vimshift.Rmd

---
title: "Variable Importance Analysis with Stochastic Interventions"
author: "[Nima Hejazi](https://nimahejazi.org), [Jeremy
  Coyle](https://github.com/jeremyrcoyle), and [Mark van der
  Laan](https://vanderlaan-lab.org)"
date: "`r Sys.Date()`"
output: rmarkdown::html_vignette
bibliography: ../inst/REFERENCES.bib
vignette: >
  %\VignetteIndexEntry{Variable Importance Analysis with Stochastic Interventions}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r options, echo=FALSE, eval=FALSE}
options(scipen=999)
```

## Introduction

Stochastic treatment regimes present a relatively simple manner in which to
assess the effects of continuous treatments by way of parameters that examine
the effects induced by the counterfactual shifting of the observed values of a
treatment of interest. Here, we present an implementation of a new algorithm for
computing targeted minimum loss-based estimates of treatment shift parameters
defined based on a shifting function $d(A,W)$. For a technical presentation of
the algorithm, the interested reader is invited to consult @diaz2018stochastic.
For additional background on Targeted Learning and previous work on stochastic
treatment regimes, please consider consulting @vdl2011targeted,
@vdl2018targeted, and @diaz2012population.

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

```{r setup, message=FALSE, warning=FALSE}
library(data.table)
library(sl3)
library(tmle3)
library(tmle3shift)
set.seed(429153)
```

## Data and Notation

Consider $n$ observed units $O_1, \ldots, O_n$, where each random variable $O =
(W, A, Y)$ corresponds to a single observational unit. Let $W$ denote baseline
covariates (e.g., age, sex, education level), $A$ an intervention variable of
interest (e.g., nutritional supplements), and $Y$ an outcome of interest (e.g.,
disease status). Though it need not be the case, let $A$ be continuous-valued,
i.e. $A \in \mathbb{R}$. Let $O_i \sim \mathcal{P} \in \mathcal{M}$, where
$\mathcal{M}$ is the nonparametric statistical model defined as the set of
continuous densities on $O$ with respect to some dominating measure. To
formalize the definition of stochastic interventions and their corresponding
causal effects, we introduce a nonparametric structural equation model (NPSEM),
based on @pearl2000causality, to define how the system changes under posited
interventions:
\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 $O = (W, A, Y)$

Letting $A$ denote a continuous-valued treatment, we assume that the
distribution of $A$ conditional on $W = w$ has support in the interval
$(l(w), u(w))$ -- for convenience, let this support be _a.e._ That is, the
minimum natural value of treatment $A$ for an individual with covariates
$W = w$ is $l(w)$; similarly, the maximum is $u(w)$. Then, a simple stochastic
intervention, based on a shift $\delta$, may be defined
\begin{equation}\label{eqn:shift}
  d(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 $0 \leq \delta \leq u(w)$ is an arbitrary pre-specified value that
defines the degree to which the observed value $A$ is to be shifted, where
possible.

### Simulate Data

```{r sim_data}
# simulate simple data for tmle-shift sketch
n_obs <- 1000 # number of observations
n_w <- 1 # number of baseline covariates
tx_mult <- 2 # multiplier for the effect of W = 1 on the treatment

# baseline covariates -- simple, binary
W <- as.numeric(replicate(n_w, rbinom(n_obs, 1, 0.5)))

# create treatment based on baseline W
A <- as.numeric(rnorm(n_obs, mean = tx_mult * W, sd = 1))

# create outcome as a linear function of A, W + white noise
Y <- A + W + rnorm(n_obs, mean = 0, sd = 0.5)
```

The above composes our observed data structure $O = (W, A, Y)$. To formally
express this fact using the `tlverse` grammar introduced by the [`tmle3`
package](https://github.com/tlverse/tmle3), we create a single data object and
specify the functional relationships between the nodes in the _directed acyclic
graph_ (DAG) via _nonparametric structural equation models_ (NPSEMs), reflected
in the node list that we set up:

