Change in personal culture over the life course

American Sociological Review 88(2): 220–251 (2023), DOI 10.1177/00031224231156456, CC BY 4.0.

• Read as the version of record: the Sage typeset PDF, 32 pages, as deposited in Humboldt University's edoc repository. The record is 18452/27281 (repository DOI 10.18452/26584), labelled "publishedVersion" and delivered by DeepGreen.
  • OpenAlex labels this copy a submitted version; the file itself is the publisher's typeset article.
  • The deposit package holds a second copy of the PDF whose extracted text is identical.
• The publisher's own PDF returned a Cloudflare bot check (HTTP 403) and was not tried again.
• Humboldt-Universität zu Berlin and DIW Berlin/SOEP. Code: osf.io/uey5h (not read). Provenance: papers/carry_on/lersch2023_personal_culture_life_course.provenance.json.

What was read

• All 1,687 lines (my count) of pdftotext -layout output (32 pages): abstract, all sections, Tables 1–7, notes 1–13, acknowledgments, data note, references, the author note, and the captions and notes of Figures 1–5.
  • The pages have two columns, which the extraction prints side by side; each column was read in turn.
• Page 17 (journal page 236) was rendered and viewed, because the extraction drops the square-root signs in the text's formulas. The same page shows the bar labels of Figure 4 (counts of preferred models), which are used below.
• Not read or not inspected:
  • Figures 1–3 and 5 (images): model diagrams, ten example trajectories, and per-survey distributions of estimates over 418 outcomes;
  • the online supplement, Parts A–E. It holds the outcome lists and case numbers (Tables A.1–A.13), the three-wave simulation (Part B), the SEM checks (Part C), the splines, fixed-effects and domain results (Tables D.1–D.8, Figures D.1–D.5) and the causal graph (Part E). It sits on the publisher's site behind the bot check;
  • the OSF code.

Question

Do adults' attitudes, beliefs, values and subjective self-descriptions ("personal culture") change persistently over the life course, or do they only fluctuate around stable baselines? The paper sets three models against each other:

• Settled disposition (SDM): a stable baseline with random fluctuation (Kiley & Vaisey 2020).
• Active updating (AUM): a first-order Markov process; the last answer plus new experience (Kiley & Vaisey 2020).
• Life course adaption (LCAM, the author's): early imprinting, persistent change through transitions, fluctuation, and dependence of change on earlier biography.

Method

• Data. Six household panels:

  • HILDA 2001–2019 (Australia);
  • BHPS 1991–2008 and UKHLS 2009–2020 (Britain);
  • SOEP 1984–2019, v36 (Germany);
  • SHP 1999–2019 (Switzerland);
  • PSID 1968–2019 (United States), with its Child Development and Transition into Adulthood supplements.

  Ages 18–79; unbalanced panels; listwise deletion.

• Outcomes. 428 variables measured in at least three waves, all standardised to z-scores.

  • They are beliefs, attitudes, values and subjective self-descriptions. Examples: gender ideology, policy issues, moral behaviour, party affiliation, interest in politics, concern about crime, national identity.
  • Gaps between measurements are 1–5 years and observation windows 3–36 years. On average an outcome has about seven waves, about three years apart.

• Models per outcome (mixed-effects, MLwiN via Stata).

  • SDM: a random intercept.
  • AUM: the lagged outcome and a fixed intercept, with an equation for the first wave.
  • LCAM: a random intercept, a random age slope (age centred at 18, in 10-year units) and their covariance.
  • All models include gender, first-generation immigrant status, birth cohort (six groups) and period fixed effects.

• Comparison by BIC (a difference below 2 counts as inconclusive): all three models on the sample with lagged observations, and SDM against LCAM on the full sample.

• Tests of the propositions (Table 1): early imprinting by the intercept variance; persistent change by the mean age slope and the slope variance; fluctuation by the residual variance; biographical experience by the intercept–slope covariance.

• One transition: parenthood, as a time-varying 0/1 indicator.

• Meta-analysis. A two-stage individual participant data meta-analysis over outcomes and surveys, with survey random intercepts.

  • Covariates: scale type, number of waves, gap between waves, first year, mean age.
  • The author states: "I consider the absolute values of the coefficients for age and the covariances between intercept and age slope as the outcomes of interest in the meta-analysis".
  • The Table 5 note reads: "variances are transformed to logged standard deviations and covariances are transformed to correlations".

