Package 'causalOT'

Description

Barycentric Projection outcome estimation

Usage

barycentric_projection(
  formula,
  data,
  weights,
  separate.samples.on = "z",
  penalty = NULL,
  cost_function = NULL,
  p = 2,
  debias = FALSE,
  cost.online = "auto",
  diameter = NULL,
  niter = 1000L,
  tol = 1e-07,
  ...
)

Arguments

Details

The barycentric projection uses the dual potentials from the optimal transport distance between the two samples to calculate projections from one sample into another. For example, in the sample of controls, we may wish to know their outcome had they been treated. In general, we then seek to minimize

$\text{argmin}_{\eta} \sum_{ij} cost(\eta_i, y_j) \pi_{ij}$

where $\pi_{ij}$ is the primal solution from the optimal transport problem.

These values can also be de-biased using the solutions from running an optimal transport problem of one sample against itself. Details are listed in Pooladian et al. (2022) https://arxiv.org/abs/2202.08919.

Value

An object of class "bp" which is a list with slots:

• potentials The dual potentials from calculating the optimal transport distance

• penalty The value of the penalty parameter used in calculating the optimal transport distance

• cost_function The cost function used to calculate the distances between units.

• cost_alg A character vector denoting if an $L_1$ distance, a squared euclidean distance, or other distance metric was used.

• p The power to which the cost matrix was raised if not using a user supplied cost function.

• debias Whether barycentric projections should be debiased.

• tensorized TRUE/FALSE denoting wether to use offline cost matrices.

• data An object of class dataHolder with the data used to calculate the optimal transport distance.

• y_a The outcome vector in the first sample.

• y_b The outcome vector in the second sample.

• x_a The covariate matrix in the first sample.

• x_b The covariate matrix in the second sample.

• a The empirical measure in the first sample.

• b The empirical measure in the second sample.

• terms The terms object from the formula.

Examples

if(torch::torch_is_installed()) {
set.seed(23483)
n <- 2^5
pp <- 6
overlap <- "low"
design <- "A"
estimate <- "ATT"
power <- 2
data <- causalOT::Hainmueller$new(n = n, p = pp,
design = design, overlap = overlap)

data$gen_data()

weights <- causalOT::calc_weight(x = data,
  z = NULL, y = NULL,
  estimand = estimate,
  method = "NNM")
  
 df <- data.frame(y = data$get_y(), z = data$get_z(), data$get_x())
  
 fit <- causalOT::barycentric_projection(y ~ ., data = df, 
    weight = weights,
    separate.samples.on = "z",
    niter = 2)
 inherits(fit, "bp")
 }

Description

Estimate causal weights

Usage

calc_weight(
  x,
  z,
  estimand = c("ATC", "ATT", "ATE"),
  method = supported_methods(),
  options = NULL,
  weights = NULL,
  ...
)

Arguments

Details

We detail some of the particulars of the function arguments below.

Causal Optimal Transport (COT)

This is the.main method of the package. This method relies on various solvers depending on the particular options chosen. Please see cotOptions() for more details.

Energy Balancing Weights (EnergyBW)

This is equivalent to COT with an infinite penalty parameter, options(lambda = Inf). Uses the same solver and options as COT, cotOptions().

Nearest Neighbor Matching with replacement (NNM)

This is equivalent to COT with a penalty parameter = 0, options(lambda = 0). Uses the same solver and options as COT, cotOptions().

Synthetic Control Method (SCM)

The SCM method is equivalent to an OT problem from a different angle. See scmOptions().

Entropy Balancing Weights (EntropyBW)

This method balances chosen functions of the covariates specified in the data argument, x. See entBWOptions() for more details. Hainmueller (2012).

Stable Balancing Weights (SBW)

Entropy Balancing Weights with a different penalty parameter, proposed by Zuizarreta (2012). See sbwOptions() for more details

Covariate Balancing Propensity Score (CBPS)

The CBPS method of Imai and Ratkovic. Options argument is passed to the function CBPS().

Logistic Regression or Probit Regression

The main methods historically for implementing inverse probability weights. Options are passed directly to the glm function from R.

Value

An object of class causalWeights

See Also

estimate_effect()

Examples

set.seed(23483)
n <- 2^5
p <- 6
#### get data ####
data <- Hainmueller$new(n = n, p = p)
data$gen_data()
x <- data$get_x()
z <- data$get_z()

if (torch::torch_is_installed()) {
# estimate weights
weights <- calc_weight(x = x,
                                 z = z, 
                                 estimand = "ATE",
                                 method = "COT",
                                 options = list(lambda = 0))
#we can also use the dataSim object directly
weightsDS <- calc_weight(x = data,
                                 z = NULL,
                                 estimand = "ATE",
                                 method = "COT",
                                 options = list(lambda = 0))
all.equal(weights@w0, weightsDS@w0)
all.equal(weights@w1, weightsDS@w1)
}

Description

causalWeights class

Details

This object is returned by the calc_weight function in this package. The slots can be accessed as any S4 object. There is no publicly accessible constructor function.

