Rounding in data deposits distorts meta-analytic heterogeneity

摘要

Meta-analysis is standard in economics, and its reporting guidelines ask for data deposits without saying at what precision. Effect sizes are computed from rounded statistics; coding files record counts formed by rounding estimated shares. Treating them as exact distorts the between-study variance through two channels: an added variance and an effect-dependent systematic shift. It is easily missed: a plug-in within-study variance from the same inputs absorbs part of it. On two dishonesty data sets whose exact counts survive, it absorbs 72 to 88 percent; correcting only the first channel recovers about seventy percent, and on a plug-in variance it overshoots.