Package 'tidyhte'

Title: Tidy Estimation of Heterogeneous Treatment Effects
Description: Estimates heterogeneous treatment effects using tidy semantics on experimental or observational data. Methods are based on the doubly-robust learner of Kennedy (n.d.) <arXiv:2004.14497>. You provide a simple recipe for what machine learning algorithms to use in estimating the nuisance functions and 'tidyhte' will take care of cross-validation, estimation, model selection, diagnostics and construction of relevant quantities of interest about the variability of treatment effects.
Authors: Drew Dimmery [aut, cre, cph]
Maintainer: Drew Dimmery <[email protected]>
License: MIT + file LICENSE
Version: 1.0.2
Built: 2024-11-07 03:45:38 UTC
Source: https://github.com/ddimmery/tidyhte

Help Index

Add an additional diagnostic to the effect model

Description

This adds a diagnostic to the effect model.

Usage

add_effect_diagnostic(hte_cfg, diag)

Arguments

hte_cfg

HTE_cfg object to update.
diag

Character indicating the name of the diagnostic to include. Possible values are "MSE", "RROC" and, for SuperLearner ensembles, "SL_risk" and "SL_coefs".

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_effect_diagnostic("RROC") -> hte_cfg

Add an additional model to the joint effect ensemble

Description

This adds a learner to the ensemble used for estimating a model of the conditional expectation of the pseudo-outcome.

Usage

add_effect_model(hte_cfg, model_name, ...)

Arguments

hte_cfg

HTE_cfg object to update.
model_name

Character indicating the name of the model to incorporate into the joint effect ensemble. Possible values use SuperLearner naming conventions. A full list is available with SuperLearner::listWrappers("SL")
...

Parameters over which to grid-search for this model class.

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_effect_model("SL.glm.interaction") -> hte_cfg

Uses a known propensity score

Description

This replaces the propensity score model with a known value of the propensity score.

Usage

add_known_propensity_score(hte_cfg, covariate_name)

Arguments

hte_cfg

HTE_cfg object to update.
covariate_name

Character indicating the name of the covariate name in the dataframe corresponding to the known propensity score.

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_known_propensity_score("ps") -> hte_cfg

Adds moderators to the configuration

Description

This adds a definition about how to display a moderators to the MCATE config. A moderator is any variable that you want to view information about CATEs with respect to.

Usage

add_moderator(hte_cfg, model_type, ..., .model_arguments = NULL)

Arguments

hte_cfg

HTE_cfg object to update.
model_type

Character indicating the model type for these moderators. Currently two model types are supported: "Stratified" for discrete moderators and "KernelSmooth" for continuous ones.
...

The (unquoted) names of the moderator variables.
.model_arguments

A named list from argument name to value to pass into the constructor for the model. See Stratified_cfg and KernelSmooth_cfg for more details.

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_moderator("Stratified", x2, x3) %>%
   add_moderator("KernelSmooth", x1, x4, x5) -> hte_cfg

Add an additional diagnostic to the outcome model

Description

This adds a diagnostic to the outcome model.

Usage

add_outcome_diagnostic(hte_cfg, diag)

Arguments

hte_cfg

HTE_cfg object to update.
diag

Character indicating the name of the diagnostic to include. Possible values are "MSE", "RROC" and, for SuperLearner ensembles, "SL_risk" and "SL_coefs".

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_outcome_diagnostic("RROC") -> hte_cfg

Add an additional model to the outcome ensemble

Description

This adds a learner to the ensemble used for estimating a model of the conditional expectation of the outcome.

Usage

add_outcome_model(hte_cfg, model_name, ...)

Arguments

hte_cfg

HTE_cfg object to update.
model_name

Character indicating the name of the model to incorporate into the outcome ensemble. Possible values use SuperLearner naming conventions. A full list is available with SuperLearner::listWrappers("SL")
...

Parameters over which to grid-search for this model class.

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_outcome_model("SL.glm.interaction") -> hte_cfg

Add an additional diagnostic to the propensity score

Description

This adds a diagnostic to the propensity score.

Usage

add_propensity_diagnostic(hte_cfg, diag)

Arguments

hte_cfg

HTE_cfg object to update.
diag

Character indicating the name of the diagnostic to include. Possible values are "MSE", "AUC" and, for SuperLearner ensembles, "SL_risk" and "SL_coefs".

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_propensity_diagnostic(c("AUC", "MSE")) -> hte_cfg

Add an additional model to the propensity score ensemble

Description

This adds a learner to the ensemble used for estimating propensity scores.

Usage

add_propensity_score_model(hte_cfg, model_name, ...)

Arguments

hte_cfg

HTE_cfg object to update.
model_name

Character indicating the name of the model to incorporate into the propensity score ensemble. Possible values use SuperLearner naming conventions. A full list is available with SuperLearner::listWrappers("SL")
...

Parameters over which to grid-search for this model class.

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   add_propensity_score_model("SL.glmnet", alpha = c(0, 0.5, 1)) -> hte_cfg

Adds variable importance information

Description

This adds a variable importance quantity of interest to the outputs.

