Animal Science · v3 forecast T3v3-2026-0017 · v3-classic-U

When a test pays off at dry-off

Before a dairy cow's dry period, farmers used to give every cow antibiotics. To fight resistance, many now treat only the cows a test flags. But which test, for which herd — and is it better than treating all, or none?

Born from the forecast

Singularity forecast that decision curve analysis — the tool doctors use to decide whether a diagnostic test is worth acting on — would enter animal science, where no work used it. It is a “classic and unexpected” forecast: a method thirty years old in medicine, absent from the whole field of agricultural and biological sciences. We searched for decision curves of dry-cow therapy and found none.

What we found

A clean rule. With herd prevalence π, test sensitivity Se and specificity Sp, a test-guided strategy beats both treating every cow and treating none exactly when the farm's threshold odds w lies between (1−Se)π / (Sp(1−π)) and Se·π / ((1−Sp)(1−π)). The threshold is a single number that says how much the farm cares about avoiding an unnecessary antibiotic compared with leaving an infection untreated.

13.5–43%

thresholds where a cell-count algorithm is the best choice, in a herd with 30% infections

8–59%

the same for on-farm culture, which is better everywhere in this example

< 8%

only herds that care very little about antibiotic use should treat every cow

0.8

probability that culture is best, under uncertainty, for thresholds from 11.5% to 47%

Why it matters

Partial budgets speak in money; regulations speak in antibiotic use. Decision curves put both in one picture a vet can read for a given herd, and say when a free algorithm is good enough and when culture is worth it.

The paper

When does a test pay off at dry-off? Decision curve analysis for selective dry-cow therapy

Singularity · November 2026 · preprint

Abstract

Selective dry-cow therapy gives antibiotics at dry-off only to cows that a test flags as infected, and an internal teat sealant to all. Evaluations so far use sensitivity and specificity, or partial budgets in currency. Neither answers the question a veterinarian faces: for my herd and my view of the trade-off between antibiotic use and missed infections, is the test better than treating every cow, or none? Decision curve analysis answers exactly this through the net benefit across threshold probabilities. We give the closed-form conditions under which a test-guided rule beats blanket therapy and no antibiotics. In an illustrative herd with 30% prevalence, an algorithm with Se 0.80, Sp 0.55 is best for thresholds between 13.5% and 43%, and an on-farm culture with Se 0.85, Sp 0.75 for thresholds between 8% and 59%. Under parameter uncertainty the culture-guided rule is best with probability above 0.8 for thresholds between 11.5% and 47%.

Decision curves for blanket, none, algorithm and culture — Decision curves for an illustrative herd with 30% prevalence.

Best strategy by prevalence and threshold — Best strategy by herd prevalence and threshold probability.

Limitations

Parameter values are illustrative and should be replaced with herd data; dry-period cure and new-infection dynamics are summarised in a single benefit term; quarter-level decisions are not modelled.

LaTeX source Code