Population- and individual-level changes in life satisfaction surrounding major life stressors

Social Psychological and Personality Science (online first, © 2017; the PDF read carries online-first pagination 1–10; issue details not checked), DOI 10.1177/1948550617727589. University of Pennsylvania and Columbia University. Read as the published PDF posted on the first author's own website (brucedore.github.io; linked there as "accepted"). The online supplement was not read. Provenance: papers/carry_on/dore2018_population_individual.provenance.json.

What was read

All 805 lines of pdftotext output in reading order (two-column layout): abstract, introduction, method, results, discussion, limitations, references and author notes. Table 1 was read from the text. Figures 1–5 are images; their captions were read, but the figures were not rendered.

Question

What is the typical life-satisfaction trajectory around divorce, spousal loss and unemployment when its shape is not fixed in advance? Do people differ in continuous ways, or fall into discrete classes such as resilient, recovering and chronic?

Method

• Data. SOEP, 1984–2011. First occurrence of each event, following earlier studies' inclusion rules.

  Event	n	Mean age	Women	Mean annual reports
  Spousal loss	1,214	61	74%	23.9
  Divorce	1,579	40	55%	14.6
  Unemployment	1,800	42	44%	17.4

  • Unemployment: registered unemployed after 3 years of full-time work, age 21–60, at least 4 later waves. The 1,800 are a random subsample of 2,461, because of memory limits.
  • Everyone was under 75 at divorce or loss.
  • All waves were modelled, not a window around the event.

• Outcome. Life satisfaction, 0–10. Time is years from the wave reporting the event.

• Models. Generalized additive mixed models (mgcv), with a data-driven smooth of time. Four nested versions:

  1. fixed curve only;
  2. • person intercepts;
  3. • person slopes;
  4. • person-specific curves.

• Per-person parameters. Intercept, slope and "typicality": the Fisher-z correlation of the person's curve with the population curve, for people with 10 or more reports.

• Tests for classes. Hartigan's dip test for unimodality, and normal-mixture models (1–10 components, bootstrap likelihood-ratio tests).

Results

• Person-specific curves fit best for all three events. Deviance explained:

  Model	Deviance explained
  Fixed curve	1.2–2.9%
  + intercepts	42–45%
  + slopes	49–52%
  + person curves	56–60%

  How the curve bends, not only its level and slope, differs between people.

• Population curves.

  • Life satisfaction declines before the reported event: for divorce up to about 6 years before, reaching a minimum of 6.27 about 9 months before.
  • It is lowest around the event and recovers only partly, stabilising within 2–4 years below the earlier baseline:

  Event	Before	4 years after
  Divorce	6.91 (6 years before)	6.65
  Spousal loss	6.98 (4 years before)	6.62
  Unemployment	6.67 (4 years before)	6.28

  • Spousal loss: drop 1.14, recovery 0.88.
  • Unemployment: drop 0.82, of which about half (0.44) recovered. It stays below the other two curves for more than 10 years.

• Individual differences are continuous and heavy-tailed.

  • Intercepts are left-skewed (−.41 to −.65; ceiling).
  • Slopes have kurtosis 4.20–4.81: more extreme risers and fallers than a normal distribution would give.
  • Typicality averages r = .60–.73 but varies widely (SD .71–.90 in z). Many people's curves are uncorrelated with, or opposite to, the population curve.
  • No multimodality (dip p = .81–1.00).
  • Mixtures choose 1–2 components, never more. For slopes, two normals share a mean and differ in variance: a low-variance and a high-variance group, not "resilient" versus "chronic" classes.
  • The three per-person parameters are nearly uncorrelated, under 2% shared variance, except slope and typicality for unemployment (about 19%).

Limits

• Three events, one outcome (a single life-satisfaction item), SOEP only, and observational data.
• The late-life decline after loss and divorce may be age (acknowledged).
• Typicality is computed over all waves, up to 28 years, so it mixes the event response with age and period trends.
• Unemployment is a random subsample.
• Inconsistencies found:
  • Table 1 swaps the n labels. Method gives spousal loss n = 1,214 and divorce n = 1,579. Table 1 heads its blocks "Divorce (n = 1,214)" and "Spousal loss (n = 1,579)".
    • Weak evidence (my computation) suggests the event labels are what is swapped: the n = 1,214 block has the larger AIC, fitting spousal loss's larger number of reports (1,214 × 23.9 ≈ 29,000 against 1,579 × 14.6 ≈ 23,000).
    • Which of Table 1's statistics belong to which event is therefore uncertain.
  • Spousal-loss arithmetic. A drop of 1.14 with recovery of 0.88 leaves 0.26 below baseline. The quoted levels (6.98 four years before, 6.62 four years after) differ by 0.36.
  • Scale. The text says "on the 1–10 scale"; the measure is 0–10.
  • Degree of recovery after spousal loss. Results: it "did not recover to prestressor baseline", with non-overlapping CIs. Discussion: "near complete recovery within 4 years".
• Checked:
  • The deviance-explained summaries (under 3%, ~44%, ~51%, ~58%) match Table 1.
  • Unemployment: 6.67 − 6.28 = 0.39 ≈ 0.82 − 0.44.
  • The percentages and ns in the method are internally consistent.

What it means for Kurisutina

• Q1: a person's reaction is a continuous deviation from the typical curve for that event, not a type.
  • No discrete resilient, recovering or chronic classes were found. Differences are continuous, heavy-tailed, and in shape as well as level (verified).
  • A replica slot should hold her own curve parameters, or the evidence for them, relative to the event's population curve, not a class label (inferred).
  • "Reacting like the average person with her views" is exactly the population curve. Many people are uncorrelated with it (verified).
• It matches the LISS T1 definition of a reaction. T1 defines a reaction as change against what the population model predicts for the same starting point, which is Doré & Bolger's "resilience relative to what is typical". Two proposals for docs/research/liss_q1_design.md:
  • Robust consistency. Slopes are heavy-tailed, so a few extreme reactors could drive a first-to-second-reaction correlation. Declare a rank correlation beside the Pearson one before loading outcomes.
  • Baseline timing by event. Divorce shows decline up to about 6 years before it is reported. "A baseline well before the event" must be event-specific, or anticipation modelled. For divorce a 1–2 year baseline already includes the decline.
• For the GSS pilot. Its lost-job stratum sits on the event with the slowest and least complete recovery (still below after 10 years). So persistent lower well-being after job loss is the population expectation, not evidence of a person-specific reaction (inferred). The pilot's population tables already carry it, which is why D-contrasts against the table are right.

Cross-references

• summaries/carry_on/infurna2016_resilience.md: the growth-mixture analyses this paper answers.
• summaries/carry_on/luhmann2012_adaptation.md: average adaptation by event.
• summaries/carry_on/luhmann2014_its_about_time.md: anticipation and timing.
• summaries/carry_on/haehner2024_event_characteristics_swb.md: log-time recovery and event characteristics.
• docs/research/liss_q1_design.md: T1 reaction definition and timing notes.