Less-is-more effect

Less-is-more effect is the phenomenon where a decision strategy or predictive model with less information outperforms one with more information. The term was popularized by Gerd Gigerenzer and the ABC Research Group in the context of ecological rationality and fast-and-frugal heuristics, but the underlying statistical principle — the bias-variance tradeoff — has been understood in machine learning and statistics since the 1970s.

In Gigerenzer's framework, the effect occurs in environments with noncompensatory cue structures: a single highly valid cue dominates all combinations of weaker cues, so that ignoring the weaker cues eliminates noise without sacrificing predictive signal. The Take-the-best heuristic exemplifies this: it uses only the first discriminating cue and ignores all others, yet it can achieve higher accuracy than multiple regression or more complex machine learning models in certain domains.

Statistically, the less-is-more effect is not paradoxical. A simpler model with fewer parameters has lower variance (less sensitivity to sampling error) but higher bias (stronger assumptions about the data-generating process). When the true environment matches the simpler model's assumptions, the reduction in variance outweighs the increase in bias, and the simpler model generalizes better. This is the same principle that underlies regularization techniques such as LASSO (L1 penalty), ridge regression, and early stopping in neural networks.

The less-is-more effect is thus domain-dependent. It does not demonstrate a universal superiority of simplicity. It demonstrates that the best model is the one whose inductive bias matches the structure of the environment. In multi-causal environments where cues interact and compensate for each other, more information and more complex integration methods are necessary for accurate prediction.