Longitudinal research on adult development studies how the same people change, remain stable, and differ from one another across repeated observations. Researchers use this approach to investigate trajectories in cognition, personality, emotional functioning, health, relationships, work, identity, and social participation from early adulthood into later life. Its central advantage is the ability to observe within-person change. Its central methodological challenge is to distinguish developmental processes from cohort and historical influences, practice effects, measurement change, selective participation, and other sources of bias. The strongest longitudinal conclusions combine a clear developmental theory with appropriate sampling, repeated valid measurements, explicit assumptions, and sensitivity analyses. Newman (2010); Schaie (2005).
What Longitudinal Research Measures
A longitudinal study observes a defined set of individuals or units at two or more occasions. A wave is one assessment period. A panel follows many of the same people through successive waves; a cohort is a group defined by a shared starting characteristic, commonly a birth interval or historical exposure. Multiple cohorts can be followed simultaneously or recruited later. Repeated observation creates an empirical record of sequences and patterns that a one-time age comparison cannot supply. Two waves establish a difference between occasions, while three or more waves provide greater leverage for studying curvature, changes in pace, and temporary departures from longer trajectories. Curran, Obeidat, and Losardo (2010); Galbraith, Bowden, and Mander (2017).
A change score expresses the difference between measured values. A trajectory describes the pattern of values over a chosen timescale. A developmental process is the underlying sequence of mechanisms generating those patterns. These concepts should be kept distinct. An observed decline in one questionnaire score can reflect a change in the underlying attribute, a changed interpretation of questionnaire items, a temporary health event, or measurement noise. Likewise, high rank-order stability—the tendency for people to maintain their relative positions—can coexist with substantial average change across a population. van de Schoot, Lugtig, and Hox (2012); Roberts, Walton, and Viechtbauer (2006).
Within-person change and between-person differences
Suppose one group of 75-year-olds scores lower on a memory test than a group of 45-year-olds. That cross-sectional contrast describes a difference between people tested at one time. It does not show how any particular 45-year-old will perform thirty years later. If the same adults complete comparable assessments repeatedly, researchers can estimate their own rates of change, investigate why those rates vary, and distinguish stable individual differences from evolving performance. The distinction between within-person dynamics and between-person associations is foundational to longitudinal interpretation. For the broader theoretical context, see Adult Developmental Psychology: Theories, Stages, and Lifespan Change. Schaie (2005); Hamaker, Kuiper, and Grasman (2015).
Population means also conceal diversity. Some adults improve through accumulated expertise or changed strategies, some remain relatively stable on a particular measure, and others experience decline associated with health, opportunities, changing responsibilities, or major life transitions. Longitudinal research is especially valuable when a psychological question concerns this heterogeneity rather than a single average curve. Age identifies a location in a life course; it is not a stand-alone explanation of psychological functioning. Newman (2010); Schaie and Willis (2010).
Research Designs: What Each Can Answer
Cross-sectional age comparisons
Cross-sectional studies measure distinct age groups at approximately the same calendar time. They efficiently describe age-patterned differences and can recruit large samples across much of adulthood. They also mix developmental and generational influences: groups who differ in age grew up during different historical conditions, received different schooling, and may have had unequal access to health care, technology, nutrition, occupations, or testing experience. Cross-sectional data are informative for describing distributions at one moment. They are a limited basis for predicting what will happen to the same people. Schaie (2005); Schaie and Caskie (2005).
Single-cohort longitudinal panels
A single-cohort panel repeatedly follows participants whose initial ages fall within a narrow range. This design directly observes within-person trajectories across the observation period and can establish whether an earlier measurement precedes a later outcome. Its disadvantages include the time needed to cover a broad age span, selective attrition, repeated-testing effects, and the historical specificity of the cohort. If a cohort reaches age 70 in a particular decade, its observations at that age also reflect that decade. The study cannot simply subtract historical experience from aging. Newman (2010); Bell and Jones (2013).
Cohort-sequential and accelerated longitudinal designs
Cohort-sequential designs recruit overlapping age cohorts and follow each over a shorter span. For example, investigators might enroll groups aged 40–49, 50–59, and 60–69 and assess them several times. Their overlapping ages can help test whether trajectories look similar across birth cohorts while allowing a wider adult age interval to be studied within a shorter calendar period. The resulting long age curve is stitched together statistically rather than observed across an entire lifetime in every participant. Its accuracy depends on comparable measures, cohort overlap, model specification, and whether cohort differences are adequately represented. Galbraith, Bowden, and Mander (2017).
