Custom functions and DeclareDesign

The declare_* functions in DeclareDesign use functions in the fabricatr, randomizr, and estimatr packages as defaults, which work great for most designs. Sometimes, however, you might want to write your own function. This advanced vignette declares a design using only custom functions.

First, we’ll write custom functions for each of the steps in the design. All functions must take a data.frame and return a data.frame, with the exception of a population step, whose inputs can be anything but whose output must be a data.frame.

# M: Model
custom_population <- function(N) {
  data.frame(u = rnorm(N))
}
custom_potential_outcomes <-
  function(data) {
    within(data,{
      Y_Z_0 <- u
      Y_Z_1 <- 0.25 + u
    })
  }

# I: Inquiry
custom_inquiry <- function(data, label) {
  data.frame(inquiry = label,
  inquiry = with(data, median(Y_Z_1 - Y_Z_0)))
}

# D: Data Strategy
custom_sampling <- function(data) {
     data$S <- rbinom(n = nrow(data),
            size = 1,
            prob = 0.1)
     data[data$S == 1, ]
}

custom_assignment <- function(data) {
  data$Z <- rbinom(n = nrow(data),
         size = 1,
         prob = 0.5)
  data
}

custom_reveal <- function(data){
  within(data, Y <- Y_Z_1 * Z + Y_Z_0 * (1 - Z))
}

# A: Answer strategy
custom_estimator <- function(data){
  data.frame(estimate = with(data, mean(Y)))
}

In order to declare the design, we pass each of the custom functions to the handler argument of each declaration step:

design <- 
  declare_model(handler = custom_population, N = 100) + 
  declare_potential_outcomes(handler = custom_potential_outcomes) + 
  declare_inquiry(handler = custom_inquiry, label = "medianTE") + 
  declare_sampling(handler = custom_sampling) + 
  declare_assignment(handler = custom_assignment) + 
  declare_reveal(handler = custom_reveal) + 
  declare_estimator(handler = tidy_estimator(custom_estimator), 
                    inquiry = "medianTE")

## Warning in tidy_estimator(custom_estimator): tidy_estimator() has been
## deprecated. Please use label_estimator() instead.

head(draw_data(design))

##       u Y_Z_1 Y_Z_0 S Z    Y
## 7  1.51  1.76  1.51 1 0 1.51
## 24 1.21  1.46  1.21 1 0 1.21
## 35 0.50  0.75  0.50 1 0 0.50
## 50 0.66  0.91  0.66 1 0 0.66
## 64 1.40  1.65  1.40 1 1 1.65
## 69 0.92  1.17  0.92 1 1 1.17

run_design(design)

##    inquiry inquiry.1 estimator estimate
## 1 medianTE      0.25 estimator     0.11

This example used very simple custom functions, but this framework is flexible enough to accommodate any design step that can be expressed as a function of data that returns data.