**3. Results**

In this section, before reporting the results related to the main hypothesized e ffects, we describe the characteristics of the by-item analysis in terms of descriptive and correlational statistics.

### *3.1. Stimuli Characteristics: Descriptive and Correlations*

For an illustration of the relationship between variables, tables are shown in the following section. Table 1 contains descriptive information of the variables under considerations, including scores of the variables measured. Table 2 shows the correlations between all variables.


**Table 1.** The descriptive information of the experimental variables.

**Table 2.** Pearson Correlations between the experimental variables.


\*\* *p* < 0.001 (2-tailed).

### *3.2. Mental Imagery Vividness and Familiarity*

The mean vividness score across all noun-cues significantly correlated with each noun-cue's familiarity rating (*r*p (50) = 0.48, *p* (two-tailed) < 0.001, ηp 2 = 0.23), as shown in Table 2. This correlation confirms the relationship between vividness and familiarity posited by previous research, suggesting that our dataset demonstrates the typical relationship observed between the two variables [7].

### *3.3. Mental Imagery Vividness and Reaction Time*

The mean vividness score across all noun-cues significantly correlated with each noun cue's mean reaction time (*r*p (50) = −0.39, *p* (two-tailed) < 0.001, ηp 2 = 0.15), as shown in Table 2. This correlation, known as the "vivid-is-fast" phenomenon, has been observed in previous research and therefore helps to validate our experimental dataset [65].

### *3.4. E*ff*ects of Resting State Networks on Learning Outcomes*

### 3.4.1. By-Item Analysis Approach

Following our hypotheses, to assess how RSN activation pattern's influence mean recall probability across noun-cues, we devised a one-factor ANOVA with three levels, each representing distinct resting state activation patterns. As was described in the introduction, when a dominant DMN activation is observed in human participants, both mental imagery vividness and object familiarity are experienced consciously as a weak stimulus. Inversely, when a dominant TPN activation is observed, both mental imagery and object familiarity are experienced consciously as a strong stimulus. Given this relationship, it is possible to infer how RSNs a ffect learning outcomes by investigating the relationship between mean familiarity, mean vividness, and mean recall probability ratings. Applying a widely-used continuous

Vividness and familiarity ratings: Min = 1 to Max = 7. RTs = Image latency times in milliseconds.

variable partitioning technique [66], we created a variable with three levels that aimed to represent the following three distinct resting state network activation patterns: 1. TPN dominant activation (composed of noun cues which have high vividness and high familiarity ratings); 2. Mixed TPN and DMN activation (composed of noun cues with high vividness/low familiarity or low vividness and high familiarity); 3. DMN dominant activation (composed of noun cues with low vividness and low familiarity ratings). The decision to create a three-leveled variable characterized by the strength of mean familiarity and mean vividness scores was influenced by the NMPH model utilized in Norman and Newman [9]. In their model, the moderate/mixed level is of critical importance to assess non-monotonic relations between the strength of a stimulus and subsequent learning outcomes. For this reason, we judged that three levels would be ideal given that this number of levels would adequately contrast differences between states of stimulus competition (high/low) and states of stimulus collaboration. The descriptive information detailing the relationship between the resulting RSN variables and recall probability is shown in Table 3.

**Table 3.** Descriptive statistics of total noun-cues recalled (in percentages) as a factor of resting state network type.


The percentage of noun-cues recalled for each of the three RSN levels were entered on a 1 (recall probability) × 3 (RSN levels: low, mixed and high activation) one-way ANOVA. This procedure yielded a significant effect of resting state network type on recall probability (*F*(2,47) = 4.51, *p* = 0.016). Additionally, the ANOVA resulted in a large effect size, <sup>η</sup>p<sup>2</sup> = 0.16 indicating that 16% of the variance in recall probability is attributable to dominant RSN activation. To investigate whether the mixed resting state activation level impaired recall probability, as previously hypothesized (i.e., our switching hypothesis), a polynomial quadratic planned contrast was further applied to the data. The contrast demonstrated a significant quadratic trend (*F*(1,47) = 7.18, *p* = 0.01, <sup>η</sup>p<sup>2</sup> = 0.13). This quadratic trend can be clearly observed as represented in the continuous connecting line in Figure 3. To investigate how well our results fit within the dualistic RSN neuropsychological model of memory consolidation, we conducted Dunnett's post-hoc testing using the DMN-Dominant level as our main comparison criteria. Dunnett's post hoc test revealed that there were no significant differences between the DMN-Dominant and TPN-Dominant levels (Dunnett's *t*(47) = 1.45, *p* = 0.154), however there was a significant difference between the DMN-Dominant and the Mixed levels (Dunnett's *t*(47) = 3.01, *p* = 0.004). These results confirm the hypothesized quadratic pattern of the dualistic RSN neuropsychological model of memory consolidation, as shown in the (Figure 3).

