# Pre-Estimation Checks

## Important Links

- [CCP Overview](../ccp.md)
- [Quick Start](quick_start.md)
- [Simulation Study](validation.md)
- [Counterfactuals](counterfactuals.md)

Read this page before fitting CCP. These checks focus on support and
conditioning because the estimator leans on first-stage choice probabilities.

Some CCP estimation risks are visible before optimization starts. Run these
checks before treating a result as structural evidence.

| Check | Why it matters |
| --- | --- |
| Feature rank | Collinear reward features cannot identify separate parameters. |
| Feature condition number | Near-collinearity inflates standard errors. |
| Transition row sums | Each transition row must be a valid probability distribution. |
| State coverage | Unvisited states require extrapolated CCPs. |
| State-action coverage | One-action states make counterfactual action values weakly supported. |
| Minimum positive CCP | Very small probabilities make the log correction unstable. |
| Reward normalization | Additive reward shifts do not change choice probabilities, and the shock scale must be normalized. |
| Transition orientation | CCP expects transition tensors in action, state, next-state order. |

The direct action-contrast rank is a conservative check used by the wrapper.
It is sufficient for the supported examples. It is not a general necessity.
A state-only feature can affect dynamic choices when actions induce different
future state distributions.

Support checks answer a different question from identification. Population
full support permits the logit inversion. Observed zeros and thin cells make
the empirical inverse unstable. Smoothing prevents a numerical zero, but it
does not create information that was absent from the panel.

## Simulation Checks

The first 20-state inference panel in the
[Simulation Study](validation.md) reports the following checks.

| Check | Value | Status |
| --- | ---: | --- |
| Reward features | 3 | pass |
| Feature rank | 3 / 3 | pass |
| Action-contrast rank | 3 / 3 | pass |
| Feature condition number | 1.000 | pass |
| Action-contrast condition number | 1.000 | pass |
| Observed states | 20 / 20 | pass |
| State-action coverage | 1.000 | pass |
| Single-action states | 0 | pass |
| States with fewer than five observations | 0 | pass |
| Action counts | 5,605 and 4,395 | pass |

The panel has full support, so the reported fit does not rely on uniform CCP
fallbacks for unvisited states.

## Common Risk Patterns

Data with many unvisited states force CCP to extrapolate the first-stage
policy. States with only one observed action make counterfactual action values
weakly supported. Too little smoothing can make log corrections unstable,
while too much smoothing biases the empirical policy toward uniform choice.
A transition tensor with the wrong orientation can have valid dimensions while
representing the wrong transition law.

Before interpreting the estimates structurally, specify the discount factor,
shock distribution and scale, reward location normalization, and utility form.
The wrapper can estimate transitions from the panel, but its reward standard
errors do not propagate uncertainty from that transition stage.