Package 'CausalState'

Description

Two sequentially doubly robust estimators for longitudinal modified treatment policies (MTPs) in settings with transitioning states - such as ICU stay, ward admission, or ED episodes - where treatment is structurally inapplicable after a state transition (e.g. discharge or death).

Sequential Doubly Robust estimator - sdr()

Implements the LMTP-SDR estimator of Diaz et al. (2021), which extends the SDR construction of Luedtke et al. (2017) to general modified treatment policies via density-ratio weighting. Starting from terminal pseudo-outcomes and working backward in time, the estimator propagates an EIF-corrected pseudo-outcome through a sequence of Q-regressions. The final estimate is the mean of these pseudo-outcomes at the first time point. The estimator is sequentially doubly robust (2^K-robust): consistent whenever, at each time point, either the treatment model g_t or the outcome model Q_t is correctly specified. Under the natural-course policy the density ratios equal one and the recursion collapses algebraically to mean(Y) regardless of model quality - a natural-course SDR run therefore cannot serve as a model calibration check.

Infinite-dimensional TMLE - itmle()

Implements the cross-validated infinite-dimensional TMLE of Luedtke et al. (2017, Section 5 and Appendix 12), adapted to general MTPs via density-ratio weighting following Diaz et al. (2021). Like SDR, the estimator fits a backward Q-regression; unlike SDR, it then applies an infinite-dimensional fluctuation (targeting step) that solves the efficient influence equation semiparametrically without restricting the fluctuation to a parametric submodel. The cross-validated variant is used throughout to prevent the targeting step from overfitting the Q estimates. iTMLE is also sequentially doubly robust. Unlike SDR, running under the natural-course policy does not collapse to mean(Y) and therefore provides a genuine calibration check on the Q models after fluctuation.

Density ratio estimation - density_ratio()

Estimates the density ratios needed by both estimators via a flexible classification-based approach. An alternative metalearner based on Wu and Benkeser combines base learners by minimising a log density-ratio loss rather than the default NNLS. Density ratios can be computed once and reused across both estimators and across multiple time horizons.

State transition model and probabilistic mixing

At each time point t, a subject in state (in_state = 1, alive = 1) faces three mutually exclusive outcomes: remain in state, exit alive (discharge), or die. Four regression components handle this, corresponding directly to the sl_ arguments of sdr(), itmle(), and qreg():

g_remain (sl_remain)

Probability of remaining in state at the next time point, given current history. Trained on all subjects currently in state.

g_death_exit (sl_death)

Probability of dying, conditional on having left the state (i.e. among exiters only). Trained on the exit subset – those subjects who leave the state at each time point.

Q_rem (sl_recursive at intermediate time points; sl_y at the terminal time point)

Expected outcome conditional on remaining in state. This is the component propagated backward through the Q-recursion.

Q_exit (sl_y)

Expected outcome conditional on exiting. In most applications this is not estimated but instead fixed to a constant for each exit type (death, discharge) via absorb_rule().

Probabilistic mixture. The backward Q-recursion assembles the expected outcome at time t as a probability-weighted mixture over the three transition branches:

Q(t) = g_remain * Q_rem
      + (1 - g_remain) * [ g_death_exit * Q_death
                          + (1 - g_death_exit) * Q_discharge ]

where g_remain and g_death_exit are the model predictions at the subject's current history, and Q_death / Q_discharge are the outcome values assigned to each exit branch (constants when absorb_rule() is used, or sl_y predictions otherwise).

Event counts and the death regression. Because g_death_exit is trained on the exit subset only, its effective sample size is often much smaller than the full cohort – and within that subset, deaths may be rare.

Before every SuperLearner fit, the code checks whether the training data are sufficient to fit a model (internal function can_fit_bin): fitting proceeds only when the training set has at least 30 observations and at least 5 events of each class (deaths and discharges within the exit subset for g_death_exit; remainers and exiters for g_remain). When either threshold is not met, the estimator substitutes the empirical proportion as a time-constant prediction instead of fitting a SuperLearner. The fallback is recorded in diagnostics$branch_cal (the ⁠p_*_const⁠ columns are non-NA when a constant was used).

Even above the floor, with few events the SuperLearner ensemble will typically collapse to SL.mean or a near-intercept logistic fit. Setting pool_g_death = TRUE fits a single g_death_exit model across all time points (with time included as a covariate), so the 30/5 thresholds apply to the full follow-up pooled rather than each time point individually. This is the recommended remedy when per-time death counts are sparse but the death hazard is roughly stable over time.

To inspect event counts: diagnostics$branch_cal contains one row per fold and time point; n_tr_exit is the exit-subset size, and the number of deaths is approximately n_tr_exit * mean_target for the g_death rows.

When absorb_rule() fixes Q_death to a constant – the common case – the impact of a poorly fitted g_death_exit is limited: misspecification shifts probability mass between the death and discharge branches of the mixture, but the sequentially doubly robust correction via density-ratio weighting partially compensates provided the treatment model is well-specified.

State machinery

absorb_rule() defines outcome overrides at absorbing states (e.g. forcing Y = 0 for subjects who die before the end of follow-up).

Custom SuperLearner wrappers

SDR-targeted learners (SL.tgt.glm(), SL.tgt.glmnet(), SL.tgt.xgboost(), etc.) for the Q-regression step, and iTMLE fluctuation learners (SL.tmle_glm(), SL.tmle_glmnet_ridge(), etc.) that carry the logit offset required for cross-validated TMLE as a data column.

Author(s)

Maintainer: Sebastiaan Blank [email protected] (ORCID) [copyright holder]

Authors:

• Sebastiaan Blank [email protected] (ORCID) [copyright holder]

References

Diaz I, Williams N, Hoffman KL, Schenck EJ (2021). Nonparametric Causal Effects Based on Longitudinal Modified Treatment Policies. JASA 118(542):846-857. doi:10.1080/01621459.2021.1955691.

Luedtke AR, Sofrygin O, van der Laan MJ, Carone M (2017). Sequential Double Robustness in Right-Censored Longitudinal Models. arXiv:1705.02459.

Rotnitzky A, Robins J, Babino L (2017). On the multiply robust estimation of the mean of the g-functional. arXiv:1705.08582.

Wu C, Benkeser D (2024). Nonparametric Efficient Estimation of Marginal Structural Models using Targeted Machine Learning. arXiv:2408.10847.

Williams NT, Diaz I (2023). lmtp: An R package for estimating the causal effects of modified treatment policies. Observational Studies.

Bang H, Robins JM (2005). Doubly robust estimation in missing data and causal inference models. Biometrics 61(4):962-973.

Haneuse S, Rotnitzky A (2013). Estimation of the effect of interventions that modify the received treatment. Statistics in Medicine 32(30):5260-5277.

Diaz Munoz I, van der Laan MJ (2012). Population intervention causal effects based on stochastic interventions. Biometrics 68(2):541-549.

See Also

Useful links:

• https://github.com/sebastiaan-blank/CausalState

• Report bugs at https://github.com/sebastiaan-blank/CausalState/issues

Description

Constructs a rule that overrides Q-model predictions for subjects in a specific absorbing state. Pass one or more rules to the absorb argument of sdr() or itmle().

Usage

absorb_rule(cond, value, branch = c("any", "death", "dc"))

Arguments

Value

A named list with elements branch, cond, and value (the latter two as unevaluated expressions via base::substitute()).

Examples

# Force outcome to 0 for patients who die at any time
absorb_rule(alive == 0, value = 0, branch = "death")

# Force outcome to 1 for discharged patients when t >= 5
absorb_rule(in_state == 0 & t >= 5, value = 1, branch = "dc")

Description

Aggregates the raw per-fold calibration diagnostics stored in diagnostics$branch_cal to one row per time step per branch, weighted by the validation-set size at each fold. Returns a tidy data.table suitable for further filtering, plotting, or data.table::dcast().

