Psychology dissertation latent growth curve modelling examines average change and individual differences in change across repeated measurements within a structural equation modelling framework. It can answer questions such as whether anxiety decreases during therapy, whether students differ in their rates of recovery, and whether baseline support predicts those trajectories. A defensible analysis begins with the timing and measurement design, not with an attractive curve fitted after seeing the data.
This guide develops a hypothetical study in which adults report social anxiety at four planned assessments during a psychological intervention. It explains how to define intercept and slope factors, test the shape of change, handle measurement and missingness, add predictors without confusing association with causation, and report results transparently.
What psychology dissertation latent growth curve modelling estimates
A latent growth curve model treats repeated outcomes as indicators of unobserved growth factors. In a basic linear model, an intercept factor represents a chosen reference level and a slope factor represents systematic change per unit of time. Their means describe the sample’s average starting level and average rate of change. Their variances describe how much participants differ in those quantities.
The covariance between intercept and slope asks whether people with higher reference levels tend to change faster or more slowly. Its interpretation depends on where time zero is placed. If the intercept is centred at baseline, the covariance relates baseline status to subsequent change. If time zero is moved to the final wave, it relates final status to the trajectory leading there.
The official lavaan growth-curve tutorial represents the intercept and slope as continuous latent variables and fixes their loadings to encode time. This structure distinguishes an estimated change process from a simple comparison of wave means.
When a latent growth curve model fits the research question
Use a growth model when the primary question concerns systematic change across at least three repeated occasions and meaningful variation between people. Four or more waves usually provide greater scope for checking nonlinearity, although adequacy depends on the intended trajectory, timing, reliability, sample, and missingness.
The outcome may be a repeatedly measured scale score, behavioural measure, task response, or latent construct. Its distribution and measurement properties must match the estimator and model. A binary event, count, ordinal response, or heavily bounded score should not automatically be treated as normally distributed merely because the software accepts the data.
| Research question | Possible approach | Important distinction |
|---|---|---|
| What is the average trajectory and how do people differ? | Latent growth curve model | Models means, variances, and covariance of growth factors |
| How do repeated observations vary within people? | Multilevel growth model | Often similar for a simple linear specification, but organised differently |
| What is the population-average association over time? | Generalised estimating equations | Targets marginal regression rather than latent growth factors |
| Do two constructs predict one another between adjacent waves? | Cross-lagged model | Lagged associations are not the same as trajectories |
| Did the mean differ between two occasions? | Paired comparison | Does not estimate a multiwave change function |
The longitudinal study guide helps align wave timing, retention, and interpretation with the psychological process. The multilevel modelling guide explains the alternative random-effects formulation.
Define time before choosing the curve
Time scores must reflect the actual scientific clock. Assessments at weeks 0, 2, 6, and 12 should not be coded 0, 1, 2, and 3 unless equal spacing is a defensible approximation. For a linear trajectory, loadings of 0, 2, 6, and 12 make the slope a change per week. Dividing all scores by four changes the slope’s unit but not the fitted trajectory.
Place zero where the intercept answers a useful question. Baseline centring is intuitive for studies of initial severity. Centring at treatment completion can make the intercept represent expected post-intervention status. Avoid centring far outside the observed period because the intercept then becomes an extrapolation.
Planned occasions and actual assessment dates
Participants rarely complete every assessment on the intended day. Small deviations may be acceptable under planned time scores, but substantial variation can distort a common loading pattern. Record actual dates and examine their distribution. Models with individually varying times or a multilevel framework may be more appropriate when deviations are scientifically important.
Build the unconditional growth model first
An unconditional model contains no predictors of the growth factors. It describes the average trajectory, between-person variation, residual variation, and overall fit. This step is not a ritual. It establishes whether there is meaningful slope variation to explain and whether the assumed time function reproduces the observed means and covariances.
Intercept loadings
Loadings from the intercept factor to every repeated measure are normally fixed to one. This makes the factor a common reference level across occasions. The intercept mean is the expected outcome when the slope loading equals zero.
