Longitudinal Design

Longitudinal Design: Concept Architecture

Orientation and learning roadmap

A longitudinal design makes time part of the study structure by linking repeated observations from the same people, organizations, practices, or other units. The sections below explain what the design can reveal, how to define its cohort and measurement schedule, why repeated observations are correlated, and how attrition, missing data, time-varying exposures, and analytical choices affect health-economic interpretation.

What makes a design longitudinal

The defining feature is not simply that data were collected over a long calendar period. A design is longitudinal when observations are linked within the same study units across two or more time points, allowing change within those units to be distinguished from differences between units. Follow-up can last days, months, or decades depending on the process being studied.

• A longitudinal design identifies each study unit consistently across measurement occasions.
• A longitudinal design specifies a meaningful time origin, such as enrolment, diagnosis, treatment initiation, discharge, or a policy change.
• A longitudinal design records repeated outcomes, exposures, covariates, or resource-use measures on a defined schedule or when events occur.
• A longitudinal design preserves the timing and ordering of measurements rather than treating repeated records as independent observations.
• A longitudinal design may be observational or experimental; repeated follow-up within a randomized trial is also longitudinal.

What longitudinal data can show

Repeated observation makes it possible to describe trajectories and establish whether a measured exposure precedes a later outcome. It can separate a population difference at one time from change occurring within the same units, and it can compare rates or patterns of change between groups. Temporal ordering strengthens some inferences but does not by itself establish causality.

Longitudinal questions commonly concern:

• how health-related quality of life changes before and after treatment;
• how costs and healthcare use accumulate over disease progression;
• whether treatment groups follow different outcome trajectories;
• how an earlier exposure is associated with a later outcome;
• when transitions, relapses, admissions, treatment changes, or deaths occur;
• whether inequalities widen, narrow, or remain stable over time.

The design can answer these questions only when measurement timing, follow-up, confounding, and missingness are aligned with the intended estimand. More waves of data add information, but they do not repair a poorly defined time origin or a comparison group that cannot support the question.

Longitudinal design compared with nearby designs

Several designs include dates or repeated samples but do not provide the same information. The key distinction is whether records are linked within the same units over time and whether the analysis targets within-unit change. Clear terminology prevents a repeated cross-sectional dataset from being interpreted as individual follow-up.

A cohort is often longitudinal, but the terms are not exact synonyms. Cohort language emphasizes membership and follow-up of a defined group, while longitudinal design emphasizes repeated linked observation and change over time. A study may also combine cohort follow-up, record linkage, repeated surveys, and event-based measurements.

Defining the cohort, time origin, and target population

The study population should be defined in relation to the question and the moment follow-up begins. Eligibility rules applied using future information can introduce selection bias, while inconsistent entry dates can create incomparable risk periods. A well-specified design distinguishes the source population, enrolled or identified cohort, analysis population, and population to which results are intended to apply.

• The eligibility date determines when each unit first qualifies for the study.
• The index date provides a reproducible time origin for baseline covariates and follow-up.
• The baseline window states how far before or around the index date measurements may be taken.
• The follow-up rule states when observation ends because of an outcome, death, disenrolment, data loss, study end, or another event.
• The target population states whose trajectories or effects the study seeks to represent.

Staggered entry is acceptable when each unit's study time is measured from a meaningful origin. Calendar time may still need adjustment because treatment availability, prices, clinical practice, coding, and policy can change while different participants are under observation.

Planning measurement occasions

Measurement timing should match the expected pace of clinical and economic change. Widely spaced observations may miss short-lived effects or misclassify event timing, while very frequent measurement can increase burden, cost, and missingness. Scheduled assessments and measurements triggered by clinical events should be distinguished because they arise through different observation processes.

The protocol should specify:

• which outcomes, exposures, confounders, resource-use items, and patient-reported measures are collected at each occasion;
• whether time is represented by visit number, time since baseline, age, calendar date, time since treatment, or another scale;
• the permitted window around each scheduled occasion and how off-schedule observations are handled;
• whether measurements are equally or unequally spaced and whether analysis uses exact elapsed time;
• how changes in measurement instruments, coding systems, prices, or data sources are harmonized;
• which events trigger additional observation and whether event-triggered data could be informative.

The design should collect important variables before they are needed analytically. A confounder measured only after exposure may be affected by that exposure and cannot automatically be treated as an ordinary baseline adjustment variable.

