Statistical analyses
All statistical analyses were carried out in RStudio (Version 4.1.2,
RStudio Team, 2021). In contrast to the pre-registration of this study,
we analyzed the data with Linear mixed models instead of rm ANOVA which
is often used in generalization research. This decision was based on
several limitations of rm ANOVA including the assumption of sphericity
which is often unmet for generalization data (Vanbrabant et al., 2015).
Instead, Linear mixed models has fewer assumptions and offer a more
reliable statistical inference (Vanbrabant et al., 2015). Furthermore,
there is a rise of Linear mixed models use in generalization research
the last years (Ginat-Frolich et al., 2019; Struyf et al., 2018).
Linear mixed models were conducted separately for each experimental
phase with SCR, ssVEPs, valence, arousal, and US-expectancy as separate
dependent variables. These models were fitted using the packages lme4
and lmerTest (Bates et al., 2015; Kuznetsova et al., 2017) and
significance is reported with the Kenward-Rogers approximation for the
degrees of freedom (Kenward & Roger, 1997). All analyses included the
intercept of the Participants as a random effect. For habituation,
Stimulus (CS+, CS-) and Group (LU, MU, HU) were fixed factors. In
acquisition, the ratings of valence and arousal were analyzed in the
same manner as in habituation. For ssVEPs, SCR, and US-expectancy,
Stimulus, Time (Acq1 for the first half of Acquisition, Acq2 for the
second half of Acquisition), and Group were fixed factors. Significant
interactions were followed up with planned contrasts on the development
of the differential stimulus responding from Acquisition 1 to
Acquisition 2 for all group comparisons (LU-HU, LU-MU, HU-MU). For the
factors Time and Stimulus, Acquisition 1 and CS- were the reference
levels, respectively. In generalization, Stimulus and Group were entered
as fixed factors but this time Stimulus had six levels (CS+, GS1, GS2,
GS3, GS4, CS-). Significant main effects for Stimulus were followed-up
with simple contrasts models with CS- as reference point (Lissek et al.,
2008). In case of significant interactions with the factor Group we
further described the shape of the gradients with trend analyses.
Specifically, we assessed whether the gradients differed in terms of
linearity or curvature across groups. To this end, two orthogonal
polynomial trend repeated measures contrasts across all test stimuli
served as fixed factors to examine the shape of generalization gradient.
Specifically, a linear trend repeated measures contrast assessed a
monotonic gradient across all test stimuli while the quadratic trend
assessed curvature gradients.
Furthermore, we quantified the strength of the generalization with a
linear deviation score (([GS1, GS2, GS3, GS4] ∕4) – ([CS+, CS-]
∕2); LDS). The LDS is a single number representing the steepness and
strength of the generalization gradient. Positive values correspond to
shallow and stronger generalization gradients while negative values
correspond to steeper and weaker generalization gradients (Berg et al.,
2021; Kaczkurkin et al., 2017). LDS scores of each group for each
measure were compared with one-way ANOVAs with LDS as dependent variable
and Group as between-subjects factor. Finally, as an exploratory
analysis, we correlated the sum of scores of the IUS with the LDS in
order to explore if dispositional intolerance of uncertainty played a
role in participants’ generalized responses. Additionally, we
investigated whether participants in the three groups differed in how
well they discriminated between CS+ and the other test stimuli.
Participant’s responses were transformed to 1 (accurate response) and 0
(inaccurate response). We then calculated the average discrimination
response per participant for the five comparisons and calculated
between-groups differences for the three groups with one-way ANOVAs. For
the analysis of the discrimination task, eight participants were further
excluded due to equipment failure, resulting in 80 participants included
in this analysis. For all statistical analyses, alpha level was set at
.05 and Bonferroni correction was used to adjust the alpha level for
multiple comparisons.