Usage

add_vimp(hte_cfg, sample_splitting = TRUE, linear_only = FALSE)

Arguments

hte_cfg

HTE_cfg object to update.
sample_splitting

Logical indicating whether to use sample splitting or not. Choosing not to use sample splitting means that inference will only be valid for moderators with non-null importance.
linear_only

Logical indicating whether the variable importance should use only a single linear-only model. Variable importance measure will only be consistent for the population quantity if the true model of pseudo-outcomes is linear.

Value

Updated HTE_cfg object

References

  • Williamson, B. D., Gilbert, P. B., Carone, M., & Simon, N. (2021). Nonparametric variable importance assessment using machine learning techniques. Biometrics, 77(1), 9-22.

  • Williamson, B. D., Gilbert, P. B., Simon, N. R., & Carone, M. (2021). A general framework for inference on algorithm-agnostic variable importance. Journal of the American Statistical Association, 1-14.

Examples

library("dplyr")
basic_config() %>%
   add_vimp(sample_splitting = FALSE) -> hte_cfg

Attach an HTE_cfg to a dataframe

Description

This adds a configuration attribute to a dataframe for HTE estimation. This configuration details the full analysis of HTE that should be performed.

Usage

attach_config(data, .HTE_cfg)

Arguments

data

dataframe
.HTE_cfg

HTE_cfg object representing the full configuration of the HTE analysis.

Details

For information about how to set up an HTE_cfg object, see the Recipe API documentation basic_config().

To see an example analysis, read vignette("experimental_analysis") in the context of an experiment, vignette("experimental_analysis") for an observational study, or vignette("methodological_details") for a deeper dive under the hood.

See Also

basic_config(), make_splits(), produce_plugin_estimates(), construct_pseudo_outcomes(), estimate_QoI()

Examples

library("dplyr")
if(require("palmerpenguins")) {
data(package = 'palmerpenguins')
penguins$unitid = seq_len(nrow(penguins))
penguins$propensity = rep(0.5, nrow(penguins))
penguins$treatment = rbinom(nrow(penguins), 1, penguins$propensity)
cfg <- basic_config() %>% 
add_known_propensity_score("propensity") %>%
add_outcome_model("SL.glm.interaction") %>%
remove_vimp()
attach_config(penguins, cfg) %>%
make_splits(unitid, .num_splits = 4) %>%
produce_plugin_estimates(outcome = body_mass_g, treatment = treatment, species, sex) %>%
construct_pseudo_outcomes(body_mass_g, treatment) %>%
estimate_QoI(species, sex)
}

Create a basic config for HTE estimation

Description

This provides a basic recipe for HTE estimation that can be extended by providing additional information about models to be estimated and what quantities of interest should be returned based on those models. This basic model includes only linear models for nuisance function estimation, and basic diagnostics.

Usage

basic_config()

Details

Additional models, diagnostics and quantities of interest should be added using their respective helper functions provided as part of the Recipe API.

To see an example analysis, read vignette("experimental_analysis") in the context of an experiment, vignette("experimental_analysis") for an observational study, or vignette("methodological_details") for a deeper dive under the hood.

Value

HTE_cfg object

See Also

add_propensity_score_model(), add_known_propensity_score(), add_propensity_diagnostic(), add_outcome_model(), add_outcome_diagnostic(), add_effect_model(), add_effect_diagnostic(), add_moderator(), add_vimp()

Examples

library("dplyr")
basic_config() %>%
   add_known_propensity_score("ps") %>%
   add_outcome_model("SL.glm.interaction") %>%
   add_outcome_model("SL.glmnet", alpha = c(0.05, 0.15, 0.2, 0.25, 0.5, 0.75)) %>%
   add_outcome_model("SL.glmnet.interaction", alpha = c(0.05, 0.15, 0.2, 0.25, 0.5, 0.75)) %>%
   add_outcome_diagnostic("RROC") %>%
   add_effect_model("SL.glm.interaction") %>%
   add_effect_model("SL.glmnet", alpha = c(0.05, 0.15, 0.2, 0.25, 0.5, 0.75)) %>%
   add_effect_model("SL.glmnet.interaction", alpha = c(0.05, 0.15, 0.2, 0.25, 0.5, 0.75)) %>%
   add_effect_diagnostic("RROC") %>%
   add_moderator("Stratified", x2, x3) %>%
   add_moderator("KernelSmooth", x1, x4, x5) %>%
   add_vimp(sample_splitting = FALSE) -> hte_cfg

Configuration of a Constant Estimator

Description

Constant_cfg is a configuration class for estimating a constant model. That is, the model is a simple, one-parameter mean model.

Super class

tidyhte::Model_cfg -> Constant_cfg

Public fields

model_class

The class of the model, required for all classes which inherit from Model_cfg.

Methods

Public methods

Method new()

Create a new Constant_cfg object.

Usage
Constant_cfg$new()
Returns

A new Constant_cfg object.

Examples
Constant_cfg$new()

Method clone()

The objects of this class are cloneable with this method.

