Choose an analysis¶

Start with the endpoint and dependence structure, then ask what quantity you need to estimate. Software availability comes after the scientific question.

Decision table¶

Scientific situation	Starting point	Typical declaration	Current calculation path
Continuous outcome; observations clustered in batches, sites, people, or devices	Gaussian LMM	`y ~ treatment + time + (1	group)`
Continuous outcome; subject-specific trajectories	Gaussian LMM	`y ~ treatment * time + (1 + time	subject)`
Binary, count, or overdispersed-count outcome with cluster effects	GLMM	`y ~ treatment + (1	site)` plus family
Repeated continuous visits where residual covariance is the focus	MMRM	fixed-effects formula plus `AR1`, `Toeplitz`, or `Unstructured` residual	ML or REML
Known-parameter one-/two-compartment concentration profile	Closed-form PK	`one_compartment_`, `two_compartment_`	Direct calculation
Dosing events, discontinuities, custom dynamics, or sensitivities	Event-aware ODE	canonical event table plus `simulate_ode`	Numerical integration
Auditable population-PK model contract	Pharmacometric DSL	parameters, ETAs, states, doses, observations	Compile/validate/IR; simulation and conditional-mode primitives
Cross-tool model/data exchange	Interoperability layer	NONMEM, PharmML, SBML, SED-ML, R-formula helpers	Supported subset with compatibility report

LMM or MMRM?¶

Both can analyze continuous longitudinal outcomes, but their dependence models answer different modeling needs.

Choose an LMM when random effects are the natural representation of between-group heterogeneity or subject-specific trajectories.
Choose MMRM when an explicit within-subject residual covariance across a finite visit axis is central to the analysis.
The current reference MMRM path does not combine its structured residual covariance with formula random effects.

LMM or GLMM?¶

Choose from the conditional distribution of the observed endpoint.

Approximately Gaussian continuous response → Gaussian LMM.
Binary response → Bernoulli/binomial GLMM with logit link.
Count response → Poisson GLMM when mean and variance assumptions are reasonable; negative-binomial-2 is available for a supported overdispersion path.

PyMixEF defines a broader catalog of probability-family objects for standalone probability calculations and model representation. Catalog presence does not mean every family can be fitted by the formula GLMM backend. The current GLMM engine fits Bernoulli, binomial, Poisson, and negative-binomial-2 with canonical links and first-order Laplace only.

Formula model or pharmacometric DSL?¶

Use the formula interface for LMM, GLMM, and MMRM workflows. Use the typed pharmacometric DSL when the model contains parameters with constraints, random effects on transformed parameters, states, differential equations, dose mappings, and observation models.

Both routes compile to a versioned ModelIR. The same representation makes capability checking, semantic hashing, comparison, serialization, and backend selection inspectable.

Estimation or simulation?¶

Closed-form PK helpers and simulate_ode calculate trajectories from supplied parameters; they do not estimate population parameters. The DSL declares and validates a model contract. Population-estimation primitives expose conditional mode and Laplace calculations, while integrated production FOCEI and SAEM workflows are not currently available.

Before committing to a method¶

Ask:

What is the observational unit and what creates dependence?
Is the endpoint continuous, binary, count, censored, or time-to-event?
Is the target population-average, cluster-conditional, subject-specific, or a contrast at a particular visit?
Which missing-data assumption is scientifically defensible?
Is visit order, dosing order, or another time axis part of the model state?
Which covariance alternatives and diagnostics will be prespecified?
Does the validation record provide enough evidence for the intended consequence?

Use the full analysis matrix for supported methods, families, outputs, and explicit boundaries.