The design is efficient under appropriate assumptions, yet overlap does not automatically remove the age–period–cohort problem. A model can estimate cohort terms, test some hypotheses, and compare sensitivity to alternative specifications; it still needs scientific assumptions that determine which components are identified. Newer simulation research illustrates that cohort variables may improve estimation under specified accelerated-design conditions, without implying universal correction of historical confounding. Bell and Jones (2013); Galbraith, Bowden, and Mander (2017).
Repeated cross-sectional surveys and refreshment samples
Repeated cross-sectional surveys draw new samples at successive calendar dates. They are valuable for monitoring population trends but cannot estimate personal trajectories without linking individual records. Refreshment samples add new participants to an existing longitudinal panel and may restore coverage of age groups or improve representativeness. Investigators must distinguish changes in the original participants from differences created when new participants enter. The Health and Retirement Study illustrates a replenished cohort design rather than a single closed sample. Health and Retirement Study, Survey Design and Methodology.
Intensive longitudinal and measurement-burst designs
Some psychological processes operate across hours, days, and weeks, while developmental change accumulates over years. Experience sampling and daily diaries capture fluctuations in affect, interactions, motivation, fatigue, or self-regulation. Measurement-burst designs collect multiple short-term observations at repeated longer-term intervals, thereby linking momentary variability with broader developmental trajectories. Short-term variability is a substantive phenomenon that may depend on context, stressors, health, and adaptation; treating every fluctuation as measurement error can conceal the process under study. Martin and Hofer (2004); Ram and Gerstorf (2009).
Age, Period, and Cohort: The Identification Problem
Age is time elapsed since birth; period is the calendar or historical moment at which observation occurs; cohort is the group defined by birth or another historical starting point. For birth cohorts, the familiar identity is period = age + birth year. As a result, conventional age, period, and cohort components are linearly dependent. Data alone cannot uniquely identify three entirely unrestricted additive trends. Researchers must introduce constraints or assumptions, exploit informative structure, or formulate narrower causal questions. Statistical sophistication cannot abolish this identification problem. Bell and Jones (2013).
Consider a study reporting that adults born in 1945 perform differently at age 70 than adults born in 1975 do at age 40. Age and birth cohort differ simultaneously, while testing occurs within an identifiable historical period. Education, social policy, occupational exposures, gender norms, migration, discrimination, and major societal events may affect these groups differently. A defensible interpretation therefore states the precise cohort, periods observed, and variables measured instead of presenting one fitted age curve as a timeless biological law. Schaie and Willis (2010).
Sequential samples and long-running studies improve the opportunity to investigate alternative explanations, because people can be observed at several ages and multiple cohorts can overlap at similar ages. Yet perfect separation still requires additional scientific structure. Researchers may use natural experiments, known policy changes, explicit cohort contrasts, historical exposure measures, or preregistered identifying assumptions. Their conclusions should be conditional on those choices, ideally supported by sensitivity analyses. A dedicated examination of age, period, and cohort identification remains a separate scientific topic; here it matters because it shapes the validity of longitudinal claims. Bell and Jones (2013).
Sampling, Recruitment, and Generalizability
Longitudinal inference begins before the first interview. Investigators define a target population, eligibility rules, recruitment frame, inclusion probabilities, recruitment period, and outcome measures. A population-representative probability sample supports different generalizations from a convenience sample of motivated volunteers. Many aging cohorts also exclude people with limited mobility, certain illnesses, institutional residence, language barriers, or no stable contact information. Those decisions may be appropriate for a specific research question, but they narrow the population to whom results apply. Newman (2010); National Institute on Aging, BLSA recruitment criteria.
Coverage matters especially across the adult life course. Life expectancy, socioeconomic position, disability, migration, race and ethnicity, rurality, digital access, and experience with institutions affect who can participate and stay in a study. A sample that represents community-dwelling adults over 50 at baseline does not automatically represent everyone who later becomes older than 80. Investigators should describe exclusions and differential retention, use suitable weights where justified, and evaluate generalization beyond the observed sample. Comparisons across countries also require culturally appropriate recruitment and harmonization of instruments. Health and Retirement Study, Survey Design and Methodology; English Longitudinal Study of Ageing, About ELSA.
