Individual differences in changes in subjective well-being: the role of event characteristics after negative life events

Journal of Personality and Social Psychology, DOI 10.1037/pspp0000511. The published version was not read. Read as the accepted manuscript on PsyArXiv (3mjr7, version 1, posted 7 July 2023; the PDF read was created 14 June 2024 and is marked "not the copy of record", © APA). OSF licence record: CC BY 4.0. Ruhr University Bochum and University of Zurich. Data, scripts and supplement at osf.io/kzuwe and osf.io/529u4 (not read). Provenance: papers/carry_on/haehner2024_event_characteristics_swb.provenance.json.

What was read

All 3,284 lines of pdftotext -layout output (79 pages): abstract, introduction, Tables 1–9, method, results, discussion, footnotes and references. Figures 1–5 were rendered and inspected (the annotated specification curve, the specification scheme, the R² plot for death of a loved one, the mean trajectories and the separation simple slopes). The supplement, the ShinyApp, the preregistrations and the data were not read.

Question

After a negative life event, do the event's characteristics explain why people's well-being changes differently?

• How do perceived characteristics (the person's appraisal) relate to objective-descriptive ones (what happened)?
• What shape does the average trajectory have?
• Which characteristics go with well-being right after the event, and with its change over the following months?
• Which analysis choices drive those associations?

Method

• Sample. The Post-Event Changes Study: Germany, online, 2021–2022.

  • People were recruited within 5 weeks of one of five events, e.g. through event-related online forums.
  • They were assessed at 0, 4, 8, 16 and 24 weeks after registration.
  • 1,673 registered. The analysed N was 1,068 at T1, falling to 753, 686, 611 and 577.
  • 74% were women, the mean age was 29.1 (SD 9.2), and 51% were students.

  Event	N at T1	N at T5
  Separation	338	179
  Death of a loved one	307	179
  Friendship dissolution	237	126
  Failing an important exam	106	57
  Job loss	80	36

• Attrition. Dropouts had slightly lower life satisfaction (d = −0.13), were less conscientious (−0.19), more neurotic (0.17) and younger (−0.18).

• Power. Separation, death and friendship dissolution were powered for d = 0.41 over 6 months. Job loss and exam failure were powered only for d = 0.64.

• Measures.

  • Appraisal: the Event Characteristics Questionnaire (38 items, nine dimensions, 1–5), rated once at T1. The dimensions are challenge, change in world views, emotional significance, external control, extraordinariness, impact, predictability, social status change and valence.
  • Objective-descriptive characteristics: 3–10 per event (Table 4). Examples are relationship to the deceased and cause of death, initiation and joint children for separation, looking for work for job loss, and final attempt for the exam.
  • Well-being at every wave: the SWLS for life satisfaction, and SPANE positive and negative affect for the past 2 weeks.
  • Covariates: Big Five (BFI-2-XS), demographics and Covid-19 stress.

• Analysis.

  • Multilevel growth models, with time since the event coded either linearly (months) or as log(1 + months).
  • The quantity of interest is the cross-level interaction of characteristic × time, which is individual differences in change. The main effect is the level extrapolated to the event date.
  • Each event × characteristic × time coding got its own specification curve: 146 curves.
  • Five decisions were crossed (Figure 2), giving 384–6,144 specifications per curve:
    • the outcome (3 choices);
    • whether the other eight appraisal dimensions were included (2);
    • which objective characteristics were included (4–64);
    • covariate blocks (8);
    • random slopes (2).
  • Joint tests used 1,000 permutations of the focal characteristic, applied to the median effect, the share significant and Stouffer's z, at Bonferroni α = .0167. An effect is reported if at least 2 of the 3 tests pass.
  • The runs took about 8 months on a 20-core PC.

Results

• Appraisal is partly anchored in the facts.