Usage

branch_cal_summary(fit)

Arguments

Value

A data.table with columns t, branch (one of "g_remain", "g_death", "q_rem", "q_exit"), n_tr, tgt_tr, pred_tr, n_vl, tgt_vl, pred_vl, cal_slope_vl; and where applicable brier_vl, auc_vl (binary branches) or rmse_vl, cor_vl (Gaussian branches).

See Also

sdr(), itmle()

Description

Takes an intervention-arm fit and a reference (from sdr() or itmle(), or the crude observed mean) and returns the risk difference (RD), risk ratio (RR), and odds ratio (OR) with standard errors derived from the per-subject influence curves.

Usage

contrast(
  fit1,
  fit0 = NULL,
  df = NULL,
  id_col = NULL,
  cluster = NULL,
  y_col = NULL
)

Arguments

Details

All standard errors use the delta method on the efficient influence curves:

• RD: $IC_{RD,i} = IC_{1,i} - IC_{0,i}$; Wald CI on the natural scale.

• RR: $IC_{\log RR,i} = IC_{1,i}/\psi_1 - IC_{0,i}/\psi_0$; SE and 95\

• OR: $IC_{\log OR,i} = IC_{1,i}/(\psi_1(1-\psi_1)) - IC_{0,i}/(\psi_0(1-\psi_0))$; SE and 95\

RR and OR are omitted for Gaussian outcomes.

When fit0 = NULL and y_col is supplied, the reference is the crude observed mean of y_col (last non-missing value per subject). The influence curve for the observed mean is $IC_{0,i} = Y_i - \bar{Y}$. This gives a simple descriptive contrast against the raw data rather than a causal comparison between two counterfactual estimates.

Value

A list of class "CausalState_contrast" with:

psi1, psi0

Point estimates.

obs_ref

Logical; TRUE when the reference is the observed mean rather than a second estimator fit.

RD, se_RD, ci_RD

Risk difference and 95 pct Wald CI.

RR, se_log_RR, ci_RR

Risk ratio, SE on log scale, and 95 pct CI (exponentiated). NULL for Gaussian outcomes.

OR, se_log_OR, ci_OR

Odds ratio, SE on log scale, and 95 pct CI (exponentiated). NULL for Gaussian outcomes.

n

Number of matched subjects.

table

Summary data frame printed by default.

See Also

sdr(), itmle()

Examples

library(SuperLearner)

sim_ex <- function(n = 2000L, tmax = 5L) {
  set.seed(42L)
  rows <- vector("list", n)
  for (i in seq_len(n)) {
    age <- round(rnorm(1, 65, 10)); L1 <- rnorm(1)
    pat <- list()
    for (t in seq_len(tmax)) {
      A     <- rbinom(1, 1, plogis(0.3 * L1 - 0.4))
      u     <- runif(1)
      p_die <- plogis(-4.0 + 0.2 * L1 - 0.1 * age / 10)
      p_dc  <- plogis(-2.5 + 0.5 * A)
      if (u < p_die) {
        alive <- 0L; in_state <- 0L
      } else if (u < p_die + p_dc) {
        alive <- 1L; in_state <- 0L
      } else {
        alive <- 1L; in_state <- 1L
      }
      Y <- rbinom(1, 1, plogis(-0.5 + 0.4 * A - 0.2 * L1))
      pat[[length(pat) + 1L]] <- data.frame(
        id = i, time = t, age = age, L1 = L1,
        A = A, alive = alive, in_state = in_state, Y = Y
      )
      if (in_state == 0L) break
      if (t < tmax) L1 <- L1 + rnorm(1, -0.1 * A, 0.3)
    }
    rows[[i]] <- do.call(rbind, pat)
  }
  do.call(rbind, rows)
}
df <- sim_ex()
sl_lib <- c("SL.mean", "SL.glm")

policy_nat <- function(D_block, t, a_names) D_block[, ..a_names, drop = FALSE]
policy_sft <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmin(D_block[[a_names[1]]] + 0.3, 1)
  out
}

dr_args <- list(
  df = df, a_names = "A", tmax = 5L, baseline = "age", tv_names = "L1",
  sl_g = sl_lib, k = 1L, inner_v = 3L, v = 3L, seed = 1L,
  id = "id", time = "time"
)
wr_nat <- do.call(density_ratio, c(dr_args, list(policy_spec_fun = policy_nat)))
wr_sft <- do.call(density_ratio, c(dr_args, list(policy_spec_fun = policy_sft)))

sdr_args <- list(
  df = df, tmax = 5L, id = "id", time = "time",
  alive = "alive", in_state = "in_state", y = "Y",
  baseline = "age", tv_names = "L1", a_names = "A",
  sl_remain = sl_lib, sl_death = sl_lib,
  sl_recursive = sl_lib, sl_y = sl_lib,
  k = 1L, inner_v = 3L, parallel = FALSE, seed = 1L
)
res_nat <- do.call(sdr, c(sdr_args,
  list(weight_object = wr_nat, policy_spec_fun = policy_nat)))
res_sft <- do.call(sdr, c(sdr_args,
  list(weight_object = wr_sft, policy_spec_fun = policy_sft)))

# Two-estimator contrast (causal RD/RR/OR)
ctr <- contrast(res_sft, res_nat)
ctr$RD; ctr$ci_RD; ctr$RR; ctr$OR

# Observed-mean reference (descriptive)
ctr_obs <- contrast(res_sft, df = df, y_col = "Y")
ctr_obs$table

Description

Fits per-time-point treatment models and computes the instantaneous density ratio $r_t = d\tilde{P}(A_t | H_t) / dP(A_t | H_t)$ comparing the modified treatment policy (MTP) to the natural course. The output is passed directly to sdr() or itmle() as the weight_object argument.

Usage

density_ratio(
  df,
  a_names,
  tmax,
  baseline = NULL,
  tv_names = character(0),
  no_lag_vars = character(0),
  policy_names = character(0),
  sl_g,
  k = 2,
  inner_v = 10L,
  cluster = NULL,
  v = 5L,
  seed = 1L,
  policy_spec_fun,
  id = "id",
  time = "trial_time",
  bounds = 1e-05,
  dr_sl = FALSE,
  drop_small_cluster_splits = TRUE,
  parallel_t = FALSE,
  t_workers = NULL,
  fold_workers = NULL,
  sl_workers = NULL,
  verbose = TRUE
)

Arguments

Details

Cross-fitting is used throughout: treatment models are trained on one fold and density ratios predicted on the held-out fold, so the same fold structure is inherited by the downstream estimator.

Value

A list with components:

weights_dt

A data.table with one row per subject-time containing Rt_t (instantaneous density ratio) and global_fold. Pass this to sdr() or itmle() as weight_object.

sl_summary

data.table of SuperLearner learner weights. One row per learner per (fold, time-point). Useful for checking which treatment models dominate across time points.

fold_diag

data.table of per-fold, per-time-point diagnostics: mean predicted probabilities, effective sample sizes (ESS) for the density ratios, and calibration summaries for each treatment model. ESS collapse (ESS much smaller than n) indicates extreme density ratios and warrants caution.

Parallelism

t_workers, fold_workers, and sl_workers can be used independently or together. Using a single level is robust with all learners. Combining two or more creates nested mclapply calls; in that case multi-threaded or GPU-based learners (e.g. xgboost with CUDA, OpenMP-based methods) may crash in the child processes and should be avoided or limited to one thread.

Sample size requirements

This package targets large longitudinal datasets – thousands of subjects – typical of ICU, ward, or emergency department cohorts. Stable density ratio estimation requires adequate treatment-variation at every time point in each cross-fitting fold. All built-in examples use a minimum of 2,000 subjects.

