Animal collocation revisited: intercohort comparison and a case study comparing call combinations between sexes in common marmosets

This paper introduces and validates a statistically rigorous method called Multiple Distinctive Collocation Analysis using Pearson residuals (MDCA-Pr) to overcome existing limitations in animal collocation analysis, enabling robust identification of non-random signal combinations and accurate comparisons between different cohorts, as demonstrated through simulations and a case study on common marmoset vocalizations.

The Gist

Imagine you are trying to figure out the secret language of a group of monkeys. You record them making sounds, and you notice they often put two sounds together, like "Peep-Howl" or "Huff-Puff." But here's the tricky part: Are they actually saying something specific by combining them, or are they just randomly making noise that happens to sound like a pattern?

For a long time, scientists had a hard time telling the difference. Their old tools were like a shaky ruler; they could guess if a pattern existed, but they couldn't tell you how sure they were, and they often got fooled by random chance.

This paper introduces a brand new, super-precise tool called MDCA-Pr to solve this problem. Here is the story of how they built it and tested it, explained simply.

1. The Problem: The "Shaky Ruler"

Think of animal communication like a giant box of Lego bricks. Scientists want to know which bricks the animals always snap together to build something special.

• The Old Way: Previous methods were like looking at the box and guessing, "Hey, those two bricks look like they go together!" But they didn't have a way to measure the strength of that connection, and they didn't account for the fact that if you look at enough bricks, you'll eventually find two that just happen to look like they fit, even if they don't.
• The Flaw: The old tools couldn't handle "nested" data (like if one monkey made 50 calls and another made 5, skewing the results) and couldn't tell you if the difference between two groups (like males vs. females) was real or just a fluke.

2. The Solution: The "Precision Scale" (MDCA-Pr)

The authors adapted a method used by human linguists and gave it a statistical upgrade. Think of MDCA-Pr as a high-tech precision scale.

• Instead of just guessing, it weighs every single combination of sounds.
• It calculates a "confidence score" (like a weather forecast saying "90% chance of rain").
• It uses a special trick called bootstrapping. Imagine you have a bag of marbles (the data). You pull a handful out, count them, put them back, and do it 10,000 times. This helps the tool understand how much the results might wiggle if you collected the data again. This solves the "shaky ruler" problem.

3. The Three Tests: Putting the Tool to Work

To prove their new tool works, the scientists ran three different experiments:

Test 1: The "Fake Monkey" Simulation

They created computer-generated monkey data.

• The Setup: They programmed some sounds to always go together (the "real" patterns) and others to be random.
• The Result: The new tool was a detective. It found the real patterns almost every time (high sensitivity) and rarely got tricked by the random noise (high selectivity).
• The Catch: When the data was very small (like a tiny bag of marbles), the tool was a bit cautious. It sometimes missed the weakest patterns to avoid making mistakes. But the scientists found a simple "filter" (a rule of thumb) to make it even more accurate.

Test 2: The "Twin vs. Stranger" Comparison

They wanted to see if the tool could spot differences between two groups of monkeys (Cohort A vs. Cohort B).

• The Setup: They made two groups of fake monkeys that were identical, and a third group that was slightly different.
• The Result: The tool correctly said, "These two groups are the same" (no false alarms). When they compared the identical group to the different one, the tool successfully spotted the difference, but only if they had enough data.
• The Lesson: If you want to compare two groups of animals, you need a decent amount of data. If you only have a few recordings, the tool will be very conservative and might say "I can't tell the difference yet" rather than guessing wrong.

Test 3: The Real Deal (Common Marmosets)

Finally, they used the tool on real data from Common Marmosets (a type of small monkey) living in captivity. They wanted to know: Do male and female marmosets combine their calls differently when they see food?

