📊 statistics

Natural direct effects of vaccines and post-vaccination behaviour

This paper argues that natural direct effects are appropriate causal estimands for quantifying vaccine efficacy while accounting for behavioral changes post-vaccination, though it highlights practical challenges in their estimation due to confounding factors like healthcare-seeking behavior and emphasizes the need for better behavioral data collection.

Original authors: Bronner P. Gonçalves, Piero L. Olliaro, Sheena G. Sullivan, Benjamin J. Cowling

Published 2026-02-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bronner P. Gonçalves, Piero L. Olliaro, Sheena G. Sullivan, Benjamin J. Cowling

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

The Big Idea: The "Superpower" Trap

Imagine you get a superhero shield (a vaccine) that protects you from a specific monster (a virus). Because you know you have this shield, you might start acting a little more recklessly. Maybe you stop wearing your helmet, you hang out in crowded monster dens, or you stop checking your six.

This paper is about a phenomenon called Risk Compensation. It asks a tricky question: If we measure how well the vaccine works in the real world, are we measuring the shield's actual power, or are we measuring the shield's power PLUS the fact that you started acting like a daredevil because you felt safe?

The authors argue that to truly understand a vaccine, we need to separate these two things:

The Biological Shield: How well the vaccine stops the virus inside your body.
The Behavioral Change: How much your behavior changes because you know you are vaccinated.

The Two Types of "Effectiveness"

The paper distinguishes between two ways of looking at vaccine success:

1. The "Real-World" Score (Total Effect)
This is what happens in real life. You get vaccinated, you feel safe, you go to a big party without a mask, and you get sick anyway.

The Analogy: Imagine a car with great brakes (the vaccine). But because the driver feels safe, they start driving 100 mph. If they crash, the "Total Effect" of the car's safety is low, even though the brakes were perfect. The crash happened because of the combination of the brakes and the reckless driving.

2. The "Pure Shield" Score (Natural Direct Effect)
This is what the authors want to calculate. They want to know: "If we could magically freeze your behavior so you acted exactly the same way as an unvaccinated person, how much better off would you be?"

The Analogy: Imagine a time-traveling scientist who gives you the brakes but also puts you in a simulator where you are forced to drive at 30 mph, no matter what. If you don't crash in the simulator, we know the brakes are 100% effective. This is the Natural Direct Effect. It isolates the vaccine's power from your risky choices.

Why Is This Hard to Measure? (The "Doctor Visit" Problem)

The paper points out a major headache in studying this: Healthcare Seeking Behavior.

The Scenario: People who are health-conscious are more likely to get vaccinated. They are also more likely to go to the doctor if they feel a tiny cough.
The Confusion: If vaccinated people go to the doctor more often, they get diagnosed with more cases of the virus (because they are looking for it). Unvaccinated people might stay home and not get tested.
The Result: It might look like the vaccine isn't working because vaccinated people are getting "diagnosed" more often, even if they are actually getting sick less often. The authors call this "confounding by healthcare seeking." It's like judging a detective's skill by how many crimes they solve, but forgetting that the detective is the only one actually looking for crimes.

How Do We Solve This?

The authors suggest a few ways to untangle this knot:

Ask the Right Questions: We need to collect data on what people are doing (social contact diaries, mask-wearing logs), not just whether they are sick.
The "Other Virus" Trick: If we don't have data on behavior, we can look at other diseases. If vaccinated people start getting more of a different virus (like the flu or a cold) that isn't targeted by the vaccine, it's a strong clue that they are behaving riskier (e.g., hanging out more). If the vaccine only protects against Virus A, but people are getting sick with Virus B more often, it's likely because they stopped wearing masks.
Blinding vs. Reality: In a clinical trial, people don't know if they got the real vaccine or a fake one (placebo). So, they don't change their behavior. The trial measures the "Pure Shield." But in the real world, everyone knows they are vaccinated, so they change their behavior. The paper argues we need to understand the gap between the trial results and the real-world results.

