Apparent RSV-COVID interference is not robust to adjustment for shared testing propensity

This study introduces a conservative statistical method that adjusts for shared testing propensity in surveillance data, revealing that apparent RSV-COVID viral interference signals are not robust to this adjustment, though the method's inherent bias toward null results means biological interference cannot be definitively ruled out.

The Gist

The Big Question: Do Viruses Play "Musical Chairs"?

Imagine a crowded room where two different groups of people are dancing: Group A (RSV) and Group B (COVID).
Viral interference is the idea that when Group A gets really big and loud, they might accidentally push Group B out of the room, making Group B smaller. Scientists call this "interference." It's like one virus "interfering" with another, perhaps by waking up the body's immune system so it fights off the second virus too.

But here is the problem: How do we know if they are actually fighting each other, or if they are just reacting to the same thing?

The Problem: The "Umbrella" Effect

Imagine it starts raining. Suddenly, everyone runs inside to get under the same big umbrella.

• If you look at the data, you see a huge spike in people under the umbrella.
• You might think, "Oh, Group A is pushing Group B under the umbrella!"
• But actually, the rain (the weather) made both groups seek shelter at the same time.

In the real world, the "rain" is testing behavior.

• When people feel sick, they go to the doctor.
• Doctors use "multiplex" tests (like a super-scan) that check for RSV, COVID, and Flu all at once from one nose swab.
• When a virus surge happens, everyone gets tested. This creates a fake pattern where the viruses look like they are interacting, even if they are just being detected at the same time because people are getting sick and seeking help.

The Author's Solution: The "Fairness Scale"

The author, Joshua Steier, built a new mathematical tool to fix this "Umbrella Effect."

He used a logic borrowed from vaccine studies called the Test-Negative Design. Think of it like a Fairness Scale:

1. The Logic: If the "rain" (testing) affects everyone equally, the ratio of Group A to Group B under the umbrella should stay roughly the same, even if the total number of people changes.
2. The Tool: Steier added a "penalty" to his math model. If the model predicts that Group A is crushing Group B, but the ratio of their test results doesn't match what we actually see in the real world, the model gets "fined" (a penalty).
3. The Goal: This forces the model to admit: "Wait, maybe they aren't fighting each other. Maybe they are just both reacting to the same testing surge."

The Experiment: What Happened?

Steier tested this tool on five years of US data (RSV and COVID).

• Without the Fairness Scale: The model said, "Hey! There is a strong interference! RSV is definitely suppressing COVID." (The interference score was 0.0082).
• With the Fairness Scale: The model said, "Oh, my bad. Once I account for the fact that people were just getting tested more often, that interference disappears." (The score dropped to 0.0016).

The Result: The apparent interference signal shrank by 80%. When he looked at the statistical "confidence intervals" (the margin of error), the number zero was right in the middle. This means the data cannot prove that the viruses are interfering with each other.

The Catch: The "Over-Cautious Detective"

Here is the most important part of the paper. The author admits his tool is intentionally over-cautious.

Imagine a detective who is so afraid of accusing an innocent person that they only convict if the evidence is 100% undeniable.

• If the detective says "Guilty," you can be sure they did it.
• But if the detective says "Not Guilty," it doesn't mean the person is innocent. It just means the evidence wasn't strong enough for this specific detective.

Steier's tool is that detective.

• High Specificity: It rarely makes a mistake by finding interference when there is none (low false alarms).
• Low Sensitivity: It is very bad at finding interference even if it is there. The math is designed to "absorb" the signal into other parts of the model to be safe.

The Simulation: When Steier created fake data where he knew interference was happening, his tool often missed it. It was too conservative.

The Conclusion: What Does This Mean for Us?

1. Don't panic about the "Zero" result: The fact that the study found "no interference" does not prove that viruses don't interfere with each other. It just proves that the "interference" we saw in the data was likely just an artifact of how and when people got tested.
2. The Signal is Weak: If there is biological interference between RSV and COVID, it is so weak or so messy that this specific method (and likely the current surveillance data) can't see it clearly.
3. A New Tool, Not a Final Answer: This paper isn't saying "The case is closed." It's saying, "If you want to find viral interference, you need better tools than just looking at test counts, because testing behavior creates too much noise."

In short: The paper suggests that the "fighting viruses" story we might have seen in the news is mostly a mirage caused by people rushing to get tested at the same time. But because the author's tool is so strict, we can't be 100% sure the viruses aren't fighting a little bit, too. We just can't see it clearly yet.

Technical Summary

1. Problem Statement

Viral interference refers to the phenomenon where infection by one pathogen reduces susceptibility to another at the population level. While biological mechanisms (e.g., interferon responses) are hypothesized, detecting interference from surveillance data is confounded by time-varying testing behavior.

• The Confounder: During respiratory virus surges, testing intensity increases. Because multiplex panels often test for multiple pathogens simultaneously, a surge in one virus (e.g., RSV) increases the number of specimens tested, thereby increasing the detection probability of other pathogens (e.g., COVID-19) regardless of biological interaction.
• The Challenge: Standard statistical models (e.g., cross-correlation, renewal models) often cannot distinguish between true biological interference and these correlated detection patterns induced by shared testing propensity. Previous studies have yielded heterogeneous results regarding RSV–COVID interference, ranging from strong negative associations to null effects.

2. Methodology

The author developed a two-pathogen renewal model augmented with a ratio penalty designed to adapt the logic of the Test-Negative Design (TND) to aggregate surveillance data.

