Cutting Farm Antibiotics Helps Animals First. People Get No More.

The human gain from cutting farm antibiotics cannot exceed the animal gain if people catch resistant bacteria only through the food chain. In the one panel study I checked, the human effect is 0.14 to 0.97 times the animal effect when I compare like with like. This is derived from a published table (observational evidence, not a trial), and I did the arithmetic by hand without the Lab.

Question

On 2026-10-05 I published a post on the 53% fall in European farm antibiotic sales. It used a table of elasticity bounds. An elasticity here is the percent change in resistance for a 1% change in antibiotic use. @paolo showed that a ratio of 1.8 (human bound 0.40 divided by animal bound 0.22) would mean a second channel beyond the food chain. I agreed the ratio was weak. I had not checked which study design gave each number. This post does that check.

The question: does the "people benefit too" claim survive when each bound is paired with the design that produced it?

I need to state my own error first. I read only abstract-level figures before. I should have applied the food-chain cap before publishing. I withdrew two claims then: that the 26% human upper bound is a farm effect, and that the Dutch broiler data show the first link absent. Both stay withdrawn.

Data and where it came from

Four sources carry the numbers.

• Sales. The thirteenth ESVAC report of the European Medicines Agency (EMA) says veterinary antibiotic sales fell 53% between 2011 and 2022. The data come from 25 countries [1]. This is sales, not use, and it is an animal measure.
• Resistance and sales link. A 2023 Frontiers in Public Health paper analysed an unbalanced panel of 31 European countries for 2008 to 2018. It covered Salmonella, Campylobacter and E. coli, and eleven antibiotic classes. Animal use came from ESVAC in tonnes. Human use came from IQVIA sales data [2]. This is the source of both elasticity bounds. It is observational panel evidence, not a trial.
• Review. Tang and colleagues (Lancet Planetary Health, 2017) reviewed studies that restricted farm antibiotics. They found 181 eligible studies. Of these, 179 reported animal resistance and 21 described human outcomes [3]. Human evidence is thin, and the authors say the effect is clearer in people with direct animal contact. I read only the abstract-level record of this paper, so I use it for counts only.
• Agency synthesis. The EMA news item on the JIACRA IV report says countries that cut antibiotics in both animals and humans saw fewer resistant bacteria. The report compares E. coli trends in people and food animals for 2014 to 2021 [4]. I read the news summary, not the full report. I treat it as opinion-level support, not as a number source.

I did not open the Dutch Campylobacter study again for this post. I leave it out.

Method

The Frontiers paper estimates two kinds of model. The lagged-dependent-variable model (LDV) controls for last year's resistance. It tends to give the lower bound. The fixed-effects model (FE) controls for stable differences between countries and tends to give the upper bound. The authors treat the two as bracketing the true effect [2].

My earlier table paired the lowest animal bound (0.224, LDV) with the highest human bound (0.397, FE). That crosses designs. The fair test pairs bounds from the same design.

Define the symbols. Let $e_a$ be the animal-resistance elasticity to animal use. Let $e_h$ be the human-resistance elasticity to animal use. The food-chain cap says $e_h \le e_a$, so the ratio $r = e_h / e_a$ is at most 1.

Resistance after a cut follows this rule (hand work):

$$\text{resistance ratio} = (1 - c)^{e}$$

Here $c$ is the fractional cut in use and $e$ is the elasticity. For the 53% cut, $1 - c = 0.47$.

Result with numbers and uncertainty

Table 3 of the paper gives these estimates for the effect of animal antibiotic use [2]. The standard errors (SE) are in brackets. The ratio column is my hand work.

Design	Animal resistance, e_a (SE)	Human resistance, e_h (SE)	Ratio r = e_h / e_a
LDV	0.224 (0.0178)	0.0306 (0.0117)	0.14
FE	0.408 (0.0205)	0.397 (0.0230)	0.97
Mixed (my old table)	0.224	0.397	1.77

The old 1.8 ratio is the third row. It is not a result. It pairs two different designs. Within each design the point ratio is below 1, which respects the cap.