Choosing observation intervals
The distance between waves must fit the process. Annual surveys may miss a short employment or caregiving transition; monthly surveys may be unnecessarily burdensome for a gradual life-course construct. Wider gaps increase the probability that events occur between observations without precise timing. Dense schedules can improve temporal resolution but raise costs, burden, reactivity, and attrition risks. Unequally spaced measurements are legitimate when their timing is recorded and the statistical model reflects actual intervals. Galbraith, Bowden, and Mander (2017); Curran, Obeidat, and Losardo (2010).
Researchers should distinguish age at assessment from years since enrollment, calendar time, time since a major event, and time before death. Retirement adjustment may be better indexed to time since retirement, for instance, while cognitive age trends may be indexed to chronological age and historical calendar period. Different choices answer different questions. Treating these time axes as interchangeable can manufacture or conceal developmental patterns. Newman (2010).
Attrition, Mortality, and Survivor Bias
Attrition occurs when previously enrolled participants fail to contribute follow-up data. Causes include refusal, lost contact, illness, changing residence, institutionalization, study burden, and death. The distinction matters. Missing an interview because a phone number changed differs from death, which makes subsequent personal outcomes undefined rather than merely unobserved. Continuing participants may become increasingly selective as the study ages. The resulting estimates can overrepresent adults with better health, higher cognitive functioning, greater resources, or fewer barriers to participation. Newman (2010); Hernandez et al. (2024).
An individual-participant-data meta-analysis across ten epidemiological longitudinal aging studies included 209,518 participants and reported an average next-wave dropout proportion of 0.26 in analyzed wave pairs, with variation across studies. Participants who later dropped out had lower cognitive performance beforehand across examined domains; the pooled standardized difference was approximately 0.37. This result documents selective attrition in these datasets. It does not imply that every longitudinal outcome is distorted by the same amount, and its definition of dropout included death. Hernandez et al. (2024).
Attrition can bias levels, trajectories, relationships, and subgroup comparisons differently. If participants with poorer baseline performance are less likely to return, a surviving sample may appear unusually healthy. Whether this translates into an underestimated rate of change depends on the dropout mechanism and the association between dropout and individual slopes. Specific analyses have found selection related more strongly to cognitive levels than to rates of change, underscoring that one broad assumption about attrition is insufficient. Salthouse (2019).
Missing-data assumptions and remedies
Missing completely at random means the probability of missing data is independent of observed and unobserved study variables. Missing at random means missingness can be explained by information in the observed data, under the analysis model. Missing not at random allows missingness to depend on unobserved outcomes even after accounting for recorded information. These are assumptions about the missingness mechanism rather than labels determined simply by the percentage missing. Testing baseline differences between completers and dropouts is informative, but it cannot prove the missing-at-random assumption. Little (1988); Newman (2010).
Full-information likelihood and multiple imputation can use observed information more efficiently than complete-case deletion when their assumptions are defensible. Inverse-probability weighting may reweight participants who remain observed using estimated retention probabilities. Pattern-mixture, selection, and joint longitudinal-survival models can probe sensitivity to informative loss and mortality. Each method can fail under poor model specification, unavailable predictors, weak overlap, or dependence on unmeasured variables. A credible paper explains missingness by wave and subgroup, records reasons for dropout, prespecifies the primary handling strategy, and compares conclusions under alternative plausible assumptions. Newman (2010); Galbraith, Bowden, and Mander (2017).
Mortality demands additional care because a measure such as daily social participation cannot be defined for the deceased. Estimands restricted to survivors differ from population-wide statements about those originally enrolled. Analyses may explicitly study time until death, consider terminal change among people approaching death, or define a composite outcome whose meaning is explained. Authors should identify which population and condition their estimates describe. Newman (2010).
Measurement Invariance: Is the Same Construct Being Measured?
Longitudinal comparisons depend on the meaning of measurement. A survey can contain identical words at every wave while respondents reinterpret them with changing roles, opportunities, health, or cultural expectations. For example, the practical meaning of autonomy for a newly employed 25-year-old may differ from that for an 85-year-old using accessible transportation or personal assistance. An item about employment-based productivity can change meaning following retirement. Without attention to this shift, changing scores risk conflating the attribute with its measurement context. van de Schoot, Lugtig, and Hox (2012).
Measurement invariance tests assess whether the structure relating observed items to a latent psychological construct is sufficiently comparable across occasions or groups. Configural invariance concerns the same general factor pattern; metric invariance concerns comparable factor loadings; scalar invariance concerns comparable intercepts or thresholds. The requirements depend on the intended comparison, and failure of one level changes what can be interpreted. Researchers may investigate partial invariance, model differential item functioning, use appropriate item-response methods, or explicitly limit comparisons. Statistical fit alone does not replace conceptual assessment of whether the construct remains meaningful. van de Schoot, Lugtig, and Hox (2012).