```{r data_nodes}
# organize data and nodes for tmle3
data <- data.table(W, A, Y)
node_list <- list(W = "W", A = "A", Y = "Y")
head(data)
```

We now have an observed data structure (`data`) and a specification of the role
that each variable in the data set plays as the nodes in a DAG.

## Methodology

### Defining a grid of counterfactual interventions

In order to specify a _grid_ of shifts $\delta$ to be used in defining a set of
stochastic intervention policies in an _a priori_ manner, let us consider an
arbitrary scalar $\delta$ that defines a counterfactual outcome $\psi_n =
Q_n(d(A, W), W)$, where, for simplicity, let $d(A, W) = A + \delta$. A
simplified expression of the auxiliary covariate for the TML estimator of $\psi$
is $H_n = \frac{g^{\star}(a \mid w)}{g(a \mid w)}$, where $g^{\star}(a \mid w)$
defines the treatment mechanism with the stochastic intervention implemented. To
ascertain whether a given choice of the shift $\delta$ is _admissable_ -- that
is, whether such an intervention may be implemented while avoiding violations of
the positivity assumption -- define a bound $C(\delta) := \frac{g^{\star}(a
\mid w)}{g(a \mid w)} \leq M$, where $g^{\star}(a \mid w)$ is a function of
$\delta$ in part, and $M$ is a potentially user-specified upper bound of
$C(\delta)$. Then, $C(\delta)$ may be interpreted as a measure of the influence
of a given observation providing a way to limit the maximum influence of a given
observation through a choice of the shift $\delta$ and the setting of the bound
$M$.

We formalize and extend the procedure to determine an acceptable set of values
for the shift $\delta$ in the sequel. Specifically, let there be a shift $d(a,
w) = a + \delta$, where the shift $\delta$ is defined
\begin{equation}
  \delta(a, w) =
    \begin{cases}
      \delta, & \delta_{\text{min}}(a,w) \leq \delta \leq
        \delta_{\text{max}}(a,w) \\
      \delta_{\text{max}}(a,w), & \delta \geq \delta_{\text{max}}(a,w) \\
      \delta_{\text{min}}(a,w), & \delta \leq \delta_{\text{min}}(a,w) \\
    \end{cases},
\end{equation}
where $$\delta_{\text{max}}(a, w) = \text{argmax}_{\left\{\delta \geq 0,
\frac{g(a - \delta \mid w)}{g(a \mid w)} \leq M \right\}} \frac{g(a - \delta
\mid w)}{g(a \mid w)}$$ and
$$\delta_{\text{min}}(a, w) = \text{argmin}_{\left\{\delta \leq 0,
\frac{g(a - \delta \mid w)}{g(a \mid w)} \leq M \right\}} \frac{g(a - \delta
\mid w)}{g(a \mid w)}.$$

The above provides a strategy for implementing a shift at the level of a given
observation $(a, w)$, thereby allowing for all observations to be shifted to an
appropriate value -- whether $\delta_{\text{min}}$, $\delta$, or
$\delta_{\text{max}}$. For the purpose of using such a shift in practice, the
present software provides the functions `shift_additive_bounded` and
`shift_additive_bounded_inv`, which define a variation of this shift:
\begin{equation}
  \delta(a, w) =
    \begin{cases}
      \delta, & C(\delta) \leq M \\
      0, \text{otherwise} \\
    \end{cases},
\end{equation}
which corresponds to an intervention in which the natural value of treatment of
a given observational unit is shifted by a value $\delta$ in the case that the
ratio of the intervened density $g^{\star}(a \mid w)$ to the natural density
$g(a \mid w)$ (that is, $C(\delta)$) does not exceed a bound $M$. In the case
that the ratio $C(\delta)$ exceeds the bound $M$, the stochastic intervention
policy does not apply to the given unit and they remain at their natural value
of treatment $a$.

### _Interlude:_ Constructing Optimal Stacked Regressions with `sl3`

To easily incorporate ensemble machine learning into the estimation procedure,
we rely on the facilities provided in the [`sl3` R
package](https://tlverse.org/sl3). For a complete guide on using the `sl3` R
package, consider consulting https://tlverse.org/sl3, or https://tlverse.org for
the [`tlverse` ecosystem](https://github.com/tlverse), of which `sl3` is a core
engine.

Using the framework provided by the [`sl3` package](https://tlverse.org/sl3),
the nuisance parameters of the TML estimator may be fit with ensemble learning,
using the cross-validation framework of the Super Learner algorithm of
@vdl2007super. To estimate the treatment mechanism (often denoted "g" in the
targeted learning literature), we must make use of learning algorithms
specifically suited to conditional density estimation; a list of such learners 
may be extracted from `sl3` by using `sl3_list_learners()`:

