Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
summary
The gist
Mendelian Randomization is widely used for causal inference but existing approaches fail to capture dynamic relationships across the life course…
In short
The paper introduces Multivariable Functional Mendelian Randomization (MV-FMR) to study multiple time-varying exposures simultaneously. It extends existing methods by combining functional principal component analysis with a data-driven cross-validation strategy. This new framework allows researchers to model complex, continuous exposure trajectories and uncover how different related exposures influence health outcomes over an individual's lifespan.
Key concepts
- Functional Principal Component Analysis (FPCA)
- FPCA is a technique used to reduce the complexity of time-varying data, like exposure trajectories. It transforms the continuous paths of multiple exposures into a smaller set of principal components that capture the most important underlying patterns. This makes it possible to model complex functional relationships while managing high-dimensional data effectively.
- Multivariable Mendelian Randomization (MV-FMR)
- MV-FMR is an advanced causal inference method designed to handle several related exposures at once. It uses genetic instruments (Mendelian randomization) combined with FPCA to simultaneously estimate the causal effects of multiple time-varying exposures, accounting for how these different factors interact and share common genetic pathways.
- Data-driven Basis Selection
- Instead of manually choosing which principal components are most important, this method uses a cross-validation procedure. This procedure tests different numbers of components by optimizing predictive performance (minimizing error). This ensures the model selects the best functional representation for each exposure trajectory, balancing accuracy and preventing overfitting.
- Disentanglement of Mediation Pathways
- This concept refers to separating the total effect of an exposure into its direct impact and any indirect impacts through other exposures. MV-FMR allows researchers to recover these distinct causal effects, which is vital when one risk factor influences the outcome both immediately and indirectly via a chain of other exposures.
Terminology used across episodes
This episode discusses
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization · Paper Radio
The paper
Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization · Read on arXiv
Nicole Fontana, Francesca Ieva, Luisa Zuccolo, Emanuele Di Angelantonio, Piercesare Secchi
MOX Department of Mathematics Politecnico di Milano Health Data Science Research Centre Human Technopole British Heart Foundation Cardiovascular Epidemiology Unit Victor Phillip Dahdaleh Heart and Lung Research Institute National Institute for Health and Care Research Blood and Transplant Research Unit in Donor Health and Behaviour University of Cambridge
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Unraveling time-varying causal effects of multiple exposures".
Jane: Here's the gist of the paper. Mendelian Randomization is widely used for causal inference but existing approaches fail to capture dynamic relationships across the life course, leading to a new framework called Multivariable Functional Mendelian Randomization (MV-FMR) that simultaneously models multiple time-varying exposures This extension allows researchers to investigate multiple related exposures, mediation pathways, and horizontal pleiotropy while preserving the continuous nature of exposure trajectories Methodological Foundation The paper introduces MV-FMR as an extension of Univariable Functional Mendelian Randomization (U-FMR) to simultaneously model multiple time-varying exposures The core methodology combines functional principal component analysis with a data-driven cross-validation strategy for basis selection and accounts for overlapping instruments and mediation effects The structural model involves representing the exposure trajectories using Functional Principal Component Analysis (FPCA) to reduce dimensionality, where the causal effect is modeled as a smooth function β(t) Estimation Strategy The estimation procedure for MV-FMR involves several key steps to handle the multivariate nature of the problem 1.
Tom: Joint modeling of time-varying exposures that share common genetic pathways is performed 2.
Jane: A data-driven cross-validation procedure is used for selecting the number of basis components for each exposure 3.
Tom: Pseudo-exposures are defined as the FPCA scores ξj,k, and these coefficients β∗j,k are estimated using a continuously updating Generalized Method of Moments (GMM) 4.
Title and authors: Tom: So we're talking about this paper today, "Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization." It sounds pretty complex, but basically, it’s trying to fix a big problem with how we use Mendelian Randomization when exposures change over time.
Jane: Right, so the core idea is that standard MR often looks at one single time point for an exposure, but real life is messy because things change throughout our lives. This paper aims to capture those changes and look at multiple related exposures all at once.
Lu: Exactly. The foundation they’re building on is Univariate Functional Mendelian Randomization, or U-FMR, which models a continuous outcome Y and a time-varying exposure X(t) using genetic instruments G > page three of that work reads: "Inference on time-varying causal effects relies on life-course variation in the associations between the genetic instruments and the exposure <ref:2512.19064#pg1>."
Meng: But if you have multiple exposures, you can’t just run separate analyses for each one, right? You need a way to handle how they might interact or share genetic pathways.
Tom: That's where this Multivariable Functional Mendelian Randomization, or MV-FMR, comes in. It lets them simultaneously model several time-varying exposures and see how they relate to the outcome while keeping track of those evolving effects.
