Skip to contents

Extract scores for the original analysis data, coefficients, and, where available, confidence intervals for score coefficients from a catefit.

Usage

# S3 method for class 'catefit'
predict(object, newdata = NULL, method = NULL, ...)

# S3 method for class 'catefit'
coef(object, method = NULL, ...)

# S3 method for class 'catefit'
confint(object, parm, level = 0.95, method = "contrastReg", ...)

Arguments

object

An object of class catefit.

newdata

New data are not currently supported; omit this argument to retrieve scores for the original analysis sample in its original order.

method

A single score to extract from the fitted object. Use "two_regression" or "contrast_regression" for those procedures, or the outcome method (such as "logistic", "boosting", "poisson", "negative_binomial", or "random_forest") for a T-learner. The established stored names "twoReg", "contrastReg", "negBin", and "randomForest" are also accepted; contrast regression can also be selected with "contrastreg". If exactly one T-learner score is present, "t_learner" selects it. If exactly one score of any kind is present, method may be omitted from predict(). For coef(), omission returns the coefficient table. For confint(), the default is "contrastReg". Coefficients and confidence intervals are available only for certain score methods.

...

Additional arguments; currently unused.

parm

Coefficient names or positions; omitted selects all coefficients.

level

Confidence level between 0 and 1; defaults to 0.95.

Value

predict() returns a numeric vector of stored scores. coef() returns the stored coefficient table or a named numeric vector for a selected method. confint() returns a matrix of Wald confidence limits for the selected coefficients.

Details

Scores for count and survival outcomes are on the log ratio scale; binary scores are on the risk-difference scale. They describe fitted scores, not validated individual treatment effects. Pass the fitted catefit object to predict(), not an individual outcome model in its fit component. When multiple T-learners are fitted, select one by its outcome method. Confidence intervals require stored coefficient standard errors, currently available for contrast regression in count and continuous fits only. They are not available for survival or binary fits. Methods that do not fit linear scores, such as random forests and boosting, do not provide score coefficients.

Author

Thomas Debray