily, or the smile being discounted. It could be some combination of all three. The rating is a single number, and the cues that fed it are not recoverable from it.
This is not merely inconvenient. Separating the contributions of multiple cues requires a measurement model built for exactly that purpose.
The MCI model
The multicue integration model estimates how much each of two cues contributed to a judgment, separately, from the pattern of responses a person actually gave.
It is a multinomial processing tree: a theory about the paths that lead to a response, written so that response counts you can observe reveal processes you cannot.
- C₁ — Use of the first cue. The probability that the cue assigned to C₁ determines the response on this trial. Which cue that is, is a modeling choice, and one the model can be refit to test.
- C₂ — Use of the second cue. The probability that the other cue determines the response, conditional on the first cue being insufficient to settle the judgment.
- g — Response bias. A tendency toward one response over the other when neither cue settles the matter.
Two features of this are worth pausing on.
The first is that C₁ and C₂ are free to vary independently. Conventional measures are relative by construction — they pit cues against each other, so an apparent trade-off may be an artifact of the measure rather than a fact about the perceiver. Because the MCI model estimates each cue's contribution on its own, how the two relate becomes something you can measure instead of something you have to assume.
The second is that the model's structure is asymmetric. C₂ is conditional on the first cue being insufficient — the tool in the next section makes that nesting visible, and it is not a throwaway detail. It turns out to be testable, which the last section returns to.
The model was introduced and validated in Klein and Sherman (2025). In a face classification study, estimated cue use behaved as it should: each cue was used more when it was relevant to the judgment at hand, and less when that cue was made perceptually ambiguous.
Work through the model
The tool below runs the model on one of the trustworthiness judgments from our own data. Faces vary in two cues at once — facial expression and race — and the two sometimes point toward the same response and sometimes pull apart.
Pick a face to see how the model predicts the response, and use the second control to reassign which cue the model treats as C₁. The trial itself never changes when you do that; what changes is which cue's contribution the model estimates unconditionally and which it estimates only when the first cue is insufficient. Fitting both ways and comparing them is how the assignment stops being a modeling assumption and becomes something the data can speak to.
The face on this trial
Which cue is assigned to C₁
P(“Trustworthy”)
P(“Untrustworthy”)
With a way to estimate each cue’s contribution separately, one long-standing assumption in person perception becomes directly testable.
The competitive cue assumption
A striking number of models in person perception assume that cues compete: that relying more on one kind of information means relying less on another. Increased stereotyping requires decreased individuation. Activating one category diminishes the others.
This assumption licenses a common inference: if a manipulation increases stereotyping, the competitive account implies individuation must have decreased. Testing that implication directly has been difficult, because conventional measures cannot separate the two contributions from one another. The MCI model offers a way to do exactly that.
What would competition actually look like?
The competitive cue assumption, stated in the model's terms, is a claim about covariation. If cues compete, then perceivers who rely heavily on one cue should rely less on the other, and the two parameters should be negatively correlated across people.
That is a claim you can look at.
WHAT WOULD EACH HYPOTHESIS LOOK LIKE IN DATA?
Correlation:
Each dot is one simulated perceiver.
Only the first of these is what person perception theory has long assumed.
Only the leftmost of those three pictures is what the field has assumed. The question is which one the data actually look like.
The test
We fit the MCI model to 14 datasets spanning three methods, six kinds of cue, and judgments ranging from split-second categorizations to deliberate selection decisions.
Across people
For each dataset we estimated the correlation between the two cue-use parameters, then combined them.
The combined estimate sits essentially at zero: r = −0.01, 95% BCI [−0.27, 0.25], with the Bayes factor favoring the null. But the aggregate is the least interesting number on the page. The heterogeneity is enormous — τ = 0.48, I² > 99% — and the individual estimates run from strongly negative to strongly positive.
That spread has structure. The negative correlations come overwhelmingly from the Judgment Bias Task, where participants make deliberate accept-or-reject decisions under an instruction to accept about half of the applicants. The positive ones come from speeded classification. So cues do sometimes compete — but competition looks like a property of the task, not a general feature of how people combine information.
We were able to test one candidate explanation directly. Running the selection task with and without the accept-half decision rule, the negative correlation did not reappear: cue use correlated positively with the rule in place (r = .28) and more strongly still without it (r = .80). Whatever produces competition in that paradigm, it is not the decision rule alone.
Within people
A second, sharper test. If cues compete, then pushing a perceiver to use one cue less should push them to use the other more — compensation within the same person. Across seven experiments we manipulated cue ambiguity and watched what happened to both parameters.
The manipulations worked — targeted cues moved by as much as 0.74. The other cue stayed where it was. Pooled across all of them, the correlation between the two changes was r = 0.00, 95% BCI [−0.10, 0.08]. Compensation appeared in 2 of 12 tests, and when cue use did shift together it was more often in the same direction than in opposite directions.
What this means
Cue competition is not a default. It happens, but it appears to depend on a host of factors, like the cues involved and the judgment at hand.
That reframes the question. Instead of asking whether cues compete, we can start asking when they compete, when they cooperate, and what makes the difference.
What else the model can ask
Because C₂ is conditional on C₁, the assignment of cues to parameters is a substantive claim about which cue is processed by default. Fit the model both ways, compare fit, and that claim becomes testable. In Klein and Sherman (2025), gender cues were better characterized as the default than emotion cues — even when the task asked participants to judge emotion.
The same framework extends to other questions: whether social categories really are processed more efficiently under cognitive load, how situational and personal cues combine in attribution, and how multiple category memberships are weighed in judgments of the same person. Any task where a perceiver chooses among discrete options, and where targets vary on two dimensions, is a candidate.
Use the model
Model files, data, commented R scripts, and instructions for implementing the MCI model are available on the Open Science Framework at https://osf.io/gxbc5/. The materials include a walkthrough of each model equation.
If you are considering applying the model to your own data, I am glad to help — get in touch.
References