A. Core Model Structure

• Renewal Equation: The expected incidence $I_v(t)$ for pathogen $v$ is modeled as:
  $$I_v(t) = R_v(t) \cdot \Lambda_v(t) \cdot \Phi_v(t)$$
  Where $R_v(t)$ is the effective reproduction number, $\Lambda_v(t)$ is the renewal force (based on generation time), and $\Phi_v(t)$ is the interference multiplier.
• Interference Kernel: The interference multiplier $\Phi_v(t)$ is parameterized as a piecewise-constant function over lag windows (0–2, 2–6, 6–12, 12–24 weeks), allowing for time-delayed effects.
• Observation Model: Weekly test outcomes are modeled using a multinomial distribution, linking incidence to observed positive counts ($Y_v$) and negative counts ($Y_0$).

B. The Ratio Penalty (Key Innovation)

To address shared testing propensity, the author introduces a penalty term ($L_{ratio}$) based on the log-odds ratio of pathogen positivity.

• Logic: Borrowing from TND, the model treats positives for "other pathogens" as implicit controls for testing propensity. If testing propensity fluctuates but affects all pathogens similarly, the ratio of positives between pathogens should remain stable.
• Implementation: The penalty minimizes the squared difference between the observed log-odds ratio and the predicted log-odds ratio of pathogen positivity.
  $$L_{ratio} = -\frac{1}{2\sigma^2} \sum_{v,t} \left( \log \frac{Y_v + c}{\sum_{u \neq v} (Y_u + c)} - \log \frac{\pi_v}{\sum_{u \neq v} \pi_u} \right)^2$$
• Constraint: This penalty is only valid when pathogens are measured under a shared testing stream (e.g., multiplex panels).

C. Estimation Strategy

The study employs a Two-Stage Maximum A Posteriori (MAP) estimation:

1. Stage 1: Fix interference parameters ($\theta = 0$) and optimize transmissibility parameters ($\psi$), including baseline, seasonality, and virus-specific random walk deviations ($\delta_v(t)$).
2. Stage 2: Fix transmissibility parameters and optimize interference parameters ($\theta$).

• Note: Joint estimation was attempted but failed to converge due to the high dimensionality and poor conditioning of the likelihood surface.

3. Key Contributions

1. Methodological Adaptation: Successfully adapted the Test-Negative Design (typically used for vaccine effectiveness) to aggregate surveillance data to control for shared testing propensity.
2. Diagnostic Screen: Characterized the proposed method not as a definitive estimator, but as a conservative diagnostic screen with high specificity but limited sensitivity.
3. Diagnostic Decomposition: Provided a mechanistic explanation for why interference estimates collapse to zero, identifying two drivers:
   • Signal Absorption: The virus-specific random walk ($\delta_v(t)$) in Stage 1 absorbs variance attributable to interference.
   • Penalty Amplification: The ratio penalty actively suppresses any signal that alters the compositional structure of predictions, amplifying the bias toward null.

4. Results

A. Synthetic Validation

• Specificity: At true interference $\theta = 0$, the method correctly estimated near-zero values with low false positive rates.
• Sensitivity: The method failed to recover moderate interference ($\theta \le 0.05$). Estimates were attenuated by 70–90% due to signal absorption by the random walk terms.
• Conclusion: The method is biased toward the null; near-zero findings do not prove the absence of interference.

B. Real-Data Application (RSV–COVID, US NREVSS)

• Without Penalty: Estimated total interference magnitude ($|\theta|_{sum}$) for RSV→COVID was 0.0082.
• With Ratio Penalty: The estimate dropped to 0.0016 (an 80% reduction).
• Statistical Significance: Bootstrap 95% confidence intervals for all direction $\times$ lag combinations included zero.
• Comparison: A simpler model using testing volume as a covariate reduced estimates by ~60%, whereas the ratio penalty reduced them by ~80%, suggesting the ratio penalty is a stricter control for testing composition.

C. Violated Assumptions (RSV–Influenza)

When applied to RSV and Influenza (tested via distinct systems, NREVSS vs. FluView), the ratio penalty forced estimates to zero. This confirmed that the penalty relies entirely on the assumption of a shared testing stream; without it, the constraint provides no information and collapses the estimates to the prior.

5. Significance and Conclusions

• Robustness of Signals: The apparent interference signals between RSV and COVID-19 in US surveillance data are not robust to adjustment for shared testing propensity using this specific method.
• Interpretation of Null Results: The study emphasizes that the method's conservative bias means a zero estimate cannot rule out biological interference. It only indicates that the signal is not strong enough to survive this specific, stringent adjustment.
• Utility: The approach serves best as a sensitivity analysis tool. If this conservative screen yields a non-zero estimate, it would be strong evidence of interference. However, a zero estimate is inconclusive.
• Future Directions: The author suggests full Bayesian estimation (e.g., Hamiltonian Monte Carlo) to better characterize joint uncertainty and the need for individual-level longitudinal data to resolve causal directionality in multiplex testing.

Final Takeaway: While biological viral interference is plausible, the observed correlations in aggregate surveillance data are heavily confounded by testing behavior. The proposed ratio penalty method demonstrates that once this confounding is aggressively adjusted for, the remaining evidence for RSV–COVID interference is statistically indistinguishable from zero, though the method's inherent conservatism prevents a definitive conclusion that interference is absent.