Common questions

Do I need R?

No. NumPy and SciPy handle fitting and prediction. R is needed only to regenerate the reference comparisons or export the original R example dataset; see development.

Which outcome model should I use?

Use Gaussian for continuous outcomes, probit (family="binomial") for zero/one outcomes, Poisson or negative binomial for counts, and survival for event times. The model guide states the likelihoods and assumptions; fitting inputs lists the required arguments.

Why does prediction require Xnew?

If a fit includes adjustment covariates, prediction at new exposure rows needs corresponding adjustment values. Supply the raw columns in training order, without the added intercept. predict_h returns only the exposure function and does not require those adjustments. Prediction explains the distinction.

How are missing values handled?

The default raises an error. missing="drop" removes rows with missing required values; for recorded subject histories, it removes the whole affected history. Dropping incomplete observations does not correct missing-data bias. See input handling.

What does lambda mean?

Default Gaussian fits store the variance ratio \(\lambda=\tau/\sigma^2\), where \(\tau\) controls Gaussian process (GP) variation and \(\sigma^2\) is residual variance. Default probit fits fix \(\sigma^2=1\). Structured and default count/survival fits store the direct GP amplitude \(\tau\). Compare priors on the stated scales; see the notation/API mapping.

Are exposure curves causal effects?

They describe fitted associations at specified values of other variables. A causal interpretation needs additional scientific assumptions. A percentile combination can also be rare in observed data, especially with correlated exposures. See contrasts and curves.

What is the difference between a mean and an observation prediction?

A mean prediction describes the fitted mean, probability or relative hazard, according to the outcome family. An observation prediction also includes new outcome variation; it is unavailable for survival outcomes. Prediction scales gives the supported combinations.

Is a survival contrast a hazard ratio?

Response-scale survival contrasts subtract relative hazards. To obtain a hazard ratio, exponentiate each latent exposure-function contrast before summarizing the draws. Survival probabilities come from predict_survival; see the survival tutorial.

Why is a group's conditional PIP NaN?

A conditional posterior inclusion probability asks how often an exposure is included when its group is selected. If the group is never selected, that fraction is undefined. Group selection permits at most one active exposure in a group. See selection priors.

Do favorable diagnostics validate the model?

No. R-hat and effective sample sizes assess sampling behavior. They cannot establish convergence of every curve, independence of daily errors, or correctness of model assumptions. See diagnostics and evidence.

Do finite GP bases preserve uncertainty from the full model?

Not in general. The predictive process omits variation remaining after its knot values are known. Random Fourier features condition inference on a fixed sampled feature map. These approximations can change uncertainty; see GP representations.

Does KernelMix reproduce the original BKMR studies?

Exact reproduction remains unresolved because original study inputs and sufficient fitting settings are missing. The ozone example and numerical software checks answer different questions. The evidence page records their scope and limits.