Resilience to major life stressors is not as common as thought

Perspectives on Psychological Science 11(2): 175–194, DOI 10.1177/1745691615621271. Read as the NIH author manuscript (PMC4800830; PMC text-mining permitted, not open access; the XML carries no copyright statement). Arizona State University; NIH grants R01AG048844 and R01DA014385.

Provenance: papers/carry_on/infurna2016_resilience.provenance.json.

What was read

• Read: all 1,029 lines of the text extracted from the PMC XML. That covers the abstract, introduction, method, results, discussion, the three footnotes, the author contributions, the acknowledgements, the full reference list (56 entries) and Tables 1–6 with their notes. Figures 1–4 are images; their captions were read.
• Not read: the online supplement, a PDF with Figures S1–S5 showing how each model is set up.
  • The PMC link serves a JavaScript download page to scripted requests.
  • Europe PMC does not serve supplements for this record, which is not open access.

Question

Is "most people are resilient" (stable, high life satisfaction around a major negative event) a finding about people, or about the model used to find it?

• Earlier papers reported resilience rates from growth mixture models on the German Socio-Economic Panel (SOEP): 58.7% for spousal loss, 71.8% for divorce and 69.0% for unemployment (Galatzer-Levy, Bonanno & Mancini 2010; Mancini, Bonanno & Clark 2011).
• This paper re-runs those models on the same panel, first as published, then with two constraints relaxed.

Method

• Data. SOEP, 28 annual waves (1984–2011); the originals used 20 (1984–2003).

  • A nationally representative panel of private households in Germany, about 50,000 residents.
  • Initial response 60–70%, attrition 3–10%.

• Outcome. One item: "How satisfied are you with your life, all things considered?", from 0 (totally unsatisfied) to 10 (totally satisfied).

• Events. All three are self-reported and aligned on the year of the event. Observations run from 5 years before to 5 years after, up to 11 per person.

  Event	n (originals)	Definition	Age at event, mean (SD)	Women	Life-satisfaction observations, mean
  Spousal loss	1,214 (464)	"spouse/partner has died"; people over 75 excluded	61 (11.09)	74%	8.40
  Divorce	1,579 (629)	"divorce"; people over 75 excluded	39.49 (9.10)	55%	8.19
  Unemployment	2,461 (774)	registered unemployed after three waves in full-time work; aged 21–60; in the study at least four years afterwards	42.38 (12.36)	44%	9.59

• Growth mixture models (Mplus, full-information maximum likelihood), each fitted with 1 to 5 classes, in three specifications. Figure 1 illustrates the difference.

  1. Part 1 (A), as published.
     • Growth model: a level and a latent-basis slope for spousal loss and divorce; three piecewise slopes for unemployment (the years before, around the event, and after).
     • Classes differ only in their means. The level variance and the residual variance are equal across classes.
     • Slope variances are fixed at 0, so everyone in a class has the same trajectory, shifted up or down.
  2. Part 2 (B). The same growth models, but slope variances and covariances are estimated and may differ between classes.
  3. Part 3 (C), multi-phase.
     • Three factors: the level, the total change from year −5 to the event year, and the total change from the event year to +5.
     • Each factor has a mean and a variance in every class.

• Choosing the number of classes.

  • Criteria: BIC; entropy (above .7 means distinct classes); two likelihood-ratio tests of k against k − 1 classes (Vuong–Lo–Mendell–Rubin and adjusted Lo–Mendell–Rubin); interpretability.
  • The chosen solution is then plotted with 95% intervals.

Results

Share of the "resilient" class. Computed as class size in the tables divided by n; the text's figure is noted where it differs.