Slots

w0

A slot with the weights for the control group with $n_0$ entries. Weights sum to 1.

w1

The weights for the treated group with $n_1$ entries. Weights sum to 1.

estimand

A character denoting the estimand targeted by the weights. One of "ATT","ATC", or "ATE".

info

A slot to store a variety of info for inference. Currently under development.

method

A character denoting the method used to estimate the weights.

penalty

A list or the selected penalty parameters, if relevant.

data

The dataHolder object containing the original data.

call

The call used to construct the weights.

Description

Extract treatment effect estimate

Usage

## S3 method for class 'causalEffect'
coef(object, ...)

Arguments

Value

A number corresponding to the estimated treatment effect

Examples

# set-up data
set.seed(1234)
data <- Hainmueller$new()
data$gen_data()

# calculate quantities
weight <- calc_weight(data, method = "Logistic", estimand = "ATE")
tx_eff <- estimate_effect(causalWeights = weight)

all.equal(coef(tx_eff), c(estimate = tx_eff@estimate))

Description

Options available for the COT method

Usage

cotOptions(
  lambda = NULL,
  delta = NULL,
  opt.direction = c("dual", "primal"),
  debias = TRUE,
  p = 2,
  cost.function = NULL,
  cost.online = "auto",
  diameter = NULL,
  balance.formula = NULL,
  quick.balance.function = TRUE,
  grid.length = 7L,
  torch.optimizer = torch::optim_rmsprop,
  torch.scheduler = torch::lr_multiplicative,
  niter = 2000,
  nboot = 100L,
  lambda.bootstrap = 0.05,
  tol = 1e-04,
  device = NULL,
  dtype = NULL,
  ...
)

Arguments

Value

A list of class cotOptions with the following slots

• lambdaThe penalty parameter for the optimal transport distance

• deltaThe constraint for the balancing functions

• opt.direction Whether to solve the primal or dual optimization problems

• debiasTRUE or FALSE if debiased optimal transport distances are used

• balance.formula The formula giving how to generate the balancing functions.

• quick.balance.function TRUE or FALSE whether quick balance functions will be run.

• grid.length The number of parameters to check in a grid search of best parameters

• p The power of the cost function

• cost.online Whether online costs are used

• cost.function The user supplied cost function if supplied.

• diameter The diameter of the covariate space.

• torch.optimizer The torch optimizer used for Sinkhorn Divergences

• torch.scheduler The scheduler for the torch optimizer

• solver.options The arguments to be passeed to the torch.optimizer

• scheduler.options The arguments to be passeed to the torch.scheduler

• osqp.options Arguments passed to the osqp function if quick balance functions are used.

• niter The number of iterations to run the solver

• nboot The number of bootstrap samples

• lambda.bootstrap The penalty parameter to use for the bootstrap hyperparameter selection.

• tol The tolerance for convergence.

• device An object of class torch_device.

• dtype An object of class torch_dtype.

Solvers and distances

The function is setup to direct the COT optimizer to run two basic methods: debiased entropy penalized optimal transport (Sinkhorn Divergences) or entropy penalized optimal transport (Sinkhorn Distances).

Sinkhorn Distances

The optimal transport problem solved is $min_w OT_\lambda(w,b)$ where

$OT_\lambda(w,b) = \sum_{ij} C(x_i, x_j) P_{ij} + \lambda \sum_{ij} P_{ij}\log(P_{ij}),$

such that the rows of the matrix $P_{ij}$ sum to $w$ and the columns sum to $b$. In this case $C(,)$ is the cost between units i and j.

Sinkhorn Divergences

The Sinkhorn Divergence solves

$min_w OT_\lambda(w,b) - 0.5 OT_\lambda(w,w) - 0.5 * OT_\lambda(b,b).$

The solver for this function uses the torch package in R and by default will use the optim_rmsprop solver. Your desired torch optimizer can be passed via torch.optimizer with a scheduler passed via torch.scheduler. GPU support is available as detailed in the torch package. Additional arguments in ... are passed as extra arguments to the torch optimizer and schedulers as appropriate.

Function balancing

There may be certain functions of the covariates that we wish to balance within some tolerance, $\delta$. For these functions $B$, we will desire

$\frac{\sum_{i: Z_i = 0} w_i B(x_i) - \sum_{j: Z_j = 1} B(x_j)/n_1}{\sigma} \leq \delta$

, where in this case we are targeting balance with the treatment group for the ATT. $\sigma$ is the pooled standard deviation prior to balancing.

Cost functions

The cost function specifies pairwise distances. If argument cost.function is NULL, the function will default to using $L_p^p$ distances with a default $p = 2$ supplied by the argument p. So for p = 2, the cost between units $x_i$ and $x_j$ will be

$C(x_i, x_j) = \frac{1}{2} \| x_i - x_j \|_2^2.$

If cost.function is provided, it should be a function that takes arguments x1, x2, and p: function(x1, x2, p){...}.