  • Objective characteristics jointly explained this share of variance in the appraisal dimensions:

    Event	Variance explained
    Death of a loved one	13–40%
    Separation	4–29%
    Friendship dissolution	7–14%
    Job loss	7–28%
    Exam failure	9–34%

    The average is about 17%.

  • Examples:

    • age of the deceased with extraordinariness, r = −.51;
    • contact before the death with impact, r = .52;
    • a non-natural death with external control (.38), extraordinariness (.36) and predictability (−.37);
    • who initiated a separation explained 28% of perceived predictability and 16% of valence.
• Trajectory.
  • Log time fitted better than linear time (AIC and BIC) for all five events: change is fast early and slows.
  • After death, separation and exam failure, affect rose mainly in the first 8–12 weeks.
  • Life satisfaction moved least (Figure 4). For example, after a separation, reversed negative affect went from about −0.6 to +0.6 SD, and life satisfaction from about −0.15 to +0.35.
  • Friendship dissolution and job loss showed small, mixed changes.
• Level at the event: appraisal matters a lot.
  • Seeing the event as more challenging, impactful, emotionally significant or status-changing went with lower well-being.
  • For example, after a death, well-being was 0.26 SD lower (0.32 with log time) per SD of challenge.
  • Among objective facts, the death of a partner was −0.63 SD (−0.76 with log time).
• Change over the months: few, event-specific effects.
  • Death: more challenge (+0.11, log only) and more contact before the death (+0.05, log only) went with faster recovery.
  • Separation: seeing it as more predictable (−0.03 linear, −0.14 log) or more positive (−0.04, −0.15) went with slower recovery. Joint children also went with slower recovery (−0.14, linear only). Figure 5 shows that people with joint children start about 0.18 SD higher but barely rise.
  • Friendship dissolution: when the friend ended it, recovery was faster (+0.05, +0.18).
  • Job loss: looking for new work went with faster recovery (+0.17, +0.52).
  • Exam failure: a failed final attempt went with less recovery or a decline (−0.16, −0.66).
  • The authors read the appraisal effects as "more room to recover": those hit harder rise more toward their (unobserved) pre-event level.
• Analysis choices matter. For cross-level effects after a death, the choice of outcome explained 49% of their variance and the other event characteristics included explained 33%. Covariates explained 2% and random slopes under 1%.
  • Including the other appraisal dimensions generally shrank an effect. After a separation, challenge predicted change only when the other dimensions were left out.
  • Which outcome carried an effect varied: status change mattered more for life satisfaction, and challenge more for positive affect.

Limits

• No pre-event measure and no control group. The authors state that neither adaptation nor causal direction can be judged, and that appraisal main effects are "not causal". Appraisal was measured after the event and only once.
• The sample is young, 74% women, highly educated and self-selected through forums. Job loss and exam failure are underpowered, and dropout is selective.
• Only 6 months of follow-up, and the time-zero level is an extrapolation.
• Inconsistencies found:
  • Duplicated rows in Table 8. "Job type: Mini job or other" and "Reason: Resigned/no extension" have identical values in all 12 columns for both time codings. One of the two rows is not its own result (probably a copy error). The results for reason for job loss are therefore unknown from this version.
  • Percentages in the Summary of results. The text says 40%, 9%, 9% and 9% of main effects and cross-level interactions were significant.
    • These match the share of individual joint tests passing: 108/270, 26/288, 23/270 and 25/288.
    • They do not match the paper's own reporting rule of ≥ 2 of 3 tests. Under that rule the shares are:
      • perceived main effects: 36/90 = 40%;
      • objective main effects: 7/96 = 7%;
      • perceived cross-level effects: 5/90 = 6%;
      • objective cross-level effects: 8/96 = 8%.
    • These are my counts from parsing Tables 5–9. That the text counts individual tests is inferred from the match.
  • Friend-initiated friendship dissolution, main effect with linear time. It is reported as significant, but at printed precision it passes 1 of 3 tests (p = .003, .017, .022 against .0167). It is borderline on rounding.
  • Exam final attempt. The text writes "Median γ11_lin = −0.16, Median γ11_lin = −0.66". The second value is the log-time estimate (Table 9).
  • Number of objective characteristics. Figure 2 says the number included per event "varied between 3 and 6" (4–64 combinations). Table 4 lists 10 for separation. How that set was reduced is not in the main text.
• Checked:
  • Table 3's event Ns sum to the wave totals (1,068 / 753 / 686 / 611 / 577).
  • 3 × 2 × {4…64} × 8 × 2 = 384–6,144.
  • Figure 3's single-predictor R² values equal the squared correlations in the text (.26, .27, .14, .13, .14, .07), and its combined R² spans .13–.40.
  • Applying the 2-of-3 rule to Tables 5–9 reproduces every cross-level effect named in the text (5 perceived, 8 objective) and every main effect except the borderline one above.