Protocol drift and mode changes
Over decades, instruments, staff, administration modes, questionnaire wording, technology, and diagnostic conventions can change. A transition from in-person testing to remote questionnaires may introduce mode effects. New hardware or altered instructions may shift scores. Researchers need calibration studies, bridge samples, recorded protocol versions, and sensitivity analyses that compare observations collected under different conditions. In long-term aging research, documenting such shifts is part of preserving construct validity, not merely a technical appendix. Newman (2010).
Retest and Practice Effects
Repeated exposure to a cognitive test can improve scores because participants recognize instructions, remember strategies, or become less anxious about the procedure. Longitudinal improvement can therefore partly reflect testing experience rather than improvement in the underlying ability. Practice effects can vary with age, retest interval, baseline ability, test type, and health. They complicate comparisons between repeated observations and a cross-sectional assessment of people taking the test for the first time. Salthouse (2010).
Common responses include alternate equivalent forms, comparison samples assessed for the first time, modeling the number of prior exposures, examining effects of retest intervals, and interpreting results alongside external indicators. These strategies have assumptions and can be difficult to implement across decades. An alternate form may differ psychometrically from the original, whereas a same-form test may produce learning. A study should state its chosen compromise and explore how much an apparent developmental trend may depend on repeated assessment. Salthouse (2010); Schaie and Caskie (2005).
Practice effects are especially important when discussing typical cognitive aging, but they do not transform an epidemiological panel into a clinical diagnosis. A longitudinal score change may suggest an area for further assessment; clinical evaluations of cognitive impairment and neurocognitive disorders rely on dedicated diagnostic standards, functional context, and clinical judgment. The methods considered here support developmental research and population-level inference rather than individual diagnosis.
Statistical Models for Developmental Change
Descriptive change and repeated-measures approaches
The simplest analyses compare scores at two occasions or examine average change among participants with both measurements. They are understandable but may lose information from additional waves, unequal observation intervals, and incomplete records. Repeated-measures approaches estimate population-level contrasts but require assumptions about dependence and missingness. Researchers should describe absolute as well as standardized differences, their uncertainty, and whether they reflect a clinically or practically important amount of change. Curran, Obeidat, and Losardo (2010).
Multilevel and mixed-effects growth models
Growth models estimate a trajectory over time and differences between participants in that trajectory. A mixed-effects model may specify a population-average intercept and slope along with participant-specific deviations. Predictors may be time-invariant, such as birth cohort, or time-varying, such as employment status or social participation at each wave. The model can be linear, nonlinear, piecewise, or based on splines. Its estimates depend on functional form, time coding, covariance structure, missing-data assumptions, and the quality of measures. Random slopes represent variation in estimated individual change, although noisy estimates should not be read as precise biographies. Curran, Obeidat, and Losardo (2010); Galbraith, Bowden, and Mander (2017).
Latent growth curves and structural equation models
Latent growth curve models represent change through latent intercepts and slopes or other trajectory factors. They can handle multiple indicators, measurement error, and links between developmental constructs within a structural equation framework. Their interpretation still depends on factor definitions, measurement invariance, fit, and appropriate control of confounding. The availability of sophisticated model outputs cannot compensate for inadequate wave timing, poor measurement, or selective follow-up. Curran, Obeidat, and Losardo (2010); van de Schoot, Lugtig, and Hox (2012).
Cross-lagged panels and within-person dynamics
Cross-lagged panel models use repeated measures of two constructs to estimate how earlier values relate to later values after accounting for previous scores. A conventional cross-lagged estimate can nevertheless combine stable differences between people with time-varying processes inside people. A longitudinal association between loneliness and well-being, for example, does not establish that a particular person's greater-than-usual loneliness would cause a later change in well-being. Random-intercept cross-lagged specifications can separate stable between-person variation from within-person deviations under their own assumptions. Neither specification creates randomization or establishes causality by itself. Hamaker, Kuiper, and Grasman (2015).
Time-to-event and joint modeling
Some research questions concern an event rather than a smooth psychological curve: first retirement, a caregiving transition, institutional admission, functional limitation, or death. Survival and event-history models investigate timing and risks, subject to censoring and competing-event assumptions. Joint models can connect repeated outcomes with event processes when participation or survival is informative. The research question should determine whether the estimand is a trajectory, an event probability, or a combined process. The same variables may play distinct roles in each analysis. Newman (2010).
Temporal Order Is Not Proof of Causation
Longitudinal observations reveal that one event or condition preceded another. Temporal precedence is necessary for many causal claims, but it is insufficient. Socioeconomic resources may influence both participation and health; an unmeasured illness may alter both mood and cognition; a historical shock may affect several outcomes together. Reverse causation may operate across multiple waves. Regression adjustments can also introduce bias if they inappropriately control variables affected by earlier exposure. Robins, Hernán, and Brumback (2000); Hamaker, Kuiper, and Grasman (2015).
The causal question should be written explicitly: whose outcome, under which exposure or intervention, during which period, compared with what alternative? Researchers can then use randomized interventions embedded in a cohort, natural experiments, target-trial emulation, instrumental-variable or fixed-effects strategies where assumptions are credible, and g-methods for certain forms of time-varying confounding. A marginal structural model, for instance, can address confounders affected by previous treatment under exchangeability, positivity, consistency, and adequately specified models. These approaches sharpen inference; they cannot guarantee identification when assumptions fail or essential variables are unmeasured. Robins, Hernán, and Brumback (2000).
Prediction is a separate goal. Baseline loneliness or functional impairment may improve prediction of later outcomes without proving that changing that predictor would change the outcome. Predictions should be evaluated in external samples and across relevant subgroups, using calibration and uncertainty, particularly when decisions could affect older adults' access to services. Evidence for an association, a prediction, and a causal effect should be reported under those respective labels.
What Major Adult-Aging Longitudinal Studies Illustrate
Seattle Longitudinal Study
The Seattle Longitudinal Study began in 1956 and developed a major body of evidence on adult cognitive development, combining successive samples and follow-up testing. Its design illustrates why cross-sectional and longitudinal cognitive patterns can diverge and why investigators must consider cohort membership, selective survival, changes in educational opportunity, and familiarity with repeated tests. Its findings cannot be condensed into a claim that all adults improve, decline, or remain stable on all abilities; different cognitive domains show different profiles, and the interpretation depends on the design and period studied. Schaie and Willis (2010); Schaie (2005).
Baltimore Longitudinal Study of Aging
The National Institute on Aging's Baltimore Longitudinal Study of Aging began in 1958 and follows volunteers through repeated assessments of physical and cognitive aging. It is well suited to detailed physiological and psychological measurement under a sustained research program. Volunteer recruitment and eligibility criteria matter when generalizing its observations to populations experiencing frailty, disability, institutional care, or difficult access to research centers. The study demonstrates the value of detailed multi-domain observation and the continuing need to interpret who the participants are. National Institute on Aging, Baltimore Longitudinal Study of Aging; National Institute on Aging, Join the BLSA.
Health and Retirement Study
The United States Health and Retirement Study began in 1992. It repeatedly surveys adults in later middle age and older adulthood, covering retirement, finances, family, psychosocial conditions, health, disability, cognition, and related exposures. It has introduced additional birth cohorts over time to maintain coverage of the aging population. This combination allows analyses of change within respondents while also examining differences across cohorts, resources, historical moments, and transitions. Its probability-sampling structure makes it valuable for population research, provided analysts use suitable design and longitudinal weights and account for nonresponse and selection. Health and Retirement Study, Design and Methodology; U.S. Office of Disease Prevention and Health Promotion, HRS data source.
English Longitudinal Study of Ageing
The English Longitudinal Study of Ageing follows people aged 50 and older in England and gathers information on health, financial circumstances, work, relationships, social participation, and cognition. Samples have been refreshed in several waves. Cross-national comparisons with U.S. cohorts offer opportunities to examine societal context, but equivalence of instruments, selection frames, wave schedules, policy environments, and analytic weights should be established before declaring that observed differences reflect country-level causes. English Longitudinal Study of Ageing, About ELSA; ELSA, Data and Documentation.
Interpreting Research Results Across Psychological Domains
Cognition and learning
Longitudinal cognition studies distinguish performance on reasoning, processing speed, vocabulary, memory, and other measures and can assess different patterns across domains. Some scores are vulnerable to practice effects, while others depend on education or cohort-specific experiences with tests. Investigators should interpret means, individual slopes, confidence intervals, and selective attrition together. A sample-average decline cannot by itself diagnose a neurocognitive disorder, and changes attributable to illness should be distinguished from typical adult variability. Salthouse (2010); Schaie and Willis (2010).