Why repeated observations are correlated

Measurements from the same individual tend to be more alike than measurements from different individuals. This within-unit correlation reduces the amount of independent information relative to a dataset with the same number of observations from unrelated units. Analyses that ignore clustering can produce incorrect standard errors, confidence intervals, and tests, and may also misrepresent the mean trajectory.

For outcome (Y_{it}) measured for unit (i) at time (t), a simple random-intercept mixed model can be written as:

[ Y_{it}=\beta_0+\beta_1 t_{it}+\beta_2 X_i+\beta_3(t_{it}X_i)+b_{0i}+\varepsilon_{it} ]

Here, (\beta_1) describes change over time in the reference group, (\beta_2) describes the baseline group difference when time is coded zero at baseline, and (\beta_3) describes the difference in trajectory associated with (X_i). The unit-specific random intercept (b_{0i}) represents persistent differences between units, while (\varepsilon_{it}) represents within-unit residual variation.

The model is illustrative rather than universal. Nonlinear trajectories, binary or count outcomes, varying slopes, irregular times, multiple levels of clustering, and informative observation processes require models suited to those features.

Population-average and unit-specific questions

Longitudinal analyses can target different kinds of effects. Marginal approaches such as generalized estimating equations commonly describe average changes or associations across the population, while mixed-effects models represent population-level coefficients together with unit-specific random effects. These approaches can answer related but not always numerically identical questions, especially for nonlinear outcome models.

Method selection should begin with the estimand, outcome distribution, timing, and causal structure rather than software availability. A statistically sophisticated model cannot compensate for an ill-defined question or measurements taken after the causal decisions they are meant to inform.

Time should be represented deliberately

Coding time as 0, 1, 2, and 3 assumes that visit numbers are equally spaced and that a one-visit difference has constant meaning. Exact elapsed time is often preferable when visits occur irregularly. Nonlinear change can be represented using transformations, polynomials, splines, piecewise functions, or clinically meaningful phases, but added flexibility should be justified and validated.

Centering time changes coefficient interpretation without changing the underlying fitted values. Setting baseline to zero makes the intercept a baseline value; centering at 12 months makes the intercept the expected value at month 12. Analysts should state the unit, origin, functional form, and interactions used for time so that model coefficients can be interpreted correctly.

A worked trajectory example

Consider a prospective study measuring health-related quality of life at baseline, 6 months, and 12 months in two treatment groups. A linear mixed model codes time in years, treatment as 0 for comparator and 1 for the new treatment, and includes a treatment-by-time interaction. Suppose the estimated coefficients are (\beta_0=0.68), (\beta_1=-0.06), (\beta_2=0.02), and (\beta_3=0.04).

The comparator group's expected value at one year is:

[ 0.68+(-0.06)(1)=0.62 ]

The new-treatment group's expected value at one year is:

[ 0.68+(-0.06)(1)+0.02+(0.04)(1)=0.68 ]

The fitted difference at one year is 0.06, composed of a 0.02 baseline difference and a 0.04 difference in annual change. The example does not establish a causal treatment effect unless treatment assignment, confounding, missingness, measurement, and model assumptions support that interpretation.

Missing observations and attrition

Missing data are especially consequential in longitudinal studies because participation and observation can change over time. Intermittent missingness occurs when a unit misses one occasion but returns later, while monotone dropout occurs when no later measurements are observed. Attrition can bias trajectories when the probability of remaining observed is related to health, treatment response, costs, or other factors relevant to the outcome.

The analysis should describe missingness by wave, outcome, treatment or exposure group, and reason where available. Complete-case analysis is rarely neutral because it changes the analysis population and is valid only under restrictive conditions. Likelihood-based models, multiple imputation, inverse-probability weighting, joint models, and sensitivity analyses each rely on assumptions that should be stated rather than hidden behind the method name.

Common missing-data frameworks include:

• Missing completely at random means missingness is unrelated to observed or unobserved data relevant to the analysis.
• Missing at random means missingness may depend on observed data but, conditional on those data, not on the missing values themselves.
• Missing not at random means missingness may depend on unobserved values even after conditioning on observed data.

The missing-at-random assumption cannot be confirmed from observed data alone. Pattern-mixture, selection-model, tipping-point, or other sensitivity analyses may be needed when departures could materially alter clinical or economic conclusions.

Time-varying exposure and confounding

In many longitudinal studies, treatment, disease severity, employment, insurance, adherence, and other variables change during follow-up. A time-varying confounder can both predict later treatment and be affected by earlier treatment, creating a feedback structure that standard adjustment may handle incorrectly. A causal diagram and explicit time ordering help determine which variables should be measured and how they may be used.