Usage
Constant_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

## ------------------------------------------------
## Method `Constant_cfg$new`
## ------------------------------------------------

Constant_cfg$new()

Construct Pseudo-outcomes

Description

construct_pseudo_outcomes takes a dataset which has been prepared with plugin estimators of nuisance parameters and transforms these into a "pseudo-outcome": an unbiased estimator of the conditional average treatment effect under exogeneity.

Usage

construct_pseudo_outcomes(data, outcome, treatment, type = "dr")

Arguments

data

dataframe (already prepared with attach_config, make_splits, and produce_plugin_estimates)
outcome

Unquoted name of outcome variable.
treatment

Unquoted name of treatment variable.
type

String representing how to construct the pseudo-outcome. Valid values are "dr" (the default), "ipw" and "plugin". See "Details" for more discussion of these options.

Details

Taking averages of these pseudo-outcomes (or fitting a model to them) will approximate averages (or models) of the underlying treatment effect.

See Also

attach_config(), make_splits(), produce_plugin_estimates(), estimate_QoI()

Configuration of Model Diagnostics

Description

Diagnostics_cfg is a configuration class for estimating a variety of diagnostics for the models trained in the course of HTE estimation.

Public fields

ps

Model diagnostics for the propensity score model.

outcome

Model diagnostics for the outcome models.

effect

Model diagnostics for the joint effect model.

params

Parameters for any requested diagnostics.

Methods

Public methods

Method new()

Create a new Diagnostics_cfg object with specified diagnostics to estimate.

Usage
Diagnostics_cfg$new(ps = NULL, outcome = NULL, effect = NULL, params = NULL)
Arguments
ps

Model diagnostics for the propensity score model.

outcome

Model diagnostics for the outcome models.

effect

Model diagnostics for the joint effect model.

params

List providing values for parameters to any requested diagnostics.

Returns

A new Diagnostics_cfg object.

Examples
Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE", "RROC"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)

Method add()

Add diagnostics to the Diagnostics_cfg object.

Usage
Diagnostics_cfg$add(ps = NULL, outcome = NULL, effect = NULL)
Arguments
ps

Model diagnostics for the propensity score model.

outcome

Model diagnostics for the outcome models.

effect

Model diagnostics for the joint effect model.

Returns

An updated Diagnostics_cfg object.

Examples
cfg <- Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE", "RROC"),
   ps = c("SL_risk", "SL_coefs")
)
cfg <- cfg$add(ps = "AUC")

Method clone()

The objects of this class are cloneable with this method.

Usage
Diagnostics_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE", "RROC"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)

## ------------------------------------------------
## Method `Diagnostics_cfg$new`
## ------------------------------------------------

Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE", "RROC"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)

## ------------------------------------------------
## Method `Diagnostics_cfg$add`
## ------------------------------------------------

cfg <- Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE", "RROC"),
   ps = c("SL_risk", "SL_coefs")
)
cfg <- cfg$add(ps = "AUC")

Estimate Quantities of Interest

Description

estimate_QoI takes a dataframe already prepared with split IDs, plugin estimates and pseudo-outcomes and calculates the requested quantities of interest (QoIs).

Usage

estimate_QoI(data, ...)

Arguments

data

data frame (already prepared with attach_config, make_splits, produce_plugin_estimates and construct_pseudo_outcomes)
...

Unquoted names of moderators to calculate QoIs for.

Details

To see an example analysis, read vignette("experimental_analysis") in the context of an experiment, vignette("experimental_analysis") for an observational study, or vignette("methodological_details") for a deeper dive under the hood.

See Also

attach_config(), make_splits(), produce_plugin_estimates(), construct_pseudo_outcomes(),

Examples

library("dplyr")
if(require("palmerpenguins")) {
data(package = 'palmerpenguins')
penguins$unitid = seq_len(nrow(penguins))
penguins$propensity = rep(0.5, nrow(penguins))
penguins$treatment = rbinom(nrow(penguins), 1, penguins$propensity)
cfg <- basic_config() %>% 
add_known_propensity_score("propensity") %>%
add_outcome_model("SL.glm.interaction") %>%
remove_vimp()
attach_config(penguins, cfg) %>%
make_splits(unitid, .num_splits = 4) %>%
produce_plugin_estimates(outcome = body_mass_g, treatment = treatment, species, sex) %>%
construct_pseudo_outcomes(body_mass_g, treatment) %>%
estimate_QoI(species, sex)
}

Configuration of Quantities of Interest

Description

HTE_cfg is a configuration class that pulls everything together, indicating the full configuration for a given HTE analysis. This includes how to estimate models and what Quantities of Interest to calculate based off those underlying models.

Public fields

outcome

Model_cfg object indicating how outcome models should be estimated.

treatment

Model_cfg object indicating how the propensity score model should be estimated.

effect

Model_cfg object indicating how the joint effect model should be estimated.

qoi

QoI_cfg object indicating what the Quantities of Interest are and providing all necessary detail on how they should be estimated.

verbose

Logical indicating whether to print debugging information.

Methods

Public methods

Method new()

Create a new HTE_cfg object with all necessary information about how to carry out an HTE analysis.

