Longitudinal data analysis methods - paired tests, mixed-effects models, growth curves, and GEE - and how to choose the one that fits your waves.
Longitudinal data analysis measures change within the same units across waves — who moved, how fast, and in what shape. Analyze a longitudinal dataset with a single paired test and you throw away the structure that made it worth collecting. For the evaluators, researchers, and analysts who have to report what changed, not just what differs.
By Unmesh Sheth · Founder & CEO, Sopact · Updated May 2026
Longitudinal data analysis is the set of statistical methods used to measure change within the same units across repeated waves of data. Unlike cross-sectional analysis, which compares different groups at one moment, it models within-unit change over time — estimating how much each unit moved, how fast, and what shape the trajectory took.
Cross-sectional analysis answers what differs. Longitudinal data analysis answers what changed. The methods exist to use the one thing a longitudinal dataset has that a cross-sectional one does not: repeated measurements on the same units, connected wave to wave.
The old model treats analysis as a batch job: collect every wave, export, then model at the end. The redefinition pulls it forward. When each wave is read on arrival against a connected record, analysis runs continuously — the change after Wave 2 is estimated at Wave 2, not at the close of the study. The end-of-study model still runs, but it confirms what the team already acted on, rather than revealing it too late.
Every method on this page assumes one thing: a dataset where the same unit's waves are connected. That connection is an architecture decision made at collection, not a cleanup task before modeling. The full case is on the pillar: longitudinal design, redefined.
Longitudinal data analysis is not one method — it is a ladder. Two waves answer a narrow question with a simple test. Each wave added past that opens a richer question and a method that can answer it.
Paired t-test and the Wilcoxon signed-rank test for continuous outcomes, McNemar's test for binary outcomes, and change scores. They tell you the group moved between two points — not the shape of the move.
Repeated-measures ANOVA, and mixed-effects (multilevel) models that fit a line per unit. The slope estimates the rate of change, and the model uses every wave a unit answered — including the incomplete ones.
Latent growth curve models with random slopes, group-based trajectory modeling to find distinct change patterns in the population, and GEE for population-average effects.
More waves do not just add precision — they change the question you are allowed to ask. Two waves can never show a trajectory.
The method is only half the analysis. These four issues sit underneath every longitudinal dataset, and ignoring any of them quietly bends the result.
Units drop out, and rarely at random. A complete-case analysis describes only the units who stayed. Mixed-effects models and maximum-likelihood estimation use the partial records; multiple imputation is the alternative. The one wrong move is deleting them silently.
A longitudinal dataset holds two kinds of variation: how units differ from each other, and how each unit changes over time. The whole point of the analysis is the second. A method that collapses them reports a cross-sectional answer from longitudinal data.
Waves rarely land on a perfect schedule. Coding time as the actual elapsed interval, not the wave number, is what lets the model estimate a real rate of change. Unequal spacing is fine for a mixed model — as long as the spacing is recorded.
Answering the same instrument repeatedly can change the answer on its own — respondents get faster, more candid, or more fatigued. A change that is really a practice effect looks like a real outcome unless the design accounts for it.
R, Stata, SPSS, Mplus — every one of them runs a mixed-effects model. None of them can run it on a dataset where the same unit's waves are not linked. The hardest part of longitudinal data analysis is not the model; it is arriving at the model with a dataset that holds together.
No matching responses by name and email after collection. No twenty to forty percent of records failing to join. The dataset that arrives in R or Stata is already long-format and connected — so the analysis time goes to the model, not the cleanup.
Bring your dataset and the question you need to answer. We will walk through which method family fits the waves you have — and whether the data is connected enough to run it.
Longitudinal data analysis is the set of statistical methods used to measure change within the same units across repeated waves of data. It models within-unit change over time, estimating how much each unit moved, how fast, and what shape the trajectory took. It differs from cross-sectional analysis, which compares different groups at a single moment.
You analyze longitudinal data by choosing a method that matches the number of waves. With two waves, a paired test compares each unit's start and end. With three or more, a mixed-effects model fits a line per unit and estimates the rate of change. With many waves, growth curve models describe the shape of change. The method must use the within-unit structure, not collapse it.
The main methods are paired comparisons for two waves, repeated-measures ANOVA and mixed-effects (multilevel) models for three or more, latent growth curve models for the shape of change, group-based trajectory modeling to find distinct change patterns, and generalized estimating equations for population-average effects. The method is set by the number of waves and the question.
A mixed-effects model, also called a multilevel or hierarchical model, fits a separate line for each unit and a line for the population at the same time. It separates within-person change from between-person difference, and it uses every wave a unit answered, so units with incomplete records still contribute to the estimate.
A growth curve model is a latent-variable form of the mixed-effects model. It estimates the shape of change over time, linear or curved, and lets each unit have its own starting point and slope. The result describes trajectories, not just averages, which is what makes it suited to longitudinal data with several waves.
GEE, or generalized estimating equations, is a method that estimates the population-average effect in longitudinal data while accounting for the correlation between repeated measurements on the same unit. It is used when the question is about the average effect across the population rather than each unit's individual trajectory.
Use a paired t-test when a longitudinal dataset has exactly two waves and a continuous outcome. It compares each unit's value at the two points and tests whether the group moved. With three or more waves, a paired test throws away the extra waves; a mixed-effects model is the better choice.
Cross-sectional analysis compares different units at one point in time and answers what differs between groups. Longitudinal data analysis compares the same units across waves and answers what changed within them. Only longitudinal analysis can estimate a rate or a trajectory of change.
Units drop out of longitudinal studies, rarely at random, so missing data must be handled, not deleted. Mixed-effects models and maximum-likelihood estimation use the partial records a unit did provide. Multiple imputation is the common alternative. A complete-case analysis, which drops any unit with a missing wave, describes only the units who stayed.
Longitudinal data analysis is done in general statistical software: R, with packages such as lme4 and nlme, plus Stata, SPSS, SAS, and Mplus for latent growth models. The software runs the method; it cannot supply the structure. The dataset must already connect the same unit's waves before any model can run.
Longitudinal analysis is the analysis of data collected from the same units at more than one point in time, to measure change within those units. It is the same idea as longitudinal data analysis, used as the broader term. It contrasts with cross-sectional analysis, which examines a single point in time.
Two waves are enough to test whether something changed. Three to four waves are needed to estimate when the change happened and how fast. Five or more waves are needed to model the shape of a trajectory and to group units by trajectory pattern. The number of waves sets the ceiling on the question the analysis can answer.
A trajectory is the path a single unit's outcome takes across waves — rising fast then plateauing, declining steadily, improving late. Growth curve models estimate a trajectory per unit; group-based trajectory modeling sorts the population into a small number of distinct trajectory shapes, such as early gainers, late gainers, and non-responders.
This page covers the analysis methods. The pillar holds the redefinition and the structural definition; the four other siblings cover the concept, the comparison, the data, and the survey instrument.
A working session, not a demo. Bring the dataset you have and the question you need to answer. We walk through how many waves you have, which method family fits, what the missing data needs, and whether the dataset is connected enough to run the model at all. You leave with the method chosen and the data checked for it.
Live walkthrough · 60 min · with Unmesh Sheth, Founder & CEO · bring your dataset, wave count, and analysis question
