Kurisutina
Library/Carrying on

Krämer et al. · 2024

Life events and life satisfaction: estimating effects of multiple life events in combined models

European Journal of Personality 39(1): 3–23 (online 2024), DOI 10.1177/08902070241231017, CC BY. Read as the accepted manuscript posted on PsyArXiv (wfg8h, CC BY 4.0). It carries the citation of the published version and the dates submitted 3 August 2023, resubmitted 11 December 2023, accepted 21 January 2024. The publisher PDF and the ZORA copy could not be fetched by script (Sage 403; ZORA anti-bot check). DIW/SOEP, FU Berlin, Zurich, Leipzig, Michigan State. Provenance: papers/carry_on/kramer2024_multiple_events.provenance.json.

What was read

All 2,525 lines of pdftotext -layout output (58 pages): title page, abstract, all sections, footnotes 1–3, Tables 1–2, figure notes and references. Most results are in Figures 2–4, which are images and were not inspected. The numbers below are those given in the text. The supplement (Table S1 literature overview, S2 eligibility rules; Figures S1–S10, including all model comparisons and the co-occurrence plots) and the OSF scripts were not read. The underlining in Table 1, which marks which repeated occurrences were modelled separately, is lost in the text extraction.

Question

Life events cluster and cause each other (partner, then cohabitation, marriage, a child). How should effects of each event on life satisfaction be estimated when others happen around it, and does it matter?

Method

  • Data. SOEP v37 (1984–2020). The main models cover 2007–2020 (death of a child was added in 2007): 40,121 people, 184,020 person-years. The abstract gives 41,402 event occurrences.
  • Events. 14 types, 30 events when repeated occurrences with at least 500 respondents are counted separately (e.g. second divorce).
    • Relationship events and childbirth come from biographical spell data, so first, second and later occurrences are known.
    • Unemployment counts only after two waves of not being registered as unemployed.
  • Outcome. Single-item life satisfaction, 0–10 (SD 1.80).
  • Model. Person fixed effects (within-person change only). Five time dummies: 2 and 1 years before; 1, 2 and 3+ years after. Each dummy is interacted with gender.
    • Controls: age, age², and a dummy for the first three panel years (initial elevation bias).
    • A non-event group restricted to people eligible for the event, e.g. parents only for the death of a child.
    • Panel-robust SEs; α = .01; the method was preregistered.
  • Five specifications per event.
    • Three individual-event models on progressively restricted samples (I1–I3).
    • C1, "total control": controls for all other events, including later ones.
    • C2, "past control": controls only for preceding and concurrent events, with later events recoded to zero. This is the favoured model.
  • The causal argument. Controlling for later events removes the part of the effect that runs through them (overcontrol) and can open collider paths. The pregnancy example: conditioning on a later childbirth makes the pregnancy coefficient compare no pregnancy with miscarriage. Not controlling for earlier events leaves confounding (undercontrol).

Results (as stated in the text)

  • Positive events. New partner, cohabitation, first marriage (rising before, peaking the year after, still positive after 2+ years) and childbirth. Childbirth effects are larger for mothers and for first births.
  • Negative events. Separation, unemployment (adaptation after one year for men, two for women), death of a partner and death of a child (the largest drops, in the year after).
  • Null or unclear. Divorce (only a small long-term rise for women; divorce here is distinct from separation), first job, retirement, a child moving out, death of father, death of mother.
  • Repeated events. Later childbirths have weaker effects; partnership, separation and cohabitation "generally" do not attenuate. Second marriage is positive only for women, in the year after. Repeated bereavement and divorce are too rare to compare.
  • Bias from other events.
    • Small for employment and bereavement events. Standardised differences: I3 against C2 mean 0.33; C1 against C2 mean 0.15.
    • Larger for relationship and fertility events (0.87 and 0.36).
    • Example, women's first separation (C2): −0.18 in the year before, −0.38 and −0.20 after. The individual model gives −0.08, −0.28 and −0.11: it underestimates, because separations often follow a new partner.
    • Controlling for later events makes separation look worse, because it conditions on staying single.
    • Bias is at most 13% of an SD of life satisfaction but up to 68% of an event's effect. It never flips a sign here. Its direction varies by event, so an individual-event model is neither an upper nor a lower bound.
  • Recommendation. Control for preceding events (the previous two years or earlier), not for later ones.

Limits

  • Average effects only. The authors note that individual differences exist, and that correlating them with third variables raises its own problems. "Scaling issues can introduce spurious effect heterogeneity, and life satisfaction scales tend to be skewed."
  • Time-varying confounders other than the listed events remain (e.g. a promotion). Only German data. The event list is limited by what SOEP asks.
  • Period effects: unemployment effects are milder in 2007–2020 than over all years. Divorce anticipation appears only in the all-years model.
  • Inconsistencies found:
    • The abstract's 41,402 event occurrences cannot be reconciled with Table 1. The final-sample totals sum to 43,711 (my computation). The difference may be occurrences not modelled separately, which the lost underlining would show.
    • Separation is described as "slightly more pronounced for the first event occurrence than for the second and third" in the results. The discussion says attenuation was "generally not the case for partnership, separation, and cohabitation".
    • Arithmetic checked: Table 1 full-sample rows sum to their totals (marriage 11,230, retirement 4,508). Unemployment sums to 13,070 against 13,071, with 6th-plus occurrences not shown.

What it means for Kurisutina

  • LISS T1: which other events to control for. This refines the note taken from Kettlewell et al. 2020.
    • When estimating a person's reaction to an event, control for events that preceded it (within about two years), not for later ones. Later events are part of what the event led to.
    • For the T1 consistency test this matters most for relationship events, where sequences are dense (partner → cohabitation → marriage → child; separation ↔ new partner). Job loss and widowhood are cleaner.
  • Design details worth copying for T1 (verified in the paper; adoption is a proposal):
    • an eligibility-restricted non-event group;
    • age and age²;
    • a first-years-of-participation dummy against initial elevation;
    • a persistence rule for unemployment (two prior waves without it; research/liss/events.py uses two months);
    • repeated occurrences coded by biographical order.
  • A trap for question 1. Skewed, bounded scales can create apparent individual differences in reactions (the authors' warning).
    • People near the top of the life-satisfaction scale cannot rise much, so their reactions to positive events are small every time. That produces spurious consistency across repeated events (inferred).
    • T1's population model conditions on pre-event answers, which absorbs much of this. The noise-floor comparison should also be run within pre-event level bands.
  • For the replica. The same event type does not have one effect: it depends on what came before (a separation after a new partner is not a separation after years alone). A replica's reaction must be conditioned on its recent event history, which the slot has to carry in order (inferred).

Cross-references

  • summaries/carry_on/kettlewell2020_life_events.md: the total-control approach this paper argues against.
  • summaries/carry_on/luhmann2014_its_about_time.md: prospective designs, controls, repeated events.
  • summaries/carry_on/luhmann2012_adaptation.md: average reactions.
  • summaries/carry_on/mancini2011_hedonic_treadmill.md, summaries/carry_on/infurna2016_resilience.md: individual differences the averages hide.
  • docs/research/liss_q1_design.md: T1; research/liss/events.py: event extraction and persistence rule.

This summary is our record of the paper, written after reading the full text and published as written; links into our own repository have been removed.