A pancultural perspective on the fading affect bias in autobiographical memory

Memory 23(2): 278–290 (2015; online 14 February 2014), DOI 10.1080/09658211.2014.884138. The published version was not read. Read as the author's accepted manuscript from the Aarhus University repository (37 pages). OpenAlex lists that copy as open access ("other-oa"); the PDF states no licence. Provenance: papers/carry_on/ritchie2015_fading_affect_pancultural.provenance.json.

What was read

All 1,749 lines of pdftotext -layout output: authors' notes, abstract, introduction, method, results, discussion, references and Tables 1–3. There are no figures.

Question

Is the fading affect bias (FAB) found outside the US? The FAB is the finding that the negative feeling attached to remembered events fades faster than the positive feeling.

Method

• Samples. Ten samples from six universities' research groups (Table 1):

  • African American, Caucasian American and Native American;
  • Ghanaian and German (the fall of the Berlin Wall);
  • New Zealanders of European, and of Māori and Pasifika, descent;
  • Caucasian British, Irish and Black British.

  562 participants and 2,439 events. Two datasets had been published before. A Puerto Rican study (Schrauf & Hoffman 2007) is re-described as a conceptual replication.

• Procedures varied by sample.

  • Recall 4–10 events (half positive, half negative), 5–10-year-old events (British sample), or one shared event (Germans).
  • In Samples 1–9, the affect "when it happened" (initial) and "when you remember it now" (current) are both rated in one session, so the initial rating is itself a memory.
  • Only Sample 10 (Black British) rated day-old events twice, two weeks apart, without seeing the first ratings.
  • Scales: −3 to +3 in Samples 1–7; 1–6 per valence in Samples 8–10.

• Analysis. Fading = |initial| − |current|, per event. A GLM gives the variance explained by a person dummy, then the added variance from event valence (ΔR²). Sample 10 uses a repeated-measures GLM.

Results

• A FAB in every sample (valence effect, all p ≤ .007). Negative affect faded by 0.73–2.12 scale points, positive by 0.07–0.84 (Table 2). Examples:
  • African American: negative 2.25 → 0.98, positive 2.51 → 1.81.
  • Black British, two weeks apart: 1.85 against 0.84.
  • Native American: positive affect barely faded (0.07).
• Valence explains about a third of fading in half the samples. ΔR² (Table 3):
  • .30–.34 for the African American, Ghanaian, Caucasian American and Native American samples;
  • .28 for European New Zealanders;
  • .13–.16 for the Caucasian British, Māori & Pasifika and Black British samples.
• Person explains much, and in one sample everything. The person dummy alone explains .10–.77 of the variance in fading.
  • In the Irish sample it explains .52, and valence then adds nothing (ΔR² .01, F(1, 292) = 0.00). The authors attribute this to which events people chose.
• Age did not account for the FAB in the seven samples that recorded it. The authors say it predicted fading itself in six of them; those results were not shown.
• From the literature, as stated in the paper:
  • the FAB appears within a day and persists for months;
  • it is weak or absent with dysphoria, anxiety, eating-disorder symptoms, narcissism, alcohol and drug use, and for regrets;
  • it arises in current-affect ratings, not initial ones;
  • it is linked to rehearsal, social disclosure and closure.

Limits

• Mostly retrospective. In Samples 1–9, "initial affect" is remembered initial affect, rated in the same session. Only Sample 10 measures change over time.
• Convenience and purposive samples, several very small (14 Ghanaians, 16 Māori and Pasifika). English only. Mostly students.
• Different recall instructions and scales across samples, so only effect sizes are comparable, not magnitudes.
• Event selection is uncontrolled, as the Irish result shows.
• Several supporting claims cite unpublished manuscripts (narcissism, eating disorders, anxiety).
• Inconsistencies found:

  • The per-sample participant counts in Table 1 sum to 560, not the stated total of 562 (also in the abstract).

  • The German sample is "180 … (86 women, 93 men)", which is 179.

  • Degrees of freedom do not match event counts for two samples:

    • Caucasian British: F(1, 236) with 244 events;
    • Irish: F(1, 391) with 395 events.

    Other samples have events − 2. Missing data are not reported.

  • Table 3 mixes metrics: Black British values are partial η², not R².

• Checked:
  • Table 2's faded affect equals initial − current for every cell, within rounding.
  • Table 3's ΔR² equals the difference of its two columns in every row.
  • Events sum to 2,439.

What it means for Kurisutina

• Q3: a direction for how the replica's feelings about its memories should age.
  • For most people, the bad feeling attached to an event fades faster than the good feeling, starting within a day. A replica recalling a stored event months later should, for most people, report the negative affect weakened more than the positive.
  • Test (proposal): log each experience with the person-appropriate initial affect at encoding. At intervals, ask the replica how remembering it feels now. Estimate the replica's FAB (negative minus positive fading) and compare it with the person's own FAB from paired diary ratings (Sample 10's design: rate, then re-rate blind two weeks later).
• The FAB is a person trait, so a universal FAB would be wrong for some people.
  • It is disrupted in dysphoria and anxiety, and the person term explains up to three quarters of the variance in fading.
  • A replica of someone with depressive symptoms should show little or no FAB. "React like Alice" applies to memory too: fading asymmetry must be calibrated per person, e.g. from the Sample 10 design with the person.
• Where the replica's default could go wrong (inferred).
  • LLM condensation keeps negative and threat content preferentially (Acerbi). A replica that consolidates memories by summarising could therefore show the opposite of a FAB.
  • MemoryBank and Mem0 have no affect in their forgetting rules at all.
  • Either way, the base model's processing would set the fading, not the person's. This is a concrete Q3 check.
• Store initial affect; don't re-derive it. The FAB is in current ratings, not initial ones.
  • The replica should keep the affect recorded at encoding as a fixed field. Its current feeling about the event should then change by a person-calibrated rule.
  • The alternative is to have the base model re-appraise the text each time, which imports its own affect (proposal).
• Retrospective ratings are a present account. Most samples rated "then" and "now" in one sitting. This matches Luhmann's note that remembered trajectories are present accounts. Human baselines for Q3 should use prospective or test–retest designs.

Cross-references

• summaries/carry_on/acerbi2023_llm_transmission.md: LLMs keep negative content when condensing.
• summaries/carry_on/zhong2023_memorybank.md, summaries/carry_on/chhikara2025_mem0.md: forgetting without affect.
• summaries/carry_on/haehner_perception_stability.md: re-appraisal of events over a year.
• summaries/carry_on/diamond2020_real_world_recall.md, summaries/carry_on/hirst2009_911_memory.md: other Q3 baselines.