Analysts should distinguish:

• baseline confounders measured before the exposure strategy begins;
• time-varying confounders updated during follow-up;
• mediators through which earlier exposure affects later outcomes;
• colliders that should not be conditioned on without a valid reason;
• time-varying treatments or exposures whose history defines the strategy of interest.

Methods such as marginal structural models, the parametric g-formula, or targeted longitudinal estimators may be appropriate for carefully defined causal questions. They require consistency, exchangeability, positivity, correct time ordering, and adequate models or estimation procedures; longitudinal data alone do not satisfy these assumptions.

Biases created by time and follow-up

The time dimension creates opportunities for bias that may not be obvious in a single-wave dataset. Eligibility, treatment classification, outcome ascertainment, and censoring must be aligned to avoid assigning follow-up incorrectly. Bias assessment should cover both ordinary observational-study concerns and mechanisms specific to repeated follow-up.

• Immortal time bias arises when a period during which the outcome could not occur for a participant to be classified as exposed is incorrectly credited to the exposed group.
• Time-window bias arises when groups have different opportunities for exposures or outcomes to be recorded.
• Informative censoring arises when loss of observation is related to the outcome process after accounting for included information.
• Regression to the mean can make unusually high or low baseline values appear to improve on repeat measurement even without a true intervention effect.
• History and secular trends can produce before-and-after changes unrelated to the exposure or intervention of interest.
• Measurement drift can create apparent change when instruments, coding, observers, or data systems change over time.
• Survivor bias can make later observations describe a progressively selected subset of the original cohort.

Design choices should prevent these problems where possible. Analytical adjustment is less reliable when the design has not captured the information needed to reconstruct comparable risk periods and observation processes.

Longitudinal evidence in health economics

Longitudinal data can connect changes in health, treatment, resource use, productivity, and quality of life within the same individuals. This makes them valuable for estimating cost and outcome trajectories, transition patterns, persistence, adherence, and heterogeneity over time. Their relevance depends on whether the observed pathway matches the decision problem and target population.

• Repeated utility measurements can support quality-adjusted life-year estimation when timing and interpolation between observations are specified appropriately.
• Repeated cost and utilization records can reveal skewed, zero-heavy, and clustered expenditure trajectories that a single annual total may conceal.
• Patient histories can inform state-transition probabilities, recurrent-event risks, treatment switching, and disease-progression models.
• Linked administrative records can extend follow-up and reduce participant burden but may introduce coding changes, incomplete capture, and linkage error.
• Longitudinal observational evidence can inform model inputs and external validity but does not inherit the protection of randomized treatment assignment.

Economic analyses should preserve correlations among repeated costs and outcomes and address censoring caused by death or loss to follow-up. Costs after death are structurally absent rather than ordinary missing values, and mortality may truncate both cost and quality-of-life trajectories in ways that require joint interpretation.

From repeated utility to quality-adjusted life-years

Repeated health-state utility measurements can be converted into quality-adjusted survival by estimating the area under the utility-versus-time curve. With measurements at adjacent times (t_j) and (t_{j+1}), the trapezoidal contribution is based on the average of the two utilities multiplied by the interval length. This calculation assumes a linear path between observations unless another method is specified.

[ QALY_j=\frac{u_j+u_{j+1}}{2}(t_{j+1}-t_j) ]

For utilities 0.70 at baseline and 0.60 at six months, with time expressed in years, the contribution over the first half-year is:

[ QALY_0=\frac{0.70+0.60}{2}(0.5)=0.325 ]

This value covers only the observed half-year interval and is not an annual QALY estimate. Death, missing utility, unequal intervals, baseline imbalance, and measurements taken after events require explicit handling, and uncertainty in utility trajectories should be propagated into the economic analysis.

A disciplined design and analysis workflow

A strong longitudinal study connects the research question, observation schedule, causal or descriptive estimand, and analysis plan before data are examined. Each stage should preserve the identity and timing of units while documenting changes in eligibility, exposure, measurement, and follow-up. The workflow remains iterative because feasibility checks may reveal that the available data cannot support the intended question.