Usage
HTE_cfg$new(
  outcome = NULL,
  treatment = NULL,
  effect = NULL,
  qoi = NULL,
  verbose = FALSE
)
Arguments
outcome

Model_cfg object indicating how outcome models should be estimated.

treatment

Model_cfg object indicating how the propensity score model should be estimated.

effect

Model_cfg object indicating how the joint effect model should be estimated.

qoi

QoI_cfg object indicating what the Quantities of Interest are and providing all necessary detail on how they should be estimated.

verbose

Logical indicating whether to print debugging information.

Examples
mcate_cfg <- MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))
pcate_cfg <- PCATE_cfg$new(
   cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)),
   model_covariates = c("x1", "x2", "x3"),
   num_mc_samples = list(x1 = 100)
)
vimp_cfg <- VIMP_cfg$new()
diag_cfg <- Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)
qoi_cfg <- QoI_cfg$new(
    mcate = mcate_cfg,
    pcate = pcate_cfg,
    vimp = vimp_cfg,
    diag = diag_cfg
)
ps_cfg <- SLEnsemble_cfg$new(
   learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)
y_cfg <- SLEnsemble_cfg$new(
   learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)
fx_cfg <- SLEnsemble_cfg$new(
   learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)
HTE_cfg$new(outcome = y_cfg, treatment = ps_cfg, effect = fx_cfg, qoi = qoi_cfg)

Method clone()

The objects of this class are cloneable with this method.

Usage
HTE_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

## ------------------------------------------------
## Method `HTE_cfg$new`
## ------------------------------------------------

mcate_cfg <- MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))
pcate_cfg <- PCATE_cfg$new(
   cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)),
   model_covariates = c("x1", "x2", "x3"),
   num_mc_samples = list(x1 = 100)
)
vimp_cfg <- VIMP_cfg$new()
diag_cfg <- Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)
qoi_cfg <- QoI_cfg$new(
    mcate = mcate_cfg,
    pcate = pcate_cfg,
    vimp = vimp_cfg,
    diag = diag_cfg
)
ps_cfg <- SLEnsemble_cfg$new(
   learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)
y_cfg <- SLEnsemble_cfg$new(
   learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)
fx_cfg <- SLEnsemble_cfg$new(
   learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)
HTE_cfg$new(outcome = y_cfg, treatment = ps_cfg, effect = fx_cfg, qoi = qoi_cfg)

Configuration for a Kernel Smoother

Description

KernelSmooth_cfg is a configuration class for non-parametric local-linear regression to construct a smooth representation of the relationship between two variables. This is typically used for displaying a surface of the conditional average treatment effect over a continuous covariate.

Kernel smoothing is handled by the nprobust package.

Super class

tidyhte::Model_cfg -> KernelSmooth_cfg

Public fields

model_class

The class of the model, required for all classes which inherit from Model_cfg.

neval

The number of points at which to evaluate the local regression. More points will provide a smoother line at the cost of somewhat higher computation.

eval_min_quantile

Minimum quantile at which to evaluate the smoother.

Methods

Public methods

Method new()

Create a new KernelSmooth_cfg object with specified number of evaluation points.

Usage
KernelSmooth_cfg$new(neval = 100, eval_min_quantile = 0.05)
Arguments
neval

The number of points at which to evaluate the local regression. More points will provide a smoother line at the cost of somewhat higher computation.

eval_min_quantile

Minimum quantile at which to evaluate the smoother. A value of zero will do no clipping. Clipping is performed from both the top and the bottom of the empirical distribution. A value of alpha would evaluate over [alpha, 1 - alpha].

Returns

A new KernelSmooth_cfg object.

Examples
KernelSmooth_cfg$new(neval = 100)

Method clone()

The objects of this class are cloneable with this method.

Usage
KernelSmooth_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

See Also

nprobust::lprobust

Examples

## ------------------------------------------------
## Method `KernelSmooth_cfg$new`
## ------------------------------------------------

KernelSmooth_cfg$new(neval = 100)

Configuration of Known Model

Description

Known_cfg is a configuration class for when a particular model is known a-priori. The prototypical usage of this class is when heterogeneous treatment effects are estimated in the context of a randomized control trial with known propensity scores.

Super class

tidyhte::Model_cfg -> Known_cfg

Public fields

covariate_name

The name of the column in the dataset which corresponds to the known model score.

model_class

The class of the model, required for all classes which inherit from Model_cfg.

Methods

Public methods

Method new()

Create a new Known_cfg object with specified covariate column.

Usage
Known_cfg$new(covariate_name)
Arguments
covariate_name

The name of the column, a string, in the dataset corresponding to the known model score (i.e. the true conditional expectation).

Returns

A new Known_cfg object.

Examples
Known_cfg$new("propensity_score")

Method clone()

The objects of this class are cloneable with this method.

Usage
Known_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

## ------------------------------------------------
## Method `Known_cfg$new`
## ------------------------------------------------

Known_cfg$new("propensity_score")

Define splits for cross-fitting

Description

This takes a dataset, a column with a unique identifier and an arbitrary number of covariates on which to stratify the splits. It returns the original dataset with an additional column .split_id corresponding to an identifier for the split.

