Task-variable geometry of population activity across the IBL Brain Wide Map
Regional differences and error-trial weakening of a ~1.5-dimensional representation
1 Original question
Does the dimensionality of population representations in the Brain Wide Map track task performance, and does it differ across brain regions?
2 Introduction
Rigotti et al. (2013) showed that neurons in monkey prefrontal cortex have nonlinear mixed selectivity: their responses depend on combinations of task variables, not on each variable separately. This makes the population representation high-dimensional. Its dimensionality is measured as the number of linearly separable dichotomies of task conditions, or the rank of the condition × neuron