Clustering Standard Errors at the "Session" Level

Session-specific features of a laboratory experiment, if those exist, do not disappear by clustering standard errors at the session level. Randomly ordering sessions, which is crucial to deal with sampling issues, cannot justify clustering the standard errors at the session level. The experimental design should primarily determine the clustering level. In a typical controlled laboratory experiment where subjects make choices in the same environment repeatedly, clustering at a participant level is inherited from the experimental design, and standard errors could be larger (that is, statistical inference can be more conservative) when clustered at the individual or decision-group level than the session level. It implies that clustering standard errors at the session level can lead to false-positive treatment effects if it is mistakenly chosen. A rule of thumb using standard deviations is introduced.