Examples

if ( torch::torch_is_installed()) {
opts1 <- cotOptions(lambda = 1e3, torch.optimizer = torch::optim_rmsprop)
opts2 <- cotOptions(lambda = NULL)
opts3 <- cotOptions(lambda = seq(0.1, 100, length.out = 7))
}

Description

CRASH3 data example

CRASH3 data example

Details

Returns the CRASH3 data. Note that gen_data() will initialize the fixed data for x and y, but z is generated from Binom(0.5).

Value

An R6 object of class DataSim

Super class

causalOT::DataSim -> CRASH3

Public fields

Methods

Public methods

• CRASH3$gen_data()

• CRASH3$gen_x()

• CRASH3$gen_y()

• CRASH3$gen_z()

• CRASH3$new()

• CRASH3$clone()

Method gen_data()

The site ID for the observations

Draws new treatment indicators. x and y data are fixed.

Usage

CRASH3$gen_data()

Method gen_x()

Sets up the covariate data. This data is fixed.

Usage

CRASH3$gen_x()

Method gen_y()

Sets up the outcome data. This data is fixed.

Usage

CRASH3$gen_y()

Method gen_z()

Sets up the treatment indicator. Drawn as Z ~ Binom(0.5)

Usage

CRASH3$gen_z()

Method new()

Initializes the CRASH3 object.

Usage

CRASH3$new(n = NULL, p = NULL, param = list(), design = NA_character_, ...)

Arguments

Examples

crash <- CRASH3$new()
crash$gen_data()
crash$get_n()
crash$site_id

Method clone()

The objects of this class are cloneable with this method.

Usage

CRASH3$clone(deep = FALSE)

Arguments

Examples

## ------------------------------------------------
## Method `CRASH3$new`
## ------------------------------------------------

crash <- CRASH3$new()
crash$gen_data()
crash$get_n()
crash$site_id

Description

dataHolder

Usage

dataHolder(x, z, y = NA_real_, weights = NA_real_)

Arguments

Details

Creates an object used internally by the causalOT package for data management.

Value

Returns an object of class dataHolder with slots

• x matrix. A matrix of confounders.

• z integer. The treatment indicator, $z_i \in \{0,1\}$.

• y numeric. The outcome data.

• n0 integer. The number of observations where z==0

• n1 integer. The number of observations where z==1

• weights numeric. The empirical distribution of the full sample.

Examples

x <- matrix(0, 100, 10)
z <- stats::rbinom(100, 1, 0.5)

# don't need to provide outcome
# function will assume each observation gets equal mass
dataHolder(x = x, z = z)

Description

R6 Data Generating Parent Class

R6 Data Generating Parent Class

Details

Can be used to make your own data simulation class. Should have the same slots listed in this class at a minimum, but you can add your own, of course. An easy way to do this is to make your class inherit from this one. See the example.

Value

An R6 object

Methods

Public methods

• DataSim$get_x()

• DataSim$get_y()

• DataSim$get_z()

• DataSim$get_n()

• DataSim$get_x1()

• DataSim$get_x0()

• DataSim$get_p()

• DataSim$get_tau()

• DataSim$gen_data()

• DataSim$clone()

Method get_x()

Gets the covariate data

Usage

DataSim$get_x()

Method get_y()

Gets the outcome vector

Usage

DataSim$get_y()

Method get_z()

Gets the treatment indicator

Usage

DataSim$get_z()

Method get_n()

Gets the number of observations

Usage

DataSim$get_n()

Method get_x1()

Gets the covariate data for the treated individuals

Usage

DataSim$get_x1()

Method get_x0()

Gets the covaraiate data for the control individuals

Usage

DataSim$get_x0()

Method get_p()

Gets the dimensionality covariate data

Usage

DataSim$get_p()

Method get_tau()

Gets the individual treatment effects

Usage

DataSim$get_tau()

Method gen_data()

Generates the data. Default is an empty function

Usage

DataSim$gen_data()

Method clone()

The objects of this class are cloneable with this method.

Usage

DataSim$clone(deep = FALSE)

Arguments

Examples

MyClass <- R6::R6Class("MyClass", 
inherit = DataSim,
public = list(),
private = list())

Description

Function to turn a data.frame into a dataHolder object.

Usage

df2dataHolder(
  treatment.formula,
  outcome.formula = NA_character_,
  data,
  weights = NA_real_
)

Arguments

Details

This will take the formulas specified and transform that data.frame into a dataHolder object that is used internally by the causalOT package. Take care if you do not specify an outcome formula that you do not include the outcome in the data.frame. If you are not careful, the function may include the outcome as a covariate, which is not kosher in causal inference during the design phase.

If both outcome.formula and treatment.formula are specified, it will assume you are in the design phase, and create a combined covariate matrix to balance on the assumed treatment and outcome models.

If you are in the outcome phase of estimation, you can just provide a dummy formula for the treatment.formula like "z ~ 0" just so the function can identify the treatment indicator appropriately in the data creation phase.

Value

Returns an object of class dataHolder()

Examples

set.seed(20348)
n <- 15
d <- 3
x <- matrix(stats::rnorm(n*d), n, d)
z <- rbinom(n, 1, prob = 0.5)
y <- rnorm(n)
weights <- rep(1/n,n)
df <- data.frame(x, z, y)
dh <- df2dataHolder(
  treatment.formula = "z ~ .",
  outcome.formula = "y ~ ." ,
  data = df,
  weights = weights)

Description

Options for the Entropy Balancing Weights

Usage

entBWOptions(delta = NULL, grid.length = 20L, nboot = 1000L, ...)

