Background and Motivation

Parkinson’s disease currently has no cure, and its global prevalence is rising faster than any other neurological disorder. This makes the search for preventive strategies more urgent than ever. Observational studies have long suggested that certain habits—like regular exercise and drinking coffee— are linked to a lower risk of developing Parkinson’s. This raises a hopeful possibility: could evidence-based lifestyle recommendations help people reduce their risk?

But there’s a catch. These are only correlations. We still don’t know whether these habits actually cause the reduction in risk, or whether healthier people are simply more likely to live in certain ways. If we want to give meaningful advice, we need to move from association to understanding the actual causal effects of the lifestyle factors we want to modify.

A Quick Dive into Causal Inference

To answer questions of the kind “What is the causal effect of an action on an outcome?”, we need causal inference. The gold standard is a Randomized Controlled Trial (RCT). Imagine randomly assigning thousands of people to either exercise daily and drink coffee, or to remain sedentary and avoid caffeine, and then see who develops Parkinson’s. Because the groups are balanced from the start, any difference in Parkinson’s incidence could be confidently attributed to the exercise itself.

For lifestyle research, this ideal is almost always impossible. It’s unethical to force people to smoke or be sedentary for years. It’s prohibitively expensive and logistically nightmarish to control the daily habits of thousands of people over decades. Consequently, we must rely on observational data—information collected as people live their lives.

This is the core challenge: without randomization, an observed link can be misleading. The observed link between exercise and lower Parkinson's risk is just one of infinitely many possible explanations. Is exercise protective, or are inherently healthier people simply more able to be active? Or could there be another hidden factor influencing both? This creates an identifiability problem: statistical evidence alone cannot distinguish between these scenarios.

To move forward, we therefore need to introduce assumptions that restrict the possible explanations to a set where the causal effect becomes identifiable. These include:

• No Unmeasured Confounding: We have measured all common factors that influence both a person’s lifestyle and their disease risk.
• Positivity: Everyone in our study had a realistic chance of following different lifestyle patterns.
• Consistency: The lifestyle we measure corresponds to a clear, well-defined “treatment.”

These assumptions are strong and often untestable from the data alone. Therefore, we must validate whether these assumptions are plausible within the specific domain [1].
 

The Time-to-Event Outcome

Successfully estimating these causal assumptions would still leave us with a major, practical complication: time. Parkinson’s doesn’t occur spontaneously. It unfolds over years or decades, until it is eventually diagnosed for a patient. This means that we cannot simplify our problem to a classification task with binary “yes/no” labels. Doing so would waste relevant information and introduce bias. We would be ignoring when the disease occurs, losing the ability to model risk trajectories. More problematically, we would have to incorrectly label participants who are censored (lost to follow-up while still healthy) as “no disease,” even though they might have developed it soon after [2].

Analyzing “time-to-event” data—where we track not just if someone gets sick, but when—is therefore essential. However, this approach introduces a new set of challenges that standard models often ignore.

Right-Censoring: The Incomplete Story

As illustrated in Figure 1, in any long-term study, people move away, withdraw, or the study ends before they develop the disease. We only know they were Parkinson’s-free up to their last check-up. If individuals with unhealthy lifestyles are more likely to drop out, the remaining “at-risk” group becomes artificially healthier, skewing our results.

Competing Risks: The Rival Events

People can die from other causes (like heart disease) before a Parkinson’s diagnosis. This is a competing risk. If a habit like heavy drinking increases overall mortality, it might falsely appear to lower Parkinson’s risk simply because it removes people from observation. Ignoring this competition misrepresents reality.

Event-Induced Selection: The Shifting Baseline

Even in a perfect study, the population “at risk” changes over time. People with a higher inherent risk tend to develop the disease earlier. As years pass, the remaining disease-free group is naturally enriched with lower-risk individuals. This systematic shift can distort the estimated effect of a lifestyle factor at different stages of life.

These are not mere statistical nuisances; they are central to interpreting real-world evidence. For example, the apparent benefit of physical activity might look strong at 5 years but fade by 15 years due to these dynamics. Traditional models often make simplistic assumptions about these processes.

Combining Causal Inference and Time-to-event Analysis

Faced with the combined challenges of confounding, time, and competing risks, we need a methodological strategy that connects established approaches in novel ways. No single existing model handles all these elements, so the solution lies in a combination of complementary frameworks.

Our approach builds upon HACSurv [3], a model specifically designed for time-to-event data. Its strength lies in using a statistical construct called a copula to explicitly model informative censoring: the scenario where a patient's reason for dropping out of a study is correlated with their underlying disease risk. Furthermore, HACSurv naturally accommodates competing risks, like death from other causes, allowing us to separate the risk of developing Parkinson's from other mortal events.

However, HACSurv is a predictive model. To answer causal questions—“Does changing lifestyle change risk?”—we must also adjust for confounding. This is where we integrate a technique called Orthogonal Regularization [4]. The core idea is to train the survival model to make predictions that are statistically independent of the part of a person's lifestyle that cannot be explained by their observed health profile.

In practice, we simultaneously train a small auxiliary network to predict a person's lifestyle (e.g., their physical activity level) from their covariates. The main survival model is then penalized if its predictions are correlated with the residual, i.e. the portion of lifestyle the auxiliary network failed to predict. This encourages the model to base its survival estimates solely on the variation in lifestyle that is independent of known confounders, edging us closer to a causal interpretation.

The result is an integrated model that simultaneously accounts for observational survival data and adjusts for confounding, providing a more robust foundation for estimating how lifestyle changes might influence Parkinson's disease risk over time.

What We Found

Among individual lifestyle factors, physical activity stands out with the strongest protective effect against Parkinson’s disease (see Figure 2). Alcohol consumption presents an interesting case: while reducing intake shows a high theoretical effect on PD risk, the achievable benefit in practice is limited, likely because most people in our cohort already drink moderately. Dietary factors demonstrate more modest effects overall.

When we optimize across multiple lifestyle factors simultaneously, a clear trade-off emerges between reducing Parkinson’s risk and reducing mortality risk, as shown in Figure 3. Aggressively optimizing for PD prevention can actually increase mortality risk, particularly for physical activity and diet interventions. These findings align with previous epidemiological studies suggesting opposing effects of certain lifestyle factors on neurodegenerative versus cardiovascular outcomes. However, we found a balanced approach that achieves near-maximal Parkinson’s risk reduction while maintaining neutral or even positive effects on mortality, demonstrating that careful, multi-objective optimization can avoid unintended trade-offs. 

Citations

[1] Feuerriegel, S., Frauen, D., Melnychuk, V., Schweisthal, J., Hess, K., Curth, A., Bauer, S., Kilbertus, N., Kohane, I., & Schaar, M. (2024). Causal machine learning for predicting treatment outcomes. Nature Medicine, 30(4), 958–968.

[2] Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American statistical association, 53(282), 457-481.

[3] Xin Liu, Weijia Zhang, and Min-Ling Zhang. HACSurv: A Hierarchical Copula-Based Approach for Survival Analysis with Dependent Competing Risks. 2025. arXiv: 2410.15180 [stat.ML]. https://arxiv.org/abs/2410.15180

[4] Tobias Hatt and Stefan Feuerriegel. “Estimating average treatment effects via orthogonal regularization”. In: Proceedings of the 30th ACM international conference on information & knowledge management. 2021, pp. 680–689.