Applications in Social, Personality, Organizational, and Health Psychology

35.1 The Terrain: Processes Where People Live

The questions of social, personality, and organizational psychology that this book can serve are questions about within-person process, and they fall into a few families that route to the tools developed earlier. The expression of a state and its contingencies, how much a person’s momentary extraversion or affect varies and what situations move it, is a distribution question answered by the variability and location-scale tools of Chapters 7 and 16. The transmission of states between people, whether one partner’s support lifts the other’s intimacy and who leads whom day to day, is a dyadic question answered by the multilevel models of Chapter 23 extended to two interdependent members and by the coupling methods of Chapter 25. The rhythms of work and recovery, the weekly cycle and the restoration of a weekend, are cyclical questions answered by the trigonometric and cyclic-smooth tools of Chapters 23 and 30. And the couplings of behavior and health across the day are dynamic questions answered by the same within-person machinery. Figure 35.1 draws the routing and Table 35.1 lists the families with estimands and exemplar literatures.

Note. Each family of within-person question routes to the tools that answer it. The design lineage runs from the Rochester Interaction Record and the event-contingent tradition of Wheeler and Reis through modern smartphone experience sampling, but the analytic families are stable, and naming the family remains the first analytic act.

35.2 Case Study A: Personality as Distributions of States

The reconciliation of states and traits that Fleeson (2001) proposed dissolves an old dispute by a change of object: a personality trait is not a person’s single characteristic value but the density distribution of that person’s momentary states, and the parameters of that distribution, the mean, the spread, and the way the state depends on the situation, are the trait. This reframing turns intensive longitudinal data into a personality-measurement instrument, and it is the content of Case A, worked on the affect_ema records reframed as a personality study: 120 persons, sampled roughly six times a day for two weeks, whose momentary positive affect, negative affect, and stress are the states from which distribution parameters are estimated.

35.2.1 Distribution Parameters as Personality

Figure 35.2 shows the raw object of study, the distribution of momentary positive affect for a spread of persons, and it makes the essential point visually: persons differ not only in where their distribution sits but in how wide it is, so that two people with the same average affect can differ sharply in their variability, and both facts are personality. Estimating the parameters properly means resisting the temptation to compute a raw within-person standard deviation and treat it as a trait, because such an index confounds true variability with measurement error and with the person’s mean, the confound Baird, Le, and Lucas (2006) documented; the location-scale model of Chapter 16 is the repair, modeling the person’s mean and the logarithm of the person’s residual variability jointly with the person’s contingent reactivity as a random slope. Fit to these data, the three parameters are all substantial and person-varying: the person means of negative affect average 2.02 with a between-person standard deviation of 0.44, the within-person variabilities average 0.56 with a between-person standard deviation of 0.10, and the stress reactivity, the degree to which a person’s negative affect rises with momentary stress, averages 0.35 with a between-person standard deviation of 0.08. Table 35.2 sets out which distribution parameter operationalizes which personality construct.

Note. Each row is one person’s distribution of momentary positive affect across two weeks of sampling, ordered by the person mean (red points). Persons differ in the location of the distribution and, independently, in its spread; both are stable personality parameters in the density-distribution view. A raw standard deviation read off these distributions would confound true spread with measurement error and with the mean, which the location-scale model of Chapter 16 corrects.

35.2.2 Stability and the Anti-Plug-In Discipline

For a distribution parameter to count as a trait it must be stable, and splitting the study into its two weeks and correlating each parameter across the halves gives the empirical verdict: the person mean is highly stable, correlating 0.90 across the halves, the within-person variability is moderately stable at 0.56, and the reactivity is the noisiest at 0.37, which is the pattern the density-distribution literature has converged on, mean level robustly trait-like and the higher-order parameters real but harder to pin down in two weeks of data. The stability numbers matter for the next step, which is carrying these parameters into trait-outcome analyses, because a parameter estimated with error is a predictor measured with error, and treating it as if it were known is the plug-in fallacy that Chapter 16 warned against. Figure 35.3 shows the consequence in a controlled demonstration: when a person-level outcome truly depends on the person’s variability, regressing the outcome on the plug-in estimate of variability attenuates the effect, here from a true 0.50 to a naive 0.45, and understates its uncertainty, because the noise in the estimated variability is ignored; propagating that estimation uncertainty, through an errors-in-variables correction that uses the known sampling variance of a standard deviation, restores the effect to 0.55 and widens the interval honestly. The discipline is simple to state and often ignored: a distribution parameter entering a between-person model carries its own uncertainty, and that uncertainty belongs in the second-stage inference.