• The Setup: They recorded males and females when food was introduced.
• The Result: Surprisingly, the males and females were very similar. They used mostly the same combinations of sounds.
• The Nuance: There were a few tiny differences (females seemed to use a few specific combos slightly more often), but overall, they were speaking the same "dialect."
• The Insight: The tool also helped them realize that some rare combinations were only made by one specific monkey, not the whole group. This is crucial! It stops scientists from thinking a whole species has a secret code when it's actually just one individual's quirk.

4. Why This Matters

This paper is like handing animal researchers a new, high-definition camera.

• Before: They were taking blurry photos and guessing what they saw.
• Now: They have a sharp lens that tells them exactly what is a real pattern and what is just noise.
• The Future: This allows scientists to finally compare different groups of animals (like males vs. females, or different populations) with confidence. It helps us understand if animals have "dialects" or if their way of combining sounds changes based on who they are.

In a nutshell: The authors built a better statistical tool to decode animal language. They proved it works on fake data, showed how to use it to compare groups, and demonstrated that male and female marmosets are actually quite similar in how they mix their calls when eating. It's a big step forward in understanding the complex "grammar" of the animal kingdom.

Technical Summary

1. Problem Statement

Animal communication research increasingly focuses on signal combinatoriality (the non-random combination of signals into sequences). While collocation analysis (derived from computational linguistics) is a standard method for identifying which signal pairs co-occur more frequently than expected by chance, existing methods applied to animal data suffer from four critical statistical limitations:

1. Lack of Uncertainty Estimates: Previous methods provide only point estimates of collocation strength without confidence intervals, making it impossible to gauge the precision of the estimates.
2. Inflated Family-Wise Error Rates: Existing approaches fail to correct for multiple comparisons when analyzing many signal types, leading to high Type I error rates (false positives).
3. Ignoring Data Non-Independence: Standard methods do not account for nested data structures (e.g., multiple calls from the same individual or bout), violating the assumption of independence.
4. Trade-off between Directionality and Sensitivity: Existing tools either detect directional transitions (A→B vs. B→A) but lack sensitivity to rare events, or are sensitive to rare events but blind to directionality.

Furthermore, these limitations prevent intercohort comparisons (e.g., comparing signal combination strategies between sexes, age classes, or populations), as researchers cannot statistically determine if observed differences are genuine behavioral variations or sampling artifacts.

2. Methodology

The authors propose and validate Multiple Distinctive Collocation Analysis using Pearson residuals (MDCA-Pr), an adaptation of a linguistic method by Gries (2023).

Core Algorithm:

• Pearson Residuals (Pr): Instead of iterative 2x2 binomial arrays, MDCA-Pr calculates collocation strength using a single global contingency table. The strength is calculated as:
  $$Pr = (O - E) / \sqrt{E}$$
  Where $O$ is the observed frequency of a bigram and $E$ is the expected frequency based on independent distribution.
• Block Bootstrapping: To address non-independence, the method employs block bootstrapping (resampling at the highest hierarchical level, e.g., by individual or bout). This generates a distribution of collocation strengths, allowing for the calculation of confidence intervals (CIs).
• Error Correction: CIs are adjusted using Bonferroni correction to control for family-wise error rates across multiple bigram comparisons.
• Directionality: The method retains directional information (A→B is distinct from B→A).

Validation Strategy (Three Studies):

1. Study 1 (Sensitivity/Selectivity): Applied MDCA-Pr to synthetic datasets of varying sizes (Small: 160 bigrams; Large: 3200 bigrams) and recombinatoriality levels (Exclusive vs. Recombinatorial). Performance was measured by True Positive Rate (sensitivity) and False Positive Rate (selectivity).
2. Study 2 (Intercohort Comparison): Simulated "cohort" data where two groups had identical bigram probabilities (control) and one group had altered probabilities (unequivalent). The method's ability to detect significant differences in collocation strength between cohorts was tested.
3. Study 3 (Real-world Application): Applied MDCA-Pr to a real dataset of captive common marmosets (Callithrix jacchus). The study compared call combinations between males and females in a feeding context.

