From the literature, we found that PGV-PD3 regressions for on-site earthquake early warning (EEW) can be quite different depending on the presumption whether or not PGV-PD3 data from different regions should be "mixable" in regression analyses. As a result, this becomes a source of epistemic uncertainty in the selection of a PGV-PD3 empirical relationship for on-site EEW. This study is aimed at examining the influence of this epistemic uncertainty on EEW decision-making, and demonstrating it with an example on the use of PGV-PD3 models developed with data from Taiwan, Japan, and Southern California. The analysis shows that using the "global" PGV-PD3 relationship for Southern California would accompany a more conservative EEW decision-making (i.e., early warning is activated more frequently) than using the local empirical model developed with the PGV-PD3 data from Southern California only. However, the influence of this epistemic uncertainty on EEW is not that obvious for the cases of Taiwan and Japan.