Narrative thinking lingers in spontaneous thought

Human thought is history-dependent: current thoughts are shaped by prior context. Prior work demonstrates priming and models of drifting mental context across semantic, spatial, and emotional domains. The authors ask which kinds of mental content most strongly influence the trajectory of spontaneous thoughts. They hypothesize that deeper, elaborative processing—especially extracting situation-level meaning and building narrative/situation models or evaluating self-relevance—leads to lingering in spontaneous thought, beyond lexical or semantic priming. They predict that coherent narratives, social and emotional information, and instances where readers feel transported into a story world will more strongly persist and bias subsequent spontaneous thought. The study tests whether narrative coherence and the depth with which individuals engage (transportation) determine how much material lingers and shapes free associations minutes after reading.

The paper situates the question within theories of semantic priming and mental context models that account for history-dependence in memory, comprehension, and decision-making. It cites work on temporal context models, context drift/shift, and the influence of semantic, spatial, and emotional context. It notes that meaningful, real-world experiences (stories, social information, emotions) often intrude on thoughts and have enduring effects, suggesting mechanisms beyond simple priming. The authors connect their hypothesis to the levels-of-processing framework, arguing that deeper, meaning-centered processing (situation models, self-referential processing) yields more persistent mental contexts. Neurocognitive literature is referenced linking coherent narratives to pronounced history-dependence in higher-order association cortices (default mode network) and the role of situation models in discourse comprehension.

Overview: Multiple online experiments assessed whether and how narrative coherence and depth of processing cause lingering in spontaneous thought, using self-report and objective language-based measures from free word association.

Participants and stimuli:

• Experiment 1 (stories): After exclusions, N=720 (80 per condition per story). Stories: Raymond Carver’s “So Much Water So Close To Home” (original, replication, rewrite) and Miranda July’s “Roy Spivey”. Conditions: Intact (coherent narrative), Sentence-scrambled, Word-scrambled (Carver original only). All conditions presented the same words; only order differed.
• Experiment 2 (word list depth manipulation): After exclusions, N=320 (80 per condition: Italic, Tangible, Theme, Story). Stimulus: a fixed list of 268 words (≈75% related to Carver story gist; ≈25% positively valenced decoys), identical across conditions.
• Experiment 3 (neutral cue): Preregistered follow-up using a neutral free-association cue to rule out cue-related artifacts.
• Experiment 4 (depth manipulation with coherent stories): Preregistered follow-up manipulating shallow vs deep processing during reading of an intact story (proofreading vs emotional valence ratings), described in SI.

Procedure (Exp 1):

1. Math distractor; 2) Pre-reading free association (5 min); 3) Self-paced reading (one sentence at a time; 2,158–2,798 words; 196–268 sentences). Word-scrambled condition included intermittent yes/no probes to ensure attention to words. 4) Post-reading free association (5 min); 5) Theme generation (participants listed up to 10 theme words for the text); 6) Narrative transportation questionnaire (13-item, 7-point scale, scored as proportion); 7) Comprehension test (26 items; 24 content, 2 catch; 4AFC); 8) Demographics/strategy; 9) Self-reported lingering (1–7 scale) and open-ended responses.

Free association task: “Word chain game” with a cue word (story-related in Exp 1; neutral in Exp 3). Participants typed any words that came to mind for five minutes; each entry briefly appeared as the new cue then faded. Instructions discouraged sentences.

Depth manipulation (Exp 2): Incidental list learning with forced-choice decisions per word:

• Italic (shallow): decide if word is italicized.
• Tangible (moderate): decide if word refers to something concrete.
• Theme (deep): decide if word belongs to a hidden common theme (75% targets; decoys 25%).
• Story (deepest): decide if word belongs to a hidden story; early sequence emphasized characters/locations/events; decoys were positively valenced unrelated words. First 15 words were non-decoys and colored blue to scaffold theme/story extraction in Theme/Story conditions. Post-task measures: free association; theme generation; transportation (adapted for hidden story); story summary; free recall; comprehension test; demographics; lingering rating.

