📊 statistics

Latent variation in pathogen strain-specific effects under multiple-versions-of-treatment theory

This paper argues that when pathogen strain-specific effects vary and strain data is unavailable, the causal interpretation of epidemiologic studies and the transportability of their findings depend critically on the population frequencies of infecting strains, thereby underscoring the importance of collecting pathogen subtype information.

Original authors: Bronner P. Gonçalves

Published 2026-04-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bronner P. Gonçalves

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: "It's Not Just About Getting Sick, It's About Which Germ You Caught"

Imagine you are a doctor trying to figure out how dangerous a specific virus is. You look at your data and see two groups of people: those who got sick and those who didn't. You calculate the average damage (like hospitalization rates) caused by the virus.

The Problem: The virus isn't just one single thing. It's like a box of chocolates, but instead of flavors, it has different "strains" (versions). Some strains are mild (like a dark chocolate truffle), some are moderate (like a caramel), and some are deadly (like a super-spicy chili pepper).

The paper argues that most studies treat the virus as if it were just "Chocolate." They don't know which specific flavor (strain) each patient got because their tests aren't detailed enough. The author, Bronner Gonçalves, uses Causal Inference (a fancy way of saying "figuring out cause and effect") to explain what these studies are actually telling us.

The Core Analogy: The "Mystery Box" of Infection

Let's break down the paper's main points using a Mystery Box analogy.

1. The "Coarsened" View (What we usually see)

Imagine you are running a study on a new "Mystery Box" delivery service.

The Box (A): You know if someone received a box (Infected) or didn't (Not Infected).
The Contents (K): Inside the box, there could be a toy car, a doll, or a puzzle. These are the different strains.
The Result (Y): The customer's reaction. Maybe the toy car breaks a leg, the doll causes a tantrum, and the puzzle causes boredom.

The Issue: Your delivery tracking system (the medical test) only tells you "Box Received" or "No Box." It doesn't tell you what is inside.

2. The "Average" Trap

If you look at the data, you might say: "People who got boxes had 10% more broken legs than people who didn't."

The paper says: That number is real, but it's misleading if you think it applies everywhere.

Why? Because the "10% broken legs" number is actually a weighted average. It depends entirely on the mix of toys inside the boxes in your specific neighborhood.

Scenario A: Your neighborhood mostly gets Toy Cars (mild strain). The "broken leg" rate is low.
Scenario B: Your neighbor's city mostly gets Super-Toys (deadly strain). Their "broken leg" rate is high.

If you take the "low broken leg" rate from your neighborhood and try to predict what will happen in your neighbor's city, you will be wrong. The average effect changes because the mix of strains changed, even if the toys themselves didn't change.

3. The "Transportability" Problem (Why you can't just copy-paste data)

The paper highlights a concept called Transportability. This is like trying to use a recipe from one country in another.

The Recipe: "If you eat this virus, you get sick."
The Catch: In Country X, the virus is 90% mild and 10% deadly. In Country Y, it's 50/50.
The Result: The "average sickness" you see in Country X will be very different from Country Y.

The author argues that if you want to apply study results from one place to another (or from one year to the next, like during the pandemic when new variants emerged), you must know the recipe (the strain composition). Without knowing the mix of strains, the "average effect" is just a snapshot of that specific time and place, not a universal truth.

The "Causal" Takeaway

The paper uses a mathematical framework (Potential Outcomes) to say:

We can still interpret the data: Even if we don't know the specific strain, the number we calculate (e.g., "Infection increases hospitalization risk by 20%") is still valid for that specific population at that specific time.
But it's a "Blended" Effect: That 20% isn't the effect of the virus itself; it's the effect of the virus plus the specific mix of strains circulating right now.
The "No Strain Data" Limit: If the mix of strains changes (e.g., a new, more dangerous variant takes over), that 20% number will change, even if the virus's biology hasn't changed.

Summary in Plain English

Think of the virus like a cocktail.

Strain 1 is mostly water.
Strain 2 is mostly juice.
Strain 3 is mostly alcohol.

If you tell people, "This cocktail makes you dizzy," you are technically correct, but the degree of dizziness depends on the recipe.

If the bar serves a "Water-heavy" cocktail, people get slightly dizzy.
If the bar switches to an "Alcohol-heavy" cocktail, people get very dizzy.

The paper is telling epidemiologists: "When you report that 'The Cocktail makes people dizzy,' you must also report what the recipe was that day. If you don't, and the recipe changes next week, your old report becomes useless for predicting the future."

The Bottom Line: To truly understand how dangerous a pathogen is, we need to stop treating all infections as identical and start paying attention to the specific "flavors" (strains) circulating in the population. Without that data, our predictions are just guesses based on a changing recipe.

1. Problem Statement

The paper addresses a critical challenge in infectious disease epidemiology: the interpretation of studies where the strain-specific effects of a pathogen vary (heterogeneity in virulence), but the specific infecting strain is unobserved (latent).