Note. A person-level outcome depends on the person’s true within-person variability, estimated here from about twenty-five usable occasions. The plug-in analysis regresses the outcome on the point estimate, attenuating the effect and reporting too narrow an interval. Propagating the estimation uncertainty, via an errors-in-variables correction using the known sampling variance of a standard deviation, restores the effect and widens the interval to its honest width. The attenuation grows as the parameter rests on fewer occasions.

35.3 Case Study B: Dyadic Daily Process

Two people in a relationship generate two interdependent streams of daily data, and the central methodological fact is that the streams cannot be analyzed apart, because each partner’s outcome depends on both partners’ predictors and because the two members of a couple are more alike than two random individuals. Case B develops the analysis on the couples_diary records, a distinguishable-dyad study of 130 couples reporting daily support and intimacy across three weeks, with a known actor-partner structure, a satisfaction moderator, a between-couple confound, a common-fate growth process, and an asymmetric cross-partner lag, so that every claim can be checked against the truth. Figure 35.4 shows the data’s anatomy, the two aligned streams with their shared couple-days, and the section builds the model that respects it.

Note. Two partners generate aligned daily streams sharing the same couple-days. Coupling takes two distinct forms: a same-day shared component, in which an event moves both partners at once, and a lagged cross-partner component, in which one partner’s state forecasts the other’s on the following day. Conflating them, or ignoring the nonindependence entirely, misstates the interdependence that is the object of study.

35.3.1 Nonindependence and the Actor-Partner Model

The nonindependence is real and quantifiable: the dyadic intraclass correlation of intimacy, the share of variance lying between couples, is 0.34, far from the zero that an analysis of individuals would assume, and it is the reason a couple-level random effect is mandatory. The model that respects the dyadic structure is the actor-partner interdependence model in its multilevel form, in which each partner’s daily intimacy is regressed on the partner’s own support, the actor effect, and on the other partner’s support, the partner effect, with the two members’ days sharing a couple random intercept. Figure 35.5 draws the specification. The subtlety that the scattered applied literature routinely gets wrong is the centering: an actor effect is a within-person effect, the degree to which a partner’s intimacy rises on days when that partner’s support is higher than usual, and it is estimated cleanly only when support is centered within person, separating the day-to-day within-person association from the between-person association of people who give more support on average. Table 35.3 is the decision table that resolves the specification, crossing distinguishability, error structure, and centering.

Note. Each partner’s intimacy depends on the partner’s own support (the actor effect) and on the other partner’s support (the partner effect, purple). Estimated in a multilevel model with support centered within person and a couple-level random intercept, the actor and partner effects are within-person daily effects, separated from the between-couple association. The paths shown are the within-person layer; the between-couple layer is a parallel set of couple-mean effects.

Fitting the properly centered model recovers the planted structure: the actor effect is 0.45 against a truth of 0.40, the partner effect is 0.19 against a truth of 0.20, and the between-couple effect of average support on intimacy is 0.59 against a truth of 0.60, so that the between-person association is genuinely larger than the within, exactly the double dissociation the centering is designed to reveal. Figure 35.6 shows these estimates against their truths and, crucially, shows the cost of the common error: an analyst who centers support only at the grand mean estimates the actor effect as 0.51, contaminated upward by the larger between-person association, an error of the same size as many published effects. Relationship satisfaction moderates the actor effect as planted, the within-person support-intimacy coupling being stronger in more satisfied couples, with a recovered interaction of 0.13 against a truth of 0.15. Table 35.4 is the glossary that keeps the four effects straight.