• Part 1 does not reproduce the published class counts. It finds 3 classes for spousal loss (published: 4), 2 for divorce (3) and 3 for unemployment (4).
  • In each case BIC is lower with more classes, but the likelihood-ratio tests for the next class are not significant:
    • spousal loss, 4 vs 3: p = .118 / .126;
    • divorce, 3 vs 2: p = .130 / .141;
    • unemployment, 4 vs 3: p = .440 / .446.
  • The classes:
    • spousal loss: resilient 75%, recovery 20%, "improvers" 5% (life satisfaction rising before the loss);
    • divorce: resilient 85%, recovery 15%;
    • unemployment: resilient 81%, recovery 15%, and 4% improving after job loss.
  • The classes' 95% intervals overlap substantially.
  • Footnote 1: re-running with the originals' number of waves gave similar results.
• Freeing the variances changes the picture.
  • BIC improves (footnote 2). One-class BIC, Part 1 → Part 2: 39,517 → 39,292 (spousal loss), 49,976 → 49,788 (divorce), 89,826 → 88,949 (unemployment).
  • Spousal loss: entropy is .644 or lower for 2–5 classes, and the authors take one class as best. The 3-class solution, shown for comparison, has 27% resilient.
  • Divorce: 2 classes, 36% resilient. The authors chose 2 over 3 on entropy (.693 against .599), although BIC and both tests favour 3.
  • Unemployment: 2 classes. 47% are resilient; 53% show a sharp drop in the year around job loss and lower levels afterwards. The authors add that there may be only one class.
• The multi-phase model finds at most two classes, and the resilient class is the smaller one in each.
  • Spousal loss: entropy .603 or lower, and one class may be best.
  • Divorce: 36% resilient against 64%.
  • Unemployment: 47% resilient against 53%.
  • The one-class multi-phase model already fits well: CFI 0.923 / 0.954 / 0.978, RMSEA 0.068 / 0.046 / 0.037 (spousal loss / divorce / unemployment). The one-class Part 1 models fit poorly: CFI 0.836–0.856, RMSEA 0.086–0.092.
• What separates the classes in Parts 2–3 (Tables 2–4):
  • The resilient class starts higher. Part 3 levels: spousal loss 7.31 vs 6.77, divorce 7.53 vs 6.34, unemployment 7.23 vs 6.35.
  • It fluctuates less from year to year: residual variance 0.59–0.75, against 1.78–4.48 in the other classes.
  • Its members' change varies less. In Part 3 the variance of pre-event change is 1.00, 1.08 and 2.10 (divorce, unemployment, spousal loss), against 2.48, 2.04 and 3.33.
  • It is not flat. In Part 3 it too declines towards the event and rises afterwards:
    • spousal loss: −0.99, then +0.81;
    • unemployment: −0.71, then +0.49;
    • divorce: −0.18 (not significant), then +0.20.
• Within a class, individuals differ about as much as the classes do. This is our computation from the Part 3 variances and covariances, in points on the 0–10 scale.
  • The SD of pre-event change is 1.00–1.82, against class means of −0.18 to −1.61.
  • The SD of post-event change is 0.84–1.41, against class means of +0.20 to +1.36.
  • Pre- and post-event change correlate −0.49 to −0.79: whoever falls further rises further.
  • Level and pre-event change correlate −0.15 to −0.55: those who start higher fall further.
• The authors' conclusions.
  • The resilient and recovery trajectories lie within each other's 95% intervals for all three events, so they may not be distinct. The authors endorse the developmental view that recovery is itself a form of resilience.
  • Growth mixture modelling is exploratory, and "rates of resilience" from one sample and one measure cannot establish population rates.
  • Resilience differs by domain:
    • unemployment (Infurna, Wiest & Luthar, then under review): 61% resilient on life satisfaction, 48% on negative affect, 20% on positive affect;
    • disability (Infurna & Wiest): 15% resilient on life satisfaction, and no resilient class at all on self-rated health.
  • The authors "strenuously argue" against declaring how many people are resilient, and against withholding interventions on that basis.

Limits

• The authors' limits.
  • One data set and one single-item measure. Footnote 3: with affect, depression or substance use as criteria, the resilient share would likely be lower still.
  • The results depend on the model specification.
  • The 2-class unemployment solution may not replicate.
• Descriptive only (our assessment). There are no covariates and no personality measures, and nothing on who ends up in which class or whose change is larger. The paper sizes the heterogeneity that question 1 asks about; it does not predict it.
• The class count rests on different criteria in different parts.
  • Part 1 lets the likelihood-ratio tests overrule BIC. Parts 2–3 let entropy overrule both BIC and significant tests.
  • Entropy measures how cleanly people are classified, not how well the model fits.
  • For spousal loss in Part 2, the paper says "the 1-class solution provided the best fit" although BIC fell by 802 from one class to two and both tests gave p < .001.
  • Many solutions needed variances or covariances fixed at 0 because estimated correlations exceeded ±1 (notes to Tables 5–6).
• Part of the heterogeneity is probably noise (inferred).
  • The classes differ most in residual variance, by a factor of 3–8 (mostly about 4).
  • A noisy reading in the event year alone would produce a negative correlation between pre- and post-event change.
  • The outcome is a single 0–10 item.
• Part 1 is not an exact replication. It covers 11 years around the event against 8 or 9, and more cases. The originals did not report fit statistics, so their choice of class number cannot be checked (the authors' point).
• Inconsistencies in the paper.
  • Spousal loss, Part 3:
    • The results text gives 36% resilient and 64% recovery.
    • The class sizes in Tables 2 and 6 (574.88 and 639.12 of 1,214) give 47% and 53%, and the discussion uses 47%.
    • The sentence looks copied from the divorce result.
  • Table 2, Part 1: the headers of the first two columns appear swapped (inferred).
    • The column headed "Resilient" has 245.40 people (20%) and the steepest slope: −1.25, about −4 points in the event year with the note's loading of 3.16.
    • The column headed "Recovery" has 907.00 (75%) and a slope of −0.23.
    • The text says 75% are resilient.
  • "Resilience represented the least common trajectory" (abstract and discussion):
    • In the 3-class Part 2 solution for spousal loss, the third class (19%) is smaller than the resilient one (27%).
    • The claim holds for the two-class solutions.
  • Unemployment:
    • Part 1 describes "two smaller classes of equal size" that are 15% and 4%.
    • Table 4 gives 370.61 for the recovery class and Table 1 gives 370.26; only the latter sums to 2,461.
    • Part 2 refers to "Part B in Figure 3" (divorce) where Figure 4B is meant.
  • The new case counts are called "percentages … somewhat higher" than the originals', although they are 2.5–3.2 times as large.
  • Figure 1's illustrative intervals are given as 1.5 versus 2 in one place and 1.5 versus 2.5 elsewhere.
  • Typos in the tables:
    • a loading of "4" for −4 (Table 4 note, Part 2);
    • "slope 2 and slope 2" as the label of a covariance;
    • a double asterisk;
    • a missing asterisk on −0.22 (0.05).

