Pre-Estimation Checks

Read this page before fitting NNES. These checks separate problems with the finite reward target from problems with the neural continuation-value approximation.

NNES can fail for reasons that are visible before the neural value path starts. Run these checks before treating a NNES result as structural evidence.

Check	Why it matters
Feature rank	Rank below the number of reward parameters means theta is not identified.
Feature condition number	A high condition number signals unstable reward estimates.
Transition model	NNES is model-based, so transition probabilities must be known or estimated before policy evaluation.
Transition row sums	NNES needs valid transition probabilities for continuation values.
Logit DDC primitives	The paper’s orthogonality result is for additive Type-I extreme-value shocks and separated transition dynamics.
State coverage	Unobserved states weaken value-network training and recovery checks.
State-action coverage	Sparse action support weakens the likelihood and counterfactual fit.
Initial CCP support	The NPL path starts from empirical or first-stage CCPs; weak support makes that start noisy.
Reward normalization	Reward level and scale need a valid anchor.
Value anchoring	High-discount problems need the value-level normalization to avoid drift.
Value-network loss	A high final loss can contaminate recovered policy and value objects.

Results Checks

These rows come from the generated NNES results file. See Simulation Study for the generator script, rendered table source, and JSON.

Check	Low-dimensional	High-dimensional primary
Feature rank	4 / 4	32 / 32
Feature condition number	4.512	1.377
Transition row error	2.42e-8	2.42e-8
Observed states	21 / 21	81 / 81
State-action coverage	1.000	0.959
Minimum action share	0.325	0.281
Outer NPL iterations	3	3
Final V loss	0.000399	0.029932

The primary cell is the high-dimensional encoded-state DGP. It checks that the neural value path still recovers known reward, policy, value, Q, and counterfactual objects.

Common Risk Patterns

A low-rank reward matrix can make several reward parameters observationally equivalent. Sparse action support can make the fitted policy look plausible while the structural parameters drift. Wrong transition orientation can produce valid-looking arrays and wrong continuation values.

An unanchored value network can drift in high-discount problems because the absolute value level is weakly identified by choice data. A neural value network with poor fit can pass data likelihood checks while failing value, Q, or counterfactual recovery checks. These risks are why the simulation table reports both structural recovery and final value-network loss.