Arguments

Value

A list of class entBWOptions with slots

• delta Delta values to try

• grid.length The number of parameters to try

• nboot Number of bootstrap samples

• solver.options A list of options passed to 'lbfgsb3c()

Function balancing

This method will balance functions of the covariates within some tolerance, $\delta$. For these functions $B$, we will desire

$\frac{\sum_{i: Z_i = 0} w_i B(x_i) - \sum_{j: Z_j = 1} B(x_j)/n_1}{\sigma} \leq \delta$

, where in this case we are targeting balance with the treatment group for the ATT. $\sigma$ is the pooled standard deviation prior to balancing.

Examples

opts <- entBWOptions(delta = 0.1)

Description

Effective Sample Size

Usage

ESS(x)

## S4 method for signature 'numeric'
ESS(x)

## S4 method for signature 'causalWeights'
ESS(x)

Arguments

Details

Calculates the effective sample size as described by Kish (1965). However, this calculation has some problems and the PSIS() function should be used instead.

Value

Either a number denoting the effective sample size or if x is of class causalWeights, then returns a list of both values in the treatment and control groups.

Methods (by class)

• ESS(numeric): default ESS method for numeric vectors

• ESS(causalWeights): ESS method for objects of class causalWeights

See Also

PSIS()

Examples

x <- rep(1/100,100)
ESS(x)

Description

Estimate treatment effects

Usage

estimate_effect(
  causalWeights,
  x = NULL,
  y = NULL,
  model.function,
  estimate.separately = TRUE,
  augment.estimate = FALSE,
  normalize.weights = TRUE,
  ...
)

Arguments

Value

an object of class causalEffect

Examples

if ( torch::torch_is_installed() ){
# set-up data
data <- Hainmueller$new()
data$gen_data()

# calculate quantities
weight <- calc_weight(data, method = "COT", 
                      estimand = "ATT",
                      options = list(lambda = 0))
tx_eff <- estimate_effect(causalWeights = weight)

# get estimate
print(tx_eff@estimate)
all.equal(coef(tx_eff), c(estimate = tx_eff@estimate))
}

Description

Hainmueller data example

Hainmueller data example

Details

Generates the data as described in Hainmueller (2012).

Value

An R6 object of class DataSim

Super class

causalOT::DataSim -> Hainmueller

Methods

Public methods

• Hainmueller$gen_data()

• Hainmueller$gen_x()

• Hainmueller$gen_y()

• Hainmueller$gen_z()

• Hainmueller$new()

• Hainmueller$get_design()

• Hainmueller$get_pscore()

• Hainmueller$clone()

Method gen_data()

Generates the data

Usage

Hainmueller$gen_data()

Method gen_x()

Generates the covaraiate data

Usage

Hainmueller$gen_x()

Method gen_y()

Generates the outcome data

Usage

Hainmueller$gen_y()

Method gen_z()

Generates the treatment indicator

Usage

Hainmueller$gen_z()

Method new()

Generates the the Hainmueller R6 class

Usage

Hainmueller$new(
  n = 100,
  p = 6,
  param = list(),
  design = "A",
  overlap = "low",
  ...
)

Arguments

Details

Design

Design "A" is the setting where the outcome is generated from a linear model, $Y(0) = Y(1) = X_1 + X_2 + X_3 - X_4 + X_5 + X_6 + \eta$ and design "B" is where the outcome is generated from the non-linear model $Y(0) = Y(1) = (X_1 + X_2 +X_5 )^2 + \eta$.

Overlap

The treatment indicator is generated from $Z = 1(X_1 + 2 X_2 - 2 X_3 - X_4 - 0.5 X_5 + X_6 + \nu > 0)$, where $\nu$ depends on the overlap selected. If overlap is "high", then $\nu \sim N(0, 100).$ If overlap is "low", then $\nu \sim N(0, 30).$ Finally, if overlap is "medium", then $\nu$ is drawn from a $\chi^2$ with 5 degrees of freedom that is scaled and centered to have mean 0.5 and variance 67.6.

Returns

An object of class DataSim.

Examples

data <- Hainmueller$new(n = 100, p = 6, design = "A", overlap = "low")
data$gen_data()
print(data$get_x()[1:2,])

Method get_design()

Returns the chosen design parameters

Usage

Hainmueller$get_design()

Method get_pscore()

Returns the true propensity score

Usage

Hainmueller$get_pscore()

Method clone()

The objects of this class are cloneable with this method.

Usage

Hainmueller$clone(deep = FALSE)

Arguments

Examples

## ------------------------------------------------
## Method `Hainmueller$new`
## ------------------------------------------------

data <- Hainmueller$new(n = 100, p = 6, design = "A", overlap = "low")
data$gen_data()
print(data$get_x()[1:2,])

Description

LaLonde data example

LaLonde data example

Details

Returns the LaLonde data as used by Dehjia and Wahba. Note the data is fixed and gen_data() will just initialize the fixed data.