The Takeaway

The authors aren't saying vaccines are bad or that risk compensation is happening everywhere. They are saying that science needs to be smarter about how it measures success.

If we only look at the "Real-World Score," we might underestimate how powerful a vaccine is because we are blaming the vaccine for people's risky behavior. By calculating the "Natural Direct Effect," we can tell policymakers:

"The vaccine is actually 90% effective biologically."
"But in the real world, it only looks like 70% effective because people are taking more risks."

This helps us design better public health messages. Instead of just saying "Get vaccinated," we might need to say, "Get vaccinated, but remember: the shield doesn't mean you can stop wearing a seatbelt!"

Summary in One Sentence

This paper argues that to truly understand how good a vaccine is, we must mathematically separate the biological protection it provides from the reckless behavior people might adopt because they feel protected, ensuring we don't blame the vaccine for the risks people take.

1. Problem Statement

Vaccination is a preventive intervention, but its real-world effectiveness may be compromised by risk compensation. This phenomenon occurs when vaccinated individuals, aware of their protection, alter their behavior (e.g., increasing social contacts or reducing mask usage) to engage in riskier activities.

The Gap: Standard vaccine effectiveness (VE) studies typically estimate the total effect ( $\tau_{rw}$ ), which conflates the biological protection of the vaccine with the behavioral changes it induces.
The Challenge: Blinded randomized trials often fail to capture this behavioral component because participants do not know their vaccination status. Conversely, observational studies in the real world capture the total effect but cannot easily disentangle the biological efficacy from the behavioral "leakage."
The Objective: The authors propose using Natural Direct Effects (NDE) as a causal estimand to quantify the effect of vaccination on disease outcomes while "blocking" the pathway through which vaccination influences behavior. This isolates the immunological protection from the behavioral risk compensation.

2. Methodology and Causal Framework

The paper utilizes Causal Inference and the Potential Outcomes Framework to define and identify these effects.

A. Causal Diagrams and Notation

Variables:
- $A$ : Vaccination status (1 = vaccinated, 0 = unvaccinated).
- $Y$ : Clinical outcome (disease).
- $B$ : Infection-related behavior (mediator, e.g., number of social contacts).
- $L$ : Confounders (e.g., age, socioeconomic status).
- $H$ : Healthcare seeking behavior (a specific confounder affecting $A$ , $B$ , and $Y$ ).
Structural Assumptions: The authors model scenarios where $A$ influences $B$ , which in turn influences $Y$ . They explicitly account for confounding by healthcare seeking behavior ( $H$ ), which can bias both the exposure-outcome and mediator-outcome relationships.

B. Definition of Estimands

Real-World Total Effect ( $\tau_{rw}$ ):
$\tau_{rw} = E[Y_{a=1} - Y_{a=0}] = E[Y_{a=1, B_{a=1}} - Y_{a=0, B_{a=0}}]$
This represents the observed difference in disease risk, including both biological protection and behavioral changes.
Natural Direct Effect (NDE):
$NDE = E[Y_{a=1, B_{a=0}} - Y_{a=0, B_{a=0}}]$
This represents the effect of vaccination on disease if the vaccination status were changed, but the behavior ( $B$ ) remained at the level it would have been in the absence of vaccination. Conceptually, this "turns off" the arrow from Vaccination to Behavior ( $A \to B$ ) while allowing behavior to vary naturally based on other factors.
Blinded Trial Estimand ( $\tau_{vt}$ ):
The authors contrast NDE with blinded trials, where perception of protection ( $P$ ) is set to uncertainty ( $P=-1$ ). In such trials, the behavioral pathway is blocked by design, making the trial estimate closer to the NDE (or a specific controlled direct effect) rather than the real-world total effect.

