Kurisutina
Library/Carrying on

Beck et al.

A longitudinal ESM network study

Published in Journal of Personality and Social Psychology (DOI 10.1037/pspp0000249); the published version was not read. Read as the accepted manuscript posted on PsyArXiv (pb92q, v1; the PDF carries "© 2019, American Psychological Association. This paper is not the copy of record"; OSF lists no licence). Washington University in St. Louis. Data and code: osf.io/fyxza, GitHub emoriebeck/Idiographic-Network-Consistency (not read). Provenance: papers/carry_on/beck2018_idiographic_networks.provenance.json.

What was read

All 2,975 lines of pdftotext -layout output (67 pages): abstract, introduction, method, results, discussion, limitations, footnotes 1–15, references, Tables 1–2 and the captions of Figures 1–8. The figures (the networks and distributions) are images and were not inspected. The online materials (between-person reference networks, alternative models in full, supplementary analyses of correlates of consistency) and the Shiny app were not read.

Question

Does a person's own personality structure, meaning how their momentary states go together and follow each other, stay the same over one to two years? Do people differ in how consistent it is?

Method

  • Data. PAIRS: college students. Experience sampling four times a day for two weeks, with 59 possible surveys, about "the last hour". Nine state items from the BFI:

    • Agreeableness (kind, rude; only asked when interacting);
    • Conscientiousness (lazy, reliable);
    • Extraversion (outgoing, quiet);
    • Neuroticism (depressed, relaxed, worried).

    No Openness items.

  • Samples. After exclusions, wave 1 has 349 people and wave 2 has 156, one or two years later: 17,715 surveys. The median was 41 surveys per person in wave 1.

  • Models. Per person and wave, a graphical VAR:

    • a contemporaneous network: which states co-occur, a "while" structure;
    • a lagged network: which states predict which states about 4 hours later, an "if … then" structure.

    Both are regularised (glasso, BIC). Robustness checks used other penalties, unregularised models, zero-order associations, GIMME, multilevel VAR (empirical Bayes) and P-technique factor analysis.

  • Consistency. Rank-order (per edge, across people) and ipsative (per person, the profile correlation of their edges across waves).

  • Reliability. Split-half (first against second week) and odd–even within each wave.

Results

  • Heterogeneous structures. People's networks differ in edges and in which states are central. Some examples:
    • about 10% feel outgoing when quiet, against the usual negative link;
    • about 25% feel more depressed when quiet, about 7% less.
  • The "while" structure is moderately consistent. The mean ipsative consistency of contemporaneous networks over one to two years is r = .62. But it ranges from −.19 to .93 (SD .42): some people are unchanged, others unlike their former selves.
    • Rank-order consistency of single edges is low: mean .10; within-trait edges .24.
    • Strength centrality profiles: ipsative .61; rank-order .19.
  • The "if … then" structure is not measurable here.
    • Lagged networks show no consistency: ipsative .03, rank-order .01.
    • But their split-half and odd–even reliability within one two-week wave is also about zero (.001–.03, SD about .15). By comparison, contemporaneous networks give .40–.50.
    • So the lagged instability is at least as much unreliability as change. The authors say the lagged results "should be interpreted cautiously" (verified).
  • Reliability is itself an individual difference. Contemporaneous split-half reliability ranges from −.34 to .96 across people, and people with reliable networks are also more consistent across years.
  • For comparison, averages of the same ESM items are very stable: rank-order .46–.74 (items) and .68–.79 (trait composites); ipsative .91 and .97.
  • Robustness. The patterns replicate across model types, except that empirical Bayes estimates (shrunk toward the group) look more consistent.

Limits

  • Students, nine items, four traits. A 4-hour lag may be too long for the dynamics of interest. Series are short (41 median against a recommended minimum of about 50). Only two waves.
  • Inconsistencies found:
    • The wave 2 survey count is "median … 33 (Wave 2; range 54 to 148)": a median outside its range.
    • Contemporaneous edge rank-order is given as "range -.10 to .34" and, in the same sentence, as "ranging from -.06 … to .35".
    • Table 2's GIMME row has a contemporaneous minimum (0.20) above its median (0.04), and a lagged minimum (0.36) above its maximum (0.23). Probably lost minus signs.
    • Footnote 12 gives nomothetic ipsative consistency as M = .97 with SD = .66. A mean of .97 for correlations bounded at 1 cannot have an SD of .66.
    • The discussion calls the "moderate rank-order consistency of contemporaneous network edges … nearly as strong as" trait inventories (.4–.6). The reported mean is .10 (max about .35).
    • Wave 2 is 207 (one year) plus 134 (two years), "collapsed" to 217; 207 + 134 = 341. The overlap is not explained.
    • Checked: 11,124 + 6,591 = 17,715 surveys, as in the abstract.

What it means for Kurisutina

  • Question 1: a person's own "while" structure is real and moderately stable. Their "if … then" dynamics could not be measured with this much data.
    • Which states go together for this person (being quiet goes with feeling depressed, or not) holds across one to two years for many people (ipsative about .6). That is slot material, and it can differ from the population pattern (Fisher 2018).
    • How one state leads to the next, the closest thing here to a "way of changing" on the hours scale, was not reliably estimable from about 40 observations per person. It is unreliable even within two weeks.
    • A slot built from a few weeks of diaries could hold co-occurrence structure, not person-specific dynamics. Capturing dynamics needs far longer series or shorter lags (inferred from the reliability results).
  • Consistency is a person-level trait. Some people are nearly unchanged over two years, others change a lot, and reliability itself differs by person.
    • A replica's rate of change should be set per person, not globally. This matches the goal doc's point that reflection schedules must be calibrated to the person's own stability (Hout's β for items).
    • It can be estimated from bursts: split-half reliability within a burst, consistency across bursts.
  • Population structure as a bad default. Between 7% and 25% of people show edges opposite to or absent from the usual pattern. A replica generating momentary states from a base model's population associations would misdescribe them.
    • A test (proposal, with Fisher 2018): compare the replica's simulated within-day co-variation with the person's own network, against the population network.

Cross-references

  • summaries/carry_on/fisher2018_group_to_individual.md: within-person against between-person structure.
  • summaries/carry_on/sliwinski2009_stress_bursts.md: bursts; stable and variable parts of reactivity.
  • summaries/carry_on/hout2016_gss_reliability.md: reliability ceilings and person stability.
  • summaries/carry_on/haehner_perception_stability.md: stability of appraisals over a year.

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.