ABSTRACT
Aggregation of data by taking an average reduces confounding. Where the different causes of an effect are fundamentally uncorrelated, so as to vary randomly around zero, averaging can completely eliminate confounding. Then the average apparent effect is the true average effect, so that averages can be used without any additional controls for confounding. Averaging across organizational subunits eliminates confounding as organizational size increases, but at a greatly decreasing rate with respect to size. Averaging avoids the need for control groups, it avoids incurring the downward bias in effect that comes from using organizational experiments. The advantage for large organizations in controlling for confounding adds to the advantage of large organizations in controlling for sampling error. These combined advantages derive from the large organization's status as a meta-machine that reduces error in the inferences that can be made from data as they flow up the hierarchy and are aggregated.
