We should produce a warning when bounded continuous is used with density type outcomes

[jpdunc23]

Param_ATT is used in the example below but this seems to affect other params as well.

library(tmle3) ## tmle3 0.1.7
library(sl3) ## sl3 1.2.0

set.seed(9519)

A <- rbinom(10, 1, 0.3)
Y_unscaled <- 30*A + 20 + rnorm(10)
bounds <- c(min(Y_unscaled), max(Y_unscaled))
Y_scaled <- (Y_unscaled - bounds[1]) / (bounds[2] - bounds[1])

df <- data.frame(A, Y_unscaled, Y_scaled)

run_example <- function(npsem) {

  lrnr_glm <- Lrnr_glm$new()
  factor_list <- list(
    define_lf(LF_fit, "A", lrnr_glm),
    define_lf(LF_fit, "Y", lrnr_glm)
  )

  likelihood_def <- Likelihood$new(factor_list)

  treatment_intervention <- define_lf(LF_static, "A", value = 1)
  control_intervention <- define_lf(LF_static, "A", value = 0)

  tmle_task <- tmle3_Task$new(df, npsem = npsem)
  likelihood <- likelihood_def$train(tmle_task)

  att <- define_param(
    Param_ATT, likelihood,
    treatment_intervention, control_intervention
  )

  updater <- tmle3_Update$new()
  updater$tmle_params <- att
  targeted_likelihood <- Targeted_Likelihood$new(likelihood, updater)

  tmle_fit <- suppressWarnings(
    fit_tmle3(tmle_task, targeted_likelihood, att, updater)
  )
  tmle_fit$estimates
}

npsem1 <- c(
  define_node('A', 'A'),
  define_node(
    'Y', 'Y_scaled', c('A')
  )
)
npsem2 <- c(
  define_node('A', 'A'),
  define_node(
    'Y', 'Y_unscaled', c('A'),
    variable_type('continuous', bounds = bounds)
  )
)
npsem3 <- c(
  define_node('A', 'A'),
  define_node(
    'Y', 'Y_unscaled', c('A'), scale = TRUE
  )
)

run_example(npsem1) ## scaled

## [[1]]
## [[1]]$psi
## [1] 0.9450154

## [[1]]$IC
##  [1] -0.019166418 -0.011467197  0.023985418  0.023000715  0.031983918
##  [6] -0.034262603 -0.027609057  0.005902324  0.019357548 -0.011724648

run_example(npsem2) ## unscaled, specify variable_type

## [[1]]
## [[1]]$psi
## [1] 0.9450154

## [[1]]$IC
##  [1]  48.63069  48.87336 -19.63742  49.95977 -19.38532 -21.47337 -21.26365
##  [8] -20.20739  49.84494  48.86525

run_example(npsem3) ## unscaled, use 'scale' arg

## [[1]]
## [[1]]$psi
## [1] 0.9450154

## [[1]]$IC
##  [1]  48.63069  48.87336 -19.63742  49.95977 -19.38532 -21.47337 -21.26365
##  [8] -20.20739  49.84494  48.86525

[jeremyrcoyle]

Thanks for the detailed example. The issue can be solved by changing the NPSEM to specify that the Y likelihood factor is a mean E(Y|A) not a density p(Y|A) as follows:

  factor_list <- list(
    define_lf(LF_fit, "A", lrnr_glm),
    define_lf(LF_fit, "Y", lrnr_glm, type="mean")
  )

Bounded continuous only really makes sense for "mean-type" likelihood factors, not "density-type" likelihood factors (which are rarely used for outcome variables anyway). I'm going to leave the issue open but change the title to reflect the fact that we should produce a warning when bounded continuous is used with density type outcomes.

[jpdunc23]

Thanks for the clarification!