Jane: So what does that actually mean for us? It suggests we can stop thinking about a single snapshot and start seeing the whole trajectory of an exposure, which makes sense when studying things like cardiovascular risk or body mass index over decades.
Lu: The methodology they use is pretty smart. They combine functional principal component analysis with a data-driven cross-validation strategy to pick the best components for each exposure's trajectory, which helps reduce the complexity of the problem.
Meng: I wonder about that selection part—how do you know which components are actually capturing the important functional effects versus just noise? It sounds like a tricky balancing act.
Tom: That’s what they tackle in their improvements section. They use data-driven k-fold cross-validation to optimize predictive performance using mean squared error, trying to balance model complexity against generalizability without overfitting, which is a big win for real data applications.
Title and authors: Jane: And they also address the issue of overlapping instruments and mediation pathways. The paper suggests joint modeling using a stacked functional principal component analysis score matrix with a generalized method of moments to estimate the coefficients, which helps them account for correlation between exposures through those overlapping genetic instruments.
Lu: That joint modeling is key because it lets them see if one exposure's effect on the outcome is being partly channeled through another exposure that they are also measuring. It’s about accounting for that correlation structure using the genetic data provided by the instruments.
Meng: If we’re thinking about deployment, running a model that needs this much joint estimation sounds computationally heavy. What about scaling this up? Does it become impossible to run on a massive dataset?
Tom: That’s a valid operational question, Meng. But they also improved how they handle binary outcomes by using a two-stage residual inclusion control function approach instead of just standard predictor substitution, because it controls for endogeneity by including first-stage residuals as additional regressors.
Jane: And that sounds like a very rigorous way to handle the measurement error in those specific scenarios. It shows they’re not taking shortcuts when dealing with different types of health outcomes.
Lu: Plus, they showed they can consistently recover nonlinear time-varying effects, including quadratic ones, which is important because many biological processes aren't just straight lines; they curve. They demonstrated that the model successfully captured these nonlinear dynamics across a wide range of exposure outcome functional forms > page two of that work reads: "The effect of body mass index on cardiovascular disease risk may differ between early adulthood and older age <ref:2512.19064#pg1>."
Meng: Recovering nonlinearity is one thing, but identifying *when* those nonlinear effects matter is another. If you’re building a predictive tool, you need to know the critical windows.
Tom: They did exactly that by identifying life course critical periods. In their UK Biobank application, they found that the critical exposure window for body mass index and systolic blood pressure in relation to coronary artery disease risk was concentrated in the sixth decade of life.
Jane: That’s a concrete finding because it moves beyond just saying "the effect changes" to pinpointing the specific time frame when that change is most influential on the health outcome. That tells you exactly where clinical intervention might be most effective.
Title and authors: Lu: They also tackled disentangling mediation pathways, which is crucial when one risk factor acts both directly and indirectly through other factors. They managed to recover direct causal effects from mediated effects for both exposures, which is important because one risk factor operates both directly on the outcome and indirectly through downstream exposures > page three of that work reads: "direct causal effects from mediated effects for both exposures <ref:2512.19064#pg1>."
Meng: So, if I were building a system that recommends interventions, knowing the direct versus indirect paths helps me decide what kind of intervention to suggest. It’s about understanding the mechanism, not just the correlation.
Tom: Exactly. This paper is really pushing us toward models that aren't just looking for a single point in time but are modeling how everything evolves together over time, integrating functional data analysis into Mendelian Randomization > page one of that work reads: "Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization" <ref:2512.19064#pg1,Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis>
Jane: So, to wrap up the main points, this paper introduces MV-FMR as a way to handle multiple time-varying exposures simultaneously by using functional principal component analysis for dimensionality reduction and joint modeling with a generalized method of moments to estimate those evolving causal effects.
Lu: The key improvements they've made are the data-driven selection of functional components, joint modeling of exposures accounting for overlapping instruments, handling binary outcomes via residual inclusion control, and successfully recovering nonlinear time-varying effects like quadratic terms > page one of that work reads: "Accounting for these time-varying genetic effects allows researchers to identify critical periods when an exposure has the greatest impact on health outcomes <ref:2512.19064#pg1>."
Meng: From an engineering standpoint, the ability to identify those life course critical periods is valuable because it lets us target interventions precisely in time. But I still have questions about the computational cost of that kind of stacked FPCA matrix estimation.
Tom: That’s a fair question, Meng. The paper doesn't necessarily solve the scaling issue completely, but it shows a path forward by using these structured methods to handle the multivariate nature robustly > page two of that work reads: "Accounting for these time-varying genetic effects allows researchers to identify critical periods when an exposure has the greatest impact on health outcomes <ref:2512.19064#pg1>."
