Hive Sound Predictability
Predicts colony survival horizons and pathogen presence / level from a hive recording, and states for every target how well that prediction held up in colony-grouped validation. Targets whose validation was not distinguishable from chance are shown with the verdict 'no evidence of predictability from sound' and their calls are greyed out. Deployed feature families: ast, hc. Trained on 1131 recordings from 112 colonies.
Gemini listening
Gemini summary
Example recordings (training recordings from the workbook: the live output is an in-sample fit; the out-of-fold column shows the validated output)
Metric: ROC-AUC for binary targets (chance = 0.50), Spearman correlation for levels (chance = 0). Brackets: colony-cluster bootstrap 95% CI. Gain: paired bootstrap of metadata+audio minus metadata alone (metadata = site, season, year, October 2023 batch). Permutation p: month-aligned colony permutation within country, so it tests for acoustic signal beyond site, season and batch; Holm-adjusted across targets in brackets. Within-country: audio-only AUC (binary targets) or Spearman correlation (level targets) and its gain over the within-country metadata model.
| Target | n (positives) | Metadata only | Best single audio family | Ensemble (primary) | Gain over metadata | Within-country | Permutation p | Verdict |
|---|---|---|---|---|---|---|---|---|
| terminal_03 Colony is Terminal (dies) within 0-3 months of the recording (1) vs Survivor (0) | 1131 (45) | 0.76 [0.66, 0.86] | hc: 0.74 [0.65, 0.84] | 0.75 [0.65, 0.86] | -0.07 [-0.16, +0.01] | NZ: 2 positive / 50 negative colonies (need 8 each); USA: audio 0.80 [0.74, 0.87], gain over metadata +0.12 [+0.03, +0.21] | 0.010 (Holm 0.158) | signal present but explained by site/season/batch |
| terminal_46 Colony is Terminal (dies) within 0-6 months (nested: every 0-3 month Terminal row is also Terminal here) of the recording (1) vs Survivor (0) | 1131 (69) | 0.71 [0.61, 0.82] | ast: 0.73 [0.67, 0.81] | 0.71 [0.63, 0.80] | -0.04 [-0.14, +0.06] | NZ: 2 positive / 50 negative colonies (need 8 each); USA: audio 0.65 [0.58, 0.74], gain over metadata +0.05 [-0.04, +0.14] | 0.010 (Holm 0.158) | signal present but explained by site/season/batch |
| terminal_4_6_only Colony dies 4-6 months after the recording (1) vs survives 6 months (0); rows already Terminal within 3 months excluded | 1086 (24) | 0.77 [0.65, 0.87] | clap: 0.66 [0.54, 0.77] | 0.64 [0.50, 0.77] | -0.15 [-0.26, -0.02] | NZ: 2 positive / 50 negative colonies (need 8 each); USA: 7 positive / 45 negative colonies (need 8 each) | 0.133 (Holm 1.000) | no evidence of predictability from sound |
| cbpv_present CBPV detected (> 0) in the recording month | 1131 (111) | 0.81 [0.76, 0.86] | ast: 0.67 [0.61, 0.74] | 0.68 [0.62, 0.75] | -0.16 [-0.23, -0.09] | NZ: audio 0.56 [0.49, 0.62], gain over metadata -0.05 [-0.14, +0.04]; USA: 2 positive / 62 negative colonies (need 8 each) | 0.436 (Holm 1.000) | signal present but explained by site/season/batch |
| cbpv_log_level log(1 + CBPV) level, all recordings | 1131 | 0.20 [0.15, 0.25] | ast: 0.12 [0.06, 0.18] | 0.13 [0.07, 0.19] | -0.07 [-0.13, -0.01] | NZ: audio +0.09 [-0.01, +0.17], gain over metadata -0.01 [-0.11, +0.10]; USA: audio +0.05 [+0.00, +0.09], gain over metadata -0.00 [-0.02, +0.01] | 0.040 (Holm 0.554) | signal present but explained by site/season/batch |
| cbpv_high Among recordings with CBPV > 0: high load (at or above the 75th percentile of positive values, 0.00301) vs low positive | 111 (28) | 0.66 [0.55, 0.76] | clap: 0.65 [0.54, 0.75] | 0.61 [0.49, 0.74] | -0.06 [-0.18, +0.06] | NZ: audio 0.63 [0.51, 0.75], gain over metadata -0.02 [-0.14, +0.09]; USA: 0 positive / 2 negative colonies (need 8 each) | 0.318 (Holm 1.000) | no evidence of predictability from sound |
| cbpv_log_level_among_positive log(1 + CBPV) among recordings where CBPV > 0 (n=111) | 111 | 0.41 [0.22, 0.56] | clap: 0.22 [0.01, 0.39] | 0.17 [-0.03, 0.36] | -0.16 [-0.35, +0.03] | NZ: audio +0.26 [+0.06, +0.42], gain over metadata -0.12 [-0.28, +0.05]; USA: 3 rows from 2 colonies (need 40 rows, 10 colonies) | 0.444 (Holm 1.000) | no evidence of predictability from sound |
| dwv_present DWV detected (> 0) in the recording month | 1131 (508) | 0.96 [0.94, 0.97] | ast: 0.87 [0.83, 0.90] | 0.89 [0.86, 0.92] | -0.08 [-0.10, -0.05] | NZ: audio 0.58 [0.50, 0.67], gain over metadata -0.10 [-0.20, -0.00]; USA: audio 0.74 [0.60, 0.89], gain over metadata -0.12 [-0.22, -0.03] | 0.277 (Holm 1.000) | signal present but explained by site/season/batch |
| dwv_log_level log(1 + DWV) level, all recordings | 1131 | 0.50 [0.46, 0.53] | ast: 0.23 [0.15, 0.31] | 0.25 [0.15, 0.34] | -0.26 [-0.34, -0.18] | NZ: audio +0.14 [+0.05, +0.24], gain over metadata -0.04 [-0.12, +0.04]; USA: audio +0.20 [+0.11, +0.30], gain over metadata +0.00 [-0.04, +0.05] | 0.228 (Holm 1.000) | signal present but explained by site/season/batch |
| dwv_high Among recordings with DWV > 0: high load (at or above the 75th percentile of positive values, 0.05338) vs low positive | 508 (127) | 0.67 [0.61, 0.73] | ast: 0.63 [0.56, 0.70] | 0.62 [0.55, 0.70] | -0.06 [-0.12, +0.01] | NZ: audio 0.53 [0.47, 0.61], gain over metadata -0.09 [-0.15, -0.02]; USA: 13 positive / 6 negative colonies (need 8 each) | 0.277 (Holm 1.000) | signal present but explained by site/season/batch |
| dwv_log_level_among_positive log(1 + DWV) among recordings where DWV > 0 (n=508) | 508 | 0.25 [0.16, 0.34] | ast: 0.22 [0.14, 0.30] | 0.19 [0.09, 0.29] | -0.06 [-0.14, +0.03] | NZ: audio +0.17 [+0.08, +0.26], gain over metadata -0.04 [-0.13, +0.05]; USA: 19 rows from 18 colonies (need 40 rows, 10 colonies) | 0.416 (Holm 1.000) | signal present but explained by site/season/batch |
| kbv_present KBV detected (> 0) in the recording month | 1131 (125) | 0.73 [0.69, 0.77] | ast: 0.69 [0.64, 0.73] | 0.70 [0.65, 0.75] | -0.04 [-0.09, +0.02] | NZ: audio 0.68 [0.59, 0.77], gain over metadata -0.12 [-0.21, -0.04]; USA: audio 0.59 [0.52, 0.66], gain over metadata -0.09 [-0.17, -0.01] | 0.040 (Holm 0.554) | signal present but explained by site/season/batch |
| kbv_log_level log(1 + KBV) level, all recordings | 1131 | 0.06 [-0.03, 0.15] | ast: 0.10 [0.03, 0.17] | 0.10 [0.03, 0.17] | +0.05 [-0.03, +0.14] | NZ: audio +0.13 [+0.00, +0.24], gain over metadata -0.10 [-0.18, -0.04]; USA: audio +0.10 [+0.01, +0.20], gain over metadata +0.01 [-0.09, +0.13] | 0.158 (Holm 1.000) | signal present but explained by site/season/batch |
| kbv_high Among recordings with KBV > 0: high load (at or above the 75th percentile of positive values, 0.01584) vs low positive | 125 (32) | 0.63 [0.47, 0.79] | hc: 0.70 [0.58, 0.81] | 0.64 [0.51, 0.75] | -0.00 [-0.17, +0.17] | NZ: 7 positive / 17 negative colonies (need 8 each); USA: audio 0.68 [0.54, 0.81], gain over metadata +0.10 [-0.11, +0.32] | 0.231 (Holm 1.000) | signal present but explained by site/season/batch |
| kbv_log_level_among_positive log(1 + KBV) among recordings where KBV > 0 (n=125) | 125 | 0.42 [0.22, 0.58] | ast: 0.25 [0.08, 0.41] | 0.24 [0.08, 0.39] | -0.17 [-0.38, +0.05] | NZ: 35 rows from 20 colonies (need 40 rows, 10 colonies); USA: audio +0.29 [+0.11, +0.49], gain over metadata +0.07 [-0.12, +0.31] | 0.167 (Holm 1.000) | signal present but explained by site/season/batch |
| any_pathogen_present Any of the measured pathogens detected (> 0); largely a site proxy in this data set | 1131 (602) | 0.90 [0.88, 0.92] | ast: 0.82 [0.78, 0.85] | 0.83 [0.80, 0.86] | -0.08 [-0.11, -0.05] | NZ: audio 0.64 [0.55, 0.73], gain over metadata +0.03 [-0.08, +0.13]; USA: audio 0.59 [0.53, 0.66], gain over metadata -0.09 [-0.16, -0.03] | 0.594 (Holm 1.000) | signal present but explained by site/season/batch |
What hive sound can and cannot predict: colony survival, viral loads and the missing Varroa data
Generated by make_report.py from the pipeline artifacts. Every number below is computed; none is typed in.