What it means for Kurisutina

• Question 1: people differ a lot in how they react to the same event, and continuously rather than by type.
  • Within a class, the SD of a person's change around the event is about 1–1.8 points on a 0–10 scale. That is the same order as the mean change.
  • How many people count as "resilient" depends on the model: 75–85% or 27–47% from the same data.
  • A slot therefore should not store a reaction type. If it stores anything about reaction, it is the person's own level, dip and recovery.
  • The paper shows that these vary. It does not show that anything measured beforehand predicts them; Yap et al. found that traits largely do not.
• Part of what marks the "resilient" is visible before the event: a higher level and less fluctuation.
  • The level is in every pilot condition (part 2 of the prompt) and in the population table.
  • Fluctuation cannot be seen from one earlier interview.
  • The feasibility check found the same thing: a person's mean absolute change carries over between intervals (rank correlation 0.31). The same caveat applies: answer noise produces it too.
  • Several interviews would let a slot carry a person's volatility. A later GSS run could give the 2→3 pairs both earlier waves; the frozen pilot does not.
• Coverage of the pilot's strata.
  • Covered: spousal loss, divorce and unemployment correspond to widowed, divorced or separated, and lost job.
  • Not covered: married and found job.
  • The outcome is global life satisfaction; neither happiness nor financial satisfaction was measured.
• The decline often starts before the event. In Part 3 this holds even for the resilient class after spousal loss and job loss.
  • For the widowed and divorced-or-separated strata, the earlier GSS interview may fall inside that decline.
  • A replica loaded from that interview starts from someone who is already changing (as in luhmann2012_adaptation).
• Score change continuously, per person and item, never as class membership.
  • X1 (changed against unchanged answers) has the right shape.
  • Changed answers will over-represent high-fluctuation people. So a gain on X1 alone does not show that the replica knows the person's reaction.
• Affect against evaluation. The unemployment example fits the X2 expectation that happiness shows more varied reactions than financial satisfaction. The figures (61% resilient on life satisfaction, 48% on negative affect, 20% on positive affect) come from a manuscript that was under review, cited second-hand and not checked.
• The pilot cannot estimate this heterogeneity. It has 40 pairs per stratum and two waves per pair. It can only test whether a slot predicts a person's deviation from the population table; it cannot model individual trajectories.

Cross-references

• summaries/carry_on/yap2012_personality_moderation.md: the same question with trait predictors; no consistent moderation.
• summaries/carry_on/luhmann2012_adaptation.md: average reactions to these events. The level before an event often already reflects it, and effect sizes vary more for affect than for life evaluations.
• summaries/carry_on/hout2016_gss_reliability.md: happy has reliability 0.60. Accurately reported volatility, as in unemployment, is read as unreliability.
• summaries/carry_on/peng2025_funhouse.md: twins built from rich personal data are under-individuated.
• summaries/carry_on/eckstein2022_context.md: a person's way of updating is local to a context.
• docs/research/carry_on_goal.md, "(b) Results": change carries over between intervals (rank correlation 0.31).
• docs/research/gss_pilot_design.md: extra analyses X1, X2 and X3.