Psychology dissertation cross-lagged panel model analysis examines whether earlier values of one construct predict later values of another while modelling their prior levels. It is often used for questions about reciprocal relations, such as whether sleep problems precede anxiety, anxiety precedes sleep problems, or both processes occur. A defensible dissertation must define which level of change the model represents, protect longitudinal measurement, align the lag with theory, and avoid treating temporal prediction as proof of causation.
This guide follows a hypothetical three-wave study of sleep problems and anxiety in university students. It explains the traditional cross-lagged panel model, the random-intercept alternative, measurement and timing decisions, missing data, power, model checks, interpretation, reporting, and sensitivity analyses. The example is illustrative, so coefficients and conclusions must come from the researcher’s actual data and design.
What a cross-lagged panel model estimates
A cross-lagged panel model, or CLPM, is a structural equation model for repeatedly measured variables. A basic bivariate model includes autoregressive paths, cross-lagged paths, within-wave covariances, and residual variances. Each part answers a different statistical question.
- Autoregressive paths relate a construct to itself at the next wave, conditional on the other predictors.
- Cross-lagged paths relate one construct to the other construct at the next wave, conditional on prior values in the model.
- Within-wave covariances represent association at the same occasion that is not given a directional interpretation.
- Residuals contain variation not explained by the lagged predictors.
In the student example, the path from wave-one sleep problems to wave-two anxiety asks whether sleep predicts later anxiety after accounting for wave-one anxiety and the other specified relations. The reverse path asks whether anxiety predicts later sleep problems. Comparing these paths can inform directional theory, but does not by itself identify causal effects.
The open-access cross-lagged panel modelling tutorial by Mackinnon and colleagues provides a worked lavaan example and stresses longitudinal measurement invariance before interpreting relations among repeated latent variables.
When psychology dissertation cross-lagged panel model analysis fits
Use a cross-lagged approach when the primary question concerns temporal ordering or reciprocal prediction between constructs measured on multiple occasions. The same constructs should be measured comparably at every wave, and the intervals should plausibly capture the process of interest.
| Research question | Possible method | Important distinction |
|---|---|---|
| Do two continuous constructs predict one another over time? | CLPM or an appropriate extension | Lagged prediction is not automatically causal |
| How do people differ in their trajectories? | Latent growth curve model | Targets continuous growth factors |
| Do people move between unobserved profiles? | Latent transition analysis | Targets categorical latent states |
| What is the population-average repeated association? | Generalised estimating equations | Targets marginal regression parameters |
| How do observations vary within people over many occasions? | Multilevel or dynamic model | Can represent intensive longitudinal processes |
The psychology dissertation longitudinal study guide helps align wave timing, retention, and claims. Compare this method with the site’s latent growth curve modelling guide and latent transition analysis guide before choosing an analysis.
Define the estimand before selecting the model
An estimand is the quantity the dissertation seeks to learn. A question about whether people who generally sleep worse are generally more anxious is between-person. A question about whether a student becomes more anxious than usual after sleeping worse than usual is within-person. These are different processes and can have different directions or magnitudes.
The traditional CLPM does not cleanly separate stable between-person differences from within-person fluctuations. Hamaker, Kuiper, and Grasman’s critique of the cross-lagged panel model introduced a random-intercept alternative that separates stable differences from occasion-specific deviations. This distinction is central to model choice, not an optional technical refinement.
Write the question at the correct level
For a traditional CLPM, use careful language about prospective associations among observed or latent scores. For a random-intercept CLPM, or RI-CLPM, a cross-lagged coefficient usually concerns whether deviation from a person’s expected level in one variable predicts deviation from that person’s expected level in another at the next wave.
Neither model automatically estimates an intervention effect. State the population, constructs, timing, conditioning set, level of analysis, and primary direction before fitting anything. If the theory concerns daily fluctuations, three yearly waves may target the wrong process even if the model converges.
