init.causalForest: Causal Effect Regression and Estimation Forests (Tree Ensembles)

Description

Build a random causal forest by fitting a user selected number of causalTree models to get an ensemble of rpart objects.

Usage

init.causalForest(
  formula,
  data,
  treatment,
  weights = FALSE,
  cost = FALSE,
  num.trees,
  ncov_sample
)
# S3 method for causalForest
predict(object, newdata, predict.all = FALSE, type = "vector", ...)
causalForest(
  formula,
  data,
  treatment,
  na.action = na.causalTree,
  split.Rule = "CT",
  double.Sample = TRUE,
  split.Honest = TRUE,
  split.Bucket = FALSE,
  bucketNum = 5,
  bucketMax = 100,
  cv.option = "CT",
  cv.Honest = TRUE,
  minsize = 2L,
  propensity,
  control,
  split.alpha = 0.5,
  cv.alpha = 0.5,
  sample.size.total = floor(nrow(data)/10),
  sample.size.train.frac = 0.5,
  mtry = ceiling(ncol(data)/3),
  nodesize = 1,
  num.trees = nrow(data),
  cost = FALSE,
  weights = FALSE,
  ncolx,
  ncov_sample
)

Value

An object of class rpart. See rpart.object.

Arguments

formula: a formula, with a response and features but no interaction terms. If this a a data frome, that is taken as the model frame (see model.frame).
data: an optional data frame that includes the variables named in the formula.
treatment: a vector that indicates the treatment status of each observation. 1 represents treated and 0 represents control. Only binary treatment supported in this version.
weights: optional case weights.
cost: a vector of non-negative costs, one for each variable in the model. Defaults to one for all variables. These are scalings to be applied when considering splits, so the improvement on splitting on a variable is divided by its cost in deciding which split to choose.
num.trees: Number of trees to be built in the causal forest
ncov_sample: Number of covariates randomly sampled to build each tree in the forest
object: a causalTree object
newdata: new data to predict
predict.all: If TRUE, return predicted individual effect for each observations. Otherwise, return the average effect.
type: the type of returned object
...: arguments to rpart.control may also be specified in the call to causalForest. They are checked against the list of valid arguments. The parameter minsize is implemented differently in causalTree than in rpart; we require a minimum of minsize treated observations and a minimum of minsize control observations in each leaf.
na.action: the default action deletes all observations for which y is missing, but keeps those in which one or more predictors are missing.
split.Rule: causalTree splitting options, one of "TOT", "CT", "fit", "tstats", four splitting rules in causalTree. Note that the "tstats" alternative does not have an associated cross-validation method cv.option; see Athey and Imbens (2016) for a discussion. Note further that split.Rule and cv.option can mix and match.
double.Sample: boolean option, TRUE or FALSE, if set to True, causalForest will build honest trees.
split.Honest: boolean option, TRUE or FALSE, used to decide the splitting rule of the trees.
split.Bucket: boolean option, TRUE or FALSE, used to specify whether to apply the discrete method in splitting the tree. If set as TRUE, in splitting a node, the observations in a leaf will be be partitioned into buckets, with each bucket containing bucketNum treated and bucketNum control units, and where observations are ordered prior to partitioning. Splitting will take place by bucket.
bucketNum: number of observations in each bucket when set split.Bucket = TRUE. However, the code will override this choice in order to guarantee that there are at least minsize and at most bucketMax buckets.
bucketMax: Option to choose maximum number of buckets to use in splitting when set split.Bucket = TRUE, bucketNum can change by choice of bucketMax.
cv.option: cross validation options, one of "TOT", "matching", "CT", "fit", four cross validation methods in causalTree. There is no cv.option for the split.Rule "tstats"; see Athey and Imbens (2016) for discussion.
cv.Honest: boolean option, TRUE or FALSE, only used for cv.option as "CT" or "fit", to specify whether to apply honest risk evalation function in cross validation. If set TRUE, use honest risk function, otherwise use adaptive risk function in cross validation. If set FALSE, the user choice of cv.alpha will be set to 1. If set TRUE, cv.alpha will default to 0.5, but the user choice of cv.alpha will be respected. Note that honest cv estimates within-leaf variances and may perform better with larger leaf sizes and/or small number of cross-validation sets.
minsize: in order to split, each leaf must have at least minsize treated cases and minsize control cases. The default value is set as 2.
propensity: propensity score used in "TOT" splitting and "TOT", honest "CT" cross validation methods. The default value is the proportion of treated cases in all observations. In this implementation, the propensity score is a constant for the whole dataset. Unit-specific propensity scores are not supported; however, the user may use inverse propensity scores as case weights if desired.
control: a list of options that control details of the rpart algorithm. See rpart.control.
split.alpha: scale parameter between 0 and 1, used in splitting risk evaluation function for "CT". When split.Honest = FALSE, split.alpha will be set as 1. For split.Rule="tstats", if split.Honest=TRUE, split.alpha is used in calculating the risk function, which determines the order of pruning in cross-validation.
cv.alpha: scale paramter between 0 and 1, used in cross validation risk evaluation function for "CT" and "fit". When cv.Honest = FALSE, cv.alpha will be set as 1.
sample.size.total: Sample size used to build each tree in the forest (sampled randomly with replacement).
sample.size.train.frac: Fraction of the sample size used for building each tree (training). For eexample, if the sample.size.total is 1000 and frac =0.5 then, 500 samples will be used to build the tree and the other 500 samples will be used the evaluate the tree.
mtry: Number of data features used to build a tree (This variable is not used presently).
nodesize: Minimum number of observations for treated and control cases in one leaf node
ncolx: Total number of covariates

Details

CausalForest builds an ensemble of CausalTrees (See Athey and Imbens, Recursive Partitioning for Heterogeneous Causal Effects (2016)), by repeated random sampling of the data with replacement. Further, each tree is built using a randomly sampled subset of all available covariates. A causal forest object is a list of trees. To predict, call R's predict function with new test data and the causalForest object (estimated on the training data) obtained after calling the causalForest function. During the prediction phase, the average value over all tree predictions is returned as the final prediction by default. To return the predictions of each tree in the forest for each test observation, set the flag predict.all=TRUE CausalTree differs from rpart function from rpart package in splitting rules and cross validation methods. Please check Athey and Imbens, Recursive Partitioning for Heterogeneous Causal Effects (2016) and Stefan Wager and Susan Athey, Estimation and Inference of Heterogeneous Treatment Effects using Random Forests for more details.

References

Breiman L., Friedman J. H., Olshen R. A., and Stone, C. J. (1984) Classification and Regression Trees. Wadsworth.

Athey, S and G Imbens (2016) Recursive Partitioning for Heterogeneous Causal Effects. http://arxiv.org/abs/1504.01132