References

Diaz I, Williams N, Hoffman KL, Schenck EJ (2021). Nonparametric Causal Effects Based on Longitudinal Modified Treatment Policies. JASA 118(542):846-857.

Wu C, Benkeser D (2024). Nonparametric Efficient Estimation of Marginal Structural Models using Targeted Machine Learning. arXiv:2408.10847.

Williams NT, Diaz I (2023). lmtp: An R package for estimating the causal effects of modified treatment policies. Observational Studies.

See Also

sdr(), itmle()

Examples

library(SuperLearner)

# ICU-like DGP: patients die, discharge, or remain in state each time point
sim_ex <- function(n = 2000L, tmax = 5L) {
  set.seed(42L)
  rows <- vector("list", n)
  for (i in seq_len(n)) {
    age <- round(rnorm(1, 65, 10)); L1 <- rnorm(1)
    pat <- list()
    for (t in seq_len(tmax)) {
      A     <- rbinom(1, 1, plogis(0.3 * L1 - 0.4))
      u     <- runif(1)
      p_die <- plogis(-4.0 + 0.2 * L1 - 0.1 * age / 10)
      p_dc  <- plogis(-2.5 + 0.5 * A)
      if (u < p_die) {
        alive <- 0L; in_state <- 0L
      } else if (u < p_die + p_dc) {
        alive <- 1L; in_state <- 0L
      } else {
        alive <- 1L; in_state <- 1L
      }
      Y <- rbinom(1, 1, plogis(-0.5 + 0.4 * A - 0.2 * L1))
      pat[[length(pat) + 1L]] <- data.frame(
        id = i, time = t, age = age, L1 = L1,
        A = A, alive = alive, in_state = in_state, Y = Y
      )
      if (in_state == 0L) break
      if (t < tmax) L1 <- L1 + rnorm(1, -0.1 * A, 0.3)
    }
    rows[[i]] <- do.call(rbind, pat)
  }
  do.call(rbind, rows)
}
df <- sim_ex()

# Policy: increase treatment probability by 0.3
policy_fn <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmin(D_block[[a_names[1]]] + 0.3, 1)
  out
}

wr <- density_ratio(
  df              = df,
  a_names         = "A",
  tmax            = 5L,
  baseline        = "age",
  tv_names        = "L1",
  sl_g            = c("SL.mean", "SL.glm"),
  k               = 1L,
  inner_v         = 3L,
  v               = 3L,
  seed            = 1L,
  id              = "id",
  time            = "time",
  policy_spec_fun = policy_fn
)
head(wr$weights_dt)


# Multi-treatment: joint density ratio for two binary treatments (A1, A2)

library(SuperLearner)

sim_multi <- function(n = 2000L, tmax = 5L) {
  set.seed(42L)
  rows <- vector("list", n)
  for (i in seq_len(n)) {
    age <- round(rnorm(1, 65, 10)); L1 <- rnorm(1)
    pat <- list()
    for (t in seq_len(tmax)) {
      A1    <- rbinom(1, 1, plogis(0.3 * L1 - 0.4))
      A2    <- rbinom(1, 1, plogis(0.2 * L1 + 0.3 * A1 - 0.3))
      u     <- runif(1)
      p_die <- plogis(-4.0 + 0.2 * L1 - 0.1 * age / 10)
      p_dc  <- plogis(-2.5 + 0.4 * A1 + 0.3 * A2)
      if (u < p_die) {
        alive <- 0L; in_state <- 0L
      } else if (u < p_die + p_dc) {
        alive <- 1L; in_state <- 0L
      } else {
        alive <- 1L; in_state <- 1L
      }
      Y <- rbinom(1, 1, plogis(-0.5 + 0.3 * A1 + 0.3 * A2 - 0.2 * L1))
      pat[[length(pat) + 1L]] <- data.frame(
        id = i, time = t, age = age, L1 = L1,
        A1 = A1, A2 = A2, alive = alive, in_state = in_state, Y = Y
      )
      if (in_state == 0L) break
      if (t < tmax) L1 <- L1 + rnorm(1, -0.1 * A1, 0.3)
    }
    rows[[i]] <- do.call(rbind, pat)
  }
  do.call(rbind, rows)
}
df2 <- sim_multi()

# Joint policy: shift both treatments upward by 0.3
policy_fn2 <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmin(D_block[[a_names[1]]] + 0.3, 1)
  out[[a_names[2]]] <- pmin(D_block[[a_names[2]]] + 0.3, 1)
  out
}

wr2 <- density_ratio(
  df              = df2,
  a_names         = c("A1", "A2"),
  tmax            = 5L,
  baseline        = "age",
  tv_names        = "L1",
  sl_g            = c("SL.mean", "SL.glm"),
  k               = 1L,
  inner_v         = 3L,
  v               = 3L,
  seed            = 1L,
  id              = "id",
  time            = "time",
  policy_spec_fun = policy_fn2
)
head(wr2$weights_dt)

Description

Estimates the mean counterfactual outcome under a modified treatment policy (MTP) using the infinite-dimensional targeted minimum loss-based estimator (iTMLE). The estimator fits a backward Q-regression to approximate the counterfactual outcome trajectory, then applies an infinite-dimensional fluctuation – a targeting step that solves the efficient influence equation without restricting the fluctuation to a finite-dimensional parametric submodel. The algorithm follows Luedtke et al. (2017, Section 5 and Appendix 12): Section 5 introduces the infinite-dimensional fluctuation, Appendix 12 the cross-validated variant used here to prevent overfitting of Q estimates. Density-ratio weights from density_ratio() adapt the estimator to general MTPs following the framework of Diaz et al. (2021). The estimator is sequentially doubly robust (2^$K$-robust): consistent whenever, at each time point t, either the treatment model $g_t$ or the outcome model $Q_t$ is consistently estimated. The targeting step uses a SuperLearner ensemble with custom wrappers (see sl_itmle) that carry the logit offset required for cross-validated TMLE as a data column.

Usage

itmle(
  df,
  weight_object,
  tmax,
  id,
  time,
  alive,
  in_state,
  y,
  baseline,
  tv_names = character(0),
  a_names = character(0),
  no_lag_vars = character(0),
  policy_names = character(0),
  sl_remain = NULL,
  sl_death = NULL,
  sl_recursive = NULL,
  sl_rec_early = NULL,
  rec_transition = NULL,
  sl_y = NULL,
  outcome_family = c("binomial", "gaussian"),
  y_bounds = NULL,
  bounds = 1e-05,
  trim = 0.99,
  absorb = list(),
  policy_spec_fun = function(D_block, t, a_names) NULL,
  k = 2,
  seed = 1,
  parallel = FALSE,
  fold_workers = NULL,
  reg_workers = NULL,
  sl_workers = NULL,
  inner_v = 5L,
  sl_tmle = NULL,
  v_target_itmle = 10L,
  v_sl_inner_itmle = 10L,
  cluster = NULL,
  cluster_se_only = FALSE,
  pool_g_death = FALSE,
  pool_q_exit = FALSE,
  pool_time = "spline",
  verbose = TRUE
)

Arguments

Details

Two properties of iTMLE worth noting relative to sdr(): (1) under the natural-course policy iTMLE does not collapse algebraically to mean(Y), so a natural-course run provides a genuine calibration check on the Q models after the fluctuation update; (2) the mean efficient influence curve is not zero by construction (targeting converges to near-zero but not exactly zero), so the SE uses the uncentered second-moment estimator $\sqrt{\text{mean}(IC^2)/n}$ rather than $sd(IC)/\sqrt{n}$ – see the se entry in Value for details.