• Size of the analysis. 2,568 models in the main analysis, which is 428 outcomes × 6 specifications (my inference).

• Robustness (supplement; only the text's summary was read): age splines by decade; singletons excluded; fixed-effects models with age splines and period dummies; first observations dropped (panel conditioning); upper age 60; AIC instead of BIC; SEM on balanced samples.

Results

• Which model fits (Figure 4 counts, text percentages).
  • Sample with lagged observations, 415 outcomes estimated: LCAM 198 (48%), AUM 113 (27%), SDM 93 (22%), inconclusive 11 (3%). 13 more had estimation problems.
  • Full sample, 417 estimated (my sum of the Figure 4 bars): LCAM beats SDM for 297 (71%); SDM 112; inconclusive 8. 11 had estimation problems.
  • By domain (note 12, sample with lagged observations), LCAM is preferred in all domains but five:
    • religion and spirituality: AUM and LCAM equally often;
    • health and morale; social life, social cohesion and trust; occupation and education: AUM more often;
    • subjective SES: SDM more often.
• Size of persistent change (Table 5, 418 outcomes).
  • The mean absolute age coefficient is .07 SD per 10 years [.06, .09].
  • Back-transformed (my computation): person-specific slopes have an SD of .14 per decade (variance .019). Baselines have an SD of .67 (variance .45). Residuals have an SD of .73 (variance .54).
  • The mean absolute intercept–slope correlation is about .25. Survey-level variance is about zero, so there are no systematic differences between the five countries.
  • Individual models: age coefficients "mostly fall between –.20 and .20"; slope variances mostly below .03, but about ten times larger in the PSID young-adult sample. Intercept variances are mostly .20–.70 and residual variances .30–.80.
  • Comparisons given in the text: cohorts born before 1940 differ from 1960–69 by .18 SD; women from men by .13; immigrants from natives by .15.
• Three HILDA examples (Tables 3–4).
  • ICCs: .34 ("working father can have as good relationship to child"), .45 ("child should start living independently by 18-20"), .82 ("importance of religion").
  • Age coefficients per decade: −.04, .10 and .07.
  • For importance of religion, 95% of person slopes lie between −.32 and .46 per decade.
• Parenthood (Table 6, 411 outcomes).
  • Parents score .07 SD [.03, .11] "differently" from non-parents on average, presumably a mean absolute difference (inferred). That is about the change of ten years of age (.08 in the same models).
  • Per the text's summary of supplement Figure D.5, the effect is mostly in "gender and family". It is present in the other domains except politics, government and economy; religion and spirituality; and national identity, ethnicity and immigration.
  • Fixed-effects models give similar results (Table D.8, not read).
• Age.
  • The spline models give smaller slopes at older ages, but still .10 per decade at 70–79.
  • The text says: "from age 20 to 60, personal culture can be expected to change by a total of about .70 standard deviations", and adds that "change is likely episodic and not continuous for individuals".
  • Fixed-effects models give larger age slopes: .22 at 18–29 down to .14 at 50–59.
• Cohorts (Table 7).
  • The spread of person slopes is 12% lower for those born before 1940 and 13% higher for 1970–79 than for 1950–59.
  • It is only slightly higher for 1980 and later. The mean age effect oscillates without trend across cohorts.
• Domains and types (note 13).
  • Change is smaller for environment and climate and for national identity, ethnicity and immigration, and larger for occupation and education.
  • It is smaller for beliefs and larger for values.
• Against Kiley & Vaisey.
  • The author's simulation (Part B, not read) finds that "three observation points over four years, which the GSS offers, are often insufficient to identify actual changes in data".
  • With long panels, AUM is preferred over SDM more often, "in contrast to Kiley and Vaisey (2020)".
  • Openness to change declines with age, which the paper says agrees with them.

