This work proposes a new approach to modeling psycholinguistic time series
– deconvolutional time series regression (DTSR) – and applies it to
existing datasets in order to show that temporal diffusion of effects may
play a major role in subject behavior. DTSR learns best-fit impulse
response functions that mediate the temporal relationship between
independent and dependent variables, providing fine-grained estimates of
the timecourse of an independent variable’s influence on the response. Our
results show that, on datasets with sufficiently long time series, DTSR
models make better predictions and learn substantially different effect
estimates than either LME or GAM.