Value

A named list with the same structure as sdr(), except where noted. Top-level elements:

psi

EIF point estimate of E[Y(d)] under the MTP, after the iTMLE targeting step.

se

Standard error from the efficient influence curve, computed as sqrt(mean(IC^2) / n) – the uncentered second-moment estimator (or its cluster-robust analogue). Unlike the standard sd(IC)/sqrt(n) used by sdr(), this does not assume E[IC] = 0: it remains conservative when the targeting loop has not fully converged to a zero-mean IC. When targeting converges well the two are essentially identical. The centered and second-moment SEs are both available in diagnostics$se_info (se_centered and se_second_moment) for comparison.

ci

95% Wald confidence interval: ⁠psi +/- 1.96 * se⁠.

Y_obs

Observed mean outcome mean(Y). Quick sanity check against psi under the natural-course policy.

ic_df

data.table with columns id and ic (per-subject influence curve values). Used by contrast().

Predictions (⁠$predictions⁠): cross-fitted Q matrices, one column per time point, one row per subject.

⁠$natural⁠

Q(t) under the natural-course policy.

⁠$shifted⁠

Q(t) under the MTP, after the targeting step.

Diagnostics (⁠$diagnostics⁠): model fit, calibration, and targeting summaries. iTMLE-specific additions are noted.

⁠$recursion_diag⁠

data.table, one row per fold x time point. Tracks g/Q predictions under natural and shifted policy, the current regression target (⁠Y_target_*⁠, the outer-step targeted Q at t+1), pre- and post-targeting Q on the training side (⁠Q_{nat,shf}_pre_*⁠, ⁠Q_{nat,shf}_post_*⁠), model-fit residuals (resid_sd, resid_q95_abs, resid_max_abs – Y_target minus shifted-mixture Q), targeting update magnitudes (⁠delta_{nat,shf}_target_{sd,q95_abs,max_abs}⁠), and pre-/post- targeting Q means on the validation set (Q_nat_vl_pre_mean, Q_shf_vl_pre_mean, Q_nat_vl_post_mean, Q_shf_vl_post_mean).

⁠$branch_cal⁠

Per-fold, per-time calibration table for g_remain, g_death, Q_rem, and Q_exit. Same structure as sdr().

⁠$target_cal⁠

(iTMLE only) data.table with one row per targeting iteration (fold x time x iteration), tracking EIF magnitude as it decreases toward convergence.

⁠$target_sl⁠

(iTMLE only) data.table of SuperLearner weights from the fluctuation (targeting) model, per fold and time point. Shows which wrappers from sl_itmle were selected.

⁠$sl_summary⁠

data.table of SuperLearner learner weights for the Q/g models. Same structure as sdr().

⁠$se_info⁠

(iTMLE only) List with SE computation details: se, n, n_eff, and cluster-adjustment bookkeeping.

Weights (⁠$weights⁠), Settings (⁠$settings⁠), and Call (⁠$call⁠) have the same structure as sdr(). ⁠$settings$start_t⁠ records the first time point used.

Natural-course diagnostic value

Unlike sdr(), iTMLE does not collapse to mean(Y) under the natural-course policy (see sdr() for why SDR does). A natural-course iTMLE run reflects the calibration of Q* (Q after the fluctuation update). To assess the Q models before targeting, use qreg() under the natural course.

Sample size requirements

This package targets large longitudinal datasets – thousands of subjects – typical of ICU, ward, or emergency department cohorts. The doubly-robust estimators require adequate observations within each branch (alive, in-state, exited) at every time point for the SuperLearner component models to be stable. pool_g_death and pool_q_exit borrow strength across time steps when exit events are sparse, but there must still be sufficient events across the pooled structure. All built-in examples use a minimum of 2,000 subjects.

Diagnostics

The returned fit carries several diagnostic tables under ⁠$diagnostics⁠:

• branch_cal – per-fold, per-time calibration for the four branches (g_remain, g_death, q_rem, q_exit); summarise with branch_cal_summary() and use to tune the g/Q SuperLearner libraries.

• recursion_diag – per-fold, per-time EIF/targeting mechanics (⁠Y_target_*⁠, ⁠Q_*_pre/post_*⁠, ⁠delta_*_target_*⁠, ⁠resid_*⁠, range-violation counts, validation-side Q means). Post-hoc sanity.

• target_cal – per outer/inner targeting iteration, tracks EIF magnitude toward convergence.

• target_sl – per (fold, t) SuperLearner coefficients from the fluctuation model; use to prune sl_tmle wrappers that never fire.

sl_summary (per-(fold, t, component) SuperLearner coefficients for the g/Q models) and ic_df (per-subject influence-curve values consumed by contrast()) are also attached. Weight diagnostics live on the density_ratio() object itself – see weight_diagnostics(). A worked workflow is in vignette("diagnostics").

References

Luedtke AR, Sofrygin O, van der Laan MJ, Carone M (2017). Sequential Double Robustness in Right-Censored Longitudinal Models. arXiv:1705.02459.

Diaz I, Williams N, Hoffman KL, Schenck EJ (2021). Nonparametric Causal Effects Based on Longitudinal Modified Treatment Policies. JASA 118(542):846-857.

See Also

density_ratio(), sdr(), contrast(), absorb_rule(), sl_itmle, weight_diagnostics(), branch_cal_summary()

Examples

library(SuperLearner)
sl_lib  <- c("SL.mean", "SL.glm")
tgt_lib <- c("SL.tmle_empty", "SL.tmle_intercept", "SL.tmle_glm")

# ---- Single binary treatment (sim_bin) ------------------------------
df <- sim_bin(n = 1000L, tmax = 3L, seed = 1L)
policy_bin <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmax(
    D_block[[a_names[1]]], as.integer(D_block[["L1"]] > 1.0)
  )
  out
}
wr <- density_ratio(
  df = df, a_names = "A", tmax = 3L,
  baseline = c("age", "sex"), tv_names = c("L1", "L2"),
  sl_g = sl_lib, k = 1L, inner_v = 3L, v = 3L, seed = 1L,
  id = "id", time = "time", policy_spec_fun = policy_bin
)
res <- itmle(
  df = df, weight_object = wr, tmax = 3L,
  id = "id", time = "time", alive = "alive", in_state = "in_state", y = "Y",
  baseline = c("age", "sex"), tv_names = c("L1", "L2"), a_names = "A",
  sl_remain = sl_lib, sl_death = sl_lib,
  sl_recursive = sl_lib, sl_y = sl_lib, sl_tmle = tgt_lib,
  k = 1L, inner_v = 2L, v_target_itmle = 2L, v_sl_inner_itmle = 2L,
  parallel = FALSE, seed = 1L, policy_spec_fun = policy_bin
)
res$psi; res$se

# ---- Single continuous treatment (sim_cont) -------------------------
df_c <- sim_cont(n = 1000L, tmax = 3L, seed = 1L)
policy_cont <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmin(D_block[[a_names[1]]] + 0.2, 2.0)
  out
}
wr_c <- density_ratio(
  df = df_c, a_names = "A", tmax = 3L,
  baseline = "age", tv_names = "L1",
  sl_g = sl_lib, k = 1L, inner_v = 3L, v = 3L, seed = 1L,
  id = "id", time = "time", policy_spec_fun = policy_cont
)
res_c <- itmle(
  df = df_c, weight_object = wr_c, tmax = 3L,
  id = "id", time = "time", alive = "alive", in_state = "in_state", y = "Y",
  baseline = "age", tv_names = "L1", a_names = "A",
  sl_remain = sl_lib, sl_death = sl_lib,
  sl_recursive = sl_lib, sl_y = sl_lib, sl_tmle = tgt_lib,
  k = 1L, inner_v = 2L, v_target_itmle = 2L, v_sl_inner_itmle = 2L,
  parallel = FALSE, seed = 1L, policy_spec_fun = policy_cont
)
res_c$psi; res_c$se

