Cognitive Algebra in Impression Formation: Additive and Averaging Models Explained

Every time you meet someone new, your brain quietly gets to work – scanning their traits, weighing their qualities, and arriving at an overall judgment within moments. But how exactly does it combine all that information? Social psychologist Norman H. Anderson spent decades trying to answer that question, and what he uncovered became the foundation of cognitive algebra – a framework that treats impression formation as a kind of mental math. At the heart of this framework are two competing models: the additive model and the averaging model. Understanding both reveals a great deal about how human judgment actually works – and where it falls short.

What is cognitive algebra? 🔗

The term “cognitive algebra” refers to the mental operations people use when combining multiple pieces of information into a single judgment. Anderson’s Information Integration Theory proposes that when we evaluate a person, we don’t just form a vague “feeling” – we effectively perform an internal calculation. Each trait we observe gets assigned a subjective value, and those values are then combined through an algebraic function to produce an overall impression. That function could be additive, averaging, or even multiplicative, depending on the situation. The two most studied and debated functions are adding and averaging.

This approach was a deliberate departure from earlier, more holistic theories – particularly Solomon Asch’s configural model, which argued that traits influence each other and combine into a unified whole. Anderson rejected this in favor of a more quantifiable, mathematical approach. His goal was to model the cognitive process with enough precision to make testable predictions.

The additive model: more is always more 🔗

Anderson’s early work, including his landmark 1962 study published in Science, first applied an additive model to impression formation. The core logic is straightforward: each trait you observe about a person carries a positive or negative value, and your overall impression is the sum of all those values.

By 1965, Anderson had formalized this into a direct test. Using a scale where traits were rated for likableness, he found that summing trait values could predict overall impressions reasonably well. Under this model, learning more positive things about someone always improves your impression of them. Every good trait adds to the total. Conversely, each negative trait subtracts from it.

The logic of the additive model 🔗

To see how this works, consider a simple example. Suppose you learn three things about a person: they are intelligent (+3), kind (+2), and reliable (+3). Under the additive model, your overall impression score would be +8. If you then discover they are also witty (+2), the score climbs to +10. The model treats every additional piece of information as a net gain or loss – each trait is independent, cumulative, and equally weighted.

This gives the additive model a clean, intuitive appeal. It predicts that the more positive traits you know about someone, the more favorably you’ll view them – regardless of the mix of traits already present.

Where the additive model breaks down 🔗

The problem emerges when you test this prediction against real behavior. In practice, adding more moderately positive information to a very positive initial impression doesn’t always make the impression better – it can actually dilute it. A person described as brilliant, creative, and inspiring may strike you as more impressive than someone described as brilliant, creative, inspiring, and also quite punctual. The fourth trait isn’t negative, but it lowers the overall standard. The additive model cannot account for this effect because it treats all positive values as purely additive.

This limitation – known as the set-size effect – was a central problem that Anderson himself identified. Adding more traits of a lower positive value than the existing average should, by additive logic, keep increasing the impression. But empirical results showed the opposite: it tended to pull the impression downward.

The averaging model: quality over quantity 🔗

To address the failures of the additive model, Anderson developed a more refined approach. In 1974, he proposed the averaging model, which holds that your overall impression is not the sum of trait values but their weighted average. The model formally appears in his broader work on cognitive algebra and information integration, where he explored how adding, averaging, and multiplying models apply differently to social judgments.

Under the averaging model, each trait still has a subjective value, but the final impression is calculated by dividing the total by the number of traits (or by their assigned weights). This means that adding new information doesn’t automatically improve the impression – it depends on whether the new trait’s value is above or below the current average.

How the averaging model works in practice 🔗

Return to the earlier example. If someone is intelligent (+3), kind (+2), and reliable (+3), the average is +2.67. If you then learn they are also somewhat shy (+1), the new average drops to +2.25. Despite “shy” being a mildly positive trait, it lowers your impression because it falls below the existing average. This is the averaging model’s key prediction – and it matches what studies consistently find in real behavior.

This also explains a counterintuitive finding: sometimes, knowing less can lead to a more favorable impression than knowing more. If the additional traits are weaker than what’s already established, they bring the average down.

Weighted averaging: accounting for trait importance 🔗

Anderson’s averaging model also accommodates the concept of differential weighting – the idea that not all traits are equally important in forming an impression. In a weighted averaging framework, a trait like “dishonest” might carry far more weight than “punctual,” even if their raw values are similar in magnitude. The differential weighted averaging model tested by researchers showed that this single integration rule could explain both additive-looking and averaging-looking patterns of data – making it a more general and powerful framework.

For example, if you’re evaluating someone as a potential business partner, traits like honesty and competence might carry much higher weight than humor or sociability. The weighted average would reflect this priority structure, producing a more realistic impression than a simple sum or unweighted average ever could.