Personality and identity
Longitudinal evidence has shown that personality mean levels can change across adulthood. In a meta-analysis aggregating 92 longitudinal samples, Roberts and colleagues examined how trait mean levels varied across the life course. Average changes in traits need not imply that all individuals change similarly, that changes follow rigid stages, or that rank-order stability disappears. Questions about the causal sources and specific trajectories of personality belong to targeted research; this methodological article concerns what repeated measurement can and cannot establish. Roberts, Walton, and Viechtbauer (2006).
Emotional well-being and relationships
Mood, social participation, self-perceived age, relationship quality, and purpose may vary across different timescales. Longitudinal designs can test whether changes tend to coincide, whether earlier values predict later ones, and how transitions interact with existing resources. Interpretation should take seriously measurement invariance, self-report bias, contextual events, cultural differences, and within-person versus between-person distinctions. Higher average life satisfaction among one subgroup cannot establish a universal developmental stage, and historical inequalities can structure both opportunities and observed outcomes. Hamaker, Kuiper, and Grasman (2015); Ram and Gerstorf (2009).
Health, functional ability, and adaptation
Repeated health and functioning assessments can clarify sequences among activity, symptoms, environmental supports, and the ability to do personally valued activities. In adult aging research, health conditions, impairment, disability, accessibility, and psychological adaptation are related but distinct variables. A design focused only on disease absence will miss adaptation, assistive supports, personal priorities, and the influence of environment on autonomy. Investigators should measure these dimensions in ways that reflect the populations studied. Newman (2010); National Institute on Aging, BLSA.
Cultural, Historical, and Equity Considerations
Longitudinal adult-development research spans generations living through different educational systems, economic crises, technological transitions, public-health conditions, and retirement institutions. It also includes people with different migration histories, gender identities, languages, disabilities, relationships, and social positions. Models fitted to one national birth cohort may describe that cohort accurately but generalize poorly to another historical or cultural setting. Strong studies clearly state inclusion boundaries, report subgroup sample sizes, test measurement comparability, and avoid presenting group averages as individual destinies. Bell and Jones (2013); Newman (2010).
Ethical research requires consent processes suitable for repeated contact, protection of sensitive longitudinal records, transparent decisions about linkage and secondary use, accessibility in interviewing, appropriate proxy or supported participation where permitted, and respectful reporting. In aging research, recontact should not depend exclusively on digital literacy or travel capacity. Investigators also need clear procedures when participants experience illness, changing communication needs, or cognitive concerns. Ethical safeguards and participant support improve scientific quality by reducing avoidable selection. These recommendations reflect study-design reasoning, not claims that one recruitment method fits every population.
How to Evaluate a Published Longitudinal Study
A reader can evaluate a developmental claim by examining a coherent sequence of questions. First, identify the target population and how participants entered the cohort. Next, ask how many waves were observed, at what ages and historical dates, and what proportion of the original sample remains at each wave. Determine whether death, missed visits, and permanent withdrawal were distinguished. Check the instrument and whether it measured the same construct comparably across waves. Inspect the model's time axis, nonlinear assumptions, cohort adjustment, treatment of practice effects, and sensitivity to missingness. Finally, assess whether the conclusion is descriptive, predictive, or causal, and whether the evidence supports that category. Newman (2010); van de Schoot, Lugtig, and Hox (2012).
A convincing paper reports trajectory variation and uncertainty rather than only a statistically significant population mean. It presents sample flow, retention by subgroup, key eligibility criteria, recorded changes in protocols, and results under alternative reasonable analyses. It explains the measurement construct in ordinary psychological terms and indicates whether effects are large enough to matter in lived functioning. Replication across cohorts, cultures, and historical periods strengthens external credibility. Agreement across multiple methods is particularly informative when methods are vulnerable to different biases. Schaie and Caskie (2005); Curran, Obeidat, and Losardo (2010).
Evidence Status: Established, Preliminary, Limited, and Contested
Established methodological findings include the fundamental distinction between within-person change and cross-sectional age differences, the need to address attrition, the importance of measurement comparability, and the risk that repeated testing changes performance. The underlying reasons are methodological and are supported by longstanding methods literature and empirical aging studies. Newman (2010); Salthouse (2010); Hernandez et al. (2024).