Linear slope loadings
Linear slope loadings equal the time scores. With waves at months 0, 1, 3, and 6, the loadings should normally be 0, 1, 3, and 6. The slope mean then estimates average change per month, assuming that a straight line is adequate over the observed period.
Residuals
Each repeated outcome has residual variance not explained by the growth factors. Equality constraints can be useful when justified, but they should not be applied only to reduce parameters. Adjacent residuals may remain associated because of occasion-specific influences or measurement continuity. Prespecify correlated residuals when theory supports them and check whether the sample can estimate the added complexity.
Choose the trajectory shape without curve chasing
Plot individual trajectories, observed means, and uncertainty before fitting a complex model. A linear model is a useful starting point, not a guaranteed truth. Psychological adaptation may be rapid early and slower later, while developmental change may accelerate or reverse.
| Shape | How it is represented | Main caution |
|---|---|---|
| Linear | One slope factor with loadings proportional to time | Can hide early gains followed by a plateau |
| Quadratic | Add a curvature factor using squared time scores | Needs enough waves and can extrapolate implausibly |
| Piecewise | Separate slopes before and after a meaningful knot | The knot should follow theory or design |
| Latent-basis | Fix selected loadings and estimate intermediate ones | Flexible loadings need careful scaling and interpretation |
| Parallel process | Estimate growth factors for two repeated outcomes | Correlated change does not establish causal direction |
The official Mplus growth-modelling examples provide separate specifications for linear, quadratic, piecewise, parallel-process, categorical, count, and individually varying-time models. These examples show why the outcome and timing determine the specification rather than one universal template.
Compare a small theory-led set of shapes using fit, residuals, parameter stability, plausibility, and useful predictions. A lower information criterion is not sufficient if the curve has an impossible interpretation or is driven by one assessment.
Protect longitudinal measurement
Change in a score is interpretable only if the construct remains comparable over time. Different instructions, response options, administration modes, translations, or item meanings can create apparent growth that reflects measurement change. Maintain consistent procedures and document unavoidable changes.
Observed totals versus multiple-indicator growth
A model using scale totals treats each wave score as an observed outcome. A multiple-indicator or second-order growth model first represents the construct with its items or parcels at each wave, then models change in the wave-specific latent factors. The latter can address measurement error more directly but is substantially more demanding.
For repeated latent constructs, evaluate longitudinal measurement invariance before interpreting factor-mean growth. The UCLA latent growth and measurement-invariance seminar demonstrates the sequence in lavaan. The site’s measurement invariance guide explains configural, loading, and intercept constraints in accessible terms.
Partial invariance may sometimes support limited comparisons when non-invariant parameters are identified and justified, but it is not permission to ignore broad measurement instability. Report each relaxation and evaluate whether substantive conclusions change.
Add predictors with a clear temporal role
A conditional growth model regresses the intercept and slope factors on covariates. A baseline characteristic such as age, treatment expectation, or prior symptom duration is time-invariant. Its slope coefficient asks whether that characteristic predicts the rate or shape of change, conditional on the other model variables.
A time-varying covariate changes across occasions, such as weekly stress or medication adherence. Its coefficient typically concerns occasion-specific deviation after accounting for the trajectory. Within-person and between-person meanings can be mixed if the variable is entered without decomposition. Centre and separate these components when the research question requires it.
Select predictors from theory, design, and temporal ordering. Univariate screening can omit confounders and retain colliders. Use the confounding variables guide to plan adjustment. Do not claim that a baseline predictor caused change merely because it predicts the slope in an observational study.
Plan sample size for the actual growth model
No universal participant rule is reliable. Information depends on the number and spacing of waves, outcome reliability, slope variance, residual structure, missingness, estimator, trajectory shape, predictor effects, and model complexity. A study may estimate an average slope reasonably but have little power to detect small individual differences in slopes.