Refinement: The authors tested an additional filter (requiring mean collocation strength to exceed one standard deviation of the global mean) to further reduce false positives identified in large, complex datasets.

3. Key Results

Study 1: Performance on Synthetic Data

• Selectivity: MDCA-Pr achieved perfect selectivity (0 false positives) in Exclusive datasets and randomized controls. In Recombinatorial datasets, initial false positives occurred in large samples, but applying the 1SD filter (mean strength > 1 SD above global mean) eliminated all false positives without sacrificing sensitivity.
• Sensitivity: Sensitivity was high in Large datasets. In Small datasets, sensitivity dropped slightly for weakly attracted bigrams (e.g., Huff→Puff at 8% probability), indicating that small sample sizes may miss weak attractions.
• Conclusion: MDCA-Pr is robust but benefits from the 1SD filter to maintain high selectivity in complex, recombinatorial data.

Study 2: Intercohort Comparisons

• Selectivity: Perfect selectivity was maintained; no false differences were detected between cohorts with identical underlying probabilities.
• Sensitivity: Sensitivity was highly dependent on sample size.
  • Large Datasets: Successfully detected probability differences as small as 5% between cohorts.
  • Small Datasets: Sensitivity was low; only large probability differences were detected.
  • Power Analysis: Doubling the sample size (either more individuals or more calls) significantly improved sensitivity, suggesting that moderate sample sizes are required for reliable intercohort comparison.

Study 3: Marmoset Case Study

• Application: Analyzed 488 pairwise transitions from 733 call productions (4 males, 3 females).
• Findings:
  • Shared Attraction: Both sexes showed significant attraction for Tsik→Ekk (mobbing) and Food Phee→Phee (food). No significant difference in attraction strength was found between sexes for these.
  • Female-Biased Attraction: Three combinations were identified as female-biased: Food Peep→Food Phee, Ekk→Tsik, and Phee→Short Phee.
  • Post-hoc Analysis: Crucially, the authors found that some "attracted" combinations were driven by only one or two individuals. This highlighted the necessity of reviewing individual-level data to distinguish between cohort-level behaviors and idiosyncratic individual habits.

4. Key Contributions

1. Statistical Rigor: Introduces a collocation method that provides uncertainty estimates (CIs) and controls for non-independence via block bootstrapping, addressing the primary statistical flaws of previous animal collocation tools.
2. Intercohort Comparison Framework: Establishes a validated pipeline for statistically comparing signal combinatoriality between different animal cohorts (e.g., sexes, populations), a capability previously unavailable due to a lack of uncertainty metrics.
3. Methodological Pipeline: Proposes a standardized workflow (Figure 3 in the paper) including:
   • Block bootstrapping for CIs.
   • Bonferroni correction for multiple testing.
   • A 1SD filter to reduce false positives in complex datasets.
   • Mandatory post-hoc review of individual contributions to avoid misinterpreting individual idiosyncrasies as cohort behaviors.
4. Empirical Validation: Demonstrates the method's efficacy on both simulated data and real-world primate vocalizations.

5. Significance

• Advancing Animal Communication Theory: MDCA-Pr allows researchers to move beyond qualitative descriptions of call combinations to rigorous, quantitative comparisons. This is essential for understanding the evolution of human language features (like syntax) in animal systems.
• Demographic Insights: The ability to compare cohorts enables investigations into how social factors (sex, age, group identity) shape communication strategies.
• Robustness: By solving the issues of false positives and non-independence, the method increases the reliability of findings in the field.
• Future Directions: The authors note that while MDCA-Pr identifies attraction, it does not explain function. However, it provides the statistical foundation necessary to link specific signal combinations to behavioral contexts or receiver responses in future studies.

Limitations Noted: The method requires sufficient sample sizes to detect subtle differences between cohorts. Additionally, while it handles nested data, researchers must still carefully design sampling protocols (e.g., stratified bootstrapping) to ensure contexts are represented equitably.