# ---- Multiple treatments -- binary + continuous (sim_multi) ---------
df_m <- sim_multi(
  n = 1000L, tmax = 3L, seed = 1L, n_binary = 1L, n_continuous = 1L
)
policy_multi <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[["A_b1"]] <- pmax(D_block[["A_b1"]], as.integer(D_block[["L1"]] > 1.0))
  out[["A_c1"]] <- pmin(D_block[["A_c1"]] + 0.2, 2.0)
  out
}
wr_m <- density_ratio(
  df = df_m, a_names = c("A_b1", "A_c1"), tmax = 3L,
  baseline = c("age", "sex"), tv_names = "L1",
  sl_g = sl_lib, k = 1L, inner_v = 2L, v = 2L, seed = 1L,
  id = "id", time = "time", policy_spec_fun = policy_multi
)
res_m <- itmle(
  df = df_m, weight_object = wr_m, tmax = 3L,
  id = "id", time = "time", alive = "alive", in_state = "in_state", y = "Y",
  baseline = c("age", "sex"), tv_names = "L1", a_names = c("A_b1", "A_c1"),
  sl_remain = sl_lib, sl_death = sl_lib,
  sl_recursive = sl_lib, sl_y = sl_lib, sl_tmle = tgt_lib,
  k = 1L, inner_v = 2L, v_target_itmle = 2L, v_sl_inner_itmle = 2L,
  parallel = FALSE, seed = 1L, policy_spec_fun = policy_multi
)
res_m$psi; res_m$se

Description

A SuperLearner metalearner that combines base learners by minimising a log-density-ratio loss rather than the default non-negative least squares (NNLS). Weights are found by BFGS optimisation over the softmax-reparameterised simplex; if BFGS fails to converge, equal weights are used as a fallback.

The metalearner operates in density-ratio space: base learners must return density ratios (not probabilities). Standard SuperLearner wrappers such as SL.glm or SL.xgboost return probabilities and are not compatible with this metalearner. Custom DR-returning wrappers are required; built-in wrappers for the WB pathway are planned for version 1.0.

Clipping note: the standard bounds clipping applied to all probability-scale predictions elsewhere in the pipeline does not apply here. DR-scale flooring is handled by dr_floor inside this function.

Usage

method.WB_dr(dr_floor = 1e-10)

Arguments

Value

A SuperLearner method list (with computeCoef and computePred slots) suitable for passing to the method argument of SuperLearner.

References

Wu and Benkeser (2021). Improved inference for vaccine-induced immune responses via shape-constrained methods.

See Also

density_ratio

Description

Compact console summaries for the main CausalState output objects.

print.sdr_fit and print.itmle_fit display the point estimate, standard error, 95% Wald CI, and the observed outcome mean.

print.qreg_fit displays the shifted estimate, SE, 95% CI, and the natural-course plug-in for comparison.

print.branch_cal_summary renders a wide table with super-column headers per branch (g_remain, g_death, q_rem, q_exit), showing fold-weighted n_vl, target mean, predicted mean, and calibration slope for each time step. The underlying data.table (returned invisibly) contains all additional metrics.

Usage

## S3 method for class 'sdr_fit'
print(x, digits = 3L, ...)

## S3 method for class 'itmle_fit'
print(x, digits = 3L, ...)

## S3 method for class 'qreg_fit'
print(x, digits = 3L, ...)

## S3 method for class 'branch_cal_summary'
print(x, digits = 3L, ...)

Arguments

Value

The input object, invisibly.

See Also

sdr, itmle, qreg, branch_cal_summary

Description

Fits sequential Q-models backwards through time and returns the plug-in (substitution) estimator $\hat\Psi = n^{-1}\sum_i \hat Q_1(H_{i1})$. No EIF update is applied, so the estimate carries first-order bias and should not be used as a real causal estimate.

Usage

qreg(
  df,
  tmax,
  id,
  time,
  alive,
  in_state,
  y,
  baseline,
  tv_names = character(0),
  a_names = character(0),
  no_lag_vars = character(0),
  policy_names = character(0),
  sl_remain = NULL,
  sl_death = NULL,
  sl_recursive = NULL,
  sl_rec_early = NULL,
  rec_transition = NULL,
  sl_y = NULL,
  outcome_family = c("binomial", "gaussian"),
  y_bounds = NULL,
  bounds = 1e-05,
  absorb = list(),
  policy_spec_fun = function(D_block, t, a_names) NULL,
  k = 2,
  seed = 1,
  parallel = FALSE,
  fold_workers = NULL,
  reg_workers = NULL,
  sl_workers = NULL,
  inner_v = 5L,
  cluster = NULL,
  pool_g_death = FALSE,
  pool_q_exit = FALSE,
  pool_time = "spline",
  weight_object = NULL,
  v = 5L,
  trim = 0.99,
  verbose = TRUE
)

Arguments

Details

The intended use is to pass the natural-course (identity) policy – that is, no shift – and compare the resulting estimate to the observed mean of y. Close agreement indicates that the Q-models are well-calibrated under the natural course, which is a prerequisite for the doubly-robust estimates from sdr() or itmle() to be trustworthy. Poor agreement signals Q-model misspecification.

Do not pass a shifted policy to qreg and interpret the result as a causal estimate. For policy evaluation use sdr() or itmle().

The reported se_naive is $s(\hat Q_1) / \sqrt{n}$, which treats the fitted Q as fixed and should be read only as a rough guide to Monte Carlo variability of the plug-in.

Value

A list with:

estimate

Point estimate $\hat\Psi$.

se_naive

Naive SE (always present).

se_eif

EIF-based SE (NULL if no weight_object).

se

Best available SE: se_eif if available, else se_naive.

ci

95% Wald CI using se.

psi_natural

Plug-in estimate under the natural course.

psi_shifted

Plug-in estimate under the MTP (= estimate).

sl_summary

data.table of SuperLearner learner weights per (fold, time-point, model component). Same structure as in sdr().

fold_diag

data.table of per-fold, per-time-point diagnostics.

diagnostics$branch_cal

Per-fold, per-time branch calibration table: empirical mean targets vs. predictions for g_remain, g_death, and Q_remain. Use this to check whether the Q-models are well-calibrated under the natural course – the primary diagnostic purpose of qreg().

diagnostics$diag_table

Additional per-fold, per-time summary statistics (Q prediction ranges, pseudo-outcome statistics).

Sample size requirements

This package targets large longitudinal datasets – thousands of subjects – typical of ICU, ward, or emergency department cohorts. The Q-models require adequate observations within each branch (alive, in-state, exited) at every time point to be stable. pool_g_death and pool_q_exit borrow strength across time steps when exit events are sparse, but there must still be sufficient events across the pooled structure. All built-in examples use a minimum of 2,000 subjects.

