Longitudinal Dyadic Modeling With Small Numbers of Dyads: A Practical Multilevel Framework and Subsampling Demonstration

Longitudinal dyadic studies examine interpersonal processes using repeated observations from members of the same dyad. Different dyadic estimands draw information from different levels of the design, so additional repeated occasions and additional independent dyads do not contribute equivalent information. This study examined how estimand definition, model structure, and design features shape precision and finite-sample performance in longitudinal dyadic models with limited numbers of dyads. Analyses combined a full-sample benchmark demonstration with 100 distinguishable dyads observed across 21 occasions, a 500-iteration subsampling analysis at 10, 20, 30, and 50 dyads, and a targeted Monte Carlo simulation with 56 conditions and 1,000 replications per condition. Simulation factors included dyad count, repeated-measure intensity, dyad-level variance, residual dependence, and between-person predictor dispersion. Performance was evaluated through bias, confidence interval coverage, standard error calibration, singularity and variance-component behavior, and fixed-effect precision. Limited dyad replication most clearly affected support for dyad-level random effects. Fixed-effect bias remained small and confidence interval coverage was close to nominal across the examined conditions. Increasing repeated occasions from 7 to 21 reduced confidence interval widths for within-person effects by approximately 42%, compared with approximately 6% for between-person effects. These findings show that dyad count and repeated-measure intensity provide different forms of information. Longitudinal dyadic designs are most informative when the target estimand is aligned with the level of information supplied by the design.