Choose between traditional CLPM and RI-CLPM
The traditional CLPM is identifiable with two waves, although two-wave evidence is weak for studying stable reciprocal processes. It mixes stable and time-varying variation. The RI-CLPM normally requires at least three waves and introduces random intercept factors that absorb stable between-person differences.
| Feature | Traditional CLPM | RI-CLPM |
|---|---|---|
| Primary lagged interpretation | Prospective relation in scores under the model | Within-person deviation predicting a later deviation |
| Stable between-person differences | Not separated cleanly | Represented by random intercept factors |
| Minimum waves for basic identification | Two | Usually three |
| Complexity | Lower | Higher, with more demanding data needs |
| Best use | Clearly justified score-level estimand or comparison model | Theory about within-person carry-over |
The RI-CLPM and extensions resource from Mulder and Hamaker provides lavaan and Mplus syntax for the basic model, time-invariant predictors and outcomes, multiple groups, and multiple indicators. The official Mplus RI-CLPM resource page links methodological papers, scripts, and categorical-outcome examples.
Do not choose RI-CLPM merely because it is newer. It may be unsuitable when the constructs have little reliable within-person variation, the sample or waves cannot support the model, the time-invariant factor assumption is implausible, or the substantive question is explicitly between-person.
Design the wave schedule around the psychological process
A cross-lagged coefficient is specific to the lag between occasions. Sleep disruption may influence next-day anxiety but show little effect across a year because intervening processes occur many times. Conversely, developmental changes may require months or years. The correct interval comes from theory, prior evidence, and practical measurement, not convenience alone.
Report planned and actual timing. When intervals vary substantially across participants, a model that treats all lags as identical may be inaccurate. Record assessment dates, plot the distribution of intervals, and consider models that can incorporate individually varying time when deviation is important.
More waves improve some questions, not every design
Additional waves can distinguish stable differences, test whether parameters are similar over time, and improve information. They do not rescue poorly timed measurement or unreliable constructs. Five observations during one examination week address a different process from five annual observations across university.
Protect longitudinal measurement
If a latent construct changes meaning across waves, its cross-lagged relations are hard to interpret. Use the same items, response scale, instructions, mode, language, and scoring where possible. Document unavoidable changes and check for floor, ceiling, and response-shift effects.
For multiple-item latent variables, test longitudinal measurement invariance before comparing structural paths. A common sequence evaluates configural, loading, intercept, and sometimes residual constraints. Partial invariance can be defensible when a limited number of parameters differ for clear reasons, but every relaxation and its effect on results should be reported.
The measurement invariance guide explains these steps. Using observed total scores reduces model complexity but does not remove measurement error or prove that the same construct is represented over time.
Specify a theory-led model
Begin with a path diagram and a written parameter table. Include both directional cross-lagged paths unless theory and design strongly justify a one-direction model. Include autoregressive paths and within-wave covariances. Decide whether means, variances, residual variances, autoregressive paths, and cross-lagged paths are freely estimated or constrained across waves.
Equality constraints can improve precision and test stationarity-like hypotheses. They should not be imposed simply to obtain significance. Compare constrained and unconstrained specifications, inspect parameter changes, and state what equality means given the time intervals.
Account for contemporaneous relations
Variables may affect one another within the same interval, while a standard panel model represents those relations as covariance rather than direction. Omitting important contemporaneous structure can change lagged estimates. When the scientific process is faster than the measurement schedule, acknowledge that cross-lagged paths may combine or miss shorter processes.
Choose covariates using a causal rationale
Age, baseline health, prior treatment, socioeconomic context, and life events may affect both constructs. Select covariates from theory, temporal ordering, and a causal diagram where appropriate. Automated screening can omit confounders and include colliders. The confounding variables guide supports principled adjustment.
Plan sample size and power for the chosen model
No universal sample rule works for cross-lagged models. Power depends on the number of waves, reliability, between-person and within-person variance, autoregressive strength, cross-lagged effect size, missingness, equality constraints, estimator, and model complexity. A sample adequate for a traditional CLPM may be inadequate for an RI-CLPM with multiple indicators.
Use Monte Carlo simulation with plausible parameters and attrition. Evaluate convergence, bias, confidence-interval coverage, Type I error, power for the primary cross-lagged contrast, and improper solutions. Simulate less favourable reliability and missingness rather than relying on one optimistic scenario.
Mulder’s open paper on power analysis for the RI-CLPM explains a model-specific simulation strategy implemented in the powRICLPM R package. The site’s broader psychology dissertation power analysis guide explains why design-matched simulation is preferable to subjects-per-parameter rules.
Handle missing data and attrition explicitly
Describe missing items, missed waves, withdrawal, and administrative loss separately. Report the number contributing at each wave and examine whether retention relates to prior scores or background characteristics. A non-significant attrition comparison does not prove that missingness is harmless.
Full-information maximum likelihood can use incomplete cases under its assumptions. Multiple imputation may also be appropriate when the imputation model respects the longitudinal, multivariate, and categorical structure. Neither method corrects all missing-not-at-random processes.
Include defensible auxiliary variables and run sensitivity analyses where attrition could alter conclusions. Avoid listwise deletion as a default, and never use last observation carried forward to manufacture stability. See the missing data guide for a fuller decision framework.
Fit and diagnose the model systematically
- Audit coding, timing, distributions, missingness, reliability, and within-person variation.
- Confirm the longitudinal measurement model when constructs are latent.
- Fit a theory-aligned baseline model and inspect convergence and identification.
- Compare CLPM and RI-CLPM when both answer relevant competing interpretations.
- Test justified equality constraints across intervals.
- Inspect global fit, local residuals, estimates, uncertainty, and influential observations.
- Run planned sensitivity analyses for timing, missingness, covariates, measurement, and model family.
- Preserve code, seeds, software versions, and decision logs.
Check warnings, gradients, residual variances, latent variances, standard errors, and factor correlations. Negative variances, correlations outside their admissible range, non-positive definite matrices, or failure to converge indicate an improper solution. A favourable comparative fit index does not make an improper solution acceptable.
Use fit indices as evidence, not a certificate
Report the model chi-square, CFI, TLI, RMSEA with interval, SRMR, and information criteria when comparisons are valid. Conventional thresholds are guides rather than universal pass marks. Inspect residual covariances and modification indices, but add parameters only when they have a defensible substantive meaning and can be tested transparently.
Interpret coefficients without causal overstatement
| Parameter | Defensible interpretation | Avoid |
|---|---|---|
| Autoregressive path | Conditional carry-over to the next occasion | Calling it total trait stability in every model |
| CLPM cross-lag | Prospective score-level association under the specification | Calling it a pure within-person effect |
| RI-CLPM cross-lag | Within-person deviation predicting a later deviation | Calling it a between-person difference |
| Within-wave covariance | Residual association at the same occasion | Assigning causal direction without support |
| Equality test | Evidence about whether selected paths differ | Comparing significance labels across paths |
Compare directional effects using an explicit parameter constraint or contrast, not because one coefficient is significant and the other is not. Report unstandardised coefficients with confidence intervals for direct model comparisons, plus standardised estimates when they aid interpretation. Standardisation can vary across waves, so specify the method.
Temporal precedence is necessary but insufficient for causation. Unmeasured time-varying confounding, measurement error, selection, contemporaneous effects, feedback at unobserved time scales, and model misspecification can produce misleading paths. Use “predicts,” “is prospectively associated with,” or “is followed by” unless the design supports stronger language.
Psychology-specific worked example
Suppose 480 students complete validated sleep-problem and anxiety scales at the beginning of term, six weeks later, and after examinations. The preregistered question concerns whether occasions of worse-than-usual sleep predict later anxiety above the student’s usual level, and the reverse. This is a within-person question, so RI-CLPM is the primary model and traditional CLPM is a comparison.
Longitudinal measurement tests support equal loadings and mostly equal intercepts, with one justified intercept freed after examination wording changes. Timing records show modest variation around the six-week target. Monte Carlo analysis had planned precision for the primary sleep-to-anxiety path under plausible reliability and attrition.