1. Data
- Workbook rows: 1134; recordings with an audio file: 1131; colonies: 112 (USA 62, NZ 50).
- Measured quantities per recording month: CBPV, DWV, KBV (the numeric columns of the workbook) and the two survival labels. The workbook has no column for Varroa mite count, colony size or strength, queen status, brood, weight, nutrition, so none of those can be modelled or tested here.
- Fraction of recordings with a detectable load, by site: NZ: CBPV 0.20, DWV 0.89, KBV 0.06; USA: CBPV 0.01, DWV 0.03, KBV 0.16.
- Terminal labels: 45 recordings (0-3 months) and 69 recordings (4-6 months) from 24 colonies. 12 colonies were recorded only once (F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12; batch USA_2023_10); all of them are Terminal.
- Label structure: within a colony the labels move monotonically S/S -> S/T -> T/T towards the month of death, i.e. the labels encode months-to-death.
2. Method
- Features per recording: handcrafted acoustic features (256), AST embedding (1536: mean and SD of 768-d window embeddings), CLAP embedding (1024), Gemini structured listening (13), recording QC (10), metadata (is_nz, month_sin, month_cos, month_sin2, month_cos2, year_offset, is_oct2023_batch).
- Validation: 5-fold colony-grouped cross-validation repeated 3 times (all recordings of a colony on one side), metrics on out-of-fold predictions; colony-cluster bootstrap (1000 resamples) 95% CI; month-aligned colony permutation test (100 permutations: colony identities permuted within country and each recording given the donor colony's label for the same calendar month, so site, season, batch and within-colony label structure are preserved and p-values test for acoustic signal beyond them).
- Models fixed in advance: PCA(64) + L2 logistic / ridge for embeddings, gradient boosting for handcrafted features, rank-average ensemble as the primary model. No per-target model selection.
- Confound controls: metadata-only baseline and paired bootstrap of the gain of audio over it; within-country evaluation with its own metadata baseline; cross-site transfer; every target re-run without the October 2023 batch; probes that ask whether audio can identify site, season and the October 2023 batch. Evaluations need at least 8 positive and 8 negative colonies.
- Verdict rule (fixed before the run): level 0 = the primary CI includes chance (the upper bound is stated); level 1 = CI excludes chance but the permutation p is at least 0.05 or the 95% CI of the paired gain over metadata does not lie above zero; level 2 = CI excludes chance, permutation p below 0.05 and the 95% CI of the paired gain over metadata lies above zero. Holm-adjusted p-values across the non-probe targets are shown in brackets.
3. Which targets are predictable from sound
Metric: ROC-AUC (chance 0.50) for binary targets, Spearman correlation (chance 0) for levels; primary model = ensemble. 'Gain' is the paired-bootstrap difference between metadata+audio and metadata alone. Verdicts follow the pre-registered rule in evaluate.py.
| target | n (pos) [positive colonies] | metadata only | ensemble | gain over metadata | within-country gain of audio over metadata | perm. p (Holm) | without Oct 2023 batch (ensemble) | verdict |
|---|---|---|---|---|---|---|---|---|
| any_pathogen_present | 1131 (602) [105] | 0.90 [0.88, 0.92] | 0.83 [0.80, 0.86] | -0.08 [-0.11, -0.05] | NZ: audio 0.64 [0.55, 0.73], gain +0.03 [-0.08, +0.13]; USA: audio 0.59 [0.53, 0.66], gain -0.09 [-0.16, -0.03] | 0.594 (1.000) | 0.83 [0.80, 0.86] | signal present but explained by site/season/batch |
| cbpv_log_level | 1131 | +0.20 [+0.15, +0.25] | +0.13 [+0.07, +0.19] | -0.07 [-0.13, -0.01] | NZ: audio +0.09 [-0.01, +0.17], gain -0.01 [-0.11, +0.10]; USA: audio +0.05 [+0.00, +0.09], gain -0.00 [-0.02, +0.01] | 0.040 (0.554) | +0.15 [+0.09, +0.20] | signal present but explained by site/season/batch |
| cbpv_present | 1131 (111) [47] | 0.81 [0.76, 0.86] | 0.68 [0.62, 0.75] | -0.16 [-0.23, -0.09] | NZ: audio 0.56 [0.49, 0.62], gain -0.05 [-0.14, +0.04]; USA: 2 positive / 62 negative colonies (need 8 each) | 0.436 (1.000) | 0.68 [0.61, 0.74] | signal present but explained by site/season/batch |
| dwv_high | 508 (127) [49] | 0.67 [0.61, 0.73] | 0.62 [0.55, 0.70] | -0.06 [-0.12, +0.01] | NZ: audio 0.53 [0.47, 0.61], gain -0.09 [-0.15, -0.02]; USA: 13 positive / 6 negative colonies (need 8 each) | 0.277 (1.000) | 0.56 [0.50, 0.63] | signal present but explained by site/season/batch |
| dwv_log_level | 1131 | +0.50 [+0.46, +0.53] | +0.25 [+0.15, +0.34] | -0.26 [-0.34, -0.18] | NZ: audio +0.14 [+0.05, +0.24], gain -0.04 [-0.12, +0.04]; USA: audio +0.20 [+0.11, +0.30], gain +0.00 [-0.04, +0.05] | 0.228 (1.000) | +0.26 [+0.16, +0.34] | signal present but explained by site/season/batch |
| dwv_log_level_among_positive | 508 | +0.25 [+0.16, +0.34] | +0.19 [+0.09, +0.29] | -0.06 [-0.14, +0.03] | NZ: audio +0.17 [+0.08, +0.26], gain -0.04 [-0.13, +0.05]; USA: 19 rows from 18 colonies (need 40 rows, 10 colonies) | 0.416 (1.000) | +0.12 [+0.04, +0.21] | signal present but explained by site/season/batch |
| dwv_present | 1131 (508) [68] | 0.96 [0.94, 0.97] | 0.89 [0.86, 0.92] | -0.08 [-0.10, -0.05] | NZ: audio 0.58 [0.50, 0.67], gain -0.10 [-0.20, -0.00]; USA: audio 0.74 [0.60, 0.89], gain -0.12 [-0.22, -0.03] | 0.277 (1.000) | 0.89 [0.85, 0.92] | signal present but explained by site/season/batch |
