📊 statistics

Joint Model for Mediation Analysis with Causally Related Longitudinal and Recurrent Event Mediators for Survival Outcome

This paper proposes a novel causal mediation analysis framework that extends joint modeling with shared random effects to quantify natural direct and indirect effects when causally related longitudinal and recurrent event mediators influence a survival outcome, thereby relaxing sequential ignorability assumptions and accounting for unmeasured confounders.

Original authors: Fang Niu, Cheng Zheng, Lei Liu

Published 2026-07-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fang Niu, Cheng Zheng, Lei Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: why did a patient's health take a turn for the worse? In the world of medical science, this is often a game of "cause and effect." Sometimes, a treatment works directly, like a shield blocking a virus. But often, the path is more winding. A treatment might change a patient's blood chemistry, which then changes how often they get sick, and that is what finally determines if they survive. This is called mediation: finding the hidden middleman that carries the message from the cause to the result.

Now, imagine the patient isn't just a static puzzle piece but a moving target. Their health is a movie, not a photo. They have repeated measures, like a daily diary of their blood sugar or immune cell counts, and recurrent events, like a series of flu shots or hospital visits that happen over and over. The tricky part is that these two things talk to each other. A drop in blood cells might cause more infections, and more infections might drop the blood cells further. For a long time, scientists had to pick just one of these moving parts to study, or they had to pretend they didn't influence each other. But in the real world, everything is connected, and ignoring those connections can lead to the wrong conclusion about what actually saves a life.

This is where a new study steps in, acting like a high-tech detective kit for these complex medical movies. The researchers, working with data from a famous HIV study, built a brand-new mathematical framework to untangle these tangled threads. They wanted to know: When a patient gets a specific treatment, how much of the survival benefit comes directly from the drug, and how much comes indirectly through the drug's ability to lower infection rates or boost immune cell counts?

The team developed a "joint model," which is essentially a super-powered calculator that watches the patient's repeated blood tests and their recurring infections simultaneously. They treated these two factors as a team rather than separate suspects. By using a clever trick involving "shared random effects" (think of them as invisible, unmeasured factors that nudge both the blood tests and the infections in the same direction), they could account for hidden influences that previous methods missed. This allowed them to relax a strict rule that usually forces scientists to assume they know everything about a patient's background, letting them get a clearer picture of the true causal pathways.

When they applied this new method to the real-world data of 467 HIV patients, they found some fascinating twists. First, they looked at two different drugs used when the standard treatment failed. Surprisingly, when they separated the direct effects from the indirect ones, they found that neither drug had a significant direct impact on survival. In previous studies, scientists thought the drugs worked directly, but this new, more detailed model suggested that the drugs' benefits were actually being carried by their effect on the patients' immune cells and infection rates. The "direct" effect was an illusion created by not looking at the middlemen closely enough.

However, the story changed when they looked at the patients' history. They found that having a prior AIDS-defining condition (a serious illness that happened before the study started) was a major red flag. This history had a strong, direct negative effect on survival. But it also worked through the middlemen: it lowered the patients' immune cell counts and increased their risk of recurring infections. Interestingly, these two middlemen pulled in opposite directions regarding how they explained the risk. While both were important, the recurring infections played a slightly larger role in explaining why these patients were at higher risk of dying.

The researchers also ran computer simulations to test their new tool. They created fake data where they knew the answers and checked if their model could find them. The results showed that their method was generally accurate and robust, though they noted that if the hidden "random effects" didn't behave exactly as they assumed, the results could get a bit wobbly at later time points. This suggests that while their new detective kit is powerful, it still needs to be used carefully and checked against the data.

Ultimately, this paper doesn't just give a number; it gives a new way of seeing. It shows that in the complex drama of chronic disease, the "direct" cause is often just the tip of the iceberg. By acknowledging that our health markers and our recurring illnesses are in a constant, causal dance with each other, we can finally stop guessing which step leads to the next and start understanding the full choreography of survival.

Technical Summary: Joint Model for Mediation Analysis with Causally Related Longitudinal and Recurrent Event Mediators for Survival Outcome

Problem Statement
Longitudinal clinical studies frequently generate two distinct types of time-dependent data: repeated measures (e.g., biomarker trajectories like CD4 counts) and recurrent events (e.g., hospitalizations or opportunistic infections). While joint models exist to analyze the correlation between these processes and a terminal event (survival), current methodologies lack robust frameworks for causal mediation analysis when multiple, causally related mediators are present. Specifically, existing approaches often assume mediator independence or fail to account for unmeasured confounding between mediators and outcomes. This gap limits the ability to disentangle how an exposure (e.g., treatment or baseline condition) affects survival through distinct pathways involving both recurrent events and longitudinal biomarkers, particularly when one mediator causally influences the other.