Value

An R6 object of class DataSim

Super class

causalOT::DataSim -> LaLonde

Methods

Public methods

• LaLonde$gen_data()

• LaLonde$get_tau()

• LaLonde$gen_x()

• LaLonde$gen_y()

• LaLonde$gen_z()

• LaLonde$new()

• LaLonde$get_design()

• LaLonde$clone()

Method gen_data()

Sets up the data

Usage

LaLonde$gen_data()

Method get_tau()

Returns the experimental treatment effect, $1794

Usage

LaLonde$get_tau()

Method gen_x()

Sets up the covariate data

Usage

LaLonde$gen_x()

Method gen_y()

Sets up the outcome data

Usage

LaLonde$gen_y()

Method gen_z()

Sets up the treatment indicator

Usage

LaLonde$gen_z()

Method new()

Initializes the LaLonde object.

Usage

LaLonde$new(n = NULL, p = NULL, param = list(), design = "NSW", ...)

Arguments

Examples

nsw <- LaLonde$new(design = "NSW")
nsw$gen_data()
nsw$get_n()

obs.study <-  LaLonde$new(design = "Full")
obs.study$gen_data()
obs.study$get_n()

Method get_design()

Returns the chosen design parameters

Usage

LaLonde$get_design()

Method clone()

The objects of this class are cloneable with this method.

Usage

LaLonde$clone(deep = FALSE)

Arguments

Examples

## ------------------------------------------------
## Method `LaLonde$new`
## ------------------------------------------------

nsw <- LaLonde$new(design = "NSW")
nsw$gen_data()
nsw$get_n()

obs.study <-  LaLonde$new(design = "Full")
obs.study$gen_data()
obs.study$get_n()

Description

This function will calculate the difference in means between treatment groups standardized by the pooled standard-deviation of the respective covariates.

Usage

mean_balance(x = NULL, z = NULL, weights = NULL, ...)

Arguments

Value

A vector of mean balances

Examples

n <- 100
p <- 6
x <- matrix(stats::rnorm(n * p), n, p)
z <- stats::rbinom(n, 1, 0.5)
weights <- calc_weight(x = x, z = z, estimand = "ATT", method = "Logistic")
mb <- mean_balance(x = x, z = z, weights = weights)
print(mb)

Description

An R6 Class for setting up measures

Usage

Measure(
  x,
  weights = NULL,
  probability.measure = TRUE,
  adapt = c("none", "weights", "x"),
  balance.functions = NA_real_,
  target.values = NA_real_,
  dtype = NULL,
  device = NULL
)

Arguments

Value

Returns a Measure object

Public fields

Active bindings

Methods

Public methods

• Measure$detach()

• Measure$get_weight_parameters()

• Measure$clone()

Method detach()

generates a deep clone of the object without gradients.

Usage

Measure$detach()

Method get_weight_parameters()

Makes a copy of the weights parameters.

Usage

Measure$get_weight_parameters()

Method clone()

The objects of this class are cloneable with this method.

Usage

Measure$clone(deep = FALSE)

Arguments

Examples

if(torch::torch_is_installed()) {
m <- Measure(x = matrix(0, 10, 2), adapt = "none")
print(m)
m$x
m$x <- matrix(1,10,2) # must have same dimensions
m$x
m$weights
m$weights <- 1:10/sum(1:10)
m$weights

# with gradients
m <- Measure(x = matrix(0, 10, 2), adapt = "weights")
m$requires_grad # TRUE
m$requires_grad <- "none" # turns off
m$requires_grad # FALSE
m$requires_grad <- "x"
m$requires_grad # TRUE
m <- Measure(matrix(0, 10, 2), adapt = "none")
m$grad # NULL
m <- Measure(matrix(0, 10, 2), adapt = "weights")
loss <- sum(m$weights * 1:10)
loss$backward()
m$grad
# note the weights gradient is on the log softmax scale
#and the first parameter is fixed for identifiability
m$grad <- rep(1,9)  
m$grad
}

Description

Optimal Transport Distance

Usage

ot_distance(
  x1,
  x2 = NULL,
  a = NULL,
  b = NULL,
  penalty,
  p = 2,
  cost = NULL,
  debias = TRUE,
  online.cost = "auto",
  diameter = NULL,
  niter = 1000,
  tol = 1e-07
)

## S3 method for class 'causalWeights'
ot_distance(
  x1,
  x2 = NULL,
  a = NULL,
  b = NULL,
  penalty,
  p = 2,
  cost = NULL,
  debias = TRUE,
  online.cost = "auto",
  diameter = NULL,
  niter = 1000,
  tol = 1e-07
)

## S3 method for class 'matrix'
ot_distance(
  x1,
  x2,
  a = NULL,
  b = NULL,
  penalty,
  p = 2,
  cost = NULL,
  debias = TRUE,
  online.cost = "auto",
  diameter = NULL,
  niter = 1000,
  tol = 1e-07
)

## S3 method for class 'array'
ot_distance(
  x1,
  x2,
  a = NULL,
  b = NULL,
  penalty,
  p = 2,
  cost = NULL,
  debias = TRUE,
  online.cost = "auto",
  diameter = NULL,
  niter = 1000,
  tol = 1e-07
)

## S3 method for class 'torch_tensor'
ot_distance(
  x1,
  x2,
  a = NULL,
  b = NULL,
  penalty,
  p = 2,
  cost = NULL,
  debias = TRUE,
  online.cost = "auto",
  diameter = NULL,
  niter = 1000,
  tol = 1e-07
)

Arguments

Value

For objects of class matrix, numeric value giving the optimal transport distance. For objects of class causalWeights, results are returned as a list for before ('pre') and after adjustment ('post').

