Anchoring Effects and Troublesome Asymmetric Transfer in Subjective Ratings.

Within-subjects experiments are prone to asymmetric transfer, which confounds results interpretation. While HCI researchers routinely test asymmetric transfer in objective data, doing so for subjective data is rare. Yet literature suggests that anchoring effects should make subjective measures particularly susceptible to asymmetric transfer. We report on four analyses of NASA-TLX data from four previously published HCI papers, with four main findings. First, asymmetric transfer is common, occurring in 42% of tests analysed. Second, the data conforms to predictions of anchoring effects. Third, the magnitude of the anchor's effect correlates with the magnitude of the difference between the interface ratings -- that is, the anchor's 'pull' correlates with the anchoring stimulus. Fourth, several of the previously published findings are changed when data are reanalysed using between-subjects treatment. We urge caution when analysing within-subjects subjective measures and recommend that researchers test for and report the occurrence of asymmetric transfer.