Usage

make_splits(data, identifier, ..., .num_splits)

Arguments

data

dataframe
identifier

Unquoted name of unique identifier column
...

variables on which to stratify (requires that quickblock be installed.)
.num_splits

number of splits to create. If VIMP is requested in QoI_cfg, this must be an even number.

Details

To see an example analysis, read vignette("experimental_analysis") in the context of an experiment, vignette("experimental_analysis") for an observational study, or vignette("methodological_details") for a deeper dive under the hood.

Value

original dataframe with additional .split_id column

See Also

attach_config(), produce_plugin_estimates(), construct_pseudo_outcomes(), estimate_QoI()

Examples

library("dplyr")
if(require("palmerpenguins")) {
data(package = 'palmerpenguins')
penguins$unitid = seq_len(nrow(penguins))
penguins$propensity = rep(0.5, nrow(penguins))
penguins$treatment = rbinom(nrow(penguins), 1, penguins$propensity)
cfg <- basic_config() %>% 
add_known_propensity_score("propensity") %>%
add_outcome_model("SL.glm.interaction") %>%
remove_vimp()
attach_config(penguins, cfg) %>%
make_splits(unitid, .num_splits = 4) %>%
produce_plugin_estimates(outcome = body_mass_g, treatment = treatment, species, sex) %>%
construct_pseudo_outcomes(body_mass_g, treatment) %>%
estimate_QoI(species, sex)
}

Configuration of Marginal CATEs

Description

MCATE_cfg is a configuration class for estimating marginal response surfaces based on heterogeneous treatment effect estimates. "Marginal" in this context implies that all other covariates are marginalized. Thus, if two covariates are highly correlated, it is likely that their MCATE surfaces will be extremely similar.

Public fields

cfgs

Named list of covariates names to a Model_cfg object defining how to present that covariate's CATE surface (while marginalizing over all other covariates).

std_errors

Boolean indicating whether the results should be returned with standard errors or not.

estimand

String indicating the estimand to target.

Methods

Public methods

Method new()

Create a new MCATE_cfg object with specified model name and hyperparameters.

Usage
MCATE_cfg$new(cfgs, std_errors = TRUE)
Arguments
cfgs

Named list from moderator name to a Model_cfg object defining how to present that covariate's CATE surface (while marginalizing over all other covariates)

std_errors

Boolean indicating whether the results should be returned with standard errors or not.

Returns

A new MCATE_cfg object.

Examples
MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))

Method add_moderator()

Add a moderator to the MCATE_cfg object. This entails defining a configuration for displaying the effect surface for that moderator.

Usage
MCATE_cfg$add_moderator(var_name, cfg)
Arguments
var_name

The name of the moderator to add (and the name of the column in the dataset).

cfg

A Model_cfg defining how to display the selected moderator's effect surface.

Returns

An updated MCATE_cfg object.

Examples
cfg <- MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))
cfg <- cfg$add_moderator("x2", KernelSmooth_cfg$new(neval = 100))

Method clone()

The objects of this class are cloneable with this method.

Usage
MCATE_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))

## ------------------------------------------------
## Method `MCATE_cfg$new`
## ------------------------------------------------

MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))

## ------------------------------------------------
## Method `MCATE_cfg$add_moderator`
## ------------------------------------------------

cfg <- MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))
cfg <- cfg$add_moderator("x2", KernelSmooth_cfg$new(neval = 100))

Base Class of Model Configurations

Description

Model_cfg is the base class from which all other model configurations inherit.

Public fields

model_class

The class of the model, required for all classes which inherit from Model_cfg.

Methods

Public methods

Method new()

Create a new Model_cfg object with any necessary parameters.

Usage
Model_cfg$new()
Returns

A new Model_cfg object.

Method clone()

The objects of this class are cloneable with this method.

Usage
Model_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

R6 class to represent data to be used in estimating a model

Description

R6 class to represent data to be used in estimating a model

R6 class to represent data to be used in estimating a model

Details

This class provides consistent names and interfaces to data which will be used in a supervised regression / classification model.

Public fields

label

The labels for the eventual model as a vector.

features

The matrix representation of the data to be used for model fitting. Constructed using stats::model.matrix.

model_frame

The data-frame representation of the data as constructed by stats::model.frame.

split_id

The split identifiers as a vector.

num_splits

The integer number of splits in the data.

cluster

A cluster ID as a vector, constructed using the unit identifiers.

weights

The case-weights as a vector.

Methods

Public methods

Method new()

Creates an R6 object to represent data to be used in a prediction model.

Usage
Model_data$new(data, label_col, ..., .weight_col = NULL)
Arguments
data

The full dataset to populate the class with.

label_col

The unquoted name of the column to use as the label in supervised learning models.

...

The unquoted names of features to use in the model.

.weight_col

The unquoted name of the column to use as case-weights in subsequent models.

Returns

A Model_data object.