Methods (by class)

• ot_distance(causalWeights): method for causalWeights class

• ot_distance(matrix): method for matrices

• ot_distance(array): method for arrays

• ot_distance(torch_tensor): method for torch_tensors

Examples

if ( torch::torch_is_installed()) {
x <- matrix(stats::rnorm(10*5), 10, 5)
z <- stats::rbinom(10, 1, 0.5)
weights <- calc_weight(x = x, z = z, method = "Logistic", estimand = "ATT")
ot1 <- ot_distance(x1 = weights, penalty = 100, 
p = 2, debias = TRUE, online.cost = "auto", 
diameter = NULL)
ot2<- ot_distance(x1 = x[z==0, ], x2 = x[z == 1,], 
a= weights@w0/sum(weights@w0), b = weights@w1,
 penalty = 100, p = 2, debias = TRUE, online.cost = "auto", diameter = NULL)

 all.equal(ot1$post, ot2) 
}

Description

Object Oriented OT Problem

Usage

OTProblem(measure_1, measure_2, ...)

Arguments

Value

An R6 object of class "OTProblem"

Public fields

Active bindings

Methods

Public methods

• OTProblem$add()

• OTProblem$subtract()

• OTProblem$multiply()

• OTProblem$divide()

• OTProblem$setup_arguments()

• OTProblem$solve()

• OTProblem$choose_hyperparameters()

• OTProblem$info()

• OTProblem$clone()

Method add()

adds o2 to the OTProblem

Usage

OTProblem$add(o2)

Arguments

Method subtract()

subtracts o2 from OTProblem

Usage

OTProblem$subtract(o2)

Arguments

Method multiply()

multiplies OTProblem by o2

Usage

OTProblem$multiply(o2)

Arguments

Method divide()

divides OTProblem by o2

Usage

OTProblem$divide(o2)

Arguments

Method setup_arguments()

Usage

OTProblem$setup_arguments(
      lambda,
      delta,
      grid.length = 7L,
      cost.function = NULL,
      p = 2,
      cost.online = "auto",
      debias = TRUE,
      diameter = NULL,
      ot_niter = 1000L,
      ot_tol = 0.001
    )

Arguments

Returns

NULL

Examples

 ot$setup_arguments(lambda = c(1000,10))
    

Method solve()

Solve the OTProblem at each parameter value. Must run setup_arguments first.

Usage

OTProblem$solve(
      niter = 1000L,
      tol = 1e-05,
      optimizer = c("torch", "frank-wolfe"),
      torch_optim = torch::optim_lbfgs,
      torch_scheduler = torch::lr_reduce_on_plateau,
      torch_args = NULL,
      osqp_args = NULL,
      quick.balance.function = TRUE
    )

Arguments

Examples

 ot$solve(niter = 1, torch_optim = torch::optim_rmsprop)
    

Method choose_hyperparameters()

Selects the hyperparameter values through a bootstrap algorithm

Usage

OTProblem$choose_hyperparameters(
      n_boot_lambda = 100L,
      n_boot_delta = 1000L,
      lambda_bootstrap = Inf
    )

Arguments

Examples

 ot$choose_hyperparameters(n_boot_lambda = 10, 
                                             n_boot_delta = 10, 
                                             lambda_bootstrap = Inf)
    

Method info()

Provides diagnostics after solve and choose_hyperparameter methods have been run.

Usage

OTProblem$info()

Returns

a list with slots

• loss the final loss values

• iterations The number of iterations run for each combination of parameters

• balance.function.differences The final differences in the balance functions

• hyperparam.metrics A list of the bootstrap evalustion for delta and lambda values

Examples

 ot$info()
    

Method clone()

The objects of this class are cloneable with this method.