Limits

• Absolute values inflate the headline. The meta-analysis averages the absolute values of estimated age coefficients.
  • Estimates scatter around their true values, so their absolute values average above zero even if every true slope were zero. A confidence interval that "does not include 0" is therefore not a test of persistent change (inferred).
  • The size of this bias depends on the standard errors, which are in the supplement (not read).
  • The .07 is a mean magnitude with no direction.
  • The same applies to the .25 correlation and to the parenthood .07 (inferred for parenthood: its text says parents score "differently").
• Age is not a purely within-person contrast. Within a person, age and period rise together.
  • With period fixed effects, the age slope is identified from age differences between people within the six cohort groups, and from the random-effects assumptions (inferred).
  • The paper says separating age, period and cohort "can only rest on (untestable) assumptions".
  • How the fixed-effects robustness models identify the linear age part is in the supplement (not read).
• Persistence of event effects is assumed, not tested. Parenthood enters as a permanent 0/1 step with no time-since-transition terms, so a fading effect would be averaged into the step (inferred). Kiley & Vaisey's question (does a given change persist or revert?) is not asked of transitions.
• Fluctuation includes measurement error. The residual variance mixes transient change with error; there is no reliability correction, though the paper notes the problem. In true-score units persistent change would be somewhat larger (inferred).
• The biographical-experience test depends on centring. Age is centred at 18. The covariance therefore compares a person's level extrapolated to age 18 with their slope, and regression to the mean also produces negative values (inferred).
• Only one transition is modelled. Marriage, separation, widowhood, job loss and retirement are not estimated. The paper argues that permanent, salient transitions (parenthood) should matter more than temporary ones (unemployment), but tests only parenthood.
• The model tournament compares misspecified models. AUM has no person intercept and SDM no slope. The BIC shares say which simplification fits each item least badly (inferred).
• The supplement was not obtainable. Case numbers, per-item results, splines, fixed-effects results and the simulation were not read.
• Inconsistencies found:
  • .70 from age 20 to 60. It does not follow from the linear .07 per decade, which gives .28 (my computation).
    • Four decades at the fixed-effects slopes the text reports (.22 at 18–29 down to .14 at 50–59) would give about .64–.80, if the two middle decades lie between those values (my computation). The text, however, attributes the .70 to the random-effects spline models.
    • Summing mean absolute slopes also assumes that each outcome keeps one direction across decades (inferred).
  • Variance or SD (Table 7). The text says "the variance in age slopes is 12 percent lower", computed as exp(−.13) − 1. The table note says variances were transformed to logged SDs, so −12% is the change in the SD; the variance would be 23% lower, and 27% higher rather than 13% for 1970–79 (my computation).
  • A correlation above 1 (Table 7). The covariance column shows 1.01 [.94, 1.07], though the note says covariances are transformed to correlations. If Fisher's z was used, Table 5's .25 is r = .245 and Table 7's 1.01 is r = .77 (my computation; Fisher's z is a guess).
  • Counts. Table 5 and Figure 5 use 418 outcomes; Figure 4 has 417 estimated full-sample models (my count). Note 11 lists 25 + 1 non-converged models and 11 without a variance–covariance matrix. Figure 4 shows 13 + 11 estimation problems, and Table 6 has 17 fewer outcomes (my computation). Not resolvable without the supplement.
  • Extraction artefact, not an error. Spreads such as "−.24 ± 1.96 × .32" are printed with a square-root sign (verified on the rendered page); with it, all four spreads in the text check.
  • Checked by script: 48 + 27 + 22 + 3 = 100; Figure 4 counts give 27%, 48%, 22%, 3% of 415 and 71% of 417; the four ± spreads; exp(−.13) − 1 = −.12 and exp(.12) − 1 = .13; 428 × 6 = 2,568; the Table 4 ICCs from the variance components (.33, .45, .84 against the printed .34, .45, .82).

What it means for Kurisutina

• Carrying on, for attitudes, mostly means holding a baseline.

  • Baselines differ between people with an SD of .67. Occasion-to-occasion fluctuation is about as large (SD .73). Persistent change averages at most .07 SD per decade, and person slopes spread with an SD of .14 per decade (my back-transforms of Table 5; the .07 is biased upwards, inferred).
  • A replica on value and ideology items should answer around her baseline, with human-sized fluctuation, and move lastingly only slowly (inferred).
  • A replica that trends by more than about .03 SD per simulated year on such items would be outside the range of about 95% of people (my computation: 1.96 × .14 per decade, ignoring the small mean slope; the use is inferred).

• Kiley & Vaisey and Lersch are compatible (inferred).

  • .07 SD per decade is about .014 SD per two-year GSS interval, against a residual SD of about .73, a ratio of about .02 (my computation). Slow drift over decades is invisible over four years, which is also the author's simulation result.
  • What neither paper measures is how often a change triggered by an event lasts. For that, Kiley & Vaisey's φ concerns any wave-to-wave change, and Lersch assumes persistence for parenthood.