| kbv_high | 125 (32) [24] | 0.63 [0.47, 0.79] | 0.64 [0.51, 0.75] | -0.00 [-0.17, +0.17] | NZ: 7 positive / 17 negative colonies (need 8 each); USA: audio 0.68 [0.54, 0.81], gain +0.10 [-0.11, +0.32] | 0.231 (1.000) | 0.50 [0.34, 0.65] | signal present but explained by site/season/batch |
| kbv_log_level | 1131 | +0.06 [-0.03, +0.15] | +0.10 [+0.03, +0.17] | +0.05 [-0.03, +0.14] | NZ: audio +0.13 [+0.00, +0.24], gain -0.10 [-0.18, -0.04]; USA: audio +0.10 [+0.01, +0.20], gain +0.01 [-0.09, +0.13] | 0.158 (1.000) | +0.07 [-0.00, +0.14] | signal present but explained by site/season/batch |
| kbv_log_level_among_positive | 125 | +0.42 [+0.22, +0.58] | +0.24 [+0.08, +0.39] | -0.17 [-0.38, +0.05] | NZ: 35 rows from 20 colonies (need 40 rows, 10 colonies); USA: audio +0.29 [+0.11, +0.49], gain +0.07 [-0.12, +0.31] | 0.167 (1.000) | +0.16 [-0.01, +0.30] | signal present but explained by site/season/batch |
| kbv_present | 1131 (125) [73] | 0.73 [0.69, 0.77] | 0.70 [0.65, 0.75] | -0.04 [-0.09, +0.02] | NZ: audio 0.68 [0.59, 0.77], gain -0.12 [-0.21, -0.04]; USA: audio 0.59 [0.52, 0.66], gain -0.09 [-0.17, -0.01] | 0.040 (0.554) | 0.70 [0.64, 0.75] | signal present but explained by site/season/batch |
| probe_batch_oct2023 | 1131 (45) [45] | 1.00 [1.00, 1.00] | 0.89 [0.85, 0.93] | -0.09 [-0.14, -0.05] | NZ: 0 positive / 50 negative colonies (need 8 each); USA: audio 0.90 [0.84, 0.94], gain -0.07 [-0.11, -0.04] | 0.178 | - | signal present but explained by site/season/batch |
| probe_country_is_nz | 1131 (552) [50] | 1.00 [1.00, 1.00] | 0.94 [0.91, 0.96] | -0.05 [-0.08, -0.04] | NZ: 50 positive / 0 negative colonies (need 8 each); USA: 0 positive / 62 negative colonies (need 8 each) | 0.307 | 0.94 [0.91, 0.97] | signal present but explained by site/season/batch |
| probe_month_cos | 1131 | +0.98 [+0.98, +0.98] | +0.50 [+0.45, +0.55] | -0.47 [-0.52, -0.42] | NZ: audio +0.69 [+0.65, +0.73], gain -0.27 [-0.32, -0.23]; USA: audio +0.65 [+0.62, +0.69], gain -0.29 [-0.32, -0.26] | 0.277 | +0.52 [+0.47, +0.57] | signal present but explained by site/season/batch |
| probe_month_sin | 1131 | +0.98 [+0.97, +0.98] | +0.47 [+0.42, +0.51] | -0.44 [-0.48, -0.40] | NZ: audio +0.61 [+0.56, +0.67], gain -0.33 [-0.38, -0.28]; USA: audio +0.55 [+0.49, +0.60], gain -0.34 [-0.39, -0.30] | 0.386 | +0.48 [+0.43, +0.53] | signal present but explained by site/season/batch |
| terminal_03 | 1131 (45) [24] | 0.76 [0.66, 0.86] | 0.75 [0.65, 0.86] | -0.07 [-0.16, +0.01] | NZ: 2 positive / 50 negative colonies (need 8 each); USA: audio 0.80 [0.74, 0.87], gain +0.12 [+0.03, +0.21] | 0.010 (0.158) | 0.69 [0.55, 0.81] | signal present but explained by site/season/batch |
| terminal_46 | 1131 (69) [24] | 0.71 [0.61, 0.82] | 0.71 [0.63, 0.80] | -0.04 [-0.14, +0.06] | NZ: 2 positive / 50 negative colonies (need 8 each); USA: audio 0.65 [0.58, 0.74], gain +0.05 [-0.04, +0.14] | 0.010 (0.158) | 0.66 [0.55, 0.76] | signal present but explained by site/season/batch |
| cbpv_high | 111 (28) [21] | 0.66 [0.55, 0.76] | 0.61 [0.49, 0.74] | -0.06 [-0.18, +0.06] | NZ: audio 0.63 [0.51, 0.75], gain -0.02 [-0.14, +0.09]; USA: 0 positive / 2 negative colonies (need 8 each) | 0.318 (1.000) | - | no evidence of predictability from sound; AUC at most 0.74 with 95% confidence |
| cbpv_log_level_among_positive | 111 | +0.41 [+0.22, +0.56] | +0.17 [-0.03, +0.36] | -0.16 [-0.35, +0.03] | NZ: audio +0.26 [+0.06, +0.42], gain -0.12 [-0.28, +0.05]; USA: 3 rows from 2 colonies (need 40 rows, 10 colonies) | 0.444 (1.000) | - | no evidence of predictability from sound; |
| terminal_4_6_only | 1086 (24) [9] | 0.77 [0.65, 0.87] | 0.64 [0.50, 0.77] | -0.15 [-0.26, -0.02] | NZ: 2 positive / 50 negative colonies (need 8 each); USA: 7 positive / 45 negative colonies (need 8 each) | 0.133 (1.000) | 0.63 [0.49, 0.78] | no evidence of predictability from sound; AUC at most 0.77 with 95% confidence |
3a. Per-family detail (point estimate [95% CI])
| target | meta | qc | gem | hc | ast | clap | ast+hc | all_audio | meta+all_audio | lab+meta | lab+meta+all_audio | ensemble |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| any_pathogen_present | 0.90 [0.88, 0.92] | 0.81 [0.78, 0.84] | 0.53 [0.49, 0.57] | 0.79 [0.76, 0.83] | 0.82 [0.78, 0.85] | 0.78 [0.74, 0.81] | 0.82 [0.78, 0.84] | 0.81 [0.78, 0.84] | 0.82 [0.79, 0.85] | 1.00 [1.00, 1.00] | 0.82 [0.79, 0.85] | 0.83 [0.80, 0.86] |
| cbpv_log_level | +0.20 [+0.15, +0.25] | +0.03 [-0.04, +0.09] | -0.04 [-0.10, +0.01] | +0.02 [-0.05, +0.09] | +0.12 [+0.06, +0.18] | +0.12 [+0.06, +0.18] | +0.10 [+0.04, +0.15] | +0.13 [+0.07, +0.19] | +0.13 [+0.07, +0.19] | +0.27 [+0.19, +0.34] | +0.13 [+0.06, +0.18] | +0.13 [+0.07, +0.19] |
| cbpv_present | 0.81 [0.76, 0.86] | 0.72 [0.67, 0.78] | 0.45 [0.40, 0.52] | 0.66 [0.60, 0.72] | 0.67 [0.61, 0.74] | 0.61 [0.54, 0.68] | 0.65 [0.58, 0.72] | 0.64 [0.57, 0.72] | 0.65 [0.58, 0.72] | 1.00 [1.00, 1.00] | 0.66 [0.59, 0.73] | 0.68 [0.62, 0.75] |
| dwv_high | 0.67 [0.61, 0.73] | 0.55 [0.48, 0.62] | 0.43 [0.38, 0.51] | 0.59 [0.52, 0.68] | 0.63 [0.56, 0.70] | 0.60 [0.53, 0.67] | 0.61 [0.54, 0.68] | 0.61 [0.54, 0.68] | 0.61 [0.54, 0.68] | 1.00 [1.00, 1.00] | 0.63 [0.56, 0.70] | 0.62 [0.55, 0.70] |
| dwv_log_level | +0.50 [+0.46, +0.53] | +0.14 [+0.07, +0.21] | -0.01 [-0.08, +0.06] | +0.18 [+0.10, +0.26] | +0.23 [+0.15, +0.31] | +0.20 [+0.09, +0.30] | +0.23 [+0.14, +0.31] | +0.23 [+0.14, +0.32] | +0.24 [+0.15, +0.33] | +0.83 [+0.78, +0.88] | +0.26 [+0.18, +0.35] | +0.25 [+0.15, +0.34] |
