Pre-Estimation Checks

Read this page before fitting TD-CCP. The goal is to catch weak feature support, weak successor-tuple support, and unstable continuation bases before the transition-density-free shortcut is interpreted structurally.

TD-CCP is most reliable when the reward features, observed choices, and panel transitions support the structural parameter target. These checks are meant to catch problems before a result is treated as recovery.

Before Fitting

Confirm that the reward is written as a finite set of known features. The features should have enough variation to identify the parameters, and one action or reward level should be normalized so the model has an anchor.

Check that every relevant state-action pair has enough observed support. TD-CCP uses observed successor tuples to learn continuation terms, so weak support in the current or next state-action data can make the recursion unstable.

Finally, check the first-stage choice probabilities. The log correction used by TD-CCP is unstable when fitted choice probabilities are too close to zero.

During Fitting

Watch the conditioning of the basis used for the continuation terms. A nearly singular basis can make the projected TD equations unstable even when the likelihood optimizer appears to run.

For locally robust inference, split folds by individual rather than by row. This keeps the held-out moment evaluation separate from the quantities learned on the other fold and matches the covariance unit used in the simulation results file.

After Fitting

Do not judge the run only by a returned parameter vector. Check that the final moment is small, the correction recursion is stable, the optimizer reached a stationary point, and reported standard errors are finite and positive.

Also keep estimation and evaluation separate. The parameter-estimation step should not silently use a transition tensor. Transition tensors may enter later for policy values, Q functions, or counterfactual evaluation.

Current Simulation Case

The simulation study uses two encoded state coordinates and a finite linear reward. Action 0 is fixed as the baseline action, leaving six reward parameters for the two non-baseline actions.

Check	Current results file
States and actions	81 states, 3 actions
Reward target	6 finite reward parameters
Reward form	Linear in encoded state features
Choice model	Logit with degree-2 state features
TD basis	Encoded semigradient basis, degree 2
Fold split	By individual
Inference	Cross-fitted locally robust standard errors
Max moment norm	7.91e-06
Max correction residual norm	0.002610
Optimizer status	Preliminary and final robust folds converged

See the simulation study page for the full results file, numerical checks, and Monte Carlo standard-error check.