The set-size effect: a critical test 🔗

One of the most revealing tests between the two models involves what’s called the set-size effect. This refers to what happens to impressions when you add more traits to a list. The additive model predicts that adding any positive trait should increase the impression – since more positive values simply add to a higher sum. The averaging model predicts the opposite: adding a moderately positive trait to a highly positive set will lower the overall impression.

Anderson’s 1967 experimental analysis directly tested this and found consistent support for the averaging prediction. Participants formed less favorable impressions when moderately positive traits were added to strongly positive profiles – precisely what the averaging model forecasts and the additive model cannot explain. This finding has since been replicated across multiple studies, cementing the averaging model’s superiority in predictive accuracy.

The primacy effect within the averaging model 🔗

An important nuance within the averaging model is that not all traits receive equal weight in practice – even when they should, mathematically. Anderson noticed that earlier information tends to carry more weight than information presented later. This is the primacy effect: first impressions dominate.

Anderson explained this through a simple attention mechanism. When you first encounter someone, each new trait is entirely unpredictable and therefore captures full attention. As more traits accumulate, each additional one commands less cognitive attention – meaning it receives less weight in the final average. A follow-up study by Anderson and Hubert found that when participants were forced to give equal attention to all traits, the primacy effect disappeared – strongly supporting this attentional explanation rather than any structural bias in the averaging process itself.

Additive vs. averaging: a direct comparison 🔗

The two models differ in their fundamental assumptions and in what they predict:

• Independence vs. interdependence: The additive model treats traits as independent – each one contributes its value in isolation. The averaging model treats them as interdependent – each trait’s influence depends on what else is present.
• Equal vs. differential weighting: The additive model weights all traits equally by default. The averaging model allows traits to carry different degrees of influence based on their relevance or salience.
• More is always better vs. quality matters: Under addition, every additional positive trait improves the impression. Under averaging, adding a weaker positive trait can dilute an otherwise strong impression.
• Predictive accuracy: Research consistently finds that the averaging model – particularly in its weighted form – provides more accurate predictions of how people actually form impressions in experimental settings.

Why this matters beyond the lab 🔗

These models aren’t just theoretical constructs – they have real implications for how judgments are made in hiring decisions, courtroom settings, medical consultations, and everyday relationships. If interviewers operate under an additive logic, they might favor candidates with the longest list of qualifications, even if some of those qualifications are mediocre. Under averaging logic, a smaller set of highly impressive qualifications would actually produce a stronger overall impression.

Similarly, in legal contexts, presenting a jury with too many pieces of moderately supporting evidence – rather than a few strong ones – could, according to the averaging model, weaken rather than strengthen the perceived case. This has been explored in judgment and decision-making research extending Anderson’s foundational framework into applied domains.

The averaging model also aligns with a broader principle in social perception: humans are not simple scorekeepers. We are sensitive to the quality and relative standing of each piece of information, not just its raw presence. This is why a single highly negative trait – like discovering someone is dishonest – can outweigh a dozen positive ones. It doesn’t just subtract; it recalibrates the entire average, often dramatically.

Limitations of both models 🔗

Despite the averaging model’s stronger empirical support, neither model fully captures the complexity of real-world impression formation. Both assume a degree of rational, systematic processing that doesn’t always reflect fast, automatic, or emotionally driven judgments. Factors like context, relationship history, cultural expectations, and nonverbal cues all shape impressions in ways that pure algebraic models struggle to represent.

The ongoing debate between algebraic modelers and contextualists in social psychology reflects this tension. Contextualists argue that impressions are too situationally specific and semantically rich to be reduced to weighted averages. Algebraic models, Anderson countered, are not meant to capture every nuance – they are approximations that reveal the core structure of the integration process.

What do you think? If you were describing yourself to a potential employer, would you list every positive quality you have – or would you carefully select only your strongest traits to avoid diluting the impression? And does knowing about the averaging model change how you interpret the way others have judged you in the past?

References

1. https://en.wikipedia.org/wiki/Information_integration_theory
2. https://pubmed.ncbi.nlm.nih.gov/17820996/
3. https://pubmed.ncbi.nlm.nih.gov/5826027/
4. https://www.researchgate.net/publication/17123240_Averaging_model_analysis_of_set_size_effect_in_impression_formation
5. https://www.sciencedirect.com/science/article/abs/pii/S0065260108600350
6. https://www.sciencedirect.com/science/article/abs/pii/S0022103173800034
7. https://pubmed.ncbi.nlm.nih.gov/6062954/
8. https://journals.sagepub.com/doi/10.2466/pms.1968.27.3f.1295
9. https://en.wikipedia.org/wiki/Impression_formation