| dwv_log_level_among_positive | +0.25 [+0.16, +0.34] | +0.11 [+0.01, +0.20] | +0.02 [-0.06, +0.11] | +0.11 [+0.01, +0.21] | +0.22 [+0.14, +0.30] | +0.19 [+0.09, +0.29] | +0.20 [+0.13, +0.29] | +0.19 [+0.09, +0.29] | +0.19 [+0.10, +0.29] | +0.93 [+0.90, +0.95] | +0.23 [+0.13, +0.33] | +0.19 [+0.09, +0.29] |
| dwv_present | 0.96 [0.94, 0.97] | 0.86 [0.84, 0.89] | 0.54 [0.50, 0.58] | 0.85 [0.81, 0.88] | 0.87 [0.83, 0.90] | 0.84 [0.81, 0.88] | 0.87 [0.83, 0.90] | 0.87 [0.83, 0.90] | 0.88 [0.85, 0.91] | 1.00 [1.00, 1.00] | 0.88 [0.85, 0.91] | 0.89 [0.86, 0.92] |
| kbv_high | 0.63 [0.47, 0.79] | 0.55 [0.44, 0.69] | 0.49 [0.37, 0.61] | 0.70 [0.58, 0.81] | 0.60 [0.49, 0.70] | 0.54 [0.41, 0.68] | 0.63 [0.52, 0.74] | 0.62 [0.51, 0.75] | 0.63 [0.52, 0.74] | 0.97 [0.91, 1.00] | 0.64 [0.53, 0.75] | 0.64 [0.51, 0.75] |
| kbv_log_level | +0.06 [-0.03, +0.15] | +0.03 [-0.04, +0.09] | +0.01 [-0.05, +0.07] | +0.08 [+0.02, +0.15] | +0.10 [+0.03, +0.17] | +0.09 [+0.01, +0.16] | +0.10 [+0.03, +0.17] | +0.11 [+0.04, +0.18] | +0.11 [+0.04, +0.18] | +0.08 [-0.01, +0.19] | +0.12 [+0.04, +0.19] | +0.10 [+0.03, +0.17] |
| kbv_log_level_among_positive | +0.42 [+0.22, +0.58] | +0.06 [-0.13, +0.24] | -0.09 [-0.26, +0.10] | +0.22 [+0.07, +0.36] | +0.25 [+0.08, +0.41] | +0.14 [-0.06, +0.34] | +0.26 [+0.09, +0.42] | +0.24 [+0.08, +0.39] | +0.24 [+0.07, +0.40] | +0.53 [+0.32, +0.69] | +0.25 [+0.07, +0.41] | +0.24 [+0.08, +0.39] |
| kbv_present | 0.73 [0.69, 0.77] | 0.69 [0.64, 0.73] | 0.52 [0.47, 0.59] | 0.65 [0.59, 0.71] | 0.69 [0.64, 0.73] | 0.66 [0.61, 0.71] | 0.68 [0.63, 0.73] | 0.68 [0.63, 0.74] | 0.69 [0.64, 0.74] | 1.00 [1.00, 1.00] | 0.70 [0.65, 0.75] | 0.70 [0.65, 0.75] |
| probe_batch_oct2023 | 1.00 [1.00, 1.00] | 0.85 [0.78, 0.90] | 0.46 [0.37, 0.56] | 0.85 [0.80, 0.90] | 0.85 [0.79, 0.90] | 0.86 [0.80, 0.91] | 0.87 [0.82, 0.92] | 0.86 [0.80, 0.92] | 0.91 [0.86, 0.95] | 1.00 [1.00, 1.00] | 0.93 [0.88, 0.96] | 0.89 [0.85, 0.93] |
| probe_country_is_nz | 1.00 [1.00, 1.00] | 0.91 [0.89, 0.93] | 0.52 [0.48, 0.55] | 0.89 [0.85, 0.93] | 0.92 [0.90, 0.94] | 0.89 [0.86, 0.92] | 0.93 [0.90, 0.95] | 0.93 [0.91, 0.95] | 0.95 [0.92, 0.96] | 1.00 [1.00, 1.00] | 0.95 [0.92, 0.96] | 0.94 [0.91, 0.96] |
| probe_month_cos | +0.98 [+0.98, +0.98] | +0.39 [+0.33, +0.44] | +0.12 [+0.05, +0.18] | +0.45 [+0.40, +0.50] | +0.46 [+0.41, +0.52] | +0.38 [+0.33, +0.44] | +0.46 [+0.40, +0.51] | +0.46 [+0.41, +0.52] | +0.51 [+0.46, +0.56] | +0.98 [+0.98, +0.98] | +0.51 [+0.46, +0.56] | +0.50 [+0.45, +0.55] |
| probe_month_sin | +0.98 [+0.97, +0.98] | +0.43 [+0.37, +0.48] | +0.05 [-0.01, +0.10] | +0.42 [+0.37, +0.48] | +0.41 [+0.36, +0.46] | +0.39 [+0.35, +0.44] | +0.44 [+0.39, +0.48] | +0.45 [+0.41, +0.50] | +0.54 [+0.50, +0.58] | +0.98 [+0.98, +0.98] | +0.54 [+0.50, +0.58] | +0.47 [+0.42, +0.51] |
| terminal_03 | 0.76 [0.66, 0.86] | 0.56 [0.46, 0.68] | 0.60 [0.53, 0.67] | 0.74 [0.65, 0.84] | 0.71 [0.61, 0.80] | 0.67 [0.56, 0.79] | 0.67 [0.55, 0.78] | 0.69 [0.58, 0.80] | 0.69 [0.58, 0.80] | 0.77 [0.68, 0.87] | 0.68 [0.57, 0.80] | 0.75 [0.65, 0.86] |
| terminal_46 | 0.71 [0.61, 0.82] | 0.60 [0.52, 0.70] | 0.61 [0.54, 0.67] | 0.65 [0.55, 0.76] | 0.73 [0.67, 0.81] | 0.67 [0.58, 0.76] | 0.70 [0.63, 0.79] | 0.67 [0.59, 0.76] | 0.67 [0.59, 0.76] | 0.73 [0.65, 0.84] | 0.67 [0.58, 0.76] | 0.71 [0.63, 0.80] |
| cbpv_high | 0.66 [0.55, 0.76] | 0.50 [0.40, 0.61] | 0.49 [0.37, 0.62] | 0.55 [0.42, 0.69] | 0.58 [0.45, 0.69] | 0.65 [0.54, 0.75] | 0.60 [0.47, 0.72] | 0.60 [0.48, 0.72] | 0.60 [0.48, 0.72] | 1.00 [1.00, 1.00] | 0.59 [0.47, 0.71] | 0.61 [0.49, 0.74] |
| cbpv_log_level_among_positive | +0.41 [+0.22, +0.56] | +0.15 [-0.05, +0.34] | -0.11 [-0.27, +0.09] | -0.05 [-0.24, +0.14] | +0.21 [+0.03, +0.36] | +0.22 [+0.01, +0.39] | +0.24 [+0.07, +0.37] | +0.25 [+0.07, +0.41] | +0.25 [+0.07, +0.40] | +0.49 [+0.28, +0.64] | +0.31 [+0.13, +0.46] | +0.17 [-0.03, +0.36] |
| terminal_4_6_only | 0.77 [0.65, 0.87] | 0.43 [0.33, 0.55] | 0.61 [0.46, 0.74] | 0.61 [0.41, 0.80] | 0.55 [0.44, 0.65] | 0.66 [0.54, 0.77] | 0.54 [0.39, 0.67] | 0.63 [0.52, 0.74] | 0.62 [0.51, 0.74] | 0.73 [0.63, 0.84] | 0.62 [0.51, 0.74] | 0.64 [0.50, 0.77] |
3b. Cross-site transfer and the October 2023 batch
| target | family | USA->NZ | NZ->USA | without Oct 2023 batch (ensemble members) |
|---|---|---|---|---|
| any_pathogen_present | all_audio | 0.41 [0.34, 0.49] | 0.45 [0.40, 0.52] | meta: 0.90 [0.88, 0.92]; all_audio: 0.82 [0.78, 0.85]; ensemble: 0.83 [0.80, 0.86] |
| cbpv_log_level | all_audio | -0.17 [-0.23, -0.10] | +0.06 [+0.00, +0.11] | meta: +0.20 [+0.14, +0.26]; all_audio: +0.14 [+0.08, +0.19]; ensemble: +0.15 [+0.09, +0.20] |
| cbpv_present | all_audio | - | - | meta: 0.81 [0.75, 0.86]; all_audio: 0.64 [0.57, 0.70]; ensemble: 0.68 [0.61, 0.74] |
| dwv_high | all_audio | - | - | meta: 0.61 [0.55, 0.67]; all_audio: 0.53 [0.46, 0.60]; ensemble: 0.56 [0.50, 0.63] |
| dwv_log_level | all_audio | +0.12 [+0.03, +0.21] | +0.15 [+0.07, +0.24] | meta: +0.54 [+0.50, +0.57]; all_audio: +0.27 [+0.18, +0.34]; ensemble: +0.26 [+0.16, +0.34] |
| dwv_log_level_among_positive | all_audio | - | - | meta: +0.22 [+0.12, +0.31]; all_audio: +0.13 [+0.04, +0.22]; ensemble: +0.12 [+0.04, +0.21] |
| dwv_present | all_audio | 0.44 [0.35, 0.52] | 0.53 [0.41, 0.66] | meta: 0.96 [0.94, 0.97]; all_audio: 0.87 [0.84, 0.90]; ensemble: 0.89 [0.85, 0.92] |
| kbv_high | all_audio | - | - | meta: 0.61 [0.40, 0.78]; all_audio: 0.50 [0.33, 0.65]; ensemble: 0.50 [0.34, 0.65] |