Methodology
The authors propose a novel causal mediation framework within a joint modeling structure that simultaneously handles:

A recurrent event mediator ( $M$ ).
A repeated measure mediator ( $W$ ).
A terminal event outcome ( $T$ ).

The methodology is grounded in the potential outcomes framework and extends the "sequential ignorability" (SI) assumption to accommodate unmeasured confounding.

Joint Model Structure: The authors define three linked sub-models:
- Recurrent Event Model: An intensity function for $M$ that depends on time, covariates, treatment, the repeated measure $W(t)$ , and shared random effects.
- Repeated Measure Model: A linear mixed-effects model for $W(t)$ incorporating treatment, covariates, time, and shared random effects.
- Survival Model: A Cox-type hazard function for the terminal event $T$ , dependent on treatment, covariates, the history of $M$ , the current value of $W$ , and shared random effects.
Relaxation of Sequential Ignorability: To address unmeasured time-independent confounders, the model incorporates shared random effects (frailties):
- $\nu_i$ : A shared frailty between the recurrent event mediator ( $M$ ) and the survival outcome ( $T$ ).
- $u_i$ : A shared frailty among the repeated measure mediator ( $W$ ), the recurrent event mediator ( $M$ ), and the survival outcome ( $T$ ).
- This structure relaxes the strict SI assumption by conditioning on these latent variables, allowing for unmeasured confounding between mediators and outcomes.
Causal Definitions: The framework defines Natural Direct Effects (NDE) and Natural Indirect Effects (NIE) for specific pathways:
- NDE: The direct effect of exposure on survival, holding mediators constant.
- $NI_{EM}$ : The indirect effect mediated through the recurrent event process $M$ (including pathways where $W$ affects $M$ ).
- $NI_{EW}$ : The indirect effect mediated solely through the repeated measure $W$ .
Estimation: Since closed-form solutions for the integrals involving infinite-dimensional mediator processes are unavailable, the authors employ Monte Carlo numerical integration combined with maximum likelihood estimation (using Gaussian quadrature) to estimate parameters and causal effects. Confidence intervals are derived via bootstrapping.

Key Contributions

Novel Framework: This is the first method to jointly model recurrent events and repeated measures as causally dependent mediators in a survival analysis setting, explicitly quantifying natural direct and indirect effects.
Handling Unmeasured Confounding: By integrating shared random effects (frailties) into the mediation analysis, the method relaxes the sequential ignorability assumption, offering a more realistic approach to causal inference in observational longitudinal data.
Decomposition of Effects: The approach allows for the separation of total effects into direct effects and distinct indirect pathways (via recurrent events vs. via longitudinal biomarkers), even when these mediators influence each other.
Computational Implementation: The paper provides a rigorous numerical approach (Monte Carlo integration) to estimate these complex causal quantities where analytical solutions are infeasible.

Results

Simulation Studies: The proposed estimators for NDE, $NI_{EM}$ , and $NI_{EW}$ demonstrated small bias and correct coverage rates under correctly specified models. However, the authors note that the estimators are sensitive to misspecification of the random effect distributions (e.g., assuming normality when the true distribution is log-gamma), particularly at later time points.
Application to CPCRA Study: The method was applied to the Terry Beirn Community Programs for Clinical Research on AIDS (CPCRA) data involving 467 HIV-infected patients.
- Treatment Comparison (ddI vs. ddC): The analysis found no significant direct or indirect effects of the treatment (ddI vs. ddC) on survival. The authors note that previous studies reporting a significant direct effect likely conflated the direct effect with the indirect effect mediated through CD4 counts; the new model successfully isolates these components.
- Prior AIDS-Defining Conditions (PADC): PADC showed a significant direct effect on survival. Both recurrent opportunistic infections (OIs) and repeated CD4 counts mediated the effect of PADC on survival, but in opposite directions. The indirect effect via OIs was larger in magnitude than that via CD4 counts, suggesting OIs play a more dominant mediating role.
- Unmeasured Confounding: The significant shared random effect ( $u$ ) between CD4 counts and survival confirmed the presence of unmeasured confounding between these variables.
- Consistency: The sum of the estimated direct and indirect effects closely matched the total effect estimated by a standard Cox regression, validating the decomposition.

Significance and Claims
The paper claims that this methodology enables a more comprehensive investigation of causal pathways in longitudinal studies with multiple mediators. By distinguishing between direct effects and specific indirect pathways (recurrent events vs. repeated measures), the framework provides deeper insights into treatment mechanisms and disease progression. The authors emphasize that this approach is particularly valuable for informing clinical decision-making, such as identifying that controlling recurrent opportunistic infections may be more critical for survival than managing CD4 trajectories alone in patients with prior AIDS-defining conditions. The study concludes that while the method is robust, careful evaluation of model fit (specifically random effect distributions) is necessary in real-world applications.

More like this