Usage

OTProblem$clone(deep = FALSE)

Arguments

Examples

## ------------------------------------------------
## Method `OTProblem(measure_1, measure_2)`
## ------------------------------------------------

if (torch::torch_is_installed()) {
  # setup measures
  x <- matrix(1, 100, 10)
  m1 <- Measure(x = x)
  
  y <- matrix(2, 100, 10)
  m2 <- Measure(x = y, adapt = "weights")
  
  z <- matrix(3,102, 10)
  m3 <- Measure(x = z)
  
  # setup OT problems
  ot1 <- OTProblem(m1, m2)
  ot2 <- OTProblem(m3, m2)
  ot <- 0.5 * ot1 + 0.5 * ot2
  print(ot)

## ------------------------------------------------
## Method `OTProblem$setup_arguments`
## ------------------------------------------------

  ot$setup_arguments(lambda = 1000)

## ------------------------------------------------
## Method `OTProblem$solve`
## ------------------------------------------------

  ot$solve(niter = 1, torch_optim = torch::optim_rmsprop)

## ------------------------------------------------
## Method `OTProblem$choose_hyperparameters`
## ------------------------------------------------

  ot$choose_hyperparameters(n_boot_lambda = 1,
                            n_boot_delta = 1, 
                            lambda_bootstrap = Inf)

## ------------------------------------------------
## Method `OTProblem$info`
## ------------------------------------------------

ot$info()
}

Description

plot.causalWeights

Usage

## S3 method for class 'causalWeights'
plot(
  x,
  r_eff = NULL,
  penalty,
  p = 2,
  cost = NULL,
  debias = TRUE,
  online.cost = "auto",
  diameter = NULL,
  niter = 1000,
  tol = 1e-07,
  ...
)

Arguments

Details

The plot method first calls summary.causalWeights on the causalWeights object. Then plots the diagnostics from that summary object.

Value

The plot method returns an invisible object of class summary_causalWeights.

See Also

summary.causalWeights()

Description

A dataset evaluating treatments for post-partum hemorrhage. The data contain treatment groups receiving misoprostol vs potential controls from other locations that received only oxytocin. The data is stored as a numeric matrix.

Usage

data(pph)

Format

A matrix with 802 rows and 17 variables

Details

The variables are as follows:

• cum_blood_20m. The outcome variable denoting cumulative blood loss in mL 20 minutes after the diagnosis of post-partum hemorrhage (650 – 2000).

• tx. The treatment indicator of whether an individual received misoprostol (1) or oxytocin (0).

• age. the mother's age in years (15 – 43).

• no_educ. whether a woman had no education (1) or some education (0).

• num_livebirth. the number of previous live births.

• cur_married. whether a mother is currently married (1 = yes, 0 = no).

• gest_age. the gestational age of the fetus in weeks (35 – 43).

• prev_pphyes. whether the woman has had a previous post-partum hemorrahge.

• hb_test. the woman's hemoglobin in mg/dL (7 – 15).

• induced_laboryes. whether labor was induced (1 = yes, 0 = no).

• augmented_laboryes. whether labor was augmented (1 = yes, 0 = no).

• early_cordclampyes. whether the umbilical cord was clamped early (1 = yes, 0 = no).

• control_cordtractionyes. whether cord traction was controlled (1 = yes, 0 = no).

• uterine_massageyes. whether a uterine massage was given (1 = yes, 0 = no).

• placenta. whether placenta was delivered before treatment given (1 = yes, 0 = no).

• bloodlossattx. amount of blood lost when treatment given (500 mL – 1800 mL)

• sitecode. Which site is the individual from? (1 = Cairo, Egypt, 2 = Turkey, 3 = Hocmon, Vietnam, 4 = Cuchi, Vietnam, and 5 Burkina Faso).

Source

Data from the following Harvard Dataverse:

• Winikoff, Beverly, 2019, "Two randomized controlled trials of misoprostol for the treatment of postpartum hemorrhage", https://doi.org/10.7910/DVN/ETHH4N, Harvard Dataverse, V1.

The data was originally analyzed in

• Blum, J. et al. Treatment of post-partum haemorrhage with sublingual misoprostol versus oxytocin in women receiving prophylactic oxytocin: a double-blind, randomised, non-inferiority trial. The Lancet 375, 217–223 (2010).

Description

Predict method for barycentric projection models

Usage

## S3 method for class 'bp'
predict(
  object,
  newdata = NULL,
  source.sample,
  cost_function = NULL,
  niter = 1000,
  tol = 1e-07,
  ...
)

Arguments

Examples

if(torch::torch_is_installed()) {
set.seed(23483)
n <- 2^5
pp <- 6
overlap <- "low"
design <- "A"
estimate <- "ATT"
power <- 2
data <- causalOT::Hainmueller$new(n = n, p = pp,
design = design, overlap = overlap)

data$gen_data()

weights <- causalOT::calc_weight(x = data,
  z = NULL, y = NULL,
  estimand = estimate,
  method = "NNM")
  
 df <- data.frame(y = data$get_y(), z = data$get_z(), data$get_x())
  
 # undebiased
 fit <- causalOT::barycentric_projection(y ~ ., data = df, 
    weight = weights,
    separate.samples.on = "z", niter = 2)
    
 #debiased
 fit_d <- causalOT::barycentric_projection(y ~ ., data = df, 
    weight = weights,
    separate.samples.on = "z", debias = TRUE, niter = 2)
 
 # predictions, without new data
 undebiased_predictions <- predict(fit,   source.sample = df$z)
 debiased_predictions   <- predict(fit_d, source.sample = df$z)
 
 isTRUE(all.equal(unname(undebiased_predictions), df$y)) # FALSE
 isTRUE(all.equal(unname(debiased_predictions), df$y)) # TRUE
 }

Description

print.dataHolder

Usage

## S3 method for class 'dataHolder'
print(x, ...)

Arguments

Description

Pareto-Smoothed Importance Sampling

Usage

PSIS(x, r_eff = NULL, ...)

