Measuring binding effects in event-based episodic representations

Remembering an experienced event in a coherent manner requires the binding of the event's constituent elements. Such binding effects manifest as a stochastic dependency of the retrieval of event elements. Several approaches for modeling these dependencies have been proposed. We compare the contingency-based approach by Horner & Burgess (Journal of Experimental Psychology: General, 142(4), 1370-1383, 2013), related approaches using Yule's Q (Yule, Journal of the Royal Statistical Society, 75(6), 579-652, 1912) or an adjusted Yule's Q (c.f. Horner & Burgess, Current Biology, 24(9), 988-992, 2014), an approach based on item response theory (IRT, Schreiner et al., in press), and a nonparametric variant of the IRT-based approach. We present evidence from a simulation study comparing the five approaches regarding their empirical detection rates and susceptibility to different levels of memory performance, and from an empirical application. We found the IRT-based approach and its nonparametric variant to yield the highest power for detecting dependencies or differences in dependency between conditions. However, the nonparametric variant yielded increasing Type I error rates with increasing dependency in the data when testing for differences in dependency. We found the approaches based on Yule's Q to yield biased estimates and to be strongly affected by memory performance. The other measures were unbiased given no dependency or differences in dependency but were also affected by memory performance if there was dependency in the data or if there were differences in dependency, but to a smaller extent. The results suggest that the IRT-based approach is best suited for measuring binding effects. Further considerations when deciding for a modeling approach are discussed.