Splitting patients into “high dose” and “low dose” groups is a common shortcut, and a costly one. Chopping up a continuous measurement this way can bend the causal conclusions drawn from it. A paper published June 24 in PLoS Computational Biology offers a way around the shortcut while still coping with datasets that carry a huge number of variables.
The authors, Y. Zhou, K. Wei, Y. Liu and colleagues, built the method for a job called causal mediation analysis. In plain terms, it asks how a treatment reaches its outcome. The total effect gets split into a direct effect and an indirect effect that runs through a go-between variable, the mediator. Think of it like tracing a message through a chain of relays to find which one actually passed it along.
Why continuous data and big datasets are hard
Most existing work on high-dimensional causal inference looks at total effects, or at mediation with yes-or-no variables. Only one earlier study, by Ye and colleagues, had brought covariate selection into mediation analysis, and it dealt with binary treatments and mediators. Real biomedical measurements rarely come in two flavors.
Observational data adds a second problem. Without randomization, confounding runs high and estimates come out biased. Researchers lean on propensity scores to correct for it. For a continuous treatment, that means the generalized propensity score, which describes how likely each treatment level is given a patient’s other characteristics. The result is sensitive to which covariates go into the model. Leave out a confounder and bias creeps in. Pile in variables that only predict the treatment and the variance inflates.
When the candidate list runs to thousands of entries, drawn from healthcare records or omics data, nobody can pick the right ones by hand.
What the method does
Zhou’s group combined two existing tools. One is the generalized outcome-adaptive LASSO, known as GOAL, which Gao and colleagues introduced for continuous treatments. LASSO methods shrink the coefficients of unimportant predictors to zero, and the outcome-adaptive version uses information about the outcome so that it keeps both confounders and variables that predict the outcome. The other tool is a weighted semiparametric estimator built on generalized propensity score weighting, proposed by Huber and colleagues.
Together they estimate natural direct and indirect effects when both treatment and mediator are continuous and the covariate pool is large. In simulations, the authors report, the approach beat other regularization-based methods on selection accuracy and estimation efficiency in most settings. It kept variables tied to the outcome and discarded noise covariates, including instrumental and spurious ones.
The paper also sets the method against machine-learning alternatives, such as a double machine learning approach from Yang and colleagues. Those tools handle nonlinear relationships and tangled confounding, but the authors say they may not actively pick out the important variables. That can cost efficiency and make the covariate structure harder to interpret.
A trial run on UK Biobank data
For a real-world test, the team turned to UK Biobank records. They examined the role of apolipoprotein B levels in the link between potential diabetes risk and cancer incidence, using a long list of candidate covariates.
The R code for the method, and for reproducing the simulations, is public on GitHub. The UK Biobank data cannot be released because of participant privacy, though other researchers can apply for access. The work was received Feb. 7, 2025 and accepted June 12, 2026. Funding came from the National Natural Science Foundation of China, Shanghai talent and science programs and a national project on noncommunicable chronic diseases. The funders played no part in the study, and the authors declared no competing interests.