To investigate if the quadratic trend was present across different percentiles of the recall probability distribution, the data were converted into proportions as detailed in Table 4. Planned quadratic proportion contrasts were performed on each third of the recall probability distribution. For both the top third (i.e., high recall probability) and middle (i.e., medium recall probability) percentile there were significant positive quadratic trend (*z* = 3.24, *p* < 0.001, <sup>η</sup>p<sup>2</sup> = 0.21; and *z* = 2.42, *p* = 0.008, <sup>η</sup>p<sup>2</sup> = 0.12, respectively). Conversely, the low recall group showed a significant negative quadratic trend (*z* = −4.3464, *p* < 0.001, <sup>η</sup>p<sup>2</sup> = 0.38). To adjust for multiple comparisons, a two-tailed Bonferroni correction was applied to each contrast using 95% confidence intervals. Given that the data being used are proportions, the confidence intervals were treated as the hypothesis test wherein if an interval crossed the 0 threshold it would signify that the null hypothesis failed to be rejected and therefore, the contrast would be interpreted as not significant. The following equation was used to test for multiple comparisons: - *p*λ ± *z* α2*g s*2*p*<sup>λ</sup>2, for further clarification on planned contrast in proportions see Rosenthal and R Rosnow (1985) and Hays (1994) [67,68]. The results of the two-tailed Bonferroni correction are as follows: High recall probability (0.445 ± (−0.301)), Medium recall probability (0.6031 ± 0.2477), Low recall probability (−1.0476 ± 0.2394). Each contrast's 95% confidence interval did not cross the 0 threshold, indicating that the quadratic contrast amongs<sup>t</sup> each level of the recall variable were significant at *p* < 0.025. The two significant positive quadratic trends in the top 2 percentiles of the recall probability distribution in addition to the significant negative quadratic trend in the bottom percentile demonstrate that noun cues associated with competing RSN activation patterns result in poor recall probability.

**Figure 3.** Percentage of words recalled in relation to the dominant resting state network activation pattern.



3.4.2. Linear Mixed Logistic Regression Approach

To assess how RSN activation patterns influence recall probability across each trial observation (*N* = 1300) we devised a linear mixed logistic regression model where recall of stimuli served as our binary target variable (1 = recalled, 0 = not recalled) and where dominant RSN activation patterns served as our fixed predictor variable. Similar to the by-item analysis, our RSN predictor variable contained three levels, each representing distinct resting state activation patterns. To demarcate each level of the predictor variable, we used the same partitioning technique used in the by-item analysis [66]. Specifically, the underling neuropsychological variable strength used to model the three RSN levels are as follow: 1. TPN dominant activation (composed of trial observations which have high vividness and high familiarity ratings); 2. Mixed TPN and DMN activation (composed of trial observations with high vividness/low familiarity or low vividness and high familiarity ratings); 3. DMN dominant activation (composed of trial observations with low vividness and low familiarity ratings). The descriptive

information detailing the relationship between the resulting RSN variables and recall proportions are shown in Table 5.


**Table 5.** Descriptive statistics of total noun-cues recalled across all trials (in proportion) as a factor of resting state network type.

> All values have been adjusted for within-subject effects.

Controlling for nested within-subjects data, we tested whether recalled noun-cues following a binomial logit link function could be predicted by the fixed e ffect categorical RSN predictor variable which reflected our three RSN levels (low, mixed, and high activation). This procedure yielded a significant fixed e ffect of resting state network type on noun-cue recall ( *F*(2,1297) = 3.30, *p* = 0.037). To investigate whether these results supported the dualistic RSN neuropsychological model, we conducted Fisher's least significant di fference (LSD) post-hoc testing. The LSD post-hoc test was used over other more conservative statistical tests for two primary reasons. First, since our predictor variable only consisted of three groups, the LSD procedure permits to preserve the experiment-wise type I error rate at nominal levels of significance, which nullifies the use of post-hoc tests that control for the family-wise error rate [69]. Second, for each level of our RSN predictor variable, the variance-covariance matrix demonstrated that covariance scores were all equal, and variance scores did not significantly di ffer among each other, suggesting that our data did not violate the assumption of compound symmetry needed for appropriate application of the LSD post-hoc test [70]. The LSD post-hoc test revealed that there were no significant di fferences between the DMN-Dominant and TPN-Dominant levels (LSD's *t*(1297) = 0.178, *p* = 0.859), however there were significant di fferences between both DMN-Dominant (LSD's *t*(1297) = 1.97, *p* = 0.049) and TPN-Dominant (LSD's *t*(1297) = 2.31, *p* = 0.021) when contrasted with the Mixed RSN level. These results confirm the hypothesized quadratic pattern of the dualistic RSN neuropsychological model of memory consolidation. Additionally, the results generated using the linear mixed logistic regression model helped to validate the findings derived from the by-item approach. Both statistical approaches demonstrated that learning outcomes are significantly predicted by the characteristic non-monotonic quadratic pattern of our dualistic RSN model of memory consolidation.

### *3.5. E*ff*ects of Resting State Networks on Mental Imagery Latency (Reaction Time)*

The mean reaction times associated with each noun-cue for each of the three resting state network levels were entered in a 1 (reaction time) × 3 (resting state network levels: low, mixed, and high activation) one-way ANOVA. Results showed a significant e ffect of resting state network type on reaction time ( *F*(2,47) = 8.06, *p* = 0.001). Additionally, the ANOVA resulted in a large e ffect size, ηp 2 = 0.26, indicating that 26% of the variance in reaction time is attributable to dominant resting state network activation. Dunnett's post hoc test revealed that visual mental image generation associated with the high strength level, reflecting TPN-Dominant activity, was significantly faster than visual mental image generation associated with both Mixed (*t*(47) = 3.936, *p* = 0.0003) and DMN-Dominant (*t*(47) = 2.299, *p* = 0.013) RSN activity levels (which did not di ffer between each other: *t* < 1). Additionally, we found equivalent results when testing whether imagery latency could be predicted by RSN activation, within and between each trial, by using a linear mixed regression model which assumed a normal distribution with an identity link (since this time our target was measured

on a continuous scale). These results confirmed that image generation latency is quicker when the relevant imagery is associated with high strength.
