Evaluating Causal Effects of Modified Treatment Policies

Nima Hejazi and David Benkeser



Stochastic treatment regimes constitute a flexible framework for evaluating the effects of continuous-valued exposures/treatments. Modified treatment policies, one such technique within this framework, examine the effects attributable to shifting the observed (“natural”) value of a treatment, usually up or down by some scalar \(\delta\). The txshift package implements algorithms for computing one-step or targeted minimum loss-based (TML) estimates of the counterfactual means induced by additive modified treatment policies (MTPs), defined by a shifting function \(\delta(A,W)\). For a technical presentation, the interested reader is invited to consult Dı́az and van der Laan (2018) or the earlier work of Dı́az and van der Laan (2012) and Haneuse and Rotnitzky (2013). For background on Targeted Learning, consider consulting van der Laan and Rose (2011), van der Laan and Rose (2018), and van der Laan et al. (2022).

To start, let’s load the packages we’ll need and set a seed for simulation:

## haldensify v0.2.3: Highly Adaptive Lasso Conditional Density Estimation
## txshift v0.3.8: Efficient Estimation of the Causal Effects of Stochastic Interventions

Data and Notation

We’ll consider \(n\) observed units \(O_1, \ldots, O_n\), where each random variable \(O = (W, A, Y)\) corresponds to the data available on a single unit. Within \(O\), \(W\) denotes baseline covariates (e.g., age, biological sex, BMI), \(A \in \mathbb{R}\) a continuous-valued exposure (e.g., dosage of nutritional supplements taken), and \(Y\) an outcome of interest (e.g., disease status). To minimize unjustifiable assumptions, we let \(O \sim \mathcal{P} \in \mathcal{M}\), where \(\mathcal{P}\) is simply any distribution within the nonparametric statistical model \(\mathcal{M}\). To formalize the definition of stochastic interventions and their corresponding causal effects, we consider a nonparametric structural equation model (NPSEM), introduced by Pearl (2000), to define how the system changes under interventions of interest: \[\begin{align*}\label{eqn:npsem} W &= f_W(U_W) \\ A &= f_A(W, U_A) \\ Y &= f_Y(A, W, U_Y), \end{align*}\] We denote the observed data structure \(O = (W, A, Y)\)

Assuming that the distribution of \(A\) conditional on \(W = w\) has support in the interval \((l(w), u(w))\) – for convenience, we assume that the minimum natural value of treatment \(A\) for an individual with covariates \(W = w\) is \(l(w)\), while, similarly, the maximum is \(u(w)\) – a simple MTP based on a shift \(\delta\), is \[\begin{equation}\label{eqn:shift} \delta(a, w) = \begin{cases} a - \delta & \text{if } a > l(w) + \delta \\ a & \text{if } a \leq l(w) + \delta, \end{cases} \end{equation}\] where \(0 \leq \delta\) is an arbitrary pre-specified value that defines the degree to which the observed value \(A\) is to be shifted, where possible.

In case-cohort studies, it is common practice to make use of outcome-dependent two-phase sampling designs, which allow for expensive measurements made on the exposure (e.g., genomic sequencing of immune markers) to be avoided. As a complication, such sampling schemes alter the observed data structure from the simpler \(O = (W, A, Y)\) to \(O = (W, \Delta A, Y, \Delta)\), where the sampling indicator \(\Delta\) may itself be a function of the variables \(\{W, Y\}\). In this revised data structure, the value of \(A\) is only observed for units in the two-phase sample, for whom \(\Delta = 1\). Hejazi et al. (2020) provide a detailed investigation of the methodological details of efficient estimation under such designs in the context of vaccine efficacy trials; their work was used in the analysis of immune correlates of protection for HIV-1 (Hejazi et al. 2020) and COVID-19 (Gilbert et al. 2021). Of course, one may also account for loss to follow-up (i.e., censoring), which the txshift package supports (through the C_cens argument of the eponymous txshift function), though we avoid this complication in our subsequent examples in the interest of clarity of exposition. Corrections for both censoring and two-phase sampling make use of inverse probability of censoring weighting (IPCW), leading to IPCW-augmented one-step and TML estimators.

