yqzhong7 / aipw Goto Github PK

• Inherit from AIPW_base class
• Something like AIPW_manual$new(A, Y, mu0, mu1, mu, raw_p_score, verbose)
• Warning about sample splitting

Supporting categorical exposure by using missing outcome mechanism. (Chapter 6, Gruber, S. and Van der Laan, M.J., 2011. tmle: An R package for targeted maximum likelihood estimation.)

• ATT
• ATC
• Output the risks by exposure status

Kennedy, E.H., Sjölander, A. and Small, D.S., 2015. Semiparametric causal inference in matched cohort studies. Biometrika, 102(3), pp.739-746.

1. Missing covariates or exposure
2. k_split
   • = 1 : no sample splitting is being used
   • = 2 : Not allowed

@yqzhong7 Could this be addressed, please?

R version 4.3.1 (2023-06-16) -- "Beagle Scouts"
Copyright (C) 2023 The R Foundation for Statistical Computing
Platform: powerpc-apple-darwin10.0.0d2 (32-bit)

> library(testthat)
> library(AIPW)
> 
> test_check("AIPW")
[ FAIL 2 | WARN 0 | SKIP 0 | PASS 206 ]

══ Failed tests ════════════════════════════════════════════════════════════════
── Error ('test-stratified_fit.R:75:3'): AIPW stratified_fit: sl3 & k_split ────
Error in `is.nan(A)`: default method not implemented for type 'list'
Backtrace:
     ▆
  1. └─aipw$stratified_fit() at test-stratified_fit.R:75:3
  2.   └─private$.f_lapply(...)
  3.     └─future.apply::future_lapply(...)
  4.       └─future.apply:::future_xapply(...)
  5.         ├─future::value(fs)
  6.         └─future:::value.list(fs)
  7.           ├─future::resolve(...)
  8.           └─future:::resolve.list(...)
  9.             └─future (local) signalConditionsASAP(obj, resignal = FALSE, pos = ii)
 10.               └─future:::signalConditions(...)
── Error ('test-tmle_support.R:53:3'): AIPW_tmle class: tmle3 ──────────────────
Error in `trim_logit(X)`: 'list' object cannot be coerced to type 'double'
Backtrace:
    ▆
 1. └─tmle3::tmle3(or_spec, data = df, node_list, learner_list) at test-tmle_support.R:53:3
 2.   └─tmle_spec$make_initial_likelihood(tmle_task, learner_list)
 3.     └─tmle3::point_tx_likelihood(tmle_task, learner_list)
 4.       └─likelihood_def$train(tmle_task)
 5.         └─delayed_fit$compute(job_type = sl3_delayed_job_type(), progress = verbose)
 6.           └─scheduler$compute()
 7.             └─self$compute_step()

[ FAIL 2 | WARN 0 | SKIP 0 | PASS 206 ]
Error: Test failures
Execution halted

• How can we have the balance summary across each covariates for the propensity score ? (To calcul the Standardized Mean Difference for example)

• Which estimator is use to compute the variance (and thus 95%CI) ?

• Output summary statistics of IP-weights
• ggplot of IP-weights

• Provide ATE on RD

Missing outcome:

• Add warning message: Missing outcome is detected, assuming missing at random (MAR)
• Outcome model: Q(A,W) = E(Y | W, A = a, delta = 1)
• Exposure model: g(W) = P(A | W) P(delta | A, W)

Packages on CRAN require that their dependencies are also on CRAN. Since sl3 and tmle3 are not on CRAN at this time, we are planning to:

• Create a separate branch without supporting sl3 and tmle3
• Submit this package to CRAN
• Revise per CRAN's comments if any
• Resubmission

• use different splits of cross-fitting for $\psi$ estimation

• summarise $\psi$ with median (Chernozhukov 2018, Section 3.4).

• Newey, W. K., & Robins, J. R. (2018). Cross-fitting and fast remainder rates for semiparametric estimation. arXiv preprint arXiv:1801.09138.

• Chernozhukov V, Chetverikov D, Demirer M, et al. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal 2018;21(1):C1–C68. doi:10.1111/ectj.12097. Publisher: Oxford Academic.

1. If only one g.bound is provided, truncation >0.5 is not allowed (provide error message)
2. If two g.bound is provided, asymmetric ps truncation is supported (e.g., g.bound = c(0.05, 0.95))
3. Not allow number of g.bound >=3

In line 278 of AIPW/R/AIPW_base.R, the std error of RR may calculate -(2*sigma_covar[1,2]/(mean(aipw_eif1)*mean(aipw_eif0))) twice, which occur NaNs in sqrt function.

Hi, thanks for the fantastic package! I have a small suggestion. Would you be able to change the text included with the results output to something like 'exposure mean', 'control mean' and 'mean difference' when the outcome is continuous? Otherwise it can be a bit confusing seeing 'risk difference' etc for a continuous outcome.

Here is a quick reprex to demonstrate the issue:

A <- rbinom(100, 1, 0.5)
W <- rnorm(100)
Y <- rnorm(100)

aipw_out <- aipw_wrapper(
  Y = Y,
  A = A,
  W = W,
  Q.SL.library = "SL.mean",
  g.SL.library = "SL.mean",
  k = 1
)

                 Estimate    SE 95% LCL 95% UCL   N
Risk of exposure  -0.2493 0.133  -0.511  0.0124  51
Risk of control   -0.0778 0.126  -0.325  0.1697  49
Risk Difference   -0.1714 0.183  -0.531  0.1879 100