The RI-CLPM separates stable student differences from within-person deviations. The sleep-to-anxiety coefficient is positive but imprecise, while the anxiety-to-sleep coefficient is near zero. A direct equality constraint does not clearly distinguish the two directions. The traditional CLPM produces larger paths, illustrating how model family changes the estimand and result.
The dissertation reports the result as limited evidence that worse-than-usual sleep may precede higher-than-usual anxiety over this interval, not proof that sleep causes anxiety. Sensitivity analyses vary equality constraints, covariates, missing-data assumptions, and the partially invariant item. Conclusions remain cautious because only three waves and one lag length were observed.
Report the cross-lagged analysis transparently
In Methods, report the design, population, recruitment, wave timing and actual intervals, measures, scoring, reliability, measurement-invariance sequence, missing-data method, estimator, CLPM type, random-intercept specification, within-wave covariances, constraints, covariates, power analysis, software and version, diagnostics, and sensitivity analyses.
In Results, report sample flow, wave summaries, correlations, missingness, measurement-model comparisons, global and local fit, every primary autoregressive and cross-lagged estimate with intervals, random-intercept variances and covariance, within-wave residual covariances, equality tests, improper-solution checks, and robustness findings. Provide a readable path diagram and reproducible syntax where permitted.
Common cross-lagged panel model mistakes
- Choosing the traditional CLPM for a within-person question without justification.
- Calling lagged association causal because one variable was measured earlier.
- Ignoring longitudinal measurement invariance.
- Using convenient waves that do not match the proposed process.
- Comparing significance labels instead of testing directional coefficients directly.
- Adding covariates or residual paths only after inspecting desired results.
- Assuming full-information estimation removes every attrition bias.
- Reporting only standardised estimates without explaining standardisation.
- Ignoring convergence warnings, negative variances, or unstable parameters.
- Presenting one model family as the uniquely true representation of the data.
Frequently asked questions
How many waves are needed for a cross-lagged panel model?
A traditional CLPM can be identified with two waves, but three or more provide stronger opportunities to evaluate reciprocal patterns and equality across intervals. An RI-CLPM normally requires at least three waves.
What is the difference between CLPM and RI-CLPM?
The traditional CLPM mixes stable between-person differences and time-varying processes. The RI-CLPM introduces random intercepts to separate stable differences, allowing lagged paths to represent within-person deviations under its assumptions.
Do cross-lagged paths establish causality?
No. They establish model-based temporal prediction after specified adjustment. Causal interpretation also requires defensible timing, measurement, confounding control, selection assumptions, and a design capable of supporting the claim.
Should autoregressive and cross-lagged paths be equal over time?
Only when theory, interval length, and evidence support equality. Test constraints against a less restricted model and report how conclusions change. Unequal intervals make simple equality especially difficult to justify.
Can I use observed scale totals?
Yes, but totals treat measurement differently from a multiple-indicator latent model and do not eliminate measurement error. Demonstrate longitudinal comparability and report reliability and scoring at every wave.
How should I compare the two directional effects?
Test a direct equality constraint or coefficient contrast within the same model. Do not infer a difference merely because one path has a smaller p-value or crosses a significance threshold.
Which software can estimate cross-lagged panel models?
Common options include lavaan in R, Mplus, OpenMx, and other SEM packages. Report the program, version, estimator, missing-data treatment, syntax, constraints, and any special boundary tests.
Conclusion
Psychology dissertation cross-lagged panel model analysis is credible when the estimand, level of change, measurement, wave spacing, and model family agree. The strongest dissertation distinguishes score-level from within-person questions, treats CLPM and RI-CLPM as different models rather than interchangeable upgrades, checks longitudinal measurement and missingness, tests directional contrasts directly, and reports uncertainty without causal overstatement.
If you need statistical tutoring, use support that helps you understand and conduct your own analysis. Protect participant data, follow ethics approval and academic-integrity requirements, retain an auditable analysis record, and remain responsible for every model choice and conclusion.
