A note on the modeling of the effects of experimental time in psycholinguistic experiments

Barr et al. (2013) proposed as gold standard for the analysis of experimental data with observations on combinations of subjects and items to fit maximally specified linear mixed effects models (LMMs). Bates et al. (2015) pointed out that such models run the risk of being overspecified, and Matuschek et al. (2017) provided detailed discussion of the balance between power and type-I error in LMMs. A study by Baayen et al. (2017b) raised another issue, namely that in psychometric data one often finds that sequences of response latencies observed over time as an experiment unfolds are not independent but are auto-correlated, often at substantial lags. Auto-correlation functions (ACF) for subject 1 from the British Lexicon Project (Keuleers et al., 2012) are presented in Figure 1. The left panel presents the ACF for the inverse-transformed reaction times. The right panel shows the ACF for the residuals of a model with lexicality (word vs. nonword) as predictor. In both cases, auto-correlations are markedly present even at lags of 40 trials.