Table 1.

Generalized mixed-effects models on individual and aggregated errors.

Effect	Estimate	Fitted Brier score	SE	t	p
(A) Individual forecasting error as a function of forecast type
Intercept	−2.14224	0.1173915	0.24230	−8.841	<2e − 16
Initial	0.62237	0.2187395	0.09040	6.884	5.81e − 12
Revised	0.69532	0.2352946	0.08947	7.772	7.73e − 15
Final	0.23849	0.1490093	0.09979	2.390	0.0169
(B) Individual forecasting error as a function of diversity and group size
Intercept	−1.96631	0.139972	0.30877	−6.368	1.91e − 10
Final	0.20997	0.1726759	0.07814	2.687	0.00720
Diverse	−0.37285	0.1189339	0.20278	−1.839	0.06595
Small	−0.20011	0.1413602	0.20094	−0.996	0.31932
Diverse:small	0.82956	0.2231896	0.29025	2.858	0.00426
(C) Aggregated forecasting error as a function of forecast type
Intercept	−1.8387	0.1590198	0.2508	−7.331	2.29e − 13
Initial	0.6815	0.3143683	0.2293	2.972	0.00296
Revised	0.5999	0.2897372	0.2301	2.607	0.00913
Final	−0.1281	0.1398955	0.2964	−0.432	0.66557
(D) Aggregated forecasting error as a function of Diversity and Modularity
Intercept	−1.76627	0.1709691	0.33428	−5.284	1.27e − 07
Final	−0.06877	0.15960632	0.15360	−0.448	0.65434
Diverse	−0.56382	0.09082084	0.23514	−2.398	0.01649
Modular	−0.82268	0.07010727	0.26515	−3.103	0.00192
Diverse:modular	0.93267	0.10137943	0.38254	2.438	0.01477

Table of analysis on forecasting errors (in Brier scores) for individual (A–B) and aggregated measures (C–D), and as a function of forecast type (A–C) and condition (B–D). Baselines for each factor: consensus, homogeneous, large/non-modular. The effect of final forecasts on individual errors (A) and the effects of composite diversity and the interaction between composite diversity and modularity (D) did not survive a Bonferroni correction. Boldface: p < 0.05; italics: p < 0.10. Tables B–C represent exploratory analyses. Hypotheses in tables A and D were preregistered. All analyses were also repeated with binarized accuracy (Supplementary Tables 11–14) and logit link function (Supplementary Table 15). For convenience, all tests refer to two-sided hypotheses and were calculated with the lmerTest package in R⁶³.