Examples
library("dplyr")
df <- dplyr::tibble(
    uid = 1:100,
    x1 = rnorm(100),
    x2 = rnorm(100),
    x3 = sample(4, 100, replace = TRUE)
) %>% dplyr::mutate(
    y = x1 + x2 + x3 + rnorm(100),
    x3 = factor(x3)
)
df <- make_splits(df, uid, .num_splits = 5)
data <- Model_data$new(df, y, x1, x2, x3)

Method SL_cv_control()

A helper function to create the cross-validation options to be used by SuperLearner.

Usage
Model_data$SL_cv_control()

Method clone()

The objects of this class are cloneable with this method.

Usage
Model_data$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

See Also

SuperLearner::SuperLearner.CV.control

Examples

## ------------------------------------------------
## Method `Model_data$new`
## ------------------------------------------------

library("dplyr")
df <- dplyr::tibble(
    uid = 1:100,
    x1 = rnorm(100),
    x2 = rnorm(100),
    x3 = sample(4, 100, replace = TRUE)
) %>% dplyr::mutate(
    y = x1 + x2 + x3 + rnorm(100),
    x3 = factor(x3)
)
df <- make_splits(df, uid, .num_splits = 5)
data <- Model_data$new(df, y, x1, x2, x3)

Prediction for an SL.glmnet object

Description

Prediction for the glmnet wrapper.

Usage

## S3 method for class 'SL.glmnet.interaction'
predict(
  object,
  newdata,
  remove_extra_cols = TRUE,
  add_missing_cols = TRUE,
  ...
)

Arguments

object

Result object from SL.glmnet
newdata

Dataframe or matrix that will generate predictions.
remove_extra_cols

Remove any extra columns in the new data that were not part of the original model.
add_missing_cols

Add any columns from original data that do not exist in the new data, and set values to 0.
...

Any additional arguments (not used).

See Also

SL.glmnet

Estimate models of nuisance functions

Description

This takes a dataset with an identified outcome and treatment column along with any number of covariates and appends three columns to the dataset corresponding to an estimate of the conditional expectation of treatment (.pi_hat), along with the conditional expectation of the control and treatment potential outcome surfaces (.mu0_hat and .mu1_hat respectively).

Usage

produce_plugin_estimates(data, outcome, treatment, ..., .weights = NULL)

Arguments

data

dataframe (already prepared with attach_config and make_splits)
outcome

Unquoted name of the outcome variable.
treatment

Unquoted name of the treatment variable.
...

Unquoted names of covariates to include in the models of the nuisance functions.
.weights

Unquoted name of weights column. If NULL, all analysis will assume weights are all equal to one and sample-based quantities will be returned.

Details

To see an example analysis, read vignette("experimental_analysis") in the context of an experiment, vignette("experimental_analysis") for an observational study, or vignette("methodological_details") for a deeper dive under the hood.

See Also

attach_config(), make_splits(), construct_pseudo_outcomes(), estimate_QoI()

Examples

library("dplyr")
if(require("palmerpenguins")) {
data(package = 'palmerpenguins')
penguins$unitid = seq_len(nrow(penguins))
penguins$propensity = rep(0.5, nrow(penguins))
penguins$treatment = rbinom(nrow(penguins), 1, penguins$propensity)
cfg <- basic_config() %>% 
add_known_propensity_score("propensity") %>%
add_outcome_model("SL.glm.interaction") %>%
remove_vimp()
attach_config(penguins, cfg) %>%
make_splits(unitid, .num_splits = 4) %>%
produce_plugin_estimates(outcome = body_mass_g, treatment = treatment, species, sex) %>%
construct_pseudo_outcomes(body_mass_g, treatment) %>%
estimate_QoI(species, sex)
}

Configuration of Quantities of Interest

Description

QoI_cfg is a configuration class for the Quantities of Interest to be generated by the HTE analysis.

Public fields

mcate

A configuration object of type MCATE_cfg of marginal effects to calculate.

pcate

A configuration object of type PCATE_cfg of partial effects to calculate.

vimp

A configuration object of type VIMP_cfg of variable importance to calculate.

diag

A configuration object of type Diagnostics_cfg of model diagnostics to calculate.

ate

Logical flag indicating whether an estimate of the ATE should be returned.

predictions

Logical flag indicating whether estimates of the CATE for every unit should be returned.

Methods

Public methods

Method new()

Create a new QoI_cfg object with specified Quantities of Interest to estimate.

Usage
QoI_cfg$new(
  mcate = NULL,
  pcate = NULL,
  vimp = NULL,
  diag = NULL,
  ate = TRUE,
  predictions = FALSE
)
Arguments
mcate

A configuration object of type MCATE_cfg of marginal effects to calculate.

pcate

A configuration object of type PCATE_cfg of partial effects to calculate.

vimp

A configuration object of type VIMP_cfg of variable importance to calculate.

diag

A configuration object of type Diagnostics_cfg of model diagnostics to calculate.

ate

A logical flag for whether to calculate the Average Treatment Effect (ATE) or not.

predictions

A logical flag for whether to return predictions of the CATE for every unit or not.

Returns

A new Diagnostics_cfg object.