Objective analyses:

• Document classification: Linear SVM (R, caret) discriminating pre- vs post-story/list free association. Input: document-term matrix of word counts per chain, with within-participant mean-centering per word and across-participant SD scaling; leave-one-participant-out CV; 500 bootstraps. Null distribution via shuffling pre/post labels in test folds (500 permutations); one-sided permutation tests.
• Word embeddings and theme similarity: GloVe (Wikipedia2014+Gigaword, 300-D). Theme words: per story, the 10 most frequent participant-generated themes (excluding the held-out participant’s own themes). Theme similarity per associate = max cosine similarity to theme vectors; averaged per chain. Time-course analysis: sliding 10-word bins (~49 s per bin) estimating Cohen’s d for post-minus-pre theme similarity with 95% CIs.

Subjective measures:

• Narrative transportation (modified Green & Brock scale), reported as proportion (1 = maximal transportation).
• Self-reported lingering (1–7) and open-ended intentionality descriptions.

Controls and variants:

• Scrambling manipulation to limit situation-level meaning extraction (sentence- and word-level).
• Neutral cue variant (Exp 3) to rule out cue priming.
• Depth manipulation during intact story reading (Exp 4; SI) to reduce deep processing via proofreading vs increase via valence ratings.

Statistics: Appropriate parametric/nonparametric tests (t-tests, ANOVAs; Mann-Whitney U; Kruskal–Wallis; permutation tests). Feature selection for predicting lingering used backward stepwise linear regression with 10-fold CV; predictors: transportation, comprehension score, experimental condition (intact/scrambled). Numerical variables z-scored before modeling.

Manipulation checks (Exp 1): Scrambling reduced deep processing.

• Narrative transportation (proportion means): Word-scrambled=0.41; Sentence-scrambled=0.52; Intact=0.64. Intact vs Word: t(158)=11.74, p<0.001, d=1.86; Sentence vs Word: t(158)=4.89, p<0.001, d=0.77; Intact vs Sentence: t(158)=6.32, p<0.001, d=1.00.
• Comprehension (proportion correct): Word-scrambled=0.45; Sentence=0.67; Intact=0.83. Intact vs Word: t(158)=14.65, p<0.001, d=2.32; Sentence vs Word: t(158)=9.87, p<0.001, d=1.56; Intact vs Sentence: t(158)=6.33, p<0.001, d=1.00.

Self-reported lingering depends on coherence:

• Carver original: Kruskal–Wallis χ²(2)=56.12, p<0.001. Medians: Intact=5; Sentence=4; Word=2. Intact>Sentence: U=2207.5, p<0.001 (r_rb=0.31); Intact>Word: U=5259, p<0.001 (r_rb=0.64); Sentence>Word: U=4560.5, p<0.001 (r_rb=0.43). Lingering often involuntary: 51% unintentional; 7% intentional; 18% both; 24% neither/unsure.

Objective lingering (document classification of free association):

• Carver original: Intact 68% (p<0.002); Sentence 68% (p<0.002); Word 52% (p=0.340).
• Replications/different stories (n=80/condition each): • Carver-Replication: Intact 74% (p<0.002); Sentence 55% (p=0.130). • Carver-Rewrite: Intact 60% (p=0.036); Sentence 62% (p=0.004). • July: Intact 81% (p<0.002); Sentence 51% (p=0.420).

Bias toward story content post-story: Post–pre odds ratios for specific content were large (e.g., Carver: “murder” OR=6.3; “funeral” OR=12; July: “loss” OR=12; “four” OR=6.5; “spy” OR=6.5).