Use Monte Carlo simulation with plausible means, variances, covariance, reliability, attrition, and non-normality. Evaluate convergence, improper estimates, bias, interval coverage, power, and precision across favourable and less favourable conditions. The power analysis guide explains why design-specific simulation is stronger than a generic subjects-per-parameter rule.
If simulation is infeasible, state that limitation and justify feasibility from prior data, pilot estimates, anticipated retention, model complexity, and the precision required for the primary growth parameter. Keep the primary model parsimonious.
Handle missing waves and attrition carefully
Distinguish a missing assessment from a true zero and document why each wave is absent. Describe missingness by occasion and baseline characteristics, and compare retention patterns without treating a non-significant test as proof of harmless attrition.
Full-information maximum likelihood can use the observed repeated outcomes under its missing-data assumptions. It does not correct attrition related to unobserved values after conditioning on model variables. Include defensible auxiliary variables when supported, consider multiple imputation that respects the longitudinal structure, and run sensitivity analyses for departures from the main assumptions.
Listwise deletion can reduce precision and change the analysed sample. Last-observation-carried-forward imposes an artificial flat trajectory and is rarely credible. The missing data guide provides a broader decision framework.
Fit and diagnose the model systematically
- Define the outcome, time origin, time unit, target population, and primary growth parameters.
- Audit dates, missingness, distributions, outliers, and individual trajectories.
- Fit a theory-led unconditional model and inspect means, variances, covariance, residuals, and convergence.
- Compare a limited set of plausible trajectory shapes.
- Evaluate longitudinal measurement when repeated latent constructs are used.
- Add prespecified predictors and clarify time-invariant and time-varying roles.
- Inspect global fit, local residuals, influence, estimates, and uncertainty.
- Run planned sensitivity analyses and preserve reproducible syntax.
Inspect warnings, gradients, residual variances, latent variances, standard errors, and correlations. A negative residual variance, impossible correlation, non-positive definite covariance matrix, or failure to converge is an improper solution, not a publishable result rescued by favourable fit indices.
Evaluate global and local fit
Consider the model chi-square, comparative fit index, Tucker-Lewis index, root mean square error of approximation, standardised root mean square residual, and information criteria where comparisons are valid. Do not turn conventional cut-offs into a binary quality certificate. Inspect residual means and covariances to locate misfit.
Columbia University’s latent growth curve analysis overview emphasises multiple fit indices and warns that good fit does not imply a causal connection. Fit supports a model’s consistency with selected features of the data, not its unique truth.
Check influential cases and distributional assumptions
Plot estimated trajectories and case-level patterns. One unusual participant can affect slope variance or curvature, especially in a modest sample. Verify the source data before considering exclusion, retain valid cases under the primary rules, and report sensitivity results. The outlier analysis guide supports transparent decisions.
Interpret growth parameters accurately
| Parameter | Useful interpretation | Avoid |
|---|---|---|
| Intercept mean | Average expected outcome at the chosen time origin | Calling it baseline when zero is elsewhere |
| Slope mean | Average change per stated time unit under the specified shape | Calling it each person’s change |
| Intercept variance | Between-person variation at the reference time | Interpreting variance as an effect |
| Slope variance | Between-person variation in rates of change | Assuming distinct trajectory classes |
| Intercept-slope covariance | Association between reference level and change | Ignoring the time origin |
| Predictor of slope | Conditional association with trajectory change | Claiming causality without design support |
Present predicted trajectories at meaningful predictor values with confidence intervals. State whether higher scores mean improvement or deterioration. A negative anxiety slope may indicate improvement, while a negative wellbeing slope may indicate worsening. Translate estimates into the instrument’s units and avoid extrapolating beyond observed time.
Psychology-specific worked example
Suppose 240 adults complete a social-anxiety scale at baseline and weeks 2, 6, and 12 during a guided intervention. The primary question is whether symptoms change and whether baseline perceived support predicts the rate of change. Actual timing is close to schedule, so loadings use 0, 2, 6, and 12.