Simulate Data

Estimating the Effects of Additive MTPs

The simplest way to compute an efficient estimator for an additive MTP is to fit each of the nuisance parameters internally. This procedure can be sped up by using generalized linear models (GLMs) to fit the outcome regression \(Q_n\). The txshift() function provides a simple interface.

est_shift <- txshift(
  W = W, A = A, Y = Y, delta = delta,
  g_exp_fit_args = list(
    fit_type = "hal", n_bins = 5,
    grid_type = "equal_mass",
    lambda_seq = exp(seq(-1, -10, length = 100))
  Q_fit_args = list(
    fit_type = "glm",
    glm_formula = "Y ~ .^2"
## Counterfactual Mean of Shifted Treatment
## Intervention: Treatment + 0.5
## txshift Estimator: tmle
## Estimate: 1.8623
## Std. Error: 0.1542
## 95% CI: [1.56, 2.1646]

Interlude: Ensemble Machine Learning with sl3

To easily incorporate ensemble machine learning into the estimation procedure, txshift integrates with the sl3 R package (Coyle, Hejazi, Malenica, et al. 2022). For a complete guide on using the sl3 R package, consider consulting the chapter on Super Learning in van der Laan et al. (2022).

Efficient Effect Estimates with Machine Learning

Using the framework provided by the sl3 package, the nuisance functions required for our efficient estimators may be fit with ensemble machine learning. The Super Learner algorithm (van der Laan, Polley, and Hubbard 2007) implemented in sl3 uses the asymptotic optimality of V-fold cross-validation (Dudoit and van der Laan 2005; van der Laan, Dudoit, and Keles 2004; van der Vaart, Dudoit, and van der Laan 2006) to select an optimal prediction functions from a library or to construct an optimal combination of prediction functions.

Estimating the Effects of Additive MTPs Under Two-Phase Sampling

In case-cohort studies in which two-phase sampling is used, the data structure takes the from \(O = (W, \Delta A, Y, \Delta)\) as previously discussed. Under such sampling, the txshift() function may still be used to estimate the causal effect of an additive MTP – only a few additional arguments need to be specified:

est_shift_ipcw <- txshift(
  W = W, A = A, Y = Y, delta = delta,
  C_samp = Delta_samp, V = c("W", "Y"),
  samp_fit_args = list(fit_type = "glm"),
  g_exp_fit_args = list(
    fit_type = "hal", n_bins = 5,
    grid_type = "equal_mass",
    lambda_seq = exp(seq(-1, -10, length = 100))
  Q_fit_args = list(
    fit_type = "glm",
    glm_formula = "Y ~ .^2"
  eif_reg_type = "glm"
## Counterfactual Mean of Shifted Treatment
## Intervention: Treatment + 0.5
## txshift Estimator: tmle
## Estimate: 1.9515
## Std. Error: 0.1398
## 95% CI: [1.6776, 2.2254]

Note that we specify a few additional arguments in the call to txshift(), including C_samp, the indicator of inclusion in the two-phase sample; V, the set of other variables that may affect the sampling decision (in this case, both the baseline covariates and the outcome); samp_fit_args, which indicates how the sampling mechanism ought to be estimated; and eif_reg_type, which indicates how a particular reduced-dimension nuisance regression ought to be estimated (see Rose and van der Laan (2011) and Hejazi et al. (2020) for details). This last argument only has options for using a GLM or the highly adaptive lasso (HAL), a nonparametric regression estimator, using the hal9001 package [Coyle, Hejazi, Phillips, et al. (2022); hejazi2020hal9001-joss]. In practice, we recommend leaving this argument to the default and fitting this nuisance function with HAL; however, this is significantly more costly computationally.