C. Identification Assumptions

To non-parametrically identify the NDE in observational studies, the authors outline four standard assumptions (conditional on confounders $L$ ):

No unmeasured confounding of $A \to Y$ : $Y_{a,b} \perp A | L$
No unmeasured confounding of $B \to Y$ : $Y_{a,b} \perp B | A, L$
No unmeasured confounding of $A \to B$ : $B_a \perp A | L$
Cross-world Independence: $Y_{a,b} \perp B_{a^*} | L$ (This is the critical assumption linking potential outcomes across different worlds).

They also discuss Positivity assumptions and the necessity of measuring $H$ (healthcare seeking) and $B$ (behavior) to satisfy these conditions.

D. Handling Complexity

Time-Varying Confounding: The paper addresses scenarios where healthcare seeking behavior ( $H$ ) changes over time and is affected by vaccination ( $A \to H_2 \to B$ ). In these cases, standard NDE identification fails, and the authors suggest sensitivity analyses, bounds, or stochastic intervention approaches.
Interference: The authors acknowledge that in infectious diseases, one person's vaccination affects others (spillover). They extend the NDE definition to include interference, defining individual-level direct effects within subgroups.
Alternative Outcomes: If direct data on behavior ( $B$ ) is missing, the authors propose using an alternative infection outcome ( $R$ , e.g., a non-targeted pathogen) as a proxy to infer vaccine effects on behavior, provided specific causal structures hold.

3. Key Results and Illustrations

Numerical Example: The authors provide a numerical simulation (Table 1) demonstrating that when vaccination induces risky behavior (increasing $B$ $B$ ), the Total Effect ( $\tau_{rw}$ $τ_{r w}$ ) underestimates the true biological protection.
- In their example, the Total Vaccine Effectiveness (VE) was 32.5%, whereas the NDE VE (biological protection alone) was 42.5%.
- This highlights that risk compensation can mask the full potential of a vaccine.
Data Requirements: The paper emphasizes that estimating NDE requires rich data sources, including:
- Behavioral Data: Social contact diaries, mask usage logs (collected post-vaccination).
- Healthcare Seeking Data: Medical registry databases, prior testing history, and screening participation (to adjust for $H$ ).
Comparison with Blinded Trials: The analysis confirms that blinded trials estimate an effect where behavior is not altered by knowledge of vaccination, effectively approximating the NDE (or a related controlled effect), whereas real-world observational studies estimate the Total Effect.

4. Significance and Contributions

Policy Relevance: By distinguishing between the biological efficacy of a vaccine and the behavioral response it triggers, policymakers can better understand the "real-world" ceiling of vaccine impact. If NDE is high but Total Effect is low due to risk compensation, interventions should focus on behavioral messaging (e.g., encouraging continued masking) rather than just vaccine deployment.
Methodological Advancement: The paper bridges the gap between causal mediation analysis and vaccine epidemiology. It formally defines the NDE in the context of infectious diseases, addressing specific confounders like healthcare seeking behavior that are often overlooked in standard VE studies.
Practical Guidance: It offers a roadmap for study design, suggesting that future vaccine studies (both observational and trial-based) should explicitly collect behavioral data to allow for the estimation of NDE.
Nuanced Interpretation: The authors clarify that while NDE is scientifically valuable for isolating immunological effects, it relies on stronger assumptions (like cross-world independence) than total effect estimates. They argue that NDE represents a plausible intervention (preventing risk compensation) more so than Controlled Direct Effects (fixing behavior to a specific level for everyone).

Conclusion

The paper argues that Natural Direct Effects are a crucial, albeit underutilized, causal estimand for vaccine research. By "blocking" the influence of vaccination on behavior, researchers can isolate the intrinsic biological protection of vaccines. This distinction is vital for interpreting discrepancies between blinded trial results and real-world effectiveness, and for designing public health strategies that mitigate risk compensation. The authors conclude that collecting granular data on both behavior and healthcare seeking is essential to move beyond total effect estimates and fully understand the dynamics of vaccination in epidemics.