Examples
mcate_cfg <- MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))
pcate_cfg <- PCATE_cfg$new(
   cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)),
   model_covariates = c("x1", "x2", "x3"),
   num_mc_samples = list(x1 = 100)
)
vimp_cfg <- VIMP_cfg$new()
diag_cfg <- Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)
QoI_cfg$new(
    mcate = mcate_cfg,
    pcate = pcate_cfg,
    vimp = vimp_cfg,
    diag = diag_cfg
)

Method clone()

The objects of this class are cloneable with this method.

Usage
QoI_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

mcate_cfg <- MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))
pcate_cfg <- PCATE_cfg$new(
   cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)),
   model_covariates = c("x1", "x2", "x3"),
   num_mc_samples = list(x1 = 100)
)
vimp_cfg <- VIMP_cfg$new()
diag_cfg <- Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)
QoI_cfg$new(
    mcate = mcate_cfg,
    pcate = pcate_cfg,
    vimp = vimp_cfg,
    diag = diag_cfg
)

## ------------------------------------------------
## Method `QoI_cfg$new`
## ------------------------------------------------

mcate_cfg <- MCATE_cfg$new(cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)))
pcate_cfg <- PCATE_cfg$new(
   cfgs = list(x1 = KernelSmooth_cfg$new(neval = 100)),
   model_covariates = c("x1", "x2", "x3"),
   num_mc_samples = list(x1 = 100)
)
vimp_cfg <- VIMP_cfg$new()
diag_cfg <- Diagnostics_cfg$new(
   outcome = c("SL_risk", "SL_coefs", "MSE"),
   ps = c("SL_risk", "SL_coefs", "AUC")
)
QoI_cfg$new(
    mcate = mcate_cfg,
    pcate = pcate_cfg,
    vimp = vimp_cfg,
    diag = diag_cfg
)

Removes variable importance information

Description

This removes the variable importance quantity of interest from an HTE_cfg.

Usage

remove_vimp(hte_cfg)

Arguments

hte_cfg

HTE_cfg object to update.

Value

Updated HTE_cfg object

Examples

library("dplyr")
basic_config() %>%
   remove_vimp() -> hte_cfg

Elastic net regression with pairwise interactions

Description

Penalized regression using elastic net. Alpha = 0 corresponds to ridge regression and alpha = 1 corresponds to Lasso. Included in the model are pairwise interactions between covariates.

See vignette("glmnet_beta", package = "glmnet") for a nice tutorial on glmnet.

Usage

SL.glmnet.interaction(
  Y,
  X,
  newX,
  family,
  obsWeights,
  id,
  alpha = 1,
  nfolds = 10,
  nlambda = 100,
  useMin = TRUE,
  loss = "deviance",
  ...
)

Arguments

Y

Outcome variable
X

Covariate dataframe
newX

Dataframe to predict the outcome
family

"gaussian" for regression, "binomial" for binary classification. Untested options: "multinomial" for multiple classification or "mgaussian" for multiple response, "poisson" for non-negative outcome with proportional mean and variance, "cox".
obsWeights

Optional observation-level weights
id

Optional id to group observations from the same unit (not used currently).
alpha

Elastic net mixing parameter, range [0, 1]. 0 = ridge regression and 1 = lasso.
nfolds

Number of folds for internal cross-validation to optimize lambda.
nlambda

Number of lambda values to check, recommended to be 100 or more.
useMin

If TRUE use lambda that minimizes risk, otherwise use 1 standard-error rule which chooses a higher penalty with performance within one standard error of the minimum (see Breiman et al. 1984 on CART for background).
loss

Loss function, can be "deviance", "mse", or "mae". If family = binomial can also be "auc" or "class" (misclassification error).
...

Any additional arguments are passed through to cv.glmnet.

Configuration for a SuperLearner Ensemble

Description

SLEnsemble_cfg is a configuration class for estimation of a model using an ensemble of models using SuperLearner.

Super class

tidyhte::Model_cfg -> SLEnsemble_cfg

Public fields

cvControl

A list of parameters for controlling the cross-validation used in SuperLearner.

SL.library

A vector of the names of learners to include in the SuperLearner ensemble.

SL.env

An environment containing all of the programmatically generated learners to be included in the SuperLearner ensemble.

family

stats::family object to determine how SuperLearner should be fitted.

model_class

The class of the model, required for all classes which inherit from Model_cfg.

Methods

Public methods

Method new()

Create a new SLEnsemble_cfg object with specified settings.

Usage
SLEnsemble_cfg$new(
  cvControl = NULL,
  learner_cfgs = NULL,
  family = stats::gaussian()
)
Arguments
cvControl

A list of parameters for controlling the cross-validation used in SuperLearner. For more details, see SuperLearner::SuperLearner.CV.control.

learner_cfgs

A list of SLLearner_cfg objects.

family

stats::family object to determine how SuperLearner should be fitted.

Returns

A new SLEnsemble_cfg object.

Examples
SLEnsemble_cfg$new(
learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)

Method add_sublearner()

Adds a model (or class of models) to the SuperLearner ensemble. If hyperparameter values are specified, this method will add a learner for every element in the cross-product of provided hyperparameter values.

