Choosing an estimator

Choose the restriction you can defend and the output you need. A flexible fit is not by itself an identification argument.

Problem	Starting point	Key restriction	Main output
Linear mean with observed covariates and additive effects	OLS/WLS/FE	Regressor exogeneity after absorption	Slopes; classical or HC0/HC1 covariance
Binary responses or counts	Logit/Poisson GLMs	Specified conditional mean/link; stronger assumptions for model-based SEs	Probabilities/count means
Endogeneity and a finite instrument vector	GMM	Valid, sufficiently informative moments	Parameters and iid/HAC sandwich
Alternative empirical weighting of moments	GEL	Valid moments and feasible tilting problem	Parameters, first-order covariance
Nonlinear structural function with conditional instruments	MMR/SMD	Conditional moments and identification of the function	Fitted function, criterion diagnostics
Several unordered actions	Static choice	Utility normalization and logit/probit restrictions	Choice probabilities
Choices that affect future states	NFXP/CCP	Markov structure, known/fixed transitions, shock assumptions	Structural utilities and policy predictions
A treated block of a balanced panel	SC/SDID	Counterfactual comparability of weighted controls	Point estimate and weights
Missing outcome cells	Matrix completion	Low-rank plus additive structure, informative support	Reconstructed outcome matrix
Binary/count panels with latent heterogeneity	Latent-factor GLM	Low-rank natural parameter and family/link	Cell means/probabilities
Discrete groups with different time profiles	Grouped FE	Informative unit embeddings; group approximation	Slopes and group assignments
Unknown error shape in linear regression	Antitonic score	Shape-constrained residual score	Slopes and fitted score
Synthetic rows for simulation	Generative models	Representative training distribution and adequate fitting	Draws, not a causal estimand

Estimation, prediction and inference are different capabilities

Dynamic choice and low-rank logit have prediction interfaces but no implemented standard errors. Neural conditional moments and factorized panel models do not provide a generic valid confidence interval merely because gradients exist. SC/SDID here return point estimates, not automatically valid placebo or bootstrap intervals. The linear estimator does not implement clustered covariance.

If the goal is a causal panel effect, distinguish an observation mask from an untreated-training mask. Matrix completion and latent-factor GLMs fit whichever cells they receive. A good held-out-cell score does not prove accurate extrapolation to a treated block.

What belongs elsewhere

Trex is not a formula parser, missing-data imputation policy, experiment randomization engine, or a complete causal-identification framework. It does not offer every estimator found in its research notebooks as a stable public class. The separate Gaussian-tabular pilot and policy-aligned CATE proposal are not documented as merged API features.

For all public entry points—including preprocessing, transition estimation and result structures—see the API index.