Statistical Inference for One-step and TML Estimators

The efficient estimators implemented in txshift are asymptotically linear; thus, the estimator \(\psi_n\) converges to the true parameter value \(\psi_0\): \[\psi_n - \psi_0 = (P_n - P_0) \cdot D(\bar{Q}_n^{\star}, g_n) + R(\hat{P}^{\star}, P_0),\] which yields \[\psi_n - \psi_0 = (P_n - P_0) \cdot D(P_0) + o_P \left( \frac{1}{\sqrt{n}} \right),\] provided the following conditions,

  1. if \(D(\bar{Q}_n^{\star}, g_n)\) converges to \(D(P_0)\) in \(L_2(P_0)\) norm;
  2. the size of the class of functions considered for estimation of \(\bar{Q}_n^{\star}\) and \(g_n\) is bounded (technically, \(\exists \mathcal{F}\) such that \(D(\bar{Q}_n^{\star}, g_n) \in \mathcal{F}\) with high probability, where \(\mathcal{F}\) is a Donsker class); and
  3. the remainder term \(R(\hat{P}^{\star}, P_0)\) decays as \(o_P \left( \frac{1}{\sqrt{n}} \right)\).

By the central limit theorem, the estimators then have a Gaussian limiting distribution, \[\sqrt{n}(\psi_n - \psi) \to N(0, V(D(P_0))),\] where \(V(D(P_0))\) is the variance of the efficient influence function (or canonical gradient).

The above implies that \(\psi_n\) is a \(\sqrt{n}\)-consistent estimator of \(\psi\), that it is asymptotically normal (as given above), and that it is locally efficient. This allows us to build Wald-type confidence intervals in a straightforward manner:

\[\psi_n \pm z_{\alpha} \cdot \frac{\sigma_n}{\sqrt{n}},\] where \(\sigma_n^2\) is an estimator of \(V(D(P_0))\). The estimator \(\sigma_n^2\) may be obtained using the bootstrap or computed directly via the following

\[\sigma_n^2 = \frac{1}{n} \sum_{i = 1}^{n} D^2(\bar{Q}_n^{\star}, g_n)(O_i).\]

Such confidence intervals may easily be created with the confint method:

(ci_est_shift <- confint(est_shift))
##   lwr_ci      est   upr_ci 
## 1.559997 1.862286 2.164576

Advanced Usage: User-Specified Nuisance Regressions

In some special cases it may be useful for the experienced user to compute the treatment mechanism, censoring mechanism, outcome regression, and sampling mechanism fits separately (i.e., outside of the txshift wrapper function), instead applying the wrapper only to construct an efficient one-step or TML estimator. In such cases, the optional arguments ending in _ext.

# compute censoring mechanism and produce IPC weights externally
pi_mech <- plogis(W)
ipcw_out <- pi_mech

# compute treatment mechanism (propensity score) externally
## first, produce the down-shifted treatment data
gn_downshift <- dnorm(A - delta, mean = tx_mult * W, sd = 1)
## next, initialize and produce the up-shifted treatment data
gn_upshift <- dnorm(A + delta, mean = tx_mult * W, sd = 1)
## now, initialize and produce the up-up-shifted (2 * delta) treatment data
gn_upupshift <- dnorm(A + 2 * delta, mean = tx_mult * W, sd = 1)
## then, initialize and produce the un-shifted treatment data
gn_noshift <- dnorm(A, mean = tx_mult * W, sd = 1)
## finally, put it all together into an object like what's produced internally
gn_out <- as.data.table(cbind(gn_downshift, gn_noshift, gn_upshift,
                              gn_upupshift))[C == 1, ]
setnames(gn_out, c("downshift", "noshift", "upshift", "upupshift"))

# compute outcome regression externally
# NOTE: transform Y to lie in the unit interval and bound predictions such that
#       no values fall near the bounds of the interval
Qn_noshift <- (W + A - min(Y)) / diff(range(Y))
Qn_upshift <- (W + A + delta - min(Y)) / diff(range(Y))
Qn_noshift[Qn_noshift < 0] <- 0.025
Qn_noshift[Qn_noshift > 1] <- 0.975
Qn_upshift[Qn_upshift < 0] <- 0.025
Qn_upshift[Qn_upshift > 1] <- 0.975
Qn_out <- as.data.table(cbind(Qn_noshift, Qn_upshift))[C == 1, ]
setnames(Qn_out, c("noshift", "upshift"))