Note. The within-centered model recovers the actor (0.45), partner (0.19), and between-person (0.59) effects near their planted truths (diamonds). The bottom row shows the actor effect when support is centered only at the grand mean: it is pulled to 0.51, contaminated by the larger between-person association, an error that inflates the reported within-person coupling. Centering is not a technicality here; it is the difference between the effect you want and an effect you did not.

35.3.2 Dyadic Growth and Who Leads Whom

Beyond the daily couplings, the couple’s intimacy has a trajectory, and the dyadic growth model asks whether the two partners’ trajectories move together. Fitting a growth model with a couple-level random slope recovers a shared upward drift of 0.015 units per day, and the couple-level slope captures the common fate, the tendency of partners to rise and fall together over the study rather than independently, which Figure 35.7 displays for a sample of couples whose two trajectories track one another. The common-fate representation, in which a single couple-level trajectory drives both partners, is the theoretically apt alternative to a parallel-process model with two separate but correlated slopes when the causal story is one of shared circumstance rather than mutual influence, and choosing between them is a choice about what the couple is, not a fit contest. The final layer is the dynamic one, the question of who leads whom, and estimating the cross-partner lag on the intimacy residual recovers the planted asymmetry: Partner 1’s intimacy on one day forecasts Partner 2’s the next with a coefficient of 0.18, while the reverse path is only 0.07, so Partner 1 leads and Partner 2 follows, the lead-lag structure Figure 35.8 displays. The caution here is the one Chapter 25 pressed: a cross-partner lag is a temporal-predictive fact, not a demonstration of influence, because a shared unmeasured driver, a common daily stressor, or a third variable can manufacture the asymmetry, and the recurrence-based coordination methods of Chapter 32 offer a complementary view without resolving the causal question.

Note. Intimacy trajectories for nine couples, each partner a separate line. Within a couple the two partners’ trajectories move together, the common-fate signature that a couple-level random slope captures. When partners rise and fall together because they share circumstances, the common-fate model is the apt representation; when the story is mutual influence, a parallel-process model with cross-partner slope couplings is the alternative.

Note. The cross-partner lagged effects on the intimacy residual, controlling each partner’s own inertia. The coupling is asymmetric, Partner 1 leading (0.18) and Partner 2 following (0.07), recovering the planted asymmetry (diamonds). The lead-lag is a predictive fact, not a proof of influence: a shared daily driver can produce the same asymmetry, which is why the caution ledger of Chapter 25 applies.

35.4 Case Study C: Workweek Rhythms and Recovery

The organizational literature on stress and recovery treats the week as a structure with meaning: strain accumulates across workdays, recovery happens in the evenings and on weekends, and the rhythm itself is the phenomenon, not a nuisance to be partialled out. Case C works these ideas on the work_week records, a two-week diary of 200 workers with a workday, evening, and morning architecture, a planted weekly cycle in vigor, a discontinuous weekend restoration, a within-person recovery effect linking the evening’s detachment to the next morning’s vigor, and a job-demands moderator, so that the theory-testing can be audited against a known process.

35.4.1 The Weekly Cycle as Theory

The weekly cycle is estimated two ways, and the contrast is instructive. A single trigonometric harmonic, a sine and cosine at the seven-day period, imposes a smooth sinusoid and recovers the planted cycle with a correlation of 0.95 to the truth; a cyclic spline, which lets the data choose the shape while enforcing that Sunday joins back to Monday, recovers it with a correlation of 0.91, theory-agnostic and nearly as efficient. Figure 35.9 shows both against the truth, a mid-week trough in vigor and a weekend peak, and Table 35.5 is the decision guide. The methodological point is that the choice between them is theoretical, not merely statistical: if theory predicts a smooth single-peaked rhythm, the trigonometric model encodes that prediction and tests it efficiently; if the shape is unknown or possibly multi-peaked, the cyclic smooth is the honest default; and treating the day of week as a set of free dummy variables, the third option, discards the cyclical structure entirely and is appropriate only when no rhythm is expected. A weekly cycle fit as decoration, a smooth added because the software offers one, is not the same as a weekly cycle fit as theory.

