Group estimates outperform individuals only when judgments are independent

What it is

Independence of judgments means members do not directly copy the same error, visible answer, or incentive signal when estimating or evaluating. A group average can outperform one person not because it is large, but because errors partly cancel.

Why it happens

Independent sources, experience, and perspective produce different errors that aggregation can reduce. Seeing one rank, leader, or others' choices correlates errors and counts one early signal repeatedly — many votes can be an echo of a single early one.

The second-order mechanism can be stated as an exact decomposition rather than a qualitative claim that "independence helps": the group's error (the squared difference between the group estimate and the true value) equals the average individual error minus the diversity among individual judgments (how spread out individual estimates are from each other). This decomposition — the diversity prediction theorem — directly explains two counterintuitive facts. First, a group with enough diversity can outperform even its single best member, without needing any assumption that members are unusually skilled; diversity is literally subtracted from group error. Second, it pinpoints exactly why conformity is harmful: conformity suppresses the diversity term, and if the loss from reduced diversity exceeds whatever modest gain individuals get in personal accuracy from referencing others, the group's overall performance gets worse, not better. Independence is a precondition rather than a nice-to-have precisely because it maps directly onto the diversity term that conformity eats away at.

Studying it

• Paradigm: compare estimate error after independent response, visible group answer, and discussion; using the decomposition above, measure and report the change in average individual error and inter-individual diversity separately across conditions, rather than only the final group error.
• Variables: information visibility, judgment correlation, group size, aggregation method, average individual error, inter-individual diversity, and confidence.
• Methodological caution: independence is not isolation from all information; distinguish evidence sharing from unsupported conformity. Reporting only the final group error hides the more diagnostic question of whether average accuracy got worse or diversity got suppressed — the two should be reported separately.

Where it stops holding

Independent ignorance does not create expertise — if the average individual error term itself is large, no amount of diversity can compensate. Coordination decisions cannot remain fully independent; preserve private input before discussion. The classic evidence for this effect comes mostly from one-shot numerical estimation tasks among strangers who only need to give a single value (the classic weight-guessing experiment). Carried over to real team decisions requiring sustained collaboration and judgment too complex to reduce to a single number, whether diversity's benefit still exceeds coordination cost lacks equally solid evidence and should not be assumed to transfer directly.

Applying it

• Collect private estimates before revealing aggregate and discussion, protecting the diversity term from being suppressed by premature conformity.
• Show source diversity and uncertainty instead of treating consensus as certain correctness; where feasible, show decision-makers how much individuals actually disagreed, not only the final aggregate number.
• Add independent review and expertise for consequential judgment.
• Verification: compare accuracy, variance, and early-answer influence before and after visibility; compute the change in average individual error and inter-individual diversity separately across the two conditions to confirm whether a drop in diversity is the main driver of any accuracy decline.

Related

• Same group: V9.07.2 Visible choices break independence and induce conformity · V9.07.3 Collaborative filtering reinforces popular items and buries the long tail · V9.07.4 Groups do not outperform experts on specialised questions · V9.07.5 A few early ratings can lock an item's long-term visibility
• Nearby: V7.07 Group polarization and echo chambers · V8.03 Content quality
• Search terms: wisdom of crowds · independence · collective intelligence · diversity prediction theorem