Title and authors: Jane: So what does this whole thing mean for the average person listening today? It means that when doctors look at your risk factors, they might not be looking at one static number, but rather understanding how those numbers move and interact as you get older.
Lu: It shifts the focus from a single measurement to a dynamic trajectory, which is fundamentally more accurate for modeling complex biological systems > page one of that work reads: "robustness to pleiotropy, provided that pleiotropic effects operate only through the modeled exposures <ref:2512.19064#pg1>."
Meng: For me, it means we need better validation on real traffic. If this model works on simulation data but breaks when we feed it actual hospital records, then it doesn't really help anyone deploy something.
Tom: That’s the next step for any complex AI system, Meng. We need to see if these theoretical improvements hold up when they are subjected to the messy reality of production data > page two of that work reads: "robustness to pleiotropy, provided that pleiotropic effects operate only through the modeled exposures <ref:2512.19064#pg1>."
Jane: It's a sophisticated tool, but it’s still a research paper, and it highlights exactly what we need in causal inference: methods that respect the continuous nature of biological reality > page one of that work reads: "To capture these dynamics, analytical methods must move beyond single time-point estimates and instead model how effects evolve over time <ref:2512.19064#pg1>."
Lu: Ultimately, this paper on "Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization" provides a framework for moving past static causal estimates in genetics toward dynamic, multi-exposure insights.
Meng: It's a lot to process, but the potential to get better at understanding disease progression over time is there. I’ll need to see some real-world performance metrics before I can really get excited about deploying this kind of complexity.
Tom: We’ll keep watching the subsequent work on this space, focusing on how these structural models translate into practical risk assessments for people > page three of that work reads: "robustness to pleiotropy, provided that pleiotropic effects operate only through the modeled exposures <ref:2512.19064#pg1>."
Jane: That's where we'll be following along. We’ve got some more papers coming soon that tackle similar challenges in high-dimensional time series modeling.
The paper's summary: Tom: So, basically, this paper is introducing Multivariable Functional Mendelian Randomization, or MV-FMR. It's an extension of what you already know about functional Mendelian Randomization that lets researchers look at several exposures that change over time all at once.
Jane: Right, so the big idea is they're moving past looking at just one snapshot in time for an exposure. They want to capture how things evolve throughout a person's life and see how multiple related things—like different health risks—interact with each other while staying true to those changing paths.
Tom: The authors use functional principal component analysis, or FPCA, to handle the complexity of those trajectories by reducing the number of time-varying exposures down to something more manageable. They treat the causal effect as a smooth function over time.
Lu: And that’s where their methodological foundation comes in—they use a data-driven cross-validation strategy to pick which components from that FPCA analysis are actually capturing meaningful functional effects, trying to balance complexity against overfitting.
Jane: The estimation part is pretty clever too. Instead of just looking at one effect, they jointly model the time-varying exposures using a generalized method of moments and a stacked score matrix. This helps them account for how those different exposures share genetic pathways or instruments, which is something you can't ignore in real data.
Tom: That joint modeling is what lets them see if one risk factor’s effect on your health outcome is being partly caused by another exposure they are also tracking through the genetic instruments. It’s about seeing the whole picture of interactions.
Jane: They also improved how they handle binary outcomes, which are those yes or no results, by using a two-stage residual inclusion control function approach instead of just standard predictor substitution. This is a more robust way to control for endogeneity by including those first-stage residuals as extra variables in the model.
Lu: One big win they show is that they can recover true nonlinear time-varying effects, even quadratic ones. This means if an exposure has a curve, the model doesn't just guess a straight line; it captures that actual shape across different exposure and outcome forms.
Tom: The real payoff for someone listening to this is the ability to identify life course critical periods. They used their UK Biobank data to show they can pinpoint exactly when exposures have their strongest impact on risk—like finding that the window for body mass index and blood pressure risk peaks in your sixth decade of life.
Jane: So, what does this mean practically? It shifts the focus from a single number at one age to understanding how your risks move and interact as you get older, which is much more accurate for modeling complex health systems.
Lu: And they even address disentangling mediation pathways. They can recover direct causal effects from mediated effects for both exposures, which is important when one risk factor affects the outcome both directly and indirectly through other factors.
Tom: It’s a sophisticated tool that moves beyond just static estimates in genetics toward dynamic insights about disease progression over time. But we still need to see if this works reliably when you feed it messy, real-world hospital data at scale.
Jane: That's the next big question for any complex AI system—does this theoretical structure hold up when it hits the actual noise of production data?
The paper's improvements: Tom: So, the paper lays out some specific upgrades to make this MV-FMR model better than just running separate analyses for each exposure one by one. It’s all about making the selection process smarter and handling those tricky data types more reliably.