Preliminary claims may involve a newly proposed pattern, a narrow cohort, a novel digital measurement strategy, or a statistical innovation validated primarily in simulations. Limited evidence may arise from small samples, two-wave data, unusually healthy volunteers, high selective attrition, weak cultural coverage, or inconsistent operational definitions. Contested questions include identification strategies for age–period–cohort decomposition, generalizations from particular cohorts to universal developmental stages, and causal interpretations of observational cross-lagged relations. Evidence status belongs to a specific claim and study design rather than to longitudinal research as a whole. Bell and Jones (2013); Hamaker, Kuiper, and Grasman (2015).
Frequently Asked Questions
Is a longitudinal study better than a cross-sectional study?
The best design depends on the question. Repeated observation is especially informative for estimating within-person trajectories, whereas cross-sectional research can efficiently describe differences in the population at a specific time. Longitudinal studies add threats from attrition and practice effects, and neither design automatically identifies causal aging effects. Mixed and sequential designs can provide complementary evidence. Schaie (2005).
How many waves are necessary to study adult development?
Two waves can estimate a difference across an interval. Three or more waves provide more information about nonlinear patterns, fluctuations, and changes in slope, although their adequacy depends on timing, measurement reliability, expected variation, and the statistical model. A longer study is not necessarily a stronger study if repeated measures are incomparable or most participants are lost. Curran, Obeidat, and Losardo (2010).
Can longitudinal studies separate aging from cohort effects?
Repeated observations and overlapping cohorts can test alternative explanations more effectively than a single cross-section. Nevertheless, unrestricted age, period, and cohort effects remain mathematically dependent, and any decomposition depends on assumptions, design structure, or external information. Results should make those assumptions explicit. Bell and Jones (2013).
Why do longitudinal results sometimes show less cognitive decline than age comparisons?
Participants can improve with familiarity or practice, while healthier or higher-performing people may be more likely to complete later waves. Cohorts may also differ in education and life experience. The contribution of each factor depends on the measure and sample and must be assessed rather than assumed. Salthouse (2010); Hernandez et al. (2024).
Does following people for decades establish cause and effect?
A long observation period can clarify temporal sequences, but unmeasured confounding, changing exposures, reverse causation, and selection remain possible. Causal inference requires a well-defined question and defensible research-design assumptions; randomized or quasi-experimental evidence may be necessary for stronger claims. Robins, Hernán, and Brumback (2000).
Are results about average aging applicable to any one person?
A population average describes the analyzed population under the stated study conditions. Individuals vary in resources, exposures, health, capacities, cultural histories, and adaptation. Individual prediction also requires separate evidence for calibration, uncertainty, and external validity. A group trajectory cannot be used as a personal diagnosis or as a deterministic life forecast. Newman (2010).
Conclusion
Longitudinal research makes adult development observable as a sequence of lived changes rather than an inference from different age groups at one moment. Its intellectual value depends on precise measurement of time, careful sampling, sustained follow-up, appropriate statistical analysis, and restraint in causal interpretation. Strong longitudinal science reveals the diversity of developmental pathways and shows how psychological change emerges within personal histories and social environments. Its conclusions become more reliable when cohorts, repeated assessment, missing data, mortality, and measurement meaning are treated as central scientific questions throughout the research process. Schaie (2005); Newman (2010).
Related Articles
Cluster Adult Development, Aging and Later Life — The lifespan perspective and adult-development research framework.
Adult Developmental Psychology: Theories, Stages, and Lifespan Change — Foundational theories and principles of change across adulthood.
Young Adulthood: Identity, Relationships, and Psychological Transitions — Developmental transitions and their social contexts.
Middle Adulthood: Developmental Changes, Roles, and Well-Being — Lifespan processes and heterogeneity at midlife.
Normal Cognitive Aging: Changes in Memory, Attention, and Learning — Population-level cognitive change and interpretation of age patterns.