| kbv_log_level | all_audio | +0.08 [-0.02, +0.17] | +0.02 [-0.06, +0.11] | meta: +0.00 [-0.09, +0.10]; all_audio: +0.06 [-0.02, +0.12]; ensemble: +0.07 [-0.00, +0.14] |
| kbv_log_level_among_positive | all_audio | - | - | meta: +0.40 [+0.20, +0.56]; all_audio: +0.19 [+0.03, +0.36]; ensemble: +0.16 [-0.01, +0.30] |
| kbv_present | all_audio | 0.67 [0.57, 0.75] | 0.56 [0.51, 0.62] | meta: 0.70 [0.66, 0.74]; all_audio: 0.70 [0.64, 0.75]; ensemble: 0.70 [0.64, 0.75] |
| probe_country_is_nz | all_audio | - | - | meta: 1.00 [1.00, 1.00]; all_audio: 0.92 [0.90, 0.95]; ensemble: 0.94 [0.91, 0.97] |
| probe_month_cos | all_audio | -0.54 [-0.60, -0.46] | -0.54 [-0.59, -0.47] | meta: +0.98 [+0.98, +0.98]; all_audio: +0.48 [+0.42, +0.53]; ensemble: +0.52 [+0.47, +0.57] |
| probe_month_sin | all_audio | -0.16 [-0.26, -0.05] | -0.09 [-0.18, -0.00] | meta: +0.98 [+0.98, +0.98]; all_audio: +0.47 [+0.42, +0.51]; ensemble: +0.48 [+0.43, +0.53] |
| terminal_03 | all_audio | - | - | meta: 0.67 [0.55, 0.81]; all_audio: 0.66 [0.52, 0.79]; ensemble: 0.69 [0.55, 0.81] |
| terminal_46 | all_audio | - | - | meta: 0.70 [0.59, 0.82]; all_audio: 0.64 [0.54, 0.74]; ensemble: 0.66 [0.55, 0.76] |
| terminal_4_6_only | all_audio | - | - | meta: 0.74 [0.59, 0.88]; all_audio: 0.62 [0.45, 0.76]; ensemble: 0.63 [0.49, 0.78] |
4. What the audio actually encodes: site, season and recording conditions
- Probe: recording made in NZ (1) vs USA (0); measures site/recording-condition detectability: best audio family all_audio 0.93 [0.91, 0.95].
- Probe: recording belongs to the single-visit October 2023 USA batch (1) vs all other recordings (0): best audio family all_audio 0.86 [0.80, 0.92].
- Probe: seasonal phase (sin) of the recording month: best audio family all_audio +0.45 [+0.41, +0.50].
- Probe: seasonal phase (cos) of the recording month: best audio family ast +0.46 [+0.41, +0.52].
Recording-quality statistics by batch (median per site-year-month; full table in artifacts/results/qc_by_batch.csv):
| batch | n | Terminal(4-6) rows | duration s | RMS dB | noise floor dB | level range dB |
|---|---|---|---|---|---|---|
| NZ_2022_08 | 47 | 0 | 120.3 | -48.8 | -52.5 | 5.7 |
| NZ_2022_09 | 46 | 1 | 120.9 | -53.4 | -57.9 | 7.4 |
| NZ_2022_10 | 48 | 1 | 120.8 | -51.3 | -54.7 | 7.4 |
| NZ_2022_11 | 47 | 2 | 120.7 | -54.3 | -57.9 | 7.0 |
| NZ_2022_12 | 47 | 2 | 120.6 | -52.4 | -57.1 | 6.3 |
| NZ_2023_01 | 48 | 2 | 120.4 | -56.9 | -60.1 | 5.5 |
| NZ_2023_02 | 48 | 2 | 120.8 | -55.7 | -59.7 | 4.3 |
| NZ_2023_03 | 47 | 1 | 120.6 | -58.6 | -61.6 | 3.7 |
| NZ_2023_04 | 43 | 1 | 120.8 | -62.4 | -63.9 | 2.5 |
| NZ_2023_05 | 43 | 0 | 120.5 | -63.4 | -64.5 | 2.4 |
| NZ_2023_06 | 41 | 0 | 120.5 | -62.5 | -63.8 | 2.3 |
| NZ_2023_07 | 46 | 0 | 120.3 | -60.3 | -62.7 | 3.7 |
| NZ_2023_11 | 1 | 0 | 121.2 | -50.6 | -52.8 | 4.5 |
| USA_2022_08 | 49 | 4 | 125.5 | -42.6 | -48.9 | 9.6 |
| USA_2022_09 | 50 | 5 | 121.2 | -51.3 | -54.4 | 5.9 |
| USA_2022_10 | 47 | 5 | 121.9 | -57.0 | -60.8 | 6.8 |
| USA_2022_11 | 47 | 6 | 121.8 | -57.4 | -61.1 | 7.5 |
| USA_2022_12 | 45 | 6 | 125.8 | -56.5 | -59.5 | 4.7 |
| USA_2023_01 | 45 | 5 | 136.8 | -53.7 | -59.2 | 6.0 |
| USA_2023_02 | 44 | 5 | 133.4 | -53.3 | -59.3 | 6.6 |
| USA_2023_03 | 43 | 4 | 120.9 | -54.6 | -59.1 | 7.2 |
| USA_2023_04 | 44 | 4 | 130.2 | -53.9 | -57.8 | 5.6 |
| USA_2023_05 | 41 | 1 | 132.5 | -54.8 | -57.4 | 4.9 |
| USA_2023_06 | 40 | 0 | 123.1 | -53.1 | -56.6 | 5.0 |
| USA_2023_07 | 39 | 0 | 121.0 | -52.1 | -54.9 | 5.1 |
| USA_2023_10 | 45 | 12 | 121.7 | -55.6 | -62.5 | 7.4 |
5. Gemini
- Structured listening (model gemini-3.8-flash (1123 recordings), 30 s excerpt, temperature 0) was run twice on 1125 recordings. Agreement between the two passes: 'is honeybee colony audio' 65% (kappa 0.26), activity level 54% (kappa 0.23), recording quality 47% (kappa 0.18). Share of recordings called honeybee audio: pass 1 62%, pass 2 64%; by site: NZ 64%, USA 59%.
- Predictive value of the Gemini listening features (family 'gem', linear model, colony-grouped CV):
- any_pathogen_present: 0.53 [0.49, 0.57]
- cbpv_log_level: -0.04 [-0.10, +0.01]
- cbpv_present: 0.45 [0.40, 0.52]
- dwv_high: 0.43 [0.38, 0.51]
- dwv_log_level: -0.01 [-0.08, +0.06]
- dwv_log_level_among_positive: +0.02 [-0.06, +0.11]
- dwv_present: 0.54 [0.50, 0.58]
- kbv_high: 0.49 [0.37, 0.61]
- kbv_log_level: +0.01 [-0.05, +0.07]
- kbv_log_level_among_positive: -0.09 [-0.26, +0.10]
- kbv_present: 0.52 [0.47, 0.59]
- probe_batch_oct2023: 0.46 [0.37, 0.56]
- probe_country_is_nz: 0.52 [0.48, 0.55]
- probe_month_cos: +0.12 [+0.05, +0.18]
- probe_month_sin: +0.05 [-0.01, +0.10]
- terminal_03: 0.60 [0.53, 0.67]
- terminal_46: 0.61 [0.54, 0.67]
- cbpv_high: 0.49 [0.37, 0.62]
- cbpv_log_level_among_positive: -0.11 [-0.27, +0.09]
- terminal_4_6_only: 0.61 [0.46, 0.74]
- The app also lets Gemini write a summary, but only from a JSON evidence object of computed numbers; the numbers are displayed next to the text so every statement can be checked.
6. Factors associated with colony death (observational)
Full tables: artifacts/results/survival_report.md. Key estimates (odds ratio per unit log(1+load) or per presence, 95% CI):
- GEE, PRIMARY: without single-visit (Oct 2023) colonies, presence indicators (n=1119, events=57 rows): pos_CBPV OR 0.95 [0.45, 2.00]; pos_DWV OR 0.86 [0.80, 0.92]; pos_KBV OR 1.11 [0.85, 1.46]; is_nz OR 0.20 [0.04, 1.01].