```{r sl3_density_lrnrs_search}
sl3_list_learners("density")
```

To proceed, we'll select two of the above learners, `Lrnr_haldensify` for using
the highly adaptive lasso for conditional density estimation, based on an
algorithm given by @diaz2011super, and `Lrnr_density_semiparametric`, an
approach for semiparametric conditional density estimation:

```{r sl3_density_lrnrs}
# learners used for conditional density regression (i.e., propensity score)
haldensify_lrnr <- Lrnr_haldensify$new(
  n_bins = 3, grid_type = "equal_mass",
  lambda_seq = exp(seq(-1, -9, length = 100))
)
hse_lrnr <- Lrnr_density_semiparametric$new(mean_learner = Lrnr_glm$new())
mvd_lrnr <- Lrnr_density_semiparametric$new(mean_learner = Lrnr_glm$new(),
                                            var_learner = Lrnr_mean$new())
sl_lrn_dens <- Lrnr_sl$new(
  learners = list(haldensify_lrnr, hse_lrnr, mvd_lrnr),
  metalearner = Lrnr_solnp_density$new()
)
```

We also required an approach for estimating the outcome regression (often
denoted "Q" in the targeted learning literature). For this, we build a Super
Learner composed of an intercept model, a GLM, and the [xgboost
algorithm](https://xgboost.ai/) for gradient boosting:

```{r sl3_regression_lrnrs}
# learners used for conditional expectation regression (e.g., outcome)
mean_lrnr <- Lrnr_mean$new()
glm_lrnr <- Lrnr_glm$new()
xgb_lrnr <- Lrnr_xgboost$new()
sl_lrn <- Lrnr_sl$new(
  learners = list(mean_lrnr, glm_lrnr, xgb_lrnr),
  metalearner = Lrnr_nnls$new()
)
```

We can make the above explicit with respect to standard notation by bundling
the ensemble learners into a `list` object below.

```{r make_lrnr_list}
# specify outcome and treatment regressions and create learner list
Q_learner <- sl_lrn
g_learner <- sl_lrn_dens
learner_list <- list(Y = Q_learner, A = g_learner)
```

The `learner_list` object above specifies the role that each of the ensemble
learners we've generated is to play in computing initial estimators to be used
in building a TMLE for the parameter of interest here. In particular, it makes
explicit the fact that our `Q_learner` is used in fitting the outcome regression
while our `g_learner` is used in fitting our treatment mechanism regression.

### Initializing `vimshift` through its `tmle3_Spec`

To start, we will initialize a specification for the TMLE of our parameter of
interest (called a `tmle3_Spec` in the `tlverse` nomenclature) simply by calling
`tmle_shift`. We specify the argument `shift_grid = seq(-1, 1, by = 1)`
when initializing the `tmle3_Spec` object to communicate that we're interested
in assessing the mean counterfactual outcome over a grid of shifts `r seq(-1,
1, by = 1)` on the scale of the treatment $A$ (note that the numerical
choice of shift is an arbitrarily chosen set of values for this example).

```{r vim_spec_init}
# what's the grid of shifts we wish to consider?
delta_grid <- seq(-1, 1, 1)

# initialize a tmle specification
tmle_spec <- tmle_vimshift_delta(shift_fxn = shift_additive_bounded,
                                 shift_fxn_inv = shift_additive_bounded_inv,
                                 shift_grid = delta_grid,
                                 max_shifted_ratio = 2)
```

As seen above, the `tmle_vimshift` specification object (like all `tmle3_Spec`
objects) does _not_ store the data for our specific analysis of interest. Later,
we'll see that passing a data object directly to the `tmle3` wrapper function,
alongside the instantiated `tmle_spec`, will serve to construct a `tmle3_Task`
object internally (see the `tmle3` documentation for details).

### Targeted Estimation of Stochastic Interventions Effects

One may walk through the step-by-step procedure for  fitting the TML estimator
of the mean counterfactual outcome under each shift in the grid, using the
machinery exposed by the [`tmle3` R package](https://tlverse.org/tmle3) (see
below); however, the step-by-step procedure is more often not of interest.

```{r fit_tmle_manual, eval=FALSE}
# NOT RUN -- SEE NEXT CODE CHUNK

# define data (from tmle3_Spec base class)
tmle_task <- tmle_spec$make_tmle_task(data, node_list)

# define likelihood (from tmle3_Spec base class)
likelihood_init <- tmle_spec$make_initial_likelihood(tmle_task, learner_list)

# define update method (fluctuation submodel and loss function)
updater <- tmle_spec$make_updater()
likelihood_targeted <- Targeted_Likelihood$new(likelihood_init, updater)

# invoke params specified in spec
tmle_params <- tmle_spec$make_params(tmle_task, likelihood_targeted)
updater$tmle_params <- tmle_params

# fit TML estimator update
tmle_fit <- fit_tmle3(tmle_task, likelihood_targeted, tmle_params, updater)

# extract results from tmle3_Fit object
tmle_fit
```

Instead, one may invoke the `tmle3` wrapper function (a user-facing convenience
utility) to fit the series of TML estimators (one for each parameter defined by
the grid delta) in a single function call:

```{r fit_tmle_auto}
# fit the TML estimator
tmle_fit <- tmle3(tmle_spec, data, node_list, learner_list)
tmle_fit
```

_Remark_: The `print` method of the resultant `tmle_fit` object conveniently
displays the results from computing our TML estimator.


### Inference with Marginal Structural Models

In the directly preceding section, we consider estimating the mean
counterfactual outcome $\psi_n$ under several values of the intervention
$\delta$, taken from the aforementioned $\delta$-grid. We now turn our attention
to an approach for obtaining inference on a single summary measure of these
estimated quantities. In particular, we propose summarizing the estimates
$\psi_n$ through a marginal structural model (MSM), obtaining inference by way
of a hypothesis test on a parameter of this working MSM. For a data structure
$O = (W, A, Y)$, let $\psi_{\delta}(P_0)$ be the mean outcome under a shift
$\delta$ of the treatment, so that we have $\vec{\psi}_{\delta} =
(\psi_{\delta}: \delta)$ with corresponding estimators $\vec{\psi}_{n, \delta}
= (\psi_{n, \delta}: \delta)$. Further, let $\beta(\vec{\psi}_{\delta}) =
\phi((\psi_{\delta}: \delta))$.

For a given MSM $m_{\beta}(\delta)$, we have that
$$\beta_0 = \text{argmin}_{\beta} \sum_{\delta}(\psi_{\delta}(P_0) -
m_{\beta}(\delta))^2 h(\delta),$$
which is the solution to
$$u(\beta, (\psi_{\delta}: \delta)) = \sum_{\delta}h(\delta)
\left(\psi_{\delta}(P_0) - m_{\beta}(\delta) \right) \frac{d}{d\beta}
m_{\beta}(\delta) = 0.$$
This then leads to the following expansion
$$\beta(\vec{\psi}_n) - \beta(\vec{\psi}_0) \approx -\frac{d}{d\beta} u(\beta_0,
\vec{\psi}_0)^{-1} \frac{d}{d\psi} u(\beta_0, \psi_0)(\vec{\psi}_n -
\vec{\psi}_0),$$
where we have
$$\frac{d}{d\beta} u(\beta, \psi) = -\sum_{\delta} h(\delta) \frac{d}{d\beta}
m_{\beta}(\delta)^t \frac{d}{d\beta} m_{\beta}(\delta)
-\sum_{\delta} h(\delta) m_{\beta}(\delta) \frac{d^2}{d\beta^2}
m_{\beta}(\delta),$$
which, in the case of an MSM that is a linear model (since
$\frac{d^2}{d\beta^2} m_{\beta}(\delta) = 0$), reduces simply to
$$\frac{d}{d\beta} u(\beta, \psi) = -\sum_{\delta} h(\delta) \frac{d}{d\beta}
m_{\beta}(\delta)^t \frac{d}{d\beta} m_{\beta}(\delta),$$
and
$$\frac{d}{d\psi}u(\beta, \psi)(\psi_n - \psi_0) = \sum_{\delta} h(\delta)
\frac{d}{d\beta} m_{\beta}(\delta) (\psi_n - \psi_0)(\delta),$$
which we may write in terms of the efficient influence function (EIF) of $\psi$
by using the first order approximation $(\psi_n - \psi_0)(\delta) =
\frac{1}{n}\sum_{i = 1}^n \text{EIF}_{\psi_{\delta}}(O_i)$,
where $\text{EIF}_{\psi_{\delta}}$ is the efficient influence function (EIF) of
$\vec{\psi}$.

Now, say, $\vec{\psi} = (\psi(\delta): \delta)$ is d-dimensional, then we may
