Common questions

Do I need R?

No. R is needed only to rerun the reference comparisons.

What does q=None mean?

The exposures are already scored. For raw measurements, the default q=4 estimates quartile boundaries. Ties can leave fewer groups. Prediction reuses training cutpoints; it does not estimate new ones.

How are missing values handled?

Array fits require complete numeric data. Supported formula fits reject missing values by default; missing="drop" removes incomplete training rows. Removing rows does not by itself prevent bias. The imputation helpers have separate assumptions; see data tools.

Are case weights the same as sign weights?

No. Case weights change a row's influence in supported fits. Sign weights describe exposure coefficients within each sign group. Neither automatically accounts for a complex survey design.

Does a confidence interval establish a causal effect?

No. A causal interpretation needs well-defined interventions, consistency, no interference, no unmeasured confounding, positivity and a correctly specified model. The software cannot verify these assumptions.

Does QuantileMix cover all of R qgcomp?

No. Each function supports a defined subset. The coverage map lists important limits. Validation distinguishes software agreement from statistical performance.

How do I report a problem?

Open a bug report with a small reproducible example, expected output, actual output, Python version and package versions. Use synthetic data rather than private records.