- GEE, sensitivity: all recordings (includes 12 single-visit colonies that all died), log levels (n=1131, events=69 rows): log1p_CBPV OR 0.68 [0.51, 0.92]; log1p_DWV OR 1.27 [1.09, 1.49]; log1p_KBV OR 1.62 [1.25, 2.10]; is_nz OR 0.08 [0.01, 0.56].
- GEE, sensitivity: all recordings (includes 12 single-visit colonies that all died), presence indicators (n=1131, events=69 rows): pos_CBPV OR 0.96 [0.26, 3.52]; pos_DWV OR 3.75 [1.39, 10.13]; pos_KBV OR 1.43 [1.13, 1.82]; is_nz OR 0.02 [0.00, 0.19].
- GEE, USA only, without single-visit colonies, log levels (n=567, events=45 rows): log1p_KBV OR 6.51 [2.14, 19.79].
- Firth (Wald CI), PRIMARY: without single-visit (Oct 2023) colonies, log levels (n=1119, event colonies=12, separation=False): log1p_CBPV OR 2.43 [0.92, 6.41]; log1p_DWV OR 0.23 [0.03, 1.67]; log1p_KBV OR 3.33 [1.44, 7.69]; is_nz OR 0.28 [0.15, 0.54].
- Firth (Wald CI), PRIMARY: without single-visit (Oct 2023) colonies, presence indicators (n=1119, event colonies=12, separation=False): pos_CBPV OR 0.87 [0.23, 3.35]; pos_DWV OR 0.39 [0.12, 1.28]; pos_KBV OR 1.49 [0.72, 3.07]; is_nz OR 0.63 [0.21, 1.91].
- Identifiability (Terminal colonies positive / Terminal colonies, outcome y46): CBPV NZ 1/2; CBPV USA 0/22; DWV NZ 2/2; DWV USA 12/22; KBV NZ 1/2; KBV USA 17/22.
- Colony level (died vs survived): CBPV ever positive 2/24 vs 45/88 (Fisher p 0.000); DWV ever positive 14/24 vs 54/88 (Fisher p 0.817); KBV ever positive 19/24 vs 54/88 (Fisher p 0.147).
- terminal_03: laboratory+metadata AUC 0.77 [0.68, 0.87]; adding audio 0.68 [0.57, 0.80]; gain -0.087 [-0.186, +0.003].
- terminal_46: laboratory+metadata AUC 0.73 [0.65, 0.84]; adding audio 0.67 [0.58, 0.76]; gain -0.064 [-0.160, +0.036].
Interpretation: the 12 single-visit colonies were each visited once and all died; the analysis treats them as selected on the outcome (an assumption: the workbook records no sampling design). They supply 12 of the 12 DWV-positive Terminal rows in the USA. The primary specification therefore excludes them; read the all-recordings odds ratios as a sensitivity run only. Varroa, the vector of DWV, is unmeasured, so the pathway Varroa -> DWV -> death cannot be tested with this workbook.
7. What to collect next
- Varroa mite counts (alcohol wash or sugar roll per 300 bees) at every recording visit. Varroa is absent from the workbook and is the vector of DWV. Verified reference: In the four-year German Bee Monitoring study (winters 2004/05 to 2007/08) of more than 1200 colonies from about 120 apiaries, high Varroa infestation in October, detection of DWV and ABPV in autumn bee samples, older queens, and lower colony strength in October were each significantly associated with overwinter colony loss, whereas Nosema infection and pesticide residues in bee bread showed no significant effect. (Genersch E, von der Ohe W, Kaatz H, Schroeder A, Otten C, Buchler R, Berg S, Ritter W, Muhlen W, Gisder S, Meixner M, Liebig G, Rosenkranz P (2010) The German bee monitoring project: a long term study to understand periodically high winter losses of honey bee colonies. Apidologie 41:332-352, DOI 10.1051/apido/2010014. URL: https://www.uni-hohenheim.de/fileadmin/einrichtungen/bienenmonitoring/Dokumente/DeBiMo_Studie_Apidologie_April2011.pdf. Verbatim quotes: "the occurrence of DWV and ABPV was significantly higher in lost than in surviving colonies" (Results 3.2); "No effects could be observed for Nosema spec. or pesticides." (Abstract)).
- Colony strength at each visit (frames of bees, brood area, adult population estimate) and queen status. These are the colony states with peer-reviewed evidence of acoustic detectability (queen loss, swarming) and the most plausible mediators of any sound-survival link.
- A recording protocol that fixes the confounds seen here: same recorder and gain at both sites, microphone position inside the hive, time of day and ambient temperature logged, and a short silence sample per visit. The audio currently identifies site and season far better than any pathogen, so recording conditions must be equalised before acoustic pathogen claims are possible.
- Pathogen sampling with matched timing and units: the workbook mixes zero-inflated loads whose units are not documented; record the assay, the unit and the detection limit, and sample the same bees that were recorded.
- Longer follow-up of the same colonies (outcome at 12 months) and more dying colonies with monthly series; of the 12 deaths that are not single-visit colonies, only 9 have a monthly series, too few for a reliable survival model of any kind.
Factors associated with colony death (observational)
Recordings: 1131, colonies: 112, colonies that died: 24 (of which 12 are single-visit October 2023 colonies).
0. Identifiability: positive Terminal rows and colonies per pathogen and country
| pathogen | outcome | country | Terminal rows positive / Terminal rows | Terminal colonies positive / Terminal colonies | Survivor rows positive / Survivor rows |
|---|---|---|---|---|---|
| CBPV | y03 | NZ | 2/6 | 1/2 | 106/546 |
| CBPV | y03 | USA | 0/39 | 0/22 | 3/540 |
| CBPV | y46 | NZ | 2/12 | 1/2 | 106/540 |
| CBPV | y46 | USA | 0/57 | 0/22 | 3/522 |
| DWV | y03 | NZ | 4/6 | 2/2 | 485/546 |
| DWV | y03 | USA | 12/39 | 12/22 | 7/540 |
| DWV | y46 | NZ | 9/12 | 2/2 | 480/540 |
| DWV | y46 | USA | 12/57 | 12/22 | 7/522 |
| KBV | y03 | NZ | 1/6 | 1/2 | 34/546 |
| KBV | y03 | USA | 18/39 | 16/22 | 72/540 |
| KBV | y46 | NZ | 1/12 | 1/2 | 34/540 |
| KBV | y46 | USA | 21/57 | 17/22 | 69/522 |
An odds ratio identified by fewer than about five positive Terminal colonies should not be interpreted.
