CUMULATIVE SEMANTIC INHIBITION
10
included in the model. N-model 5 reports the estimates when only the significant variables
are considered in the model3 . The combined estimated effects of ordinal position within
co-category and ordinal positions within supra-category are plotted in the right-center panel
of Figure 2. In summary, this third step establishes that the ordinal position within the
supra-category has a significant inhibitory effect over and above a significant effect of ordinal
position within the (original) category.
Discussion
The analysis of single trial data using linear mixed-effects models has equipped us
with sufficient statistical power to analyze a subset of the original data with models similar
to those used for the complete dataset. These analysis are based on the plausible pairings
of categories summarized in Table 2. They indicate that, at least for the items under
consideration, the semantic cumulative inhibition effect is not restricted within categories.
The plots on the leftmost and rightmost panels of Figure 2 illustrate the cumulative
inhibition effect for models that consider a single level of category (respectively: when the
only the co-category or only the supra-category levels are considered). The two middle
panels plot models where the dependency between co-categories is captured, in addition to
within category effects, by either a relative category position predictor or a supra-category
ordinal predictor. Our analyses enable us discarding both single category models in terms
of their predictive accuracy. The predictions are significantly better when both representa-
tional levels are considered.
Statistics alone are not sufficient to disentangle the two remaining models, however,
as they both produce equally good predictions for this dataset. One way to resolve this issue
would be to consider a hypothetical experiment where co-categories would be intermixed.
The definition and predictions of the model with two ordinal factors at the two levels are
straightforward, and our interpretation of the results described above leads to a distinct
prediction. The middle panel of Figure 3 should be the better description of the results,
over and above the left and right panels. The prediction depicted in this panel is an ordinal
inhibition effect between items of any co-category plus an ordinal inhibition effect of different
magnitude between successive trials belonging to two different co-categories. In contrast,
defining which co-category comes first when co-categories are intermixed in the experiment
would require additional assumptions. For this reason, the model with two ordinal factors
may be preferred. Pending further evidence, however, we will simply draw the important
conclusion that the dependency between co-categories is not reductible to supra-category.
General discussion
The cumulative inhibition effect reported by Howard et al. (2006) is present across
categories. This confirms the robustness of their finding when random variation is explicitly
taken into account. On top of this, our analyses have added some facts that were not
previously considered. We found a significant variation in the magnitude of the cumulative
effect. This variation is independent of the variation in overall speed across categories. In the
3 As it was the case in step two, the variable magnitude of the cumulative inhibition effect does not reach
significance in this analysis