Longitudinal Dyadic Modeling with Limited Numbers of Dyads
Aligning estimands, information sources, and multilevel structure
Published paper + active program Methods in Psychology Liu, L.
Repeated observations and independent dyads contribute different kinds of information. This program studies how target estimands, multilevel structure, role parameterization, random effects, and design features determine what longitudinal dyadic data can support when the number of dyads is limited.
Published Paper
Liu, L. (2026). Longitudinal dyadic modeling with limited numbers of dyads: Aligning estimands, information sources, and multilevel structure. Methods in Psychology, 15, 100281. DOI
The paper develops an estimand-information alignment framework: each target dyadic estimand should be evaluated in relation to the level of the design that supplies its statistical information. The study distinguishes fixed-effect precision, repeated-sampling calibration, model completion, optimizer convergence, and support for random-effect structure rather than treating them as a single notion of model performance.
More repeated observations do not substitute for more independent dyads.
Research Design
The published study combines three complementary components.
The Monte Carlo design varies dyad count, repeated-measure intensity, dyad-level variance, residual dependence, and between-person predictor dispersion. Across the simulation, 56,000 model fits were requested. Performance was evaluated using bias, RMSE, confidence-interval coverage, standard-error calibration, Type I error, singularity, variance-component behavior, and fixed-effect precision.
Key Findings
Limited dyad replication most clearly reduced support for dyad-level random-effect structure.
Bias remained small and confidence-interval coverage was close to nominal across the examined conditions.
Increasing occasions from 7 to 21 reduced interval widths by about 42%.
The same increase in occasions reduced interval widths by only about 6%.
The practical implication is that more repeated measurements cannot be treated as a substitute for more independent dyads. Repeated occasions add substantial information for within-person estimands, while between-dyad parameters and higher-level random effects depend more strongly on independent dyad replication.
Code
The published article includes a public simulated pedagogical benchmark and supplementary analysis materials. View the article · OSF materials
Contribution to my research program
This program extends my broader work on multilevel inference under limited higher-level information into longitudinal interpersonal data. It links design, estimands, random-effect support, and practical modeling decisions rather than treating sample size as a single undifferentiated quantity.