Jane: Right, so they are talking about ways to improve how the system chooses which functional components are actually important for each exposure's trajectory. They propose using a data-driven k-fold cross-validation procedure that optimizes predictive performance based on mean squared error, trying to find the right balance between complexity and overfitting.
Lu: That's smart because it stops them from just picking the most complicated model possible; they are actively optimizing for how well the model performs on new, unseen data. It’s about capturing meaningful functional effects without just memorizing the existing noise.
Tom: And then there's handling those binary outcomes, which are those yes or no results in health studies. They suggest using a two-stage residual inclusion control function approach instead of just standard predictor substitution because that method controls for endogeneity by including those first-stage residuals as extra variables.
Jane: That sounds like a solid fix for bias in those specific types of data, making the estimates more trustworthy when you are dealing with outcomes that aren't continuous. But what about the curves themselves?
Lu: They also showed that this system consistently recovers true nonlinear effects, including quadratic ones. This means if an exposure has a curve—like an inverted U shape—the model doesn't just guess a straight line; it captures that actual shape across different exposure and outcome forms.
Tom: So, they’re not just looking for linear relationships anymore; they can capture those curved biological processes, which is definitely more realistic for how health factors work.
Jane: And this leads to something really useful: identifying life course critical periods. They showed you can analyze the estimated effect functions over time to figure out exactly when an exposure has its strongest impact on risk, like finding that a specific window for body mass index and blood pressure risk peaks in your sixth decade of life.
Lu: That’s concrete because it tells you precisely where clinical intervention might be most effective, not just that the effect changes somewhere down the road.
Tom: They also tackled disentangling mediation pathways better. They showed how to recover direct causal effects from mediated effects for both exposures, which is crucial when one risk factor acts both directly on the outcome and indirectly through other factors.
Jane: That ability to separate those paths is vital because it helps you figure out if an intervention needs to target a direct cause or work through a secondary pathway.
Lu: It’s about understanding the mechanism, not just the correlation between two things that are both influencing your health.
Tom: So, these improvements show they are serious about building a tool that can actually handle the messy reality of biological data and give you actionable insights into *when* and *how* risks change over time. But we still need to see if this whole setup runs smoothly when you try to deploy it on a massive set of real hospital records.
Conclusion: Tom: So to wrap up, this paper on "Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization" shows how to model several changing risk factors simultaneously using functional data analysis and joint modeling.
Jane: Basically, it gives us a much more complete picture than just looking at one factor at one time point; it lets us see the whole life course trajectory of multiple exposures interacting with each other.
Lu: The key is that they’ve shown how to handle the complexity by using functional principal component analysis to simplify those trajectories and then use joint modeling with generalized method of moments for estimation.
Tom: The main implication is that we can finally pinpoint exactly when an exposure has its biggest impact on health, not just somewhere in the middle of a decade.
Jane: That means better, more targeted advice for people—knowing when to focus on risk management based on their age and the specific exposures involved.
Lu: And they’ve also shown how to separate direct causal effects from indirect ones through mediation pathways, which is crucial for understanding the true mechanism of a disease.
Tom: It’s a strong framework for getting causal answers in complex biological systems, provided you can handle the computational load of that stacked functional principal component analysis matrix.
Jane: It’s definitely more sophisticated than what we usually see in basic Mendelian Randomization, but it pushes the field toward models that respect the continuous nature of biological reality.
Lu: For me, it’s about pushing representation learning beyond simple snapshots to capture true temporal dynamics in genetic data structures.
Meng: From an engineering standpoint, I still have concerns about the cost of running such a model on a massive scale; if this gets deployed, we need to know the operational budget for that kind of computation.
Tom: That’s fair, Meng; scaling these structural models is always a hurdle when you move from research to production deployment.
Jane: So while this paper shows the potential for more nuanced insights into disease progression over time, we still need rigorous validation on real-world health data before it becomes standard clinical practice.
Lu: I'm looking forward to seeing how these structural models integrate with other learning techniques in future work, especially as we try to handle even higher dimensional time series data.
Meng: I’ll be watching the performance metrics closely; if the error scales worse than we expect when you run it against live traffic, then it doesn't really help anyone deploy this kind of complexity.
Lalam: Based on the literature and what I’ve processed, this approach for "Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization" is highly promising because it directly addresses the limitations of static models by incorporating temporal smoothness.
Tom: So, that’s the word from the model itself; it's looking at how things evolve over time instead of just a single point.
Jane: It really does shift our thinking toward dynamic risk assessment rather than static snapshots.
Lu: It opens up new avenues for understanding complex trait genetics where the relationship between exposures is inherently time-dependent.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates