Technical Summary: Targeted Highly Adaptive Lasso Minimum Loss Estimation of Target Functions

1. Problem Statement

The paper addresses the statistical challenge of estimating non-pathwise differentiable functional parameters, specifically focusing on the confounder-adjusted Dose-Response Curve (DRC) for continuous exposures. The DRC is defined as $\psi_0(a) = E_0[Y(a)]$, representing the expected outcome under a specific treatment dose $a$, adjusted for baseline covariates $W$.

Standard Targeted Minimum Loss Estimation (TMLE) is a robust framework for pathwise differentiable parameters (e.g., average treatment effects). However, it does not directly extend to non-pathwise differentiable targets like the full DRC function. Existing approaches, such as the plug-in Highly Adaptive Lasso (HAL) Maximum Likelihood Estimator (MLE), suffer from two primary limitations:

1. Inefficient Bias Reduction: The plug-in HAL-MLE targets the high-dimensional conditional outcome regression $\bar{Q}_0(A, W)$. To reduce bias for the lower-dimensional target $\psi_0$, it requires global undersmoothing of $\bar{Q}_0$, which is inefficient and can degrade the estimation of the nuisance parameter.
2. Smoothness Constraints: The convergence rate of the plug-in estimator is determined by the smoothness of the high-dimensional outcome regression $\bar{Q}_0$. If $\bar{Q}_0$ is rough (e.g., discontinuous in $W$) but the target DRC $\psi_0$ is smooth, the plug-in estimator fails to achieve the optimal minimax rate for the target function.

2. Methodology

The authors propose the Targeted Highly Adaptive Lasso Minimum Loss Estimator (T-HAL-MLE). This approach constructs an estimator that achieves optimal convergence rates for the target function by decoupling the smoothness requirements of the nuisance parameter from those of the target.

Core Strategy

The methodology proceeds in three conceptual stages:

1. Sieve Construction and Projection:
   Instead of estimating the full infinite-dimensional $\bar{Q}_0$ and marginalizing, the method introduces a sieve (a finite-dimensional working model) for the target function $\psi_0$. This working model, denoted $D^{(k)}(R_\lambda)$, is spanned by $k$-th order spline basis functions. The true target $\psi_0$ is projected onto this working model to define a pathwise differentiable approximation $\Psi_\lambda(P)$.

   • The projection is defined via an $L_2$-minimization (weighted by a density $\omega(a)$) of the target function onto the spline basis.
   • This projection parameter $\alpha_\lambda(P)$ is pathwise differentiable, allowing standard TMLE theory to apply.

2. Data-Adaptive Sieve via LASSO:
   Constructing a fixed sieve manually is difficult for high-dimensional targets. The authors propose a Targeted HAL-MLE that avoids explicit sieve construction:

   • Start with a large initial working model (rich set of basis functions) derived from the HAL framework.
   • Perform an initial fit $\bar{Q}_n$ using HAL-MLE.
   • Execute a targeting step using LASSO (Least Absolute Shrinkage and Selection Operator) rather than standard Maximum Likelihood Estimation (MLE). The LASSO minimizes the empirical risk over the fluctuation parameter $\epsilon$ with an $L_1$-norm constraint.
   • This LASSO step data-adaptively selects the relevant basis functions for the target function, effectively constructing a sparse working model $D^{(k)}(R_{\lambda,n})$ without requiring the user to pre-specify the sieve dimension or growth rate.

3. Theoretical Guarantees:
   The paper proves that the resulting estimator targets the projection $\Psi_\lambda(P_0)$ and, under appropriate undersmoothing (choosing a sieve dimension $n(\lambda)$ that grows with sample size $n$), achieves asymptotic normality with respect to the true target $\psi_0$.

   • Convergence Rate: The estimator achieves a rate of $O_+(n^{-k^*/(2k^*+1)})$, where $k^* = k+1$ and $k$ is the smoothness of the target function $\psi_0$, not the nuisance $\bar{Q}_0$. This is a dimension-free rate (up to log factors).
   • Asymptotic Normality: The estimator is pointwise asymptotically normally distributed, enabling valid Wald-type inference and simultaneous confidence bands.

3. Key Contributions

• Novel Estimator for Non-Pathwise Differentiable Targets: The paper provides the first formal construction of a TMLE for non-pathwise differentiable functional parameters (like the DRC) that achieves minimax optimal rates determined solely by the target's smoothness.
• Data-Adaptive Sieve via LASSO: It introduces a method to avoid manual sieve specification by using a LASSO step within the TMLE framework. This allows the method to adaptively select the working model complexity, balancing bias and variance without requiring the user to prescribe the sieve dimension.
• Theoretical Proof of Dimension-Free Rates: The authors prove that the T-HAL-MLE achieves convergence rates dependent only on the smoothness of the target function $\psi_0$, even if the underlying outcome regression $\bar{Q}_0$ is less smooth or higher dimensional. This overcomes a key limitation of plug-in HAL estimators.
• Inference Validity: The paper establishes pointwise asymptotic normality and provides a framework for constructing simultaneous confidence bands for the entire function.

4. Simulation Results

The authors conducted a Monte Carlo simulation study comparing the T-HAL-MLE against a plug-in HAL-MLE and a T-HAL-MLE (Plugin) variant (which applies the targeting step but skips the final projection onto the selected basis).

• Performance Metrics: The study evaluated Root Mean Squared Error (RMSE), bias, and confidence interval coverage across various data-generating processes (DGPs), including those with discontinuities and high-frequency oscillations.
• Bias Reduction: The T-HAL-MLE consistently demonstrated lower bias and RMSE compared to the plug-in HAL-MLE. The targeting step successfully reduced the bias inherent in the plug-in approach.
• Confidence Intervals:
  • Under uniform treatment distributions, the T-HAL-MLE achieved near-nominal 95% coverage.
  • Under normal treatment distributions (which create sparsity in the tails), the T-HAL-MLE showed some undercoverage in challenging settings (e.g., multiple discontinuities), primarily due to residual bias in sparse regions. However, it still outperformed the plug-in HAL-MLE, which suffered from wider intervals and poor coverage.
• Variance: The T-HAL-MLE produced narrower confidence intervals than the plug-in HAL-MLE because its variance estimator reflects the lower-dimensional working model rather than the full complexity of the fitted outcome model.

5. Significance and Claims

The paper claims that the T-HAL-MLE offers a fully data-adaptive approach to inference on functional parameters that does not require parametric assumptions or manual sieve specification.

• Efficiency: By targeting the projection of the target function rather than the full outcome regression, the method avoids the inefficiency of global undersmoothing. It achieves optimal rates even when the target function is significantly smoother than the underlying data-generating mechanism.
• Robustness: The use of LASSO in the targeting step ensures that the sectional variation norm of the estimator is controlled, preserving the conditions necessary for asymptotic normality while selecting the most relevant basis functions.
• Limitations: The authors acknowledge that the method may be more sensitive to inverse probability weighting (due to the $1/g(A|W)$ term in the efficient influence curve) compared to plug-in estimators, potentially affecting finite-sample robustness in regions of positivity violation. They suggest that A-TMLE or targeted bootstrap methods could be future extensions to address this.

In summary, the paper establishes a rigorous theoretical framework and practical algorithm for estimating complex, non-pathwise differentiable causal functions, demonstrating that targeted estimation can overcome the limitations of standard plug-in nonparametric methods.