1. Define the estimand. State the population, exposure or intervention strategy, outcome, time horizon, comparison, and summary measure of change or effect.
2. Choose the time origin. Align eligibility, baseline covariates, exposure classification, and the start of outcome follow-up.
3. Specify the observation process. Set scheduled waves, permissible windows, event-triggered measurements, and the end-of-follow-up rule.
4. Map the causal and temporal structure. Identify baseline and time-varying confounders, mediators, competing events, and variables affected by prior exposure.
5. Plan retention and linkage. Minimize avoidable attrition, maintain identifiers securely, and document record-linkage quality.
6. Pre-specify the analysis. Define the time scale, functional form, correlation structure, estimand-specific model, and missing-data strategy.
7. Check data integrity. Test identifiers, duplicate records, date ordering, impossible transitions, measurement units, and consistency across waves.
8. Describe follow-up. Report observation counts, person-time, missingness, attrition, reasons for loss, and the characteristics of retained and lost units.
9. Estimate trajectories or effects. Use a method that reflects within-unit correlation, outcome distribution, clustering, and the intended interpretation.
10. Evaluate robustness. Examine plausible missing-not-at-random mechanisms, time functions, covariance structures, exposure definitions, and censoring assumptions.
11. Connect results to decisions. Translate estimates into the model inputs or policy quantities required, preserving uncertainty and population relevance.
12. Report transparently. State the design, cohort construction, timing, measurements, methods, limitations, and deviations from the plan.

Validation checks for longitudinal datasets

Longitudinal validation must examine relationships across records, not only the validity of each row. A dataset can contain plausible individual values while still having impossible time order, duplicate waves, or broken identifiers. Checks should run before analysis and after every material linkage or reshaping step.

• Each unit should have one stable identifier and no accidental merging of different people or splitting of one person across identifiers.
• Dates should follow a valid sequence relative to eligibility, index date, treatment, outcome, death, and end of follow-up.
• Each unit and scheduled occasion should have the permitted number of records, with duplicates resolved using a documented rule.
• Time-varying values should respect logical constraints, such as no resource use after a verified death unless the record represents post-death administrative cost.
• Measurement units, instruments, coding versions, price years, and data provenance should be harmonized or explicitly modelled.
• Attrition tables should reconcile the baseline cohort with observed, censored, deceased, withdrawn, and otherwise missing units at each wave.
• Derived durations and cumulative totals should reconcile with source dates and event records.
• Reshaping between wide and long formats should preserve row counts, keys, values, and missingness patterns.

Reporting the design and results

Transparent reporting allows readers to judge whether the observed trajectories answer the stated question. The report should identify the design early, describe participant or unit flow, and state how repeated measurements and losses were handled. Relevant reporting guidance should be used for the underlying study type, such as cohort or randomized-trial guidance, rather than treating “longitudinal” as a complete reporting label.

At minimum, report the setting and dates, eligibility rules, recruitment or data-source procedures, time origin, measurement schedule, variable definitions, sample-size basis, statistical methods, missing-data assumptions, attrition, follow-up time, and sensitivity analyses. Present estimates with uncertainty and show the number observed at relevant occasions. Graphs of individual or group trajectories can improve understanding when they do not conceal sparse follow-up or selective retention.

Common errors and safeguards

Longitudinal datasets often look information-rich because they contain many rows, but the effective evidence depends on the number of independent units, follow-up quality, and observation process. Treating rows as independent or time as an ordinary untimed covariate can create precise-looking but invalid results. Design documentation and analysis diagnostics should remain linked to the decision question.

• Do not describe repeated samples of different people as individual longitudinal follow-up.
• Do not count every row as an independent participant when calculating precision or sample size.
• Do not adjust for post-exposure variables without considering whether they are mediators, colliders, or time-varying confounders.
• Do not interpret temporal ordering alone as proof of causation.
• Do not treat dropout as harmless because baseline characteristics appear similar among completers and non-completers.
• Do not assume missing-at-random mechanisms are verified by the observed data.
• Do not code irregular visits as equally spaced waves when exact timing matters.
• Do not average trajectories across survivors without explaining how death and censoring were handled.
• Do not apply ordinary least squares with independent errors to repeated measurements without a justified correlation strategy.
• Do not ignore changes in instruments, coding, clinical practice, or prices across calendar time.

Interpreting longitudinal findings responsibly

Longitudinal evidence can reveal change and establish the order in which measured events occurred, but its interpretation remains conditional on selection, measurement, confounding, missingness, and model assumptions. A fitted trajectory is a summary of observed and modelled information, not a record of what would have happened under an alternative exposure unless a causal estimand and its identifying assumptions are justified. Results should distinguish descriptive change, prediction, association, and causal effect.

For health-economic decisions, the strongest longitudinal analysis connects clinically meaningful time to costs and health outcomes while retaining uncertainty and population relevance. When attrition, time-varying confounding, or observation gaps could alter the decision, those issues should be represented through sensitivity analysis rather than relegated to a general limitation statement.