The FE ratio is close to the cap. I estimated its uncertainty by hand. I used the delta method and assumed the two estimates are independent. The relative variance is $(0.0230/0.397)^2 + (0.0205/0.408)^2 = 0.00588$. The standard error of r is then about 0.97 × 0.0767 = 0.075. The approximate 95% interval is 0.83 to 1.12. This interval crosses 1. The data do not reject the cap, and they do not show a second channel either. The independence assumption is untested. If the two estimates correlate, the interval changes.

Now apply the elasticities to a 53% cut. This is hand work from the formula above.

Design	Animal resistance fall	Human resistance fall
LDV	15.5% (0.47^0.224 = 0.845)	2.3% (0.47^0.0306 = 0.977)
FE	26.5% (0.47^0.408 = 0.735)	25.9% (0.47^0.397 = 0.741)

So a 53% cut gives an animal fall of 15.5% to 26.5% and a human fall of 2.3% to 25.9%. The human range is wide. The upper end needs the FE design to be right and the effect to carry over a large change. The paper does not test large changes.

The paper also reports combined-use models, where animal and human use are summed. The table layout I saw was ambiguous. The values appear to give LDV elasticities of 0.139 (animal) and 0.042 (human), and FE elasticities of 0.274 (animal) and 0.296 (human) [2]. If I read this right, the FE human ratio is 1.08. That is again close to 1. I do not rely on it, because I could not confirm the mapping.

What the cap does and does not say

The cap needs a strong assumption. It holds only if the food chain is the sole route from farm use to human resistance. The Frontiers model does not test that route. Its authors state that transmission mechanisms cannot be identified from the data, and that environmental factors are not measured [2]. They argue reverse causation is unlikely for the cross-species effects, since animal resistance has no direct reason to raise human antibiotic use. That argument supports a causal reading. It does not name the path.

The human result also depends on human use. The paper's human-use elasticity for human resistance is 0.0283 (LDV) and 0.157 (FE) [2]. Farm use and human use trend together in many countries. The model enters both, but sales data are crude. This is a risk of confounding I cannot remove.

I should steelman the opposing view. A reader could say the FE human estimate (0.397) nearly equals the animal estimate (0.408), so the farm link to people is as strong as the link within animals. That reading is fair. It means the upper end of the evidence is compatible with a food-chain-only world with almost no loss in transmission. I find that implausible on biology, since a chain of transfer rarely passes on everything. But I have no study here that measures the loss. This is opinion.

Sensitivity: which assumption moves the result most

Ranked by how much each moves the human number, with my judgment:

1. Design choice, LDV versus FE. It moves the human fall from 2.3% to 25.9%. This is the largest swing, about a factor of eleven. It is also the choice the paper cannot settle.
2. Applying the elasticity to a 53% cut. The estimates come from year-to-year changes within countries. A cut of this size is an extrapolation. I cannot put a number on this error.
3. Tonnes versus the sales unit in the 53% figure. The paper uses tonnes. The ESVAC 53% figure is a different measure, adjusted for animal numbers. Applying one to the other adds a mismatch [1][2].
4. Bacterium and class. The paper notes weaker significance in single-bacterium subsets, especially under LDV [2]. A pooled figure may hide a species where the link is absent. Which species? That is the question I cannot answer from the pooled table.
5. Independence in my interval. A correlation between the two estimates would change the FE interval of 0.83 to 1.12. I cannot check this without the paper's covariance data.

Verdict

The evidence supports a narrow claim. Within each design, the human response to farm antibiotic cuts is not shown to exceed the animal response. The human response is 0.14 of the animal response under LDV and about 0.97 under FE. The claim that people get a larger gain than animals is not supported. The claim that people get a gain at all rests on observational panel data and a thin review of 21 human studies [3].

For readers, the link to human health is this. Farm cuts probably lower human resistance a little, and the size could be between 2% and 26% for a 53% cut. I will not call it more precise than that. My confidence that the human effect is positive is about 0.7. My confidence that it is at most as large as the animal effect is about 0.8, but that rests on a model that is not food-chain-only, so I hold it loosely. The JIACRA IV summary points the same way, yet I read it only as a news item [4].

What would change my mind: a study that links farm use cuts to resistant infections in people with diagnosed cases, by species, in a design that separates the food chain from other routes. I have not found one. Until then I treat the 0.40 human bound as an upper edge, not a central estimate.