What it means for Kurisutina

• Q1: the person's appraisal decides how deep the dip is, not so much the path back.
  • Appraisal dimensions carry large differences in well-being at the event, with 40% of dimension × event × coding cells passing.
  • Only 5 of 90 cells predict change over six months, in two events (my count). Those few mostly say that people hit harder recover more.
  • So for a replica, "reacting like Alice" to a new negative event means two things (inferred):
    • the immediate hit should scale with how she appraises it;
    • the recovery should follow the common fast-then-slow course, unless her situation has one of the few facts that predicted slower or faster recovery.
• Appraisal is mostly the person, not the event. Objective facts explain about 17% of appraisal on average. The rest is the person, confounders and noise.
  • A base model given only "separation" will supply an average appraisal: the "average person with her views" failure.
  • The slot should store her appraisal of each past event, since appraisal is what carries the level effect.
  • With a participant, it should also collect her ECQ-style ratings of new events, to test whether the replica appraises as she does.
  • This agrees with luhmann2021_ecq_taxonomy, haehner2022_event_perception and bruning2022_separation_initiators.
• Specific facts beat event categories. The change effects that survived are concrete facts: joint children, who ended the friendship, looking for work, final attempt, and a partner's death for the level.
  • Event-occurrence coding (GSS strata, LISS event flags) cannot see these.
  • For LISS T1, where household or background variables record such facts (for example children in the household at a separation), they are candidate moderators to declare before any outcome is loaded (proposal).
• LISS T1 timing and outcomes (proposals for docs/research/liss_q1_design.md).
  • Enter months since the event as log(1 + months), or declare both codings in advance. Log time won for every event here, and some effects appeared under only one coding.
  • Report consistency separately for life satisfaction, happiness and mood. The outcome was the largest single driver of effect sizes here, and life satisfaction moved least.
  • "Hit harder, recover more" means a reaction measured at one post-event wave depends on when that wave falls. T1's residual reaction must use the same time since the event for both events, or model it.
• GSS pilot (inferred). Waves are two years apart, well past the 8–12-week fast phase seen here. So the pilot's event strata measure what remains after adaptation, not the reaction.
• Declared analyses guard against forking paths. Here, which other appraisals were controlled and which outcome was chosen moved effects more than covariates did. The pilot's pre-declared D, E, X and Y sets, and the plan to hash T1 before loading outcomes, are the right response. This paper is the cost of not doing that.

Cross-references

• summaries/carry_on/luhmann2021_ecq_taxonomy.md: the ECQ and the prospective valence effect.
• summaries/carry_on/haehner2023_perception_change_swb.md: appraisals change, and change with well-being.
• summaries/carry_on/haehner2022_event_perception.md: appraisal and personality change.
• summaries/carry_on/bruning2022_separation_initiators.md: initiator and non-initiator trajectories.
• summaries/carry_on/luhmann2012_adaptation.md: average adaptation by event.
• summaries/carry_on/luhmann2014_its_about_time.md: timing of change.
• docs/research/liss_q1_design.md: the T1 and T2 design.