Data, prediction and results

Convert exposure measurements to ordered category scores.

Examples:

>>> quantize([0, 1, 2, 3], q=2).scores.ravel().tolist()
[0.0, 0.0, 1.0, 1.0]

Choose a supported additive quantile g-computation fit by family.

Standard error of a named linear combination of coefficients.

names gives covariance row and column order. By default each selected exposure receives coefficient one; grad may be a scalar for all selected exposures or a full vector in names order.

Covariance of a named coefficient and an exposure contrast.

The off-diagonal uses the supplied gradient. R qgcomp 2.19.6 does not apply grad to that entry; the default unweighted behavior agrees.

Randomly partition rows, with optional whole-cluster assignment.

The training fraction is rounded to the nearest integer. A seed makes the Python split reproducible; it does not reproduce R's random-number stream. With cluster, each distinct cluster is assigned wholly to one side. This corrects an R 2.19.6 implementation difference: its cluster helper samples rows inside each cluster despite its group-preserving description. Missing cluster IDs and a split with an empty side are rejected.

Select coefficient signs in training rows and estimate in held-out rows.

This interface needs an independently chosen train/validation split. family is "gaussian", "binomial" or "poisson". For each nonempty sign group, its held-out fit treats the other exposures as adjustment variables. With raw exposures, only the selected group is quantized in held-out fits, matching R's expnms convention. The default reuses training cutpoints. fixed_breaks=False computes new held-out cutpoints. The binary result is a conditional log odds ratio; the count result is a conditional log mean ratio for equal observation periods. Cross-fitting, bootstrap, case weights, Poisson time offsets, other families and inference for a data-dependent split are outside this interface. Keeping clusters on one side with split_data does not make these model-based intervals cluster-robust.

Compare each joint exposure score with a selected score level.

Compare marginal fitted means at each score against one reference score.

pointwiseref is one-based, matching R. In R qgcomp 2.19.6, the Poisson bootstrap covariance is calculated from marginal-model predictions on the log scale. The displayed fitted means are calculated on the response scale, then logged for the linear-predictor column.

Bound a fitted additive marginal curve at every joint score.

Pointwise limits use the bootstrap standard deviation at each score. Simultaneous limits follow R qgcomp 2.19.6: retain resampled marginal coefficient vectors inside a chi-square ellipsoid around the original coefficients, then take the minimum and maximum fitted curve among those retained vectors at each score. These are reference-method bounds, not a guaranteed finite-sample or causal coverage statement.

Return pointwise and R-method simultaneous bounds for Gaussian EE.

Like R qgcomp 2.19.6, this samples 5,000 marginal coefficient vectors from the fitted normal approximation. seed makes Python's draws reproducible; the same integer does not give the same draws in R. These bounds are sampling-based and do not establish repeated-data coverage.

Predict from a fitted bootstrap marginal structural model.

Report the fitted intercept and joint effect in a small named table.

Only additive non-bootstrap Gaussian, binomial, Poisson and conditional Cox fits are supported. Cox has no intercept row. By default the confidence level is the fit's 1-alpha. quick=True returns only term names and estimates. Exponentiation is available for log or logit links; standard errors and test statistics remain on the link scale.

Unlike R qgcomp 2.19.6's tidy method, the upper exponentiated limit uses the upper limit, and an intercept has its own p-value instead of a recycled joint-effect p-value. These are explicit correctness repairs. The Gaussian joint p-value retains R qgcomp's reference degrees of freedom; its intercept uses the fitted model's residual degrees of freedom. Normal confidence limits are used for both.

Give one row of likelihood and fit measures for a conditional model.

Supports additive Gaussian, Bernoulli-logit and Poisson-log fits. nobs counts rows, df_model counts slope coefficients, and df_residual is the number of rows minus all fitted coefficients. deviance and null_deviance are weighted when case weights were used; Gaussian null deviance is the weighted total sum of squares. The null Poisson model retains each row's observation time. AIC and BIC use the maximized full likelihood for Gaussian and the model likelihood for the two generalized linear models. Gaussian AIC/BIC count the error variance as one additional parameter; BIC uses the number of rows.

Zero residual sum of squares makes a Gaussian maximum-likelihood variance degenerate. In that case likelihood-based measures are None instead of infinite. An undefined R-squared or F test is also None. Bootstrap, estimating-equation, Cox and other result classes are not yet supported; they must not inherit likelihood claims from a different fit.

Draw descriptive positive and negative exposure weights.

Draw saved standardized responses at joint exposure scores.

Draw positive values below a limit of detection (LOD).

y is one measured variable. ry marks observed values; selected left-censored rows have ry=False and are marked by wy (by default, all unobserved rows). Observed values must be finite and positive. x contains complete numeric predictors without an intercept; an intercept is added. lod is a positive scalar or one positive value per row. If omitted, the smallest observed value is used for every censored row.

A lognormal regression uses both observed densities and censored lower- tail probabilities. One parameter vector is drawn from the fitted normal approximation, then each selected value is sampled from its lognormal distribution truncated above at its LOD. The return contains imputed values in selected-row order, as in R's MICE helper. Assign them to those rows in an external imputation workflow; this package does not silently pool multiply imputed analyses.

This is a Python interface to the R qgcomp 2.19.6 left-censored lognormal method, not an R MICE callback. Its random stream and optimizer differ from R's survreg implementation. The normal parameter draw is an asymptotic approximation and can be unreliable in small or weakly identified samples. impute_tobit is an alias for this same algorithm.

Generate balanced quantized exposures and one synthetic outcome.

outcometype is "continuous", "logistic", "survival" or "multinomial". Each exposure column has scores from zero through q-1. corr gives signed template-copy probabilities for exposures after the first, not exact target Pearson correlations. A positive value copies the first exposure's shuffled score template; a negative value copies its reverse. Other rows are shuffled within that exposure.

coef is a vector with one coefficient per exposure, except for multinomial outcomes, where it is a matrix of shape (exposures, nonreference_categories). The multinomial outcome y uses one-based category labels. Survival returns time and d; d=1 means the event preceded administrative censoring at censtime. The returned mapping includes x1, x2, ... and outcome columns; the known coefficients are separate attributes, not data columns.

NumPy's local random generator makes Python runs repeat for the same seed. Its draws are not bitwise equal to R's random stream. Two R 2.19.6 source errors are corrected deliberately: when ncheck is true, n is increased to the next multiple of q (R adds the remainder), and survival baseline times use scale0 (R passes shape0 as scale). This helper generates synthetic examples; it does not validate a method.