Research/Earth
Elephant name calls cannot yet be separated from day effects
In a pre-registered test on the public data behind the 2024 elephant names study, a null that keeps each caller's day fixed could move only 11 of 469 calls. The test was valid but caught a matched names effect in 1 of 20 simulations, so these data cannot settle the dispute.
Nov 20, 2026 · 7 min read

In 2024, Pardo and colleagues reported in Nature Ecology & Evolution that wild African elephants address one another with individually specific, name-like calls. In 2026 a critique in Animal Behaviour argued that the evidence could come from caller identity and from calls drifting over time, and a reply now under double-blind review (Zenodo record 23284924) added corrected timestamps and new controls. We ran the test that neither side had run: shuffle who each call was addressed to, but only among one caller's calls on one day to the same kind of partner, and refit the whole analysis for every shuffle. The answer is that the public data cannot support this test. Only 4 caller-days hold calls to more than one receiver of the same kind, with 11 of the 469 calls in them. At an effect size that reproduces the published contrast, the test detected a pure names effect in 1 of 20 simulated data sets.
11 of 469
calls that a day-blocked shuffle can move
1 of 20
simulated names effects detected (power)
20 of 20
drift-only worlds that fool the naive shuffle
The question
The original analysis trains a random forest to predict the receiver of each rumble from 94 acoustic features, then asks whether calls from one caller to the same receiver are more alike than calls to different receivers. The worry is that a caller's calls to one partner often come from the same day or the same stretch of weeks. If a caller's voice drifts slowly, calls to the same partner can look alike for that reason alone. Permuting receiver labels only within one caller, one date and one relationship type keeps every day-level and caller-level effect in place and breaks only who the call was for. If the same-receiver similarity survived that null, drift could not explain it.
What we did
We fixed the design, the decision rules and eight predictions in a pre-registration before computing any statistic, and published an independent grader before the job returned. The statistic follows the analysis code released with that reply: RF-GAP proximities from a forest of 300 trees, and the coefficient of same receiver on the proximity rank of same-caller pairs, adjusted for same context and same date. The data are the public files of Zenodo record 23284924, checked by md5. We ran three nulls: the primary one blocked by caller, date and relationship (999 shuffles), one blocked by caller and date, and the naive one blocked by caller and relationship with all dates pooled. To measure power and false alarms, we built two synthetic worlds on the real label structure: a names world with a pure receiver effect, and a drift world in which each caller's voice wanders over a 60-day time scale with no receiver effect at all. We tuned each to reproduce the observed contrast and ran 20 data sets per world.
How much each null can move
| Blocks | Blocks with 2+ receivers | Calls in them | Pairs with a movable call |
|---|---|---|---|
| Caller, date, relationship (primary) | 4 | 11 | 36 |
| Caller, date | 29 | 109 | 248 |
| Caller, relationship (naive) | 22 | 131 | 391 |
What we found
The observed contrast was positive, T = +0.11584. None of the three nulls rejected it at 0.05. The primary null gave p = 0.3830, but the reason is that it has almost nothing to shuffle: about 48 distinct relabelings exist. The test kept its size in the drift world (0 of 20 false alarms), so it is valid, just nearly blind. Under our pre-registered rules, that is the verdict UNDERPOWERED.
p value of the same-receiver contrast under each null
Two side readings matter for the wider debate. First, even the naive within-caller null, which ignores days, gave p = 0.0550 in our version of the pipeline. Before the run we had put 0.75 on it rejecting. Second, the drift world reproduced the observed contrast at a modest drift size, and the naive null called that drift a names effect in 20 of 20 simulated data sets. In this label structure, a shuffle within caller that ignores days cannot tell names from slow drift.
Calibration on synthetic elephants
of 20
Mean contrast produced by each synthetic world
What it means
This result does not say elephants lack names, and it does not confirm them. It says that the recordings behind the claim cannot separate a receiver effect from day-level structure, because a caller almost never addresses two different partners of the same kind on the same day in these data. The reply handles time with date as a fixed effect and lag-binned tests. Our calibration shows why that question matters: drift alone can produce a contrast of the observed size, and a shuffle that pools days is fooled by it every time. A decisive test needs new recordings in which the same caller addresses two or more partners of the same kind within a day. The pre-registered grader and every number here are public, so anyone can check the verdict or rerun the job.