Theme similarity (GloVe) increases with coherence:

• Three-way ANOVA (Phase×Condition×Story): Phase×Condition F(1,632)=11.20, p<0.001, η²=0.007.
• Intact: Pre=0.260, Post=0.286; t(319)=8.42, p<0.001, d=0.56. Sentence: Pre=0.263, Post=0.275; t(319)=4.06, p<0.001, d=0.28. Word: Pre=0.272, Post=0.271; t(79)=-0.08, p=0.93.
• Theme similarity change correlated with self-reported lingering: r=0.25, t(718)=6.97, p<0.001.

Time course: Post–pre theme similarity remained elevated longest after intact stories: ~3 bins (~147 s) Intact; ~1 bin (~49 s) Sentence; 0 bins Word.

Transportation predicts lingering beyond objective coherence:

• Self-reported lingering model (cross-validated stepwise): R²=0.51; F(2,1437)=739.6, p<0.001. Transportation b=0.669, t=31.92, p<0.001; Comprehension b=0.085, t=4.04, p<0.001.
• Theme similarity change: final model included transportation only; F(1,1438)=5.14, p=0.024; b=0.06, t=2.3, p=0.024.

Depth-of-processing with word lists (Exp 2):

• Self-reported lingering: Kruskal–Wallis χ²(3)=78.57, p<0.001. Medians: Story=5; Theme=4.5; Tangible=3; Italic=2. Pairwise: Story vs Theme p=0.12; all other contrasts p<0.001 with moderate-to-large r_rb.
• Document classification: not above chance—Story 57% (p=0.056); Theme 54% (p=0.176); Tangible 56% (p=0.072); Italic 44% (p=0.842).
• Theme similarity ANOVA (Phase×Condition): F(3,316)=3.40, p=0.020. Paired t-tests: Theme Pre=0.270, Post=0.283; t(79)=4.06, p=0.009, d=0.32. Story Pre=0.278, Post=0.292; t(79)=2.51, p=0.014, d=0.32. Tangible and Italic ns.

Additional: A cover task encouraging shallow processing during intact reading (proofreading) reduced lingering (SI). A neutral free-association cue replicated lingering (SI).

Findings support the hypothesis that deeper, situation-level processing produces a lingering mental context that biases spontaneous thought for minutes after an experience. Coherent narratives reliably increased self-reported lingering and objectively altered the content and semantics of post-reading free associations, with effects persisting over roughly two to three dozen seconds and strongest immediately after reading. The degree of subjective transportation—a proxy for engagement in constructing a situation model—predicted both subjective lingering and semantic carryover beyond the objective coherence of the text and comprehension of verifiable details. Results align with hierarchical discourse comprehension models in which situation models represent deep meaning, and with neurocognitive evidence that higher-order association cortices (default mode network) exhibit history dependence during coherent narratives. The data suggest that lingering in spontaneous thought may be a behavioral correlate of robust encoding and possibly memory consolidation processes, with implications for how meaningful experiences shape subsequent cognition.

The study demonstrates that when people extract situation-level meaning—especially within coherent narratives—the experience is more likely to linger and shape subsequent spontaneous thought. Both subjective (self-reported lingering) and objective (document classification, theme similarity) measures show stronger and longer-lasting effects after intact stories than after scrambled sentences or words. Individual differences in narrative transportation best predict lingering, surpassing objective coherence and item-specific comprehension. Deep processing of non-narrative material (word lists) can also induce lingering, though effects are weaker. These findings advance a depth-of-processing account of lingering mental context and link spontaneous thought dynamics to narrative engagement. Future research should test causal links to memory consolidation, use richer linguistic and neural indices to quantify processing depth and lingering content, and examine how lingering influences real-world beliefs and behaviors.

Objective measures relied on word-level free association, which may not fully capture the richer mental representations (events, situations, emotions) that subjectively linger; correspondence between theme similarity and self-report was modest (r=0.25). The study lacks an objective neural index of processing depth, a long-standing limitation of levels-of-processing approaches. Future work could employ more unconstrained think-aloud paradigms, richer language embeddings grounded in agency, and brain-based measures (e.g., default mode network sensitivity to situation-level meaning) to index deep processing and improve alignment with subjective lingering.