write the efficient influence function of the MSM parameter $\beta$ (assuming a
linear MSM) as follows
$$\text{EIF}_{\beta}(O) = \left(\sum_{\delta} h(\delta) \frac{d}{d\beta}
m_{\beta}(\delta) \frac{d}{d\beta} m_{\beta}(\delta)^t \right)^{-1} \cdot
\sum_{\delta} h(\delta) \frac{d}{d\beta} m_{\beta}(\delta)
\text{EIF}_{\psi_{\delta}}(O),$$ where the first term is of dimension
$d \times d$ and the second term is of dimension $d \times 1$.

In an effort to generalize still further, consider the case where
$\psi_{\delta}(P_0) \in (0, 1)$ -- that is, $\psi_{\delta}(P_0)$ corresponds
to the probability of some event of interest. In such a case, it would be more
natural to consider a logistic MSM
$$m_{\beta}(\delta) = \frac{1}{1 + \exp(-f_{\beta}(\delta))},$$
where $f_{\beta}$ is taken to be linear in $\beta$ (e.g.,
$f_{\beta} = \beta_0 + \beta_1 \delta + \ldots$). In such a case, we have the
parameter of interest
$$\beta_0 = \text{argmax}_{\beta} \sum_{\delta} \left(\psi_{\delta}(P_0)
\text{log} m_{\beta}(\delta) + (1 - \psi_{\delta}(P_0))\log(1 -
m_{\beta}(\delta))\right)h(\delta),$$
where $\beta_0$ solves the following
$$
\sum_{\delta} h(\delta) \frac{d}{d\beta} f_{\beta}(\delta) (\psi_{\delta}(P_0)
- m_{\beta}(\delta)) = 0.$$

Inference from a working MSM is rather straightforward. To wit, the limiting
distribution for $m_{\beta}(\delta)$ may be expressed
$$\sqrt{n}(\beta_n - \beta_0) \to N(0, \Sigma),$$
where $\Sigma$ is the empirical covariance matrix of $\text{EIF}_{\beta}(O)$.

#### Directly Targeting the MSM Parameter $\beta$

Note that in the above, a working MSM is fit to the individual TML estimates of
the mean counterfactual outcome under a given value of the shift $\delta$ in the
supplied grid. The parameter of interest $\beta$ of the MSM is asymptotically
linear (and, in fact, a TML estimator) as a consequence of its construction from
individual TML estimators. In smaller samples, it may be prudent to perform a
TML estimation procedure that targets the parameter $\beta$ directly, as opposed
to constructing it from several independently targeted TML estimates. An
approach for constructing such an estimator is proposed in the sequel.

Let $C = \left(\sum_{\delta} h(\delta) \frac{d}{d\beta} m_{\beta}(\delta)
\frac{d}{d\beta} m_{\beta}(\delta)^t \right)$, then
$$\text{EIF}_{\beta}(O) = C^{-1} \cdot \sum_{\delta} h(\delta)
\frac{d}{d\beta} m_{\beta}(\delta)(Y - \overline{Q}(A,W) + C^{-1} \sum_{\delta}
h(\delta) \frac{d}{d\beta} m_{\beta}(\delta) \left(\int \overline{Q}(a,w)
g_{\delta}^0(a \mid w) - \Psi_{\delta}\right).$$

Suppose a simple working MSM $\mathbb{E}Y_{g^0_{\delta}} = \beta_0 + \beta_1
\delta$, then a TML estimator targeting $\beta_0$ and $\beta_1$ may be
constructed as
$$\overline{Q}_{n, \epsilon}(A,W) = \overline{Q}_n(A,W) + \epsilon (H_1(g),
H_2(g),$$ for all $\delta$, where $H_1(g)$ is the auxiliary covariate for
$\beta_0$ and $H_2(g)$ is the auxiliary covariate for $\beta_1$.

To construct a targeted maximum likelihood estimator that directly targets the
parameters of the working marginal structural model, we may use the
`tmle_vimshift_msm` Spec (instead of the `tmle_vimshift_delta` Spec that
appears above):

```{r vim_targeted_msm_fit}
# what's the grid of shifts we wish to consider?
delta_grid <- seq(-1, 1, 1)

# initialize a tmle specification
tmle_msm_spec <- tmle_vimshift_msm(shift_fxn = shift_additive_bounded,
                                   shift_fxn_inv = shift_additive_bounded_inv,
                                   shift_grid = delta_grid,
                                   max_shifted_ratio = 2)

# fit the TML estimator and examine the results
tmle_msm_fit <- tmle3(tmle_msm_spec, data, node_list, learner_list)
tmle_msm_fit
```

## References