Note. Panel (a): the weekly vigor cycle recovered by a single trigonometric harmonic and by a cyclic spline, both tracking the planted truth (a mid-week trough, a weekend peak). The trigonometric model is efficient when the shape is a known sinusoid; the cyclic smooth is the theory-agnostic default. Panel (b): the within-person recovery effect, next-morning vigor rising with the previous evening’s psychological detachment, the recovery process at the heart of the stress-recovery literature.

35.4.2 Recovery, Weekends, and Demands

The recovery hypothesis is that detaching from work in the evening restores the resources that the next day draws on, and it is a within-person lagged claim that the diary is built to test: regressing next-morning vigor on the previous evening’s psychological detachment, centered within person and crossing the night boundary with the care that Chapter 5 pressed, recovers a coefficient of 0.33 against a planted 0.35, so that a worker’s own better-than-usual detachment forecasts that worker’s own higher vigor. The weekend adds a discontinuity, not merely a low point on a smooth cycle but a genuine jump, and modeling it as a piecewise shift recovers a weekend vigor boost of 0.51 against a truth of 0.50, which Figure 35.10 displays as the within-week restoration it is. Job demands moderate the system as planted: workers reporting higher demands detach less, with a demands-detachment coefficient of \(-0.36\) against a truth of \(-0.30\), and report lower baseline vigor, so the recovery process that protects vigor is least available to the workers who need it most, a substantive finding the model surfaces. Finally, the day’s interruptions are a count, and a Poisson mixed model recovers their structure: interruptions rise with job demands, the log-rate coefficient being 0.51 against a truth of 0.50, and peak in the middle of the week, the count outcome handled by the generalized linear mixed model of Chapter 15 rather than forced into a linear mold.

Note. Mean vigor across the two-week diary, with workdays and weekends distinguished. Vigor jumps on the weekend and falls back when the workweek resumes, a within-week discontinuity of about half a scale point that a piecewise model captures and a smooth weekly cycle would blur. The weekend is not merely the low-strain end of a gradient; it is a qualitatively different state.

35.5 Domain Synthesis

The three cases share a discipline and a set of recurring objections that a reviewer for these fields will raise. The discipline is that within-person structure is the object, that a distribution parameter, an actor effect, a recovery lag are all within-person quantities that must be estimated with the centering, the uncertainty propagation, and the causal caution their nature demands, and that summarizing them into a between-person or trait network without laundering their uncertainty is the standard the plug-in fallacy fails. The recurring objections are three. The first is the demand for an experiment, the reviewer who reads any within-person association as a causal claim and asks where the manipulation is; the honest answer is that these designs estimate within-person association and temporal prediction, which constrain causal hypotheses without establishing them, and that overclaiming influence from a partner effect or a lead-lag is the error the caution ledger exists to prevent. The second is the misreading of sample size, the reviewer who calls eighty couples or two hundred workers a small sample; the answer is level-specific power literacy, that the relevant sample for a within-person effect is the number of occasions times persons, that the between-person sample governs only the between-person effects, and that the design’s power must be assessed at the level of the estimand, as Chapter 4 sets out. The third is the dismissal of the rhythm or the weekday as a nuisance control, which mistakes the theory for noise. Table 35.6 collects the reporting obligations for these domains, including the translation of within-person findings into the language applied audiences use, and the anchor literatures orient the reader who would go deeper: the density-distribution program in Fleeson and in Fleeson and Jayawickreme, the dyadic canon in Kenny, Kashy, and Cook with the diary craft in Laurenceau and Bolger, and the recovery and organizational-sampling literatures in Sonnentag, in Sonnentag and Fritz, and in Beal and in Gabriel and colleagues. These are the domains where the within-person revolution began, and the craft they reward is the one this chapter has practiced: to model the process where it lives, and to say no more than the within-person data support.