Abstract
Causal variable selection in time-varying treatment settings is challenging due to evolving confounding effects. Existing methods mainly focus on time-fixed exposures and are not directly applicable to time-varying scenarios. We propose a novel two-step procedure for variable selection when modeling the treatment probability at each time point. We first introduce a novel approach to longitudinal confounder selection using a Longitudinal Outcome Adaptive LASSO (LOAL) that will data-adaptively select covariates with theoretical justification of variance reduction of the estimator of the causal effect. We then propose an Adaptive Fused LASSO that can collapse treatment model parameters over time points with the goal of simplifying the models in order to improve the efficiency of the estimator while minimizing model misspecification bias compared with naive pooled logistic regression models. Our simulation studies highlight the need for and usefulness of the proposed approach in practice. We implemented our method on data from the Nicotine Dependence in Teens study to estimate the effect of the timing of alcohol initiation during adolescence on depressive symptoms in early adulthood.
Introduction
In longitudinal research, determining whether "X causes Y" over extended periods often encounters the "positivity violation." This occurs when data becomes sparse at certain time points—for instance, if nearly all participants in a study are either receiving a treatment or not at a specific year. In such cases, standard statistical models used for causal inference, such as Marginal Structural Models (MSMs) and Inverse Probability Weighting (IPW), become unstable, leading to high variance and unreliable results.
Mireille Schnitzer and colleagues address this challenge by introducing a two-step "sparsening and smoothing" procedure. This method stabilizes treatment models by leveraging outcome information and temporal similarities, making longitudinal causal inference more robust even when data is thin.
Methodology: The Sparsening and Smoothing Pipeline
The proposed method moves away from the traditional choice between overly flexible stratified models and potentially biased pooled models. Instead, it employs a data-driven two-phase approach.
Phase 1: Longitudinal Outcome Adaptive LASSO (LOAL)
Standard variable selection methods like the LASSO typically prioritize variables that predict the treatment. However, adjusting for "instruments"—variables that predict treatment but have no relationship with the outcome—actually increases the variance of the causal estimate without reducing bias.
As shown in the causal diagram above, LOAL prioritizes variables associated with the outcome ($Y$). By focusing on confounders ($C_0, C_1$) and discarding noisy instruments ($I_0, I_1$), the method builds a more efficient treatment model tailored for causal estimation.
Phase 2: Adaptive Fused LASSO
Once variables are selected, the Adaptive Fused LASSO determines how to handle them over time. It examines the coefficients for the same variable at successive time points. If the data suggests these coefficients are similar, the algorithm "fuses" them into a single value. This adaptively smooths the model—pooling where it makes sense and remaining stratified where the effects truly differ.
Results and Applications
The authors validated their approach using both simulated scenarios and real-world data from the Nicotine Dependence in Teens (NDIT) study.
Simulation Studies
In complex simulations ranging from simple 2-time-point models to high-dimensional 5-time-point models with over 30 covariates, LOAL and Fused LOAL consistently demonstrated lower Mean Squared Error (MSE) than "full" models that adjusted for all variables. The method successfully traded a negligible amount of potential bias for a significant reduction in variance.
The NDIT Study
The method was applied to a longitudinal study of 1,294 students in Montreal to estimate how the timing of alcohol initiation during adolescence affects depressive symptoms at age 20.
- The initial treatment model was highly complex, with 135 parameters.
- LOAL reduced this to 37 parameters.
- The Fusion step further collapsed it to just 23 parameters. This 83% reduction in model complexity halved the estimated standard errors compared to standard Longitudinal Targeted Maximum Likelihood Estimation (LTMLE), making the findings significantly more robust.
Discussion and Future Work
The proposed "sparsening and smoothing" procedure provides a principled way to handle the bias-variance trade-off in longitudinal causal inference. By focusing on outcome-relevant variables and adaptively pooling temporal effects, researchers can obtain more stable estimates from sparse data.
While the current implementation relies on parametric logistic regression and uses bootstrap methods for inference, future work could incorporate non-parametric machine learning, such as the "Highly Adaptive Lasso," to handle non-linear relationships. Additionally, while the two-step regularization process is faster than some alternatives, it still requires careful hyperparameter tuning. Valid post-selection inference remains a challenge, though the use of bootstrap methods provides a practical path forward.
Summary updated