See Also

sdr(), itmle(), density_ratio()

Examples

library(SuperLearner)

# ICU-like DGP: patients die, discharge, or remain in state each time point
sim_ex <- function(n = 2000L, tmax = 5L) {
  set.seed(42L)
  rows <- vector("list", n)
  for (i in seq_len(n)) {
    age <- round(rnorm(1, 65, 10)); L1 <- rnorm(1)
    pat <- list()
    for (t in seq_len(tmax)) {
      A     <- rbinom(1, 1, plogis(0.3 * L1 - 0.4))
      u     <- runif(1)
      p_die <- plogis(-4.0 + 0.2 * L1 - 0.1 * age / 10)
      p_dc  <- plogis(-2.5 + 0.5 * A)
      if (u < p_die) {
        alive <- 0L; in_state <- 0L
      } else if (u < p_die + p_dc) {
        alive <- 1L; in_state <- 0L
      } else {
        alive <- 1L; in_state <- 1L
      }
      Y <- rbinom(1, 1, plogis(-0.5 + 0.4 * A - 0.2 * L1))
      pat[[length(pat) + 1L]] <- data.frame(
        id = i, time = t, age = age, L1 = L1,
        A = A, alive = alive, in_state = in_state, Y = Y
      )
      if (in_state == 0L) break
      if (t < tmax) L1 <- L1 + rnorm(1, -0.1 * A, 0.3)
    }
    rows[[i]] <- do.call(rbind, pat)
  }
  do.call(rbind, rows)
}
df <- sim_ex()

# Natural-course policy (no shift) -- the correct use of qreg
policy_nat <- function(D_block, t, a_names) D_block[, ..a_names, drop = FALSE]

sl_lib <- c("SL.mean", "SL.glm")

# Q-model calibration check: estimate should be close to mean(df$Y)
res <- qreg(
  df              = df,
  tmax            = 5L,
  id              = "id",
  time            = "time",
  alive           = "alive",
  in_state        = "in_state",
  y               = "Y",
  baseline        = "age",
  tv_names        = "L1",
  a_names         = "A",
  sl_remain       = sl_lib,
  sl_death        = sl_lib,
  sl_recursive    = sl_lib,
  sl_y            = sl_lib,
  k               = 1L,
  inner_v         = 3L,
  parallel        = FALSE,
  seed            = 1L,
  policy_spec_fun = policy_nat
)

# Compare plug-in to observed mean: close agreement = well-calibrated Q
res$estimate
mean(df$Y)


# Multi-treatment: plug-in estimate under a joint policy over two treatments

library(SuperLearner)

sim_multi <- function(n = 2000L, tmax = 5L) {
  set.seed(42L)
  rows <- vector("list", n)
  for (i in seq_len(n)) {
    age <- round(rnorm(1, 65, 10)); L1 <- rnorm(1)
    pat <- list()
    for (t in seq_len(tmax)) {
      A1    <- rbinom(1, 1, plogis(0.3 * L1 - 0.4))
      A2    <- rbinom(1, 1, plogis(0.2 * L1 + 0.3 * A1 - 0.3))
      u     <- runif(1)
      p_die <- plogis(-4.0 + 0.2 * L1 - 0.1 * age / 10)
      p_dc  <- plogis(-2.5 + 0.4 * A1 + 0.3 * A2)
      if (u < p_die) {
        alive <- 0L; in_state <- 0L
      } else if (u < p_die + p_dc) {
        alive <- 1L; in_state <- 0L
      } else {
        alive <- 1L; in_state <- 1L
      }
      Y <- rbinom(1, 1, plogis(-0.5 + 0.3 * A1 + 0.3 * A2 - 0.2 * L1))
      pat[[length(pat) + 1L]] <- data.frame(
        id = i, time = t, age = age, L1 = L1,
        A1 = A1, A2 = A2, alive = alive, in_state = in_state, Y = Y
      )
      if (in_state == 0L) break
      if (t < tmax) L1 <- L1 + rnorm(1, -0.1 * A1, 0.3)
    }
    rows[[i]] <- do.call(rbind, pat)
  }
  do.call(rbind, rows)
}
df2 <- sim_multi()

policy_fn2 <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmin(D_block[[a_names[1]]] + 0.3, 1)
  out[[a_names[2]]] <- pmin(D_block[[a_names[2]]] + 0.3, 1)
  out
}

sl_lib <- c("SL.mean", "SL.glm")

res2 <- qreg(
  df              = df2,
  tmax            = 5L,
  id              = "id",
  time            = "time",
  alive           = "alive",
  in_state        = "in_state",
  y               = "Y",
  baseline        = "age",
  tv_names        = "L1",
  a_names         = c("A1", "A2"),
  sl_remain       = sl_lib,
  sl_death        = sl_lib,
  sl_recursive    = sl_lib,
  sl_y            = sl_lib,
  k               = 1L,
  inner_v         = 3L,
  parallel        = FALSE,
  seed            = 1L,
  policy_spec_fun = policy_fn2
)
res2$estimate
res2$se

Description

Estimates the mean counterfactual outcome under a modified treatment policy (MTP) using the LMTP-SDR estimator of Diaz et al. (2021), which extends the SDR construction of Luedtke et al. (2017) to general MTPs via density-ratio weighting. Starting from terminal pseudo-outcomes and working backward in time, the estimator propagates an EIF-corrected pseudo-outcome through a sequence of Q-regressions; the final estimate is the mean of the pseudo-outcomes at the first time point. The estimator is sequentially doubly robust (2^$K$-robust): consistent whenever, at each time point t, either the treatment model $g_t$ or the outcome model $Q_t$ is consistently estimated. Requires pre-computed density ratio weights from density_ratio().

Usage

sdr(
  df,
  weight_object,
  tmax,
  id,
  time,
  alive,
  in_state,
  y,
  baseline,
  tv_names = character(0),
  a_names = character(0),
  no_lag_vars = character(0),
  policy_names = character(0),
  sl_remain = NULL,
  sl_death = NULL,
  sl_recursive = NULL,
  sl_rec_early = NULL,
  rec_transition = NULL,
  sl_y = NULL,
  outcome_family = c("binomial", "gaussian"),
  y_bounds = NULL,
  bounds = 1e-05,
  trim = 0.99,
  absorb = list(),
  policy_spec_fun = function(D_block, t, a_names) NULL,
  k = 2,
  seed = 1,
  parallel = FALSE,
  fold_workers = NULL,
  reg_workers = NULL,
  sl_workers = NULL,
  inner_v = 5L,
  cluster = NULL,
  cluster_se_only = FALSE,
  pool_g_death = FALSE,
  pool_q_exit = FALSE,
  pool_time = "spline",
  verbose = TRUE
)

Arguments

Value

A named list. Top-level elements:

psi

EIF-corrected point estimate of E[Y(d)] under the MTP.

se

Standard error from the efficient influence curve, computed as sd(IC) / sqrt(n) (or the cluster-robust sandwich equivalent when a cluster variable is supplied). This is the standard centered variance estimator and assumes E[IC] = 0, which holds exactly under the natural course and approximately under an MTP when the Q and g models are well-specified.

ci

95% Wald confidence interval: ⁠psi +/- 1.96 * se⁠.

Y_obs

Observed mean outcome mean(Y) across all subjects. Quick sanity check: under the natural-course policy psi should be close to Y_obs; a large gap suggests a data or model issue.

ic_df

data.table with columns id and ic (per-subject influence curve values). Used by contrast().

Predictions (⁠$predictions⁠): cross-fitted Q matrices, one column per time point, one row per subject.

⁠$natural⁠

Q(t) under the natural-course policy.

⁠$shifted⁠

Q(t) under the MTP.

Decomposition (⁠$decomposition⁠): plug-in components explaining the EIF correction relative to the raw plug-in.

⁠$psi_plugin_nat⁠

Plug-in estimate under the natural course.

⁠$psi_plugin_shf⁠

Plug-in estimate under the MTP.

⁠$psi_plugin_diff⁠

Plug-in risk difference (psi_plugin_shf - psi_plugin_nat).

⁠$psi_eif_gap⁠

EIF correction: psi - psi_plugin_shf. A large value means the pseudo-outcome recursion shifted the estimate substantially beyond the raw plug-in.

Diagnostics (⁠$diagnostics⁠): model fit and calibration summaries.

⁠$recursion_diag⁠

data.table, one row per fold x time point. Tracks g/Q predictions (training side) under natural and shifted policy, the current regression target (⁠Y_target_*⁠) and the EIF-updated pseudo-outcome (⁠pseudo_post_*⁠), EIF update magnitude (delta_sd, delta_q95_abs, delta_max_abs), model-fit residuals (resid_sd, resid_q95_abs, resid_max_abs – Y_target minus shifted-mixture Q), and validation-side mixture Q means (Q_nat_vl_mean, Q_shf_vl_mean) for comparing training vs. validation trajectories.