Usage
SLEnsemble_cfg$add_sublearner(learner_name, hps = NULL)
Arguments
learner_name

Possible values use SuperLearner naming conventions. A full list is available with SuperLearner::listWrappers("SL")

hps

A named list of hyper-parameters. Every element of the cross-product of these hyper-parameters will be included in the ensemble. cfg <- SLEnsemble_cfg$new( learner_cfgs = list(SLLearner_cfg$new("SL.glm")) ) cfg <- cfg$add_sublearner("SL.gam", list(deg.gam = c(2, 3)))

Method clone()

The objects of this class are cloneable with this method.

Usage
SLEnsemble_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

SLEnsemble_cfg$new(
learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)

## ------------------------------------------------
## Method `SLEnsemble_cfg$new`
## ------------------------------------------------

SLEnsemble_cfg$new(
learner_cfgs = list(SLLearner_cfg$new("SL.glm"), SLLearner_cfg$new("SL.gam"))
)

Configuration of SuperLearner Submodel

Description

SLLearner_cfg is a configuration class for a single sublearner to be included in SuperLearner. By constructing with a named list of hyperparameters, this configuration allows distinct submodels for each unique combination of hyperparameters. To understand what models and hyperparameters are available, examine the methods listed in SuperLearner::listWrappers("SL").

Public fields

model_name

The name of the model as passed to SuperLearner through the SL.library parameter.

hyperparameters

Named list from hyperparameter name to a vector of values that should be swept over.

Methods

Public methods

Method new()

Create a new SLLearner_cfg object with specified model name and hyperparameters.

Usage
SLLearner_cfg$new(model_name, hp = NULL)
Arguments
model_name

The name of the model as passed to SuperLearner through the SL.library parameter.

hp

Named list from hyperparameter name to a vector of values that should be swept over. Hyperparameters not included in this list are left at their SuperLearner default values.

Returns

A new SLLearner_cfg object.

Examples
SLLearner_cfg$new("SL.glm")
SLLearner_cfg$new("SL.gam", list(deg.gam = c(2, 3)))

Method clone()

The objects of this class are cloneable with this method.

Usage
SLLearner_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

## ------------------------------------------------
## Method `SLLearner_cfg$new`
## ------------------------------------------------

SLLearner_cfg$new("SL.glm")
SLLearner_cfg$new("SL.gam", list(deg.gam = c(2, 3)))

Configuration for a Stratification Estimator

Description

Stratified_cfg is a configuration class for stratifying a covariate and calculating statistics within each cell.

Super class

tidyhte::Model_cfg -> Stratified_cfg

Public fields

model_class

The class of the model, required for all classes which inherit from Model_cfg.

covariate

The name of the column in the dataset which corresponds to the covariate on which to stratify.

Methods

Public methods

Method new()

Create a new Stratified_cfg object with specified number of evaluation points.

Usage
Stratified_cfg$new(covariate)
Arguments
covariate

The name of the column in the dataset which corresponds to the covariate on which to stratify.

Returns

A new Stratified_cfg object.

Examples
Stratified_cfg$new(covariate = "test_covariate")

Method clone()

The objects of this class are cloneable with this method.

Usage
Stratified_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

Examples

## ------------------------------------------------
## Method `Stratified_cfg$new`
## ------------------------------------------------

Stratified_cfg$new(covariate = "test_covariate")

Configuration of Variable Importance

Description

VIMP_cfg is a configuration class for estimating a variable importance measure across all moderators. This provides a meaningful measure of which moderators explain the most of the CATE surface.

Public fields

estimand

String indicating the estimand to target.

sample_splitting

Logical indicating whether to use sample splitting in the calculation of variable importance.

linear

Logical indicating whether the variable importance assuming a linear model should be estimated.

Methods

Public methods

Method new()

Create a new VIMP_cfg object with specified model configuration.

Usage
VIMP_cfg$new(sample_splitting = TRUE, linear_only = FALSE)
Arguments
sample_splitting

Logical indicating whether to use sample splitting in the calculation of variable importance. Choosing not to use sample splitting means that inference will only be valid for moderators with non-null importance.

linear_only

Logical indicating whether the variable importance should use only a single linear-only model. Variable importance measure will only be consistent for the population quantity if the true model of pseudo-outcomes is linear.

Returns

A new VIMP_cfg object.

Examples
VIMP_cfg$new()

Method clone()

The objects of this class are cloneable with this method.

Usage
VIMP_cfg$clone(deep = FALSE)
Arguments
deep

Whether to make a deep clone.

References

  • Williamson, B. D., Gilbert, P. B., Carone, M., & Simon, N. (2021). Nonparametric variable importance assessment using machine learning techniques. Biometrics, 77(1), 9-22.

  • Williamson, B. D., Gilbert, P. B., Simon, N. R., & Carone, M. (2021). A general framework for inference on algorithm-agnostic variable importance. Journal of the American Statistical Association, 1-14.

Examples

VIMP_cfg$new()

## ------------------------------------------------
## Method `VIMP_cfg$new`
## ------------------------------------------------

VIMP_cfg$new()