## S4 method for signature 'numeric'
PSIS(x, r_eff = NULL, ...)

## S4 method for signature 'causalWeights'
PSIS(x, r_eff = NULL, ...)

## S4 method for signature 'list'
PSIS(x, r_eff = NULL, ...)

PSIS_diag(x, ...)

## S4 method for signature 'numeric'
PSIS_diag(x, r_eff = NULL)

## S4 method for signature 'causalWeights'
PSIS_diag(x, r_eff = NULL)

## S4 method for signature 'causalPSIS'
PSIS_diag(x, ...)

## S4 method for signature 'list'
PSIS_diag(x, r_eff = NULL)

## S4 method for signature 'psis'
PSIS_diag(x, r_eff = NULL)

Arguments

Details

Acts as a wrapper to the psis() function from the loo package. It is built to handle the data types found in this package. This method is preferred to the ESS() function in causalOT since the latter is prone to error (infinite variances) but will not give good any indication that the estimates are problematic.

Value

For PSIS(), returns a list. See psis() from loo for a description of the outputs. Will give the log of the smoothed weights in slot log_weights, and in the slot diagnostics, it will give the pareto_k parameter (see the pareto-k-diagnostic page) and the n_eff estimates. PSIS_diag() returns the diagnostic slot from an object of class "psis".

Methods (by class)

• PSIS(numeric): numeric weights

• PSIS(causalWeights): object of class causalWeights

• PSIS(list): list of weights

• PSIS_diag(numeric): numeric weights

• PSIS_diag(causalWeights): object of class causalWeights diagnostics

• PSIS_diag(causalPSIS): diagnostics from the output of a previous call to PSIS

• PSIS_diag(list): a list of objects

• PSIS_diag(psis): output of PSIS function

See Also

ESS()

Examples

x <- runif(100)
w <- x/sum(x)

res <- PSIS(x = w, r_eff = 1)
PSIS_diag(res)

Description

Options for the SBW method

Usage

sbwOptions(delta = NULL, grid.length = 20L, nboot = 1000L, ...)

Arguments

Value

A list of class sbwOptions with slots

• delta Delta values to try

• grid.length The number of parameters to try

• sumto1 Forced to be TRUE. Weights will always sum to 1.

• nboot Number of bootstrap samples

• solver.optionsA list with arguments passed to osqpSettings()

Function balancing

This method will balance functions of the covariates within some tolerance, $\delta$. For these functions $B$, we will desire

$\frac{\sum_{i: Z_i = 0} w_i B(x_i) - \sum_{j: Z_j = 1} B(x_j)/n_1}{\sigma} \leq \delta$

, where in this case we are targeting balance with the treatment group for the ATT. $\sigma$ is the pooled standard deviation prior to balancing.

Examples

opts <- sbwOptions(delta = 0.1)

Description

Options for the SCM Method

Usage

scmOptions(...)

Arguments

Details

Options for the solver used in the optimization of the Synthetic Control Method of Abadie and Gardeazabal (2003).

Value

A list with arguments to pass to osqpSettings()

Examples

opts <- scmOptions()

Description

Summary diagnostics for causalWeights

print.summary_causalWeights

plot.summary_causalWeights

Usage

## S3 method for class 'causalWeights'
summary(
  object,
  r_eff = NULL,
  penalty,
  p = 2,
  cost = NULL,
  debias = TRUE,
  online.cost = "auto",
  diameter = NULL,
  niter = 1000,
  tol = 1e-07,
  ...
)

## S3 method for class 'summary_causalWeights'
print(x, ...)

## S3 method for class 'summary_causalWeights'
plot(x, ...)

Arguments

Value

The summary method returns an object of class "summary_causalWeights".

Functions

• print(summary_causalWeights): print method

• plot(summary_causalWeights): plot method

Examples

if(torch::torch_is_installed()) {
n <- 2^6
p <- 6
overlap <- "high"
design <- "A"
estimand <- "ATE"

#### get simulation functions ####
original <- Hainmueller$new(n = n, p = p, 
                            design = design, overlap = overlap)
original$gen_data()
weights <- calc_weight(x = original, estimand = estimand, method = "Logistic")
s <- summary(weights)
plot(s)
}

Description

Supported Methods

Usage

supported_methods()

Value

A character list with supported methods. Note "COT" is the same as "Wasserstein". We provide the second name for backwards compatibility.

Examples

supported_methods()

Description

Get the variance of a causalEffect

Usage

## S3 method for class 'causalEffect'
vcov(object, ...)

Arguments

Value

The variance of the treatment effect as a matrix

Examples

# set-up data
set.seed(1234)
data <- Hainmueller$new()
data$gen_data()

# calculate quantities
weight <- calc_weight(data, estimand = "ATT", method = "Logistic")
tx_eff <- estimate_effect(causalWeights = weight)

vcov(tx_eff)