⁠$branch_cal⁠

Per-fold, per-time calibration table: empirical mean target vs. mean prediction for g_remain, g_death, Q_rem, and Q_exit. The Q_rem target is the pseudo-outcome mean – directly interpretable as a calibration check only under the natural-course policy. qexit_type records whether Q_exit is a binomial or Gaussian regression.

⁠$sl_summary⁠

data.table of SuperLearner learner weights, one row per learner per (fold, time, component). Columns: fold, t, component (g_remain, g_death, Q_rem, Q_exit), plus one numeric column per learner. Consistently zero-weight learners can be dropped from the library.

Weights (⁠$weights⁠): density ratio inputs used by the estimator.

⁠$weights_dt⁠

data.table copy of the density ratio object (trimmed to tmax, Rt_cum column removed).

⁠$trim⁠

Trim quantile applied before forming cumulative density ratios.

Settings (⁠$settings⁠): parameters used for the fit.

⁠$start_t⁠

First time point included in the estimation.

⁠$outcome_family⁠

"binomial" or "gaussian".

⁠$pool_g_death⁠

Whether g_death was pooled across time.

⁠$pool_q_exit⁠

Whether Q_exit was pooled across time.

⁠$variable_info⁠

Named list of variable names and settings: id, time, alive, in_state, cluster, baseline, time_varying, treatment, outcome, tmax, k.

Call (⁠$call⁠): matched call expression, for reproducibility.

Parallelism

fold_workers, reg_workers, and sl_workers can be used independently or together. Using a single level is robust with all learners. Combining two or more creates nested mclapply calls; in that case multi-threaded or GPU-based learners (e.g. xgboost with CUDA, OpenMP-based methods) may crash in the child processes and should be avoided or limited to one thread.

Natural-course caution

Under the natural-course policy the cumulative density ratios equal one and the SDR pseudo-outcome recursion collapses algebraically to mean(Y), regardless of Q-model quality. A natural-course SDR run cannot detect model misspecification. To assess model calibration use diagnostics$branch_cal, fold_diag, and sl_summary, or run itmle() under the natural course (iTMLE does not collapse in the same way).

Sample size requirements

This package targets large longitudinal datasets – thousands of subjects – typical of ICU, ward, or emergency department cohorts. The doubly-robust estimators require adequate observations within each branch (alive, in-state, exited) at every time point for the SuperLearner component models to be stable. pool_g_death and pool_q_exit borrow strength across time steps when exit events are sparse, but there must still be sufficient events across the pooled structure. All built-in examples use a minimum of 2,000 subjects.

Diagnostics

Two diagnostic tables are attached to the returned object under ⁠$diagnostics⁠:

• branch_cal – per-fold, per-time calibration for the four branches (g_remain, g_death, q_rem, q_exit); summarise with branch_cal_summary() and use to tune the SuperLearner libraries.

• recursion_diag – per-fold, per-time EIF-mechanics diagnostics (⁠Y_target_*⁠, ⁠pseudo_post_*⁠, ⁠delta_*⁠, ⁠resid_*⁠, range-violation counts, validation-side Q means). Post-hoc sanity checks.

The fit also carries sl_summary (per-(fold, t, component) SuperLearner coefficients) and ic_df (per-subject influence-curve values consumed by contrast()). Weight diagnostics live on the density_ratio() object itself – see weight_diagnostics(). A worked workflow is in vignette("diagnostics").

References

Diaz I, Williams N, Hoffman KL, Schenck EJ (2021). Nonparametric Causal Effects Based on Longitudinal Modified Treatment Policies. JASA 118(542):846–857.

Luedtke AR, Sofrygin O, van der Laan MJ, Carone M (2017). Sequential Double Robustness in Right-Censored Longitudinal Models. arXiv:1705.02459.

See Also

density_ratio(), itmle(), contrast(), absorb_rule(), weight_diagnostics(), branch_cal_summary()

Examples

library(SuperLearner)
sl_lib <- c("SL.mean", "SL.glm")

# ---- Single binary treatment (sim_bin) ------------------------------
df <- sim_bin(n = 1000L, tmax = 3L, seed = 1L)
policy_bin <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmax(
    D_block[[a_names[1]]], as.integer(D_block[["L1"]] > 1.0)
  )
  out
}
wr <- density_ratio(
  df = df, a_names = "A", tmax = 3L,
  baseline = c("age", "sex"), tv_names = c("L1", "L2"),
  sl_g = sl_lib, k = 1L, inner_v = 2L, v = 2L, seed = 1L,
  id = "id", time = "time", policy_spec_fun = policy_bin
)
res <- sdr(
  df = df, weight_object = wr, tmax = 3L,
  id = "id", time = "time", alive = "alive", in_state = "in_state", y = "Y",
  baseline = c("age", "sex"), tv_names = c("L1", "L2"), a_names = "A",
  sl_remain = sl_lib, sl_death = sl_lib,
  sl_recursive = sl_lib, sl_y = sl_lib,
  k = 1L, inner_v = 2L, parallel = FALSE, seed = 1L,
  policy_spec_fun = policy_bin
)
res$psi; res$se

# ---- Single continuous treatment (sim_cont) -------------------------
df_c <- sim_cont(n = 1000L, tmax = 3L, seed = 1L)
policy_cont <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[[a_names[1]]] <- pmin(D_block[[a_names[1]]] + 0.2, 2.0)
  out
}
wr_c <- density_ratio(
  df = df_c, a_names = "A", tmax = 3L,
  baseline = "age", tv_names = "L1",
  sl_g = sl_lib, k = 1L, inner_v = 2L, v = 2L, seed = 1L,
  id = "id", time = "time", policy_spec_fun = policy_cont
)
res_c <- sdr(
  df = df_c, weight_object = wr_c, tmax = 3L,
  id = "id", time = "time", alive = "alive", in_state = "in_state", y = "Y",
  baseline = "age", tv_names = "L1", a_names = "A",
  sl_remain = sl_lib, sl_death = sl_lib,
  sl_recursive = sl_lib, sl_y = sl_lib,
  k = 1L, inner_v = 2L, parallel = FALSE, seed = 1L,
  policy_spec_fun = policy_cont
)
res_c$psi; res_c$se

# ---- Multiple treatments -- binary + continuous (sim_multi) ---------
df_m <- sim_multi(
  n = 1000L, tmax = 3L, seed = 1L, n_binary = 1L, n_continuous = 1L
)
policy_multi <- function(D_block, t, a_names) {
  out <- D_block[, ..a_names, drop = FALSE]
  out[["A_b1"]] <- pmax(D_block[["A_b1"]], as.integer(D_block[["L1"]] > 1.0))
  out[["A_c1"]] <- pmin(D_block[["A_c1"]] + 0.2, 2.0)
  out
}
wr_m <- density_ratio(
  df = df_m, a_names = c("A_b1", "A_c1"), tmax = 3L,
  baseline = c("age", "sex"), tv_names = "L1",
  sl_g = sl_lib, k = 1L, inner_v = 2L, v = 2L, seed = 1L,
  id = "id", time = "time", policy_spec_fun = policy_multi
)
res_m <- sdr(
  df = df_m, weight_object = wr_m, tmax = 3L,
  id = "id", time = "time", alive = "alive", in_state = "in_state", y = "Y",
  baseline = c("age", "sex"), tv_names = "L1", a_names = c("A_b1", "A_c1"),
  sl_remain = sl_lib, sl_death = sl_lib,
  sl_recursive = sl_lib, sl_y = sl_lib,
  k = 1L, inner_v = 2L, parallel = FALSE, seed = 1L,
  policy_spec_fun = policy_multi
)
res_m$psi; res_m$se

Description

Generates a longitudinal panel of ICU-style trajectories with one binary treatment (A), competing exit events (death vs discharge), and a binary outcome (Y). Intended for vignettes, examples, and quick smoke tests.

