Predict oriented hub-spoke comparison probabilities

Predict oriented hub-spoke comparison probabilities

Usage

predict_link(result, pairs)

Arguments

result

A valid common result returned by fit_link().

pairs

Data frame with observation_id, A_set, A_item, B_set, and B_item, using the identities in prepare_link_input(). No outcomes are accepted. A and B encode presentation orientation; both directions are legal.

Value

A double vector of probabilities that presented A wins, aligned with the input rows. The estimator's prediction hook integrates its uncertainty; the dispatcher never silently substitutes probabilities at posterior means.

Details

Conditional probabilities are (1-epsilon) * plogis(theta_A-theta_B+beta) + epsilon/2. Reversing presentation is complementary only when beta is zero. Prediction requires no refit or provider calls. E1 integrates over delta using the serialized posterior quadrature nodes. It refines numerical integration for a requested pair when necessary, without updating the posterior or modifying the result. There is no plug-in-mean prediction mode. Integration failure raises pairwiseLLM_e1_numerical_error, rather than substituting another prediction. E2 integrates the oriented Gaussian pair contrast under its joint Laplace approximation, including all item/offset cross-covariances. Deterministic adaptive quadrature over the standard Normal variable splits at zero and the logistic transition (bounded to plus/minus eight SDs to avoid excessively long finite intervals); both infinite tails are integrated. Local relative and absolute error budgets divide the requested global tolerances by the segment count and, successively, 4, 16, 64, and 256. The first valid attempt whose summed error estimate meets the original global absolute/relative tolerance is returned; exhaustion raises a numerical error. Refinement does not change saved controls or fit state. Degenerate zero-variance contrasts use the exact conditional probability. Repeated calls do not sample or mutate state. Integration failure raises pairwiseLLM_e2_numerical_error. E3 MAP uses the same integration with pairwiseLLM_e3_numerical_error on failure. E3-MCMC averages the conditional probability over its retained raw draws, without new sampling.

See also

prepare_link_input(), fit_link(), predict_link(), start_link_session()

Other linking: fit_link(), prepare_link_input(), save_link_session(), start_link_session(), summary.pairwiseLLM_link_result()