Choosing between lme/nlme, brms, KFAS, and ctsem

Decision axes

Ask:

1. Population vs single-series focus — Do you need inference for a population of persons (fixed/random effects, WP/BP) or a structured model for one long series (latent level/trend)?
2. Multilevel vs state-space — Is the scientific target between- and within-person structure with residual temporal correlation, or explicit latent dynamics (Kalman smoothing/filtering) for one trajectory?
3. Frequentist vs Bayesian — ild_lme() / nlme vs ild_brms() (Stan); KFAS in tidyILD is likelihood / ML-oriented for the wrapped fits.
4. Diagnostics — Frequentist/brms: ild_diagnostics() / bundle from ild_diagnose() for standard models; KFAS: ild_diagnose() emphasizes innovations, ACF, and bundle sections tuned for state space (see ?ild_diagnose).

Comparison table

	ild_lme() (lme4 / nlme)	ild_brms()	ild_kfas() (KFAS)	ild_ctsem() (ctsem)
Typical use	Multilevel regression, AR1/CAR1 on residuals	Hierarchical Bayes, flexible priors	Univariate latent level (v1), single .ild_id per fit	Continuous-time latent dynamics with irregular timing
ID structure	Multiple .ild_id, random effects	Multiple .ild_id, partial pooling	One person per fit (v1); not pooled latent state across IDs	v1 wrapper is conservative (single-series focus)
Temporal structure	Residual AR1/CAR1; spacing-informed choice	Flexible time / residual modeling	Discrete-time local level on observation order	Continuous-time drift/diffusion parameterization
Priors	N/A (frequentist)	Yes (user-specified)	Implicit via likelihood / ML	Depends on ctsem fit mode (ctStanFit / ctFit)
Predictive checks	Via bundle / extensions	PPC via brms (pp_check)	Forecast / innovation plots (not PPC); see ild_autoplot(bundle) for KFAS	Bundle diagnostics + ctsem residual/fitted views
Provenance / report	ild_provenance, ild_report	Same	Same; KFAS stores state spec, smoothing, fit_context	Same; ctsem fit type and scaling recorded

Important: ild_kfas() is not a drop-in replacement for a multilevel model with person-level random effects. If you fit separate state-space models per person, use fit_context = "independent_series_per_id" when that matches your design and read GR_KFAS_UNMODELED_BETWEEN_PERSON_HETEROGENEITY and related guardrails.

Short pointers to other vignettes

• Temporal dynamics and model choice (lags, AR, TVEM vs latent): vignette("temporal-dynamics-model-choice", package = "tidyILD")
• Overall workflow: vignette("tidyILD-workflow", package = "tidyILD")
• Report-style analysis: vignette("ild-analysis-report", package = "tidyILD")
• State-space concepts (KFAS): vignette("kfas-state-space-modeling", package = "tidyILD")
• Irregular timing + KFAS: vignette("kfas-irregular-timing-spacing", package = "tidyILD")
• Continuous-time ctsem backend: vignette("ctsem-continuous-time-dynamics", package = "tidyILD")
• Spacing and WP/BP: vignette("ild-decomposition-and-spacing", package = "tidyILD")

See also

inst/dev/KFAS_V1_BACKEND.md (backend scope), ?ild_kfas, ?ild_ctsem, ?ild_lme, ?ild_brms.