• Q1: a personal way of changing exists at the slow time scale, but it is expensive to measure.

  • Person slopes spread about twice as widely as the mean slope magnitude (.14 against .07; my comparison). The spread is wider in younger cohorts, and slopes relate to baselines (|r| about .25; centring-dependent, inferred).
  • So the average person's drift is a poor model of one person's drift. Estimating one person's slope needs many waves over years, which a few interviews cannot supply (inferred).
  • Over the short horizons of our tests, the slot's baseline answers matter far more than any personal drift rate (inferred).

• LISS (docs/research/liss_q1_design.md). The LISS core modules, Politics and Values among them, are fielded yearly from 2008 (per the design doc). A possible slow-change add-on, not part of T1 as designed: fit person slopes on the early waves of the trust and left–right items, and test whether they forecast later waves better than the population trend (proposal).

• Which transitions move attitudes. Only parenthood was estimated: about .07 SD, mostly on gender and family attitudes, not on politics, religion or national identity (verified from the text). Marriage, divorce, widowhood and job loss, the pilot's events, have no estimates here.

• GSS pilot (frozen; nothing to change). Expected movement of the target items after the pilot's events, over two years. Every row is inferred or a guess.

  • The φ values are from the kiley2020 summary and the reliabilities from the hout2016 summary.
  • Mapping GSS items to this paper's domains is a guess, because the outcome lists are in the unread supplement.

  Item	Evidence	Expected event effect
  satfin, finalter	Circumstance items; φ .64 and .71, updating in the recession; knowing about a job loss improved Brier in the feasibility check	The items where events should show, mostly lost or found job (inferred)
  happy	Reliability .60; the appraisal coupling with affect fades within months (haehner2023 summary)	Transient; after two years mostly faded except for lasting changes such as widowhood; the person-specific part is unrecorded (inferred)
  health	φ .59; reliability .78; no pilot stratum is a health event	Little (guess)
  trust	φ .56, no updating above 30; here AUM is preferred more often for social life, social cohesion and trust (note 12)	Near zero from the pilot's events (guess)
  polviews, partyid	Parenthood does not reach politics; slow drift of about .014 SD per two years (my computation); φ .56 and .67	Near zero. Movement between waves is fluctuation or slow drift (inferred)
  attend	Parenthood does not reach religion; φ .66	Possibly after marriage or widowhood, as behaviour. Not covered here (guess)
  confinan	No updating (kiley2020); reliability .67; "hardly any" confidence rose from 14.6% to 41.5% over the 2006 panel's waves (goal doc)	Small individual effects under a common period shift (guess)

  • How much. The one transition estimated moved attitudes by about .07 SD. On one item with 40 pairs per stratum, the standard error of a mean change is about .10–.16 SD, taking a 2-year retest correlation of .5–.8 (my computation). An effect of that size is invisible in the pilot (inferred).
  • What a pilot result can claim. A three-wave panel cannot tell persistent from transient change (the author's simulation, not read). The pilot's next-wave forecasts are unaffected, but the pilot cannot show that a replica "carries on" in the lasting sense (inferred).

• As a measure for Q2 (drift): proposal, not started.

  • Fit the same decomposition to a replica's repeated answers over a long simulated run: per item, baseline, slope and residual variance. Compare with the human profile above: baseline and residual variances of similar size (.45 and .54, my back-transforms) and a mean slope magnitude under .1 SD per decade.
  • The model tournament can label the replica's answer process. A replica whose context accumulates its own answers should look like AUM (Markov) on most items, which people are for 27% of items (proposal; mapping simulated time to years is unsolved).
  • This complements the φ test proposed in the kiley2020 summary, which needs only three simulated waves.

Cross-references

• summaries/carry_on/kiley2020_personal_culture.md: persistent change over 2–4 years on the pilot's panels; per-item φ.
• summaries/carry_on/hout2016_gss_reliability.md, summaries/carry_on/gss_mr119.md: reliability and stability of the pilot's items.
• summaries/carry_on/haehner2023_perception_change_swb.md: re-appraisal and SWB change; timing.
• summaries/carry_on/kramer2024_multiple_events.md, summaries/carry_on/luhmann2014_its_about_time.md: event effects and designs for well-being outcomes.
• summaries/carry_on/beck2018_idiographic_networks.md: consistency and change as person-level properties.
• docs/research/gss_pilot_design.md, docs/research/liss_q1_design.md.