1. Recording-level GEE logistic regression (clustered by colony)
PRIMARY: without single-visit (Oct 2023) colonies, log levels - outcome y46, n=1119, events=57, colonies=100
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| GEE did not converge within 200 iterations (statsmodels reports converged = False); see the Firth estimates in 1b, which also report whether separation was detected. |
PRIMARY: without single-visit (Oct 2023) colonies, presence indicators - outcome y46, n=1119, events=57, colonies=100
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| pos_CBPV | 0.95 | 0.45 to 2.00 | 0.897 |
| pos_DWV | 0.86 | 0.80 to 0.92 | 0.000 |
| pos_KBV | 1.11 | 0.85 to 1.46 | 0.446 |
| is_nz | 0.20 | 0.04 to 1.01 | 0.052 |
| month_sin | 1.46 | 1.15 to 1.86 | 0.002 |
| month_cos | 1.30 | 1.16 to 1.44 | 0.000 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), log levels - outcome y46, n=1131, events=69, colonies=112
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| log1p_CBPV | 0.68 | 0.51 to 0.92 | 0.012 |
| log1p_DWV | 1.27 | 1.09 to 1.49 | 0.003 |
| log1p_KBV | 1.62 | 1.25 to 2.10 | 0.000 |
| is_nz | 0.08 | 0.01 to 0.56 | 0.011 |
| month_sin | 1.36 | 1.11 to 1.67 | 0.003 |
| month_cos | 1.29 | 1.18 to 1.41 | 0.000 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), presence indicators - outcome y46, n=1131, events=69, colonies=112
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| pos_CBPV | 0.96 | 0.26 to 3.52 | 0.956 |
| pos_DWV | 3.75 | 1.39 to 10.13 | 0.009 |
| pos_KBV | 1.43 | 1.13 to 1.82 | 0.003 |
| is_nz | 0.02 | 0.00 to 0.19 | 0.001 |
| month_sin | 1.31 | 1.05 to 1.64 | 0.019 |
| month_cos | 1.29 | 1.17 to 1.43 | 0.000 |
USA only, without single-visit colonies, log levels - outcome y46, n=567, events=45, colonies=50
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| log1p_KBV | 6.51 | 2.14 to 19.79 | 0.001 |
| month_sin | 1.47 | 1.12 to 1.92 | 0.005 |
| month_cos | 1.26 | 1.11 to 1.43 | 0.000 |
| log1p_CBPV | not estimable (only 3 non-zero rows) | ||
| log1p_DWV | not estimable (only 7 non-zero rows) |
NZ only, without single-visit colonies, log levels - outcome y46, n=552, events=12, colonies=50
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| GEE did not converge within 200 iterations (statsmodels reports converged = False); see the Firth estimates in 1b, which also report whether separation was detected. |
PRIMARY: without single-visit (Oct 2023) colonies, log levels - outcome y03, n=1119, events=33, colonies=100
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| GEE did not converge within 200 iterations (statsmodels reports converged = False); see the Firth estimates in 1b, which also report whether separation was detected. |
PRIMARY: without single-visit (Oct 2023) colonies, presence indicators - outcome y03, n=1119, events=33, colonies=100
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| pos_CBPV | 1.70 | 0.52 to 5.53 | 0.377 |
| pos_DWV | 0.47 | 0.29 to 0.75 | 0.001 |
| pos_KBV | 1.52 | 0.85 to 2.71 | 0.154 |
| is_nz | 0.32 | 0.06 to 1.80 | 0.198 |
| month_sin | 1.82 | 1.14 to 2.92 | 0.012 |
| month_cos | 1.00 | 0.79 to 1.28 | 0.974 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), log levels - outcome y03, n=1131, events=45, colonies=112
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| log1p_CBPV | 0.59 | 0.38 to 0.92 | 0.021 |
| log1p_DWV | 1.47 | 1.15 to 1.89 | 0.002 |
| log1p_KBV | 1.96 | 1.35 to 2.84 | 0.000 |
| is_nz | 0.06 | 0.01 to 0.56 | 0.014 |
| month_sin | 1.54 | 1.08 to 2.19 | 0.017 |
| month_cos | 1.13 | 0.91 to 1.39 | 0.259 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), presence indicators - outcome y03, n=1131, events=45, colonies=112
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| pos_CBPV | 2.12 | 0.34 to 13.37 | 0.422 |
| pos_DWV | 11.27 | 3.35 to 37.94 | 0.000 |
| pos_KBV | 2.54 | 1.46 to 4.43 | 0.001 |
| is_nz | 0.01 | 0.00 to 0.07 | 0.000 |
| month_sin | 1.44 | 0.86 to 2.40 | 0.166 |
| month_cos | 1.10 | 0.80 to 1.52 | 0.556 |
USA only, without single-visit colonies, log levels - outcome y03, n=567, events=27, colonies=50
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| log1p_KBV | 25.66 | 2.05 to 320.74 | 0.012 |
| month_sin | 1.73 | 1.05 to 2.86 | 0.031 |
| month_cos | 0.96 | 0.73 to 1.26 | 0.779 |
| log1p_CBPV | not estimable (only 3 non-zero rows) | ||
| log1p_DWV | not estimable (only 7 non-zero rows) |
NZ only, without single-visit colonies, log levels - outcome y03, n=552, events=6, colonies=50
| term | odds ratio | 95% CI | p |
|---|---|---|---|
| log1p_CBPV | 18.38 | 0.18 to 1877.37 | 0.217 |
| log1p_DWV | 0.00 | 0.00 to 92.01 | 0.095 |
| log1p_KBV | 336.34 | 1.06 to 107002.26 | 0.048 |
| month_sin | 11.70 | 2.61 to 52.48 | 0.001 |
| month_cos | 2.30 | 0.69 to 7.67 | 0.176 |
1b. Firth penalised logistic regression (finite under separation)
PRIMARY: without single-visit (Oct 2023) colonies, log levels - outcome y46, n=1119, events=57, colonies=100
Event colonies: 12. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_CBPV | 2.43 | 0.92 to 6.41 | 0.18 to 2122.44 | 0.90 |
| log1p_DWV | 0.23 | 0.03 to 1.67 | 0.00 to 1.65 | 0.11 |
| log1p_KBV | 3.33 | 1.44 to 7.69 | 1.35 to 818.00 | 0.99 |
| is_nz | 0.28 | 0.15 to 0.54 | 0.00 to 0.91 | 0.01 |
| month_sin | 1.22 | 0.83 to 1.80 | 0.60 to 2.29 | 0.69 |
| month_cos | 2.29 | 1.47 to 3.55 | 1.62 to 3.02 | 1.00 |
PRIMARY: without single-visit (Oct 2023) colonies, presence indicators - outcome y46, n=1119, events=57, colonies=100
Event colonies: 12. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| pos_CBPV | 0.87 | 0.23 to 3.35 | 0.11 to 2.57 | 0.42 |
| pos_DWV | 0.39 | 0.12 to 1.28 | 0.16 to 0.66 | 0.00 |
| pos_KBV | 1.49 | 0.72 to 3.07 | 0.53 to 3.80 | 0.76 |
| is_nz | 0.63 | 0.21 to 1.91 | 0.05 to 1.59 | 0.19 |
| month_sin | 1.16 | 0.79 to 1.70 | 0.54 to 2.15 | 0.60 |
| month_cos | 2.20 | 1.43 to 3.39 | 1.61 to 2.91 | 1.00 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), log levels - outcome y46, n=1131, events=69, colonies=112
Event colonies: 24. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_CBPV | 0.78 | 0.33 to 1.82 | 0.08 to 460.58 | 0.46 |
| log1p_DWV | 1.27 | 0.97 to 1.65 | 0.84 to 5.05 | 0.92 |
| log1p_KBV | 2.05 | 1.33 to 3.16 | 1.45 to 2164.86 | 1.00 |
| is_nz | 0.19 | 0.10 to 0.37 | 0.00 to 0.68 | 0.00 |
| month_sin | 1.07 | 0.73 to 1.55 | 0.47 to 1.87 | 0.55 |
| month_cos | 2.50 | 1.63 to 3.83 | 1.84 to 3.42 | 1.00 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), presence indicators - outcome y46, n=1131, events=69, colonies=112
Event colonies: 24. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| pos_CBPV | 0.78 | 0.20 to 3.09 | 0.08 to 2.61 | 0.41 |
| pos_DWV | 5.69 | 2.09 to 15.54 | 1.85 to 22.35 | 1.00 |
| pos_KBV | 2.34 | 1.26 to 4.33 | 1.08 to 4.90 | 0.99 |
| is_nz | 0.06 | 0.02 to 0.18 | 0.00 to 0.33 | 0.00 |
| month_sin | 1.00 | 0.70 to 1.44 | 0.43 to 1.68 | 0.50 |
| month_cos | 2.38 | 1.54 to 3.68 | 1.74 to 3.29 | 1.00 |
USA only, without single-visit colonies, log levels - outcome y46, n=567, events=45, colonies=50
Event colonies: 10. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_KBV | 45.31 | 3.26 to 630.29 | 2.51 to 5148.24 | 0.99 |
