Longitudinal data analysis

A longitudinal analysis is needed when the same units of analysis are measured repeatedly over time and the question is how they change, because those measurements are not independent of one another. It applies to any follow-up design, such as a clinical registry, a research cohort, an intervention study with successive visits or a population panel, rather than to any particular source. The sign that a cross-sectional analysis falls short is that observations from the same unit are correlated: treating them as isolated data points overstates confidence in the results and wastes the temporal information in the follow-up. Longitudinal analysis models that dependence instead of ignoring it, and separates the average trajectory from the variability between units. If your data are a single snapshot, a cross-sectional analysis is enough; if they are a film of the same units, they need longitudinal methods.

Years of systematic data collection are an investment that can become publishable evidence and informed decisions, rather than remaining an underused repository. This applies when an organisation has invested resources in following the same units in a structured way and the scientific return has not been proportionate to that effort. The analysis starts from the data already collected to describe the natural history of a disease, study the changes associated with interventions in real-world practice and, when the design allows, estimate their effectiveness with the caveats that apply to observational studies, or to produce publications. Collecting data for years without an analysis plan produces work, not evidence; the service closes that gap. The dataset goes from underused repository to scientific asset, with defensible questions, analyses and results.

Not only registries. The service applies to any follow-up structure in which the same units of analysis are measured repeatedly: clinical registries from routine care, research cohorts set up for a specific question, intervention studies with pre, post and follow-up measurements, and population panels with successive waves. The units need not be patients: they can be participants, centres, areas, devices or populations. What these sources share, and what defines the service, is the structure of the data, not where they come from: repeated measurements, dependence between observations from the same unit and, often, loss to follow-up and recurring events. The clinical registry is the most common source in health research, but the same methods support an intervention study with follow-up or a pragmatic trial with repeated visits.

At both stages, and each one addresses a different need. At the design stage, the population, variables, follow-up frequency and quality checks are defined before collection begins, which avoids structural limitations (poorly defined variables, an unsuitable follow-up frequency) that become irreversible once the data have been collected. With follow-up already under way or mature, the existing data are analysed and, often, limitations that can still be corrected for future collection are identified in time; the first step is to assess completeness, longitudinal consistency and fit with the question. The service designs the structure and the quality criteria, analyses and interprets; routine data entry, technical maintenance of the platform and systematic collection belong to the team that runs the follow-up. We do not take on the day-to-day administration of the system.

By first characterising the missingness mechanism and then applying a method consistent with it, not by imputing values without an explicit assumption about why the data are missing. This applies to most longitudinal designs, where dropout and missed measurements are inevitable over time. The first step is to analyse the pattern of missingness: MCAR if the data are missing completely at random, MAR if missingness depends on observed variables, and MNAR if it is not at random. Each assumption calls for different handling. From there, techniques such as multiple imputation or models estimated by maximum likelihood are used, depending on the design and the plausible mechanism, with sensitivity analyses that test whether the conclusions hold under different assumptions. Analysing complete cases only can bias the results without the bias being visible.

Models that respect the fact that observations from the same unit are not independent, not analyses that treat each measurement as an isolated data point. This applies whenever follow-up tracks the same units and the question is how they change. For trajectories of repeated variables, linear and generalised linear mixed-effects models and growth curve models are used; they treat time as continuous and characterise the average trajectory and its variability. When the design has fixed visits, MMRM models time as a categorical factor. Generalised estimating equations (GEE) are an alternative when the interest lies in the population-average effect. For events that can recur in the same unit, such as admissions or relapses, recurrent event models such as Andersen-Gill can be used. The choice depends on the nature of the variable and on the question.

They are different frameworks for data that change over time. Longitudinal analysis follows several units of analysis measured repeatedly and models how they change and how they differ from one another. Time series analysis works on one or a few aggregate metrics ordered in time (monthly incidence, total hospital admissions) and focuses on trend, seasonality and forecasting. Survival analysis measures the time until an event occurs and the factors that speed it up or delay it. Real-world evidence (RWE) is not a data structure but a type of evidence generated from real-world data. RWE can draw on longitudinal data, but is defined by the source and the context in which the data are reused, not by the shape of the data. They can be combined in the same project, but they do not answer the same question.

Results and interpretation ready to go into a manuscript or a report: interpreted statistical results, tables, figures and a reading of the findings, not just software output. This applies both if you aim to publish in high-impact journals and if you need to support a funding application, or the renewal of a follow-up programme, with evidence of its scientific output. The work is planned from the outset around the methodological standards that journals demand, which helps anticipate reviewers' objections, and extends to preparing the response to reviewers with a technical argument. The research team authors and leads the publication; the service provides the design, the analysis and the statistical interpretation that support it. What you receive is material ready for scientific use, not tables that have to be reinterpreted.