# construct efficient estimator by applying wrapper function 
est_shift_spec <- txshift(
  W = W, A = A, Y = Y, delta = delta,
 samp_fit_args = NULL,
 samp_fit_ext = ipcw_out,
 g_exp_fit_args = list(fit_type = "external"),
 Q_fit_args = list(fit_type = "external"),
 gn_exp_fit_ext = gn_out,
 Qn_fit_ext = Qn_out


Coyle, Jeremy R, Nima S Hejazi, Ivana Malenica, Rachael V Phillips, and Oleg Sofrygin. 2022. sl3: Modern Machine Learning Pipelines for Super Learning. https://doi.org/10.5281/zenodo.1342293.

Coyle, Jeremy R, Nima S Hejazi, Rachael V Phillips, Lars W van der Laan, and Mark J van der Laan. 2022. hal9001: The Scalable Highly Adaptive Lasso. https://doi.org/10.5281/zenodo.3558313.

Dı́az, Iván, and Mark J van der Laan. 2012. “Population Intervention Causal Effects Based on Stochastic Interventions.” Biometrics 68 (2): 541–49.

———. 2018. “Stochastic Treatment Regimes.” In Targeted Learning in Data Science: Causal Inference for Complex Longitudinal Studies, 167–80. Springer Science & Business Media.

Dudoit, Sandrine, and Mark J van der Laan. 2005. “Asymptotics of Cross-Validated Risk Estimation in Estimator Selection and Performance Assessment.” Statistical Methodology 2 (2): 131–54.

Gilbert, Peter B, Youyi Fong, David Benkeser, Jessica Andriesen, Bhavesh Borate, Marco Carone, Lindsay N Carpp, et al. 2021. “CoVPN Covid-19 Vaccine Efficacy Trial Immune Correlates Statistical Analysis Plan.” https://doi.org/10.6084/m9.figshare.13198595.

Haneuse, Sebastian, and Andrea Rotnitzky. 2013. “Estimation of the Effect of Interventions That Modify the Received Treatment.” Statistics in Medicine 32 (30): 5260–77.

Hejazi, Nima S, Mark J van der Laan, Holly E Janes, Peter B Gilbert, and David C Benkeser. 2020. “Efficient Nonparametric Inference on the Effects of Stochastic Interventions Under Two-Phase Sampling, with Applications to Vaccine Efficacy Trials.” Biometrics 77 (4): 1241–53. https://doi.org/10.1111/biom.13375.

Pearl, Judea. 2000. Causality. Cambridge University Press.

Rose, Sherri, and Mark J van der Laan. 2011. “A Targeted Maximum Likelihood Estimator for Two-Stage Designs.” International Journal of Biostatistics 7 (1): 1–21.

van der Laan, Mark J, Jeremy R Coyle, Nima S Hejazi, Ivana Malenica, Rachael V Phillips, and Alan E Hubbard. 2022. Targeted Learning in R: A Causal Data Science Handbook. CRC Press. https://tlverse.org/tlverse-handbook/.

van der Laan, Mark J, Sandrine Dudoit, and Sunduz Keles. 2004. “Asymptotic Optimality of Likelihood-Based Cross-Validation.” Statistical Applications in Genetics and Molecular Biology 3 (1): 1–23.

van der Laan, Mark J, Eric C Polley, and Alan E Hubbard. 2007. “Super Learner.” Statistical Applications in Genetics and Molecular Biology 6 (1).

van der Laan, Mark J, and Sherri Rose. 2011. Targeted Learning: Causal Inference for Observational and Experimental Data. Springer Science & Business Media.

———. 2018. Targeted Learning in Data Science: Causal Inference for Complex Longitudinal Studies. Springer Science & Business Media.

van der Vaart, Aad W, Sandrine Dudoit, and Mark J van der Laan. 2006. “Oracle Inequalities for Multi-Fold Cross Validation.” Statistics & Decisions 24 (3): 351–71.