Emerging Adulthood: Definition, Age Range, and Research Debates
Age, Period, and Cohort Effects in Aging Research
Selective Optimization With Compensation: The Baltes Model of Aging
References
Bell, A., and Jones, K. (2013). The impossibility of separating age, period and cohort effects. Social Science and Medicine, 93, 163–165. https://doi.org/10.1016/j.socscimed.2013.04.029
Curran, P. J., Obeidat, K., and Losardo, D. (2010). Twelve frequently asked questions about growth curve modeling. Journal of Cognition and Development, 11(2), 121–136. https://doi.org/10.1080/15248371003699969
English Longitudinal Study of Ageing. (n.d.). About ELSA. https://www.elsa-project.ac.uk/about-elsa
English Longitudinal Study of Ageing. (n.d.). Data and documentation. https://www.elsa-project.ac.uk/data-and-documentation
Galbraith, S., Bowden, J., and Mander, A. (2017). Accelerated longitudinal designs: An overview of modelling, power, costs and handling missing data. Statistical Methods in Medical Research, 26(1), 374–398. https://doi.org/10.1177/0962280214547150
Hamaker, E. L., Kuiper, R. M., and Grasman, R. P. P. P. (2015). A critique of the cross-lagged panel model. Psychological Methods, 20(1), 102–116. https://doi.org/10.1037/a0038889
Health and Retirement Study. (n.d.). Survey design and methodology. Institute for Social Research, University of Michigan. https://hrsonline.isr.umich.edu/index.php?jumpfrom=II&p=sdesign
Hernandez, R., Jin, H., Lee, P.-J., Schneider, S., Junghaenel, D. U., Stone, A. A., Meijer, E., Gao, H., Maupin, D., and Zelinski, E. M. (2024). Attrition from longitudinal ageing studies and performance across domains of cognitive functioning: An individual participant data meta-analysis. BMJ Open, 14(3), e079241. https://doi.org/10.1136/bmjopen-2023-079241
Little, R. J. A. (1988). A test of missing completely at random for multivariate data with missing values. Journal of the American Statistical Association, 83(404), 1198–1202. https://doi.org/10.1080/01621459.1988.10478722
Martin, M., and Hofer, S. M. (2004). Intraindividual variability, change, and aging: Conceptual and analytical issues. Gerontology, 50(1), 7–11. https://doi.org/10.1159/000074382
National Institute on Aging. (n.d.). The Baltimore Longitudinal Study of Aging. https://www.nia.nih.gov/research/labs/blsa
National Institute on Aging. (n.d.). Join the BLSA! https://www.nia.nih.gov/research/labs/blsa/join-blsa
Newman, A. B. (2010). An overview of the design, implementation, and analyses of longitudinal studies on aging. Journal of the American Geriatrics Society, 58(Suppl. 2), S287–S291. https://doi.org/10.1111/j.1532-5415.2010.02916.x
Ram, N., and Gerstorf, D. (2009). Time-structured and net intraindividual variability: Tools for examining the development of dynamic characteristics and processes. Psychology and Aging, 24(4), 778–791. https://doi.org/10.1037/a0017915
Roberts, B. W., Walton, K. E., and Viechtbauer, W. (2006). Patterns of mean-level change in personality traits across the life course: A meta-analysis of longitudinal studies. Psychological Bulletin, 132(1), 1–25. https://doi.org/10.1037/0033-2909.132.1.1
Robins, J. M., Hernán, M. A., and Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550–560. https://doi.org/10.1097/00001648-200009000-00011
Salthouse, T. A. (2010). Influence of age on practice effects in longitudinal neurocognitive change. Neuropsychology, 24(5), 563–572. https://doi.org/10.1037/a0019026
Salthouse, T. A. (2019). Attrition in longitudinal data is primarily selective with respect to level rather than rate of change. Journal of the International Neuropsychological Society, 25(6), 618–623. https://doi.org/10.1017/S135561771900016X
Schaie, K. W. (2005). What can we learn from longitudinal studies of adult development? Research in Human Development, 2(3), 133–158. https://pmc.ncbi.nlm.nih.gov/articles/PMC1350981/
Schaie, K. W., and Caskie, I. L. (2005). Methodological issues in aging research. In Handbook of Research Methods in Developmental Science. https://doi.org/10.1002/9780470756676.ch2
Schaie, K. W., and Willis, S. L. (2010). The Seattle Longitudinal Study of adult cognitive development. ISSBD Bulletin, 57(1), 24–29. https://pmc.ncbi.nlm.nih.gov/articles/PMC3607395/
U.S. Office of Disease Prevention and Health Promotion. (n.d.). Health and Retirement Study (HRS). Healthy People 2030. https://odphp.health.gov/healthypeople/objectives-and-data/data-sources-and-methods/data-sources/health-and-retirement-study-hrs
van de Schoot, R., Lugtig, P., and Hox, J. (2012). A checklist for testing measurement invariance. European Journal of Developmental Psychology, 9(4), 486–492. https://doi.org/10.1080/17405629.2012.686740