Usage

sim_bin(n = 2000L, tmax = 5L, seed = 1L, apply_policy = FALSE)

Arguments

Value

A data.frame with one row per subject-time. Columns: id, time, age, sex, L1, L2, A, alive, in_state, Y.

See Also

sim_cont(), sim_multi(), sdr(), itmle()

Description

Longitudinal ICU-style trajectories with one continuous treatment (A, bounded roughly in ⁠[0, 2]⁠), competing exit events, and a binary outcome. Continuous-treatment analogue of sim_bin().

Usage

sim_cont(n = 2000L, tmax = 5L, seed = 1L, dose_shift = 0)

Arguments

Value

A data.frame with columns id, time, age, L1, A, alive, in_state, Y.

See Also

sim_bin(), sim_multi(), sdr(), itmle()

Description

Longitudinal ICU-style trajectories with an arbitrary number of binary and/or continuous treatments, competing exit events, and a binary outcome. Generalises sim_bin() / sim_cont() to multi-treatment MTPs.

Usage

sim_multi(
  n = 2000L,
  tmax = 5L,
  seed = 1L,
  n_binary = 1L,
  n_continuous = 1L,
  apply_policy = FALSE
)

Arguments

Details

Binary treatments are named ⁠A_b1, A_b2, ...⁠ (first n_binary columns); continuous treatments are named ⁠A_c1, A_c2, ...⁠. Each treatment enters the exit-event and outcome models with a coefficient scaled by 1 / sqrt(k) where k is the treatment index within its type, so adding more treatments does not blow up effect sizes.

Value

A data.frame with columns id, time, age, sex, L1, ⁠A_b1..A_b{n_binary}⁠, ⁠A_c1..A_c{n_continuous}⁠, alive, in_state, Y.

See Also

sim_bin(), sim_cont(), sdr(), itmle()

Description

A family of SuperLearner-compatible learner functions designed for the iTMLE targeting step. They differ from standard wrappers in one critical way: the logit offset is passed as a column in X (named ._sl_offset) rather than via the offset argument. This is necessary because SuperLearner's internal cross-validation subsetting drops the offset vector, whereas a column in X is correctly subset.

Each learner:

1. Extracts ._sl_offset from X/newX via extract_offset().

2. Fits a fluctuation model with that offset.

3. Returns predictions on the probability scale, clipped to ⁠[bounds, 1-bounds]⁠.

Learners available:

SL.tgt.empty

No fluctuation: returns expit(offset) unchanged. Acts as the "no update" option.

SL.tgt.intercept

One-parameter intercept fluctuation (standard TMLE update).

SL.tgt.glm

Main-terms logistic GLM fluctuation with offset.

SL.tgt.glmnet

Penalised logistic regression (elastic net) with offset via glmnet::cv.glmnet().

SL.tgt.xgboost

Gradient boosted trees with offset via base_margin. CUDA-capable.

Pre-configured variants (⁠SL.tmle_*⁠) expose specific hyperparameter choices and are the recommended building blocks for sl_tmle:

SL.tmle_empty

No fluctuation (pass-through).

SL.tmle_intercept

One-parameter intercept (standard TMLE update).

SL.tmle_glm

Main-terms logistic GLM.

SL.tmle_glmnet_ridge

Ridge regression (alpha = 0).

SL.tmle_glmnet_enet

Elastic net (alpha = 0.5).

SL.tmle_glmnet_lasso

Lasso (alpha = 1).

SL.tmle_xgb_d1

XGBoost depth 1, CPU, 50 rounds.

SL.tmle_xgb_d3

XGBoost depth 3, CPU, 50 rounds.

SL.tmle_xgb_d6

XGBoost depth 6, CPU, 50 rounds.

The default sl_tmle vector follows Luedtke et al. (2017) Algorithm 4: empty, intercept, glm, and the three XGBoost depths. The glmnet variants are available but must be specified explicitly via sl_tmle.

Usage

SL.tgt.empty(Y, X, newX, family, obsWeights, id, bounds = 1e-05, ...)

SL.tgt.intercept(Y, X, newX, family, obsWeights, id, bounds = 1e-05, ...)

SL.tgt.glm(Y, X, newX, family, obsWeights, id, bounds = 1e-05, ...)

SL.tgt.glmnet(
  Y,
  X,
  newX,
  family,
  obsWeights,
  id,
  alpha = 0.5,
  nfolds = 3L,
  s_select = "lambda.min",
  bounds = 1e-05,
  ...
)

SL.tgt.xgboost(
  Y,
  X,
  newX,
  family,
  obsWeights,
  id,
  nrounds = 100L,
  max_depth = 2L,
  eta = 0.1,
  subsample = 0.8,
  colsample_bytree = 0.8,
  max_delta_step = 0,
  nthread = 1L,
  use_cuda = TRUE,
  bounds = 1e-05,
  ...
)

SL.tmle_empty(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_intercept(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_glm(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_glmnet_ridge(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_glmnet_enet(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_glmnet_lasso(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_xgb_d1(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_xgb_d3(Y, X, newX, family, obsWeights, id, ...)

SL.tmle_xgb_d6(Y, X, newX, family, obsWeights, id, ...)

Arguments

Value

A list with elements pred (numeric predictions on the probability scale) and fit (a fitted object with a predict method).

See Also

itmle()

Examples

# Default targeting stack (Luedtke et al. Algorithm 4)
sl_tmle

# Lightweight stack for examples and testing (no XGBoost required)
sl_tmle_light <- c("SL.tmle_empty", "SL.tmle_intercept", "SL.tmle_glm")

# Add glmnet variants explicitly when desired
sl_tmle_penalised <- c(
  "SL.tmle_empty", "SL.tmle_intercept", "SL.tmle_glm",
  "SL.tmle_glmnet_ridge", "SL.tmle_glmnet_enet",
  "SL.tmle_xgb_d1", "SL.tmle_xgb_d3", "SL.tmle_xgb_d6"
)

Description

A character vector naming the default library of learners used in the iTMLE targeting step. Contains empty, intercept-only, GLM, and XGBoost wrappers at three tree depths. All wrappers carry the logit offset required for cross-validated TMLE as a data column (the “offset-as-column” convention) and clip internal probability predictions at 1e-5.

Pass this vector to the sl_target argument of itmle or construct your own library from the SL.tmle_* wrappers exported by this package.

Usage

sl_tmle

Format

A character vector of SuperLearner learner names.

Value

A character vector of SuperLearner learner name strings to be passed to the sl_target argument of itmle.

See Also

itmle, SL.tmle_glm, SL.tmle_glmnet_ridge, SL.tmle_xgb_d1

Description

Summarises the instantaneous density ratios and their cumulative products at each time step, restricted to subjects at risk at that time. This matches how the weights enter the SDR and iTMLE corrections: the EIF term at time t uses the product of ratios from t=1 through t, and only at-risk subjects contribute non-zero correction terms.

Usage

weight_diagnostics(weight_object, trim = 1)

Arguments

Details

Call with different trim values to assess trim sensitivity before passing the weight_object to sdr() or itmle().

Value

A data.table with one row per time step and columns: t, n_risk; instantaneous ratio stats (Rt_mean, Rt_median, Rt_p05, Rt_p95, Rt_max, Rt_ess); cumulative product stats (cum_mean, cum_median, cum_p05, cum_p95, cum_max, cum_ess). ESS uses the Kish approximation: $(\sum w)^2 / \sum w^2$.

See Also

density_ratio(), sdr(), itmle()