Context: Many pathogens (e.g., SARS-CoV-2, Streptococcus, malaria parasites) exist in multiple strains or variants with differing impacts on health outcomes (e.g., hospitalization risk, severity).
The Gap: Standard epidemiological studies often detect the presence of a pathogen ( $A=1$ ) without distinguishing between strains ( $K$ ). Consequently, researchers analyze the effect of "infection" generally, unaware that the observed effect is a composite of different strain-specific effects.
Theoretical Issue: In causal inference, this scenario violates the consistency condition (or the "no multiple versions of treatment" assumption). If the outcome depends on the specific strain, but the exposure is defined only as "infected" vs. "uninfected," the standard definition of the potential outcome $Y_a$ becomes ambiguous because $a$ (infection) can be realized by multiple versions ( $k$ ) with different causal mechanisms.

2. Methodology

The author employs Causal Inference Theory, specifically the framework of Potential Outcomes and Multiple Versions of Treatment (as developed by VanderWeele and Hernán).

Notation and Definitions:
- $A$ : Observed exposure (1 = pathogen detected, 0 = no detection).
- $K$ : The "version" of the exposure. $K=0$ is no infection; $K \in \Phi$ represents specific strains. $A$ is a coarsened version of $K$ .
- $Y$ : The outcome (e.g., hospitalization).
- $Y_k$ : The potential outcome had the individual been exposed to strain $k$ .
- $L$ : Confounders (e.g., age) affecting both the strain distribution and the outcome.
Causal Structure:
- The paper assumes a causal diagram where the strain ( $K$ ) precedes the observed exposure ( $A$ ).
- It assumes Exchangeability: $Y_k \perp K | L$ (conditional on confounders, the potential outcome is independent of the specific strain received).
- It explicitly assumes Violation of Exposure Variation Irrelevance: $Y_{a,k} \neq Y_{a,k'}$ ; the outcome depends on the specific strain $k$ .
Mathematical Derivation:
- The author derives the interpretation of the standard epidemiological estimand (the difference in mean outcomes between infected and uninfected groups) without observing $K$ .
- The standard quantity is: $\sum_l E[Y|A=1, L=l]Pr(L=l) - \sum_l E[Y|A=0, L=l]Pr(L=l)$ .
- By applying the law of total probability and the exchangeability assumption, this quantity is decomposed into a weighted average of strain-specific effects.

3. Key Contributions

The paper provides a rigorous mathematical interpretation of "compound exposures" in infectious diseases where strain data is missing.

Reinterpretation of the Standard Estimand: The paper demonstrates that the commonly reported effect of "infection" is not a single causal effect of the pathogen, but rather a weighted average of strain-specific effects.
- The weight for each strain is determined by the population frequency of that strain among the infected individuals within specific strata of confounders ( $L$ ).
- Formulaically, the observed effect equals:
  $\sum_{l,k} \{E[Y_k|L=l] - E[Y_0|L=l]\} \times Pr(K=k|A=1, L=l) \times Pr(L=l)$
Clarification of Transportability: The paper highlights that even if the biological effect of a specific strain ( $E[Y_k - Y_0]$ ) is constant across populations, the observed average effect will differ between populations if the distribution of strains ($Pr(K|A=1)$) differs. This complicates the generalizability (transportability) of study results.
Temporal Dynamics: The framework explains why the estimated effect of an infection might change over time (e.g., during a pandemic) even if the intrinsic virulence of strains remains constant, simply because the circulating strain composition changes.

4. Results and Illustrations

The author uses hypothetical scenarios (illustrated in Figure 1 of the paper) to demonstrate the theoretical findings:

Scenario Setup: A pathogen with three strains ( $S_1, S_2, S_3$ $S_{1}, S_{2}, S_{3}$ ).
- $S_1$ and $S_3$ cause adverse outcomes (hospitalization).
- $S_2$ has no effect on the outcome.
Comparison of Populations:
- Population A: High frequency of $S_2$ (the non-virulent strain).
- Population B: Low frequency of $S_2$ (high frequency of virulent strains).
Outcome: Despite the biological effects of $S_1$ and $S_3$ being identical in both populations, the observed average effect of "infection" ( $E[Y|A=1] - E[Y|A=0]$ ) is significantly lower in Population A than in Population B.
Conclusion: The observed effect is a function of both the biological virulence of the strains and the epidemiological prevalence of those strains.

5. Significance and Implications

Interpretation of Existing Literature: The paper provides a coherent framework for interpreting studies that lack strain-level data. It warns against interpreting the "effect of infection" as a fixed biological constant; it is a population-specific parameter dependent on strain prevalence.
Policy and Public Health:
- Data Collection: It underscores the critical value of collecting pathogen subtype/strain data. Without this, effect estimates are difficult to compare across regions or time periods.
- Vaccine and Treatment Evaluation: When evaluating interventions, the "average effect" observed in a trial may not apply to a new setting if the strain composition has shifted (e.g., the emergence of a new variant).
Theoretical Advancement: It extends the "multiple versions of treatment" theory from general epidemiology to the specific context of infectious diseases, clarifying that the "consistency" assumption is often violated in pathogen studies due to latent strain heterogeneity.

In summary, Gonçalves argues that in the absence of strain data, the estimated effect of an infection is a mixture parameter reflecting the local strain distribution. Therefore, causal claims about the "effect of infection" must be qualified by the specific population and time period in which the strain distribution was observed.