Plots suggest rapid early improvement followed by slower change. The researcher compares prespecified linear and latent-basis models. The latent-basis model fits the observed means and covariances better without improper estimates. The slope mean is negative, indicating average improvement, and the slope variance is positive with a sufficiently precise interval, indicating meaningful differences in change.
Baseline support predicts a more negative slope after adjustment for baseline age and prior treatment, but the result is described as an association. Predicted curves show the expected pattern at low, typical, and high support values. Sensitivity analyses using a linear curve and alternative missing-data assumptions preserve the direction but vary the magnitude.
The dissertation reports assessment timing, attrition, score reliability, time loadings, estimator, fit, residual checks, predictor coding, intervals, sensitivity analyses, and limits arising from self-selection and unmeasured treatment engagement.
Report latent growth curve modelling transparently
In Methods, report the design, sample flow, wave schedule and actual timing, outcome construction, measurement invariance procedure where relevant, missing-data approach, time origin and loadings, trajectory form, residual constraints, estimator, growth-factor predictors, software and version, diagnostics, and sensitivity analyses.
In Results, report observed wave summaries, correlations, missingness, model comparison rationale, fit indices, intercept and slope means, their variances and covariance, residuals, predictor coefficients with confidence intervals, predicted trajectories, improper-solution checks, and robustness findings. Provide syntax and a labelled path diagram when permitted.
Common latent growth curve mistakes
- Coding unequally spaced waves as equally spaced without justification.
- Calling the intercept baseline after centring time elsewhere.
- Choosing curvature only because it improves one fit statistic.
- Ignoring longitudinal measurement invariance for repeated latent constructs.
- Adding many slope predictors when slope variation is weak or imprecise.
- Confusing time-varying within-person effects with between-person differences.
- Treating full-information estimation as protection from every attrition mechanism.
- Ignoring negative variances, extreme correlations, or convergence warnings.
- Interpreting average growth as every participant’s trajectory.
- Claiming that predictors caused change in an observational design.
Frequently asked questions
How many waves do I need for a latent growth curve model?
At least three repeated occasions are normally needed to identify a simple linear trajectory, but four or more provide better scope for testing shape and residual assumptions. Requirements rise with curvature, multiple processes, or latent measurement models.
Is latent growth curve modelling the same as multilevel modelling?
They can yield similar estimates for equivalent simple models. The growth-curve approach uses an SEM representation and readily combines measurement and structural models, while multilevel modelling often handles irregular timing and long-format data naturally.
What does a significant slope variance mean?
It indicates evidence of between-person variation in the specified change parameter. It does not prove that discrete trajectory classes exist or explain why people differ.
Should I always fit a quadratic curve?
No. Fit a shape supported by theory, timing, plots, wave count, and precision. Quadratic models can be unstable and can imply unrealistic change outside the observed period.
Can I use latent growth modelling with missing waves?
Yes, suitable likelihood or imputation methods can use incomplete records under stated assumptions. Investigate attrition, include defensible predictors of missingness, and assess sensitivity to plausible departures.
Do good fit indices prove the growth model is correct?
No. They assess selected discrepancies under one specification. Inspect local fit, estimates, assumptions, alternatives, theory, and prediction, and do not infer causation from fit.
Which software can estimate latent growth curves?
Common options include lavaan in R, Mplus, OpenMx, Amos, and other SEM software. Report the program, version, estimator, syntax, constraints, and any treatment of missing or non-normal data.
Conclusion
Psychology dissertation latent growth curve modelling is strongest when the research question, measurement schedule, time coding, and trajectory form are aligned before estimation. A credible dissertation distinguishes average change from individual differences, protects longitudinal measurement, treats missingness explicitly, checks local as well as global fit, and communicates predicted trajectories without causal overstatement.
If you need statistical tutoring, choose support that helps you understand and conduct your own analysis. Protect participant data, follow ethics approval and academic-integrity rules, keep an auditable analysis record, and remain responsible for every model choice and conclusion.