| month_sin | 1.22 | 0.79 to 1.90 | 0.45 to 2.63 | 0.71 |
| month_cos | 2.10 | 1.28 to 3.45 | 1.48 to 3.25 | 1.00 |
| log1p_CBPV | not estimable (only 3 non-zero rows) | |||
| log1p_DWV | not estimable (only 7 non-zero rows) |
NZ only, without single-visit colonies, log levels - outcome y46, n=552, events=12, colonies=50
Event colonies: 2. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_CBPV | 1.35 | 0.68 to 2.67 | - | - |
| log1p_DWV | 0.89 | 0.45 to 1.74 | - | - |
| log1p_KBV | 1.61 | 0.66 to 3.93 | - | - |
| month_sin | 1.24 | 0.54 to 2.83 | - | - |
| month_cos | 3.00 | 1.18 to 7.65 | - | - |
PRIMARY: without single-visit (Oct 2023) colonies, log levels - outcome y03, n=1119, events=33, colonies=100
Event colonies: 12. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_CBPV | 3.20 | 1.15 to 8.86 | 0.16 to 5119.65 | 0.88 |
| log1p_DWV | 0.18 | 0.02 to 1.44 | 0.00 to 1.85 | 0.24 |
| log1p_KBV | 4.14 | 1.73 to 9.92 | 1.15 to 1140.82 | 0.98 |
| is_nz | 0.23 | 0.10 to 0.57 | 0.00 to 0.69 | 0.00 |
| month_sin | 1.76 | 1.06 to 2.91 | 0.69 to 5.64 | 0.88 |
| month_cos | 1.37 | 0.82 to 2.30 | 0.85 to 2.49 | 0.88 |
PRIMARY: without single-visit (Oct 2023) colonies, presence indicators - outcome y03, n=1119, events=33, colonies=100
Event colonies: 12. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| pos_CBPV | 1.94 | 0.46 to 8.17 | 0.21 to 6.99 | 0.69 |
| pos_DWV | 0.24 | 0.05 to 1.06 | 0.12 to 0.34 | 0.00 |
| pos_KBV | 1.91 | 0.81 to 4.49 | 0.53 to 5.19 | 0.86 |
| is_nz | 0.72 | 0.18 to 2.84 | 0.12 to 2.69 | 0.29 |
| month_sin | 1.60 | 0.99 to 2.58 | 0.60 to 4.67 | 0.83 |
| month_cos | 1.28 | 0.77 to 2.11 | 0.85 to 2.28 | 0.87 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), log levels - outcome y03, n=1131, events=45, colonies=112
Event colonies: 24. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_CBPV | 0.75 | 0.31 to 1.82 | 0.20 to 1408.30 | 0.47 |
| log1p_DWV | 1.47 | 1.10 to 1.97 | 0.98 to 8.45 | 0.97 |
| log1p_KBV | 2.21 | 1.40 to 3.48 | 1.43 to 563.50 | 1.00 |
| is_nz | 0.12 | 0.05 to 0.30 | 0.00 to 0.41 | 0.00 |
| month_sin | 1.32 | 0.84 to 2.08 | 0.48 to 2.76 | 0.78 |
| month_cos | 1.60 | 0.98 to 2.60 | 1.05 to 2.69 | 0.98 |
sensitivity: all recordings (includes 12 single-visit colonies that all died), presence indicators - outcome y03, n=1131, events=45, colonies=112
Event colonies: 24. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| pos_CBPV | 1.74 | 0.39 to 7.74 | 0.15 to 6.92 | 0.68 |
| pos_DWV | 12.38 | 4.18 to 36.64 | 4.31 to 57.19 | 1.00 |
| pos_KBV | 3.13 | 1.55 to 6.31 | 1.46 to 7.74 | 1.00 |
| is_nz | 0.02 | 0.01 to 0.08 | 0.00 to 0.12 | 0.00 |
| month_sin | 1.27 | 0.81 to 1.97 | 0.43 to 2.59 | 0.71 |
| month_cos | 1.41 | 0.85 to 2.34 | 0.93 to 2.61 | 0.94 |
USA only, without single-visit colonies, log levels - outcome y03, n=567, events=27, colonies=50
Event colonies: 10. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_KBV | 54.30 | 3.71 to 794.37 | 4.17 to 1884.55 | 1.00 |
| month_sin | 1.50 | 0.86 to 2.59 | 0.50 to 7.57 | 0.81 |
| month_cos | 1.21 | 0.68 to 2.15 | 0.69 to 2.41 | 0.73 |
| log1p_CBPV | not estimable (only 3 non-zero rows) | |||
| log1p_DWV | not estimable (only 7 non-zero rows) |
NZ only, without single-visit colonies, log levels - outcome y03, n=552, events=6, colonies=50
Event colonies: 2. Separation detected: False.
| term | odds ratio | Wald 95% CI | cluster-bootstrap 95% CI | bootstrap fraction OR > 1 |
|---|---|---|---|---|
| log1p_CBPV | 1.81 | 0.88 to 3.75 | - | - |
| log1p_DWV | 1.35 | 0.71 to 2.57 | - | - |
| log1p_KBV | 1.20 | 0.36 to 3.99 | - | - |
| month_sin | 4.81 | 1.12 to 20.71 | - | - |
| month_cos | 2.48 | 0.68 to 9.04 | - | - |
2. Colony-level comparisons (died vs survived)
| pathogen | max log1p level, median died | median survived | Mann-Whitney p | ever positive, died | ever positive, survived | Fisher p |
|---|---|---|---|---|---|---|
| CBPV | 0.000 | 0.000 | 0.001 | 2/24 | 45/88 | 0.000 |
| DWV | 0.057 | 0.011 | 0.875 | 14/24 | 54/88 | 0.817 |
| KBV | 0.011 | 0.000 | 0.003 | 19/24 | 54/88 | 0.147 |
Median recordings per colony: died 2, survived 12 (dying colonies stop being recorded, so their maxima are taken over fewer months).
The pooled table above compares countries as much as death status (dying colonies are mostly USA, survivors half NZ) and its maxima are taken over unequal numbers of months. Length-independent comparison, within country, colonies with at least 6 recordings, summary = first 6 recordings:
| country | died / survived | pathogen | mean6 log1p, median died | median survived | Mann-Whitney p | positive in first 6, died | positive in first 6, survived | Fisher p |
|---|---|---|---|---|---|---|---|---|
| NZ | 2 / 48 | CBPV | 0.001 | 0.000 | 0.673 | 1/2 | 40/48 | 0.331 |
| NZ | 2 / 48 | DWV | 0.001 | 0.112 | 0.015 | 2/2 | 48/48 | 1.000 |
| NZ | 2 / 48 | KBV | 0.000 | 0.000 | 0.333 | 0/2 | 17/48 | 0.542 |
| USA | 5 / 40 | CBPV | 0.000 | 0.000 | 0.649 | 0/5 | 2/40 | 1.000 |
| USA | 5 / 40 | DWV | 0.000 | 0.000 | 0.559 | 0/5 | 3/40 | 1.000 |
| USA | 5 / 40 | KBV | 0.000 | 0.000 | 0.395 | 3/5 | 22/40 | 1.000 |
3. Timing within dying colonies
Dying colonies with a monthly series: 1, 17, 30, 32, 37, 49, 64, 66, 7.
One value per colony per period (mean log1p load). Paired: last 3 months before death vs the colony's earlier months. Unpaired, within country: dying colonies' last-3-month value vs colony-level survivors (never Terminal) in the same calendar months.
| pathogen | dying colonies (by country) | mean log1p last 3 months | mean log1p earlier | Wilcoxon p | country | dying / survivor colonies | median dying | median survivors | Mann-Whitney p |
|---|---|---|---|---|---|---|---|---|---|
| CBPV | 9 (NZ 2, USA 7) | 0.008 | 0.000 | 1.000 | NZ | 2 / 47 | 0.036 | 0.000 | - (too few) |
| CBPV | 9 (NZ 2, USA 7) | 0.008 | 0.000 | 1.000 | USA | 7 / 40 | 0.000 | 0.000 | 0.579 |
| DWV | 9 (NZ 2, USA 7) | 0.001 | 0.000 | 1.000 | NZ | 2 / 47 | 0.004 | 0.008 | - (too few) |
| DWV | 9 (NZ 2, USA 7) | 0.001 | 0.000 | 1.000 | USA | 7 / 40 | 0.000 | 0.000 | 0.405 |
| KBV | 9 (NZ 2, USA 7) | 0.035 | 0.008 | 0.344 | NZ | 2 / 47 | 0.001 | 0.000 | - (too few) |
| KBV | 9 (NZ 2, USA 7) | 0.035 | 0.008 | 0.344 | USA | 7 / 40 | 0.000 | 0.000 | 0.634 |
4. Does sound add to the laboratory values? (colony-grouped AUC)
| target | meta | lab+meta | audio only | lab+meta+audio | gain of audio over lab+meta (95% CI) |
|---|---|---|---|---|---|
| terminal_03 | 0.756 [0.66, 0.86] | 0.770 [0.68, 0.87] | 0.693 [0.58, 0.80] | 0.684 [0.57, 0.80] | -0.087 [-0.186, +0.003] |
| terminal_46 | 0.712 [0.61, 0.82] | 0.732 [0.65, 0.84] | 0.674 [0.59, 0.76] | 0.668 [0.58, 0.76] | -0.064 [-0.160, +0.036] |
Reading these tables
Outcome y46 means death within 0-6 months (it contains every y03 = death within 0-3 months row). Odds ratios above 1 mean higher odds of death within the horizon per unit of log(1 + load). Events are counted as rows; the number of independent death events is the number of event colonies. The data contain no Varroa mite counts, colony size, queen status or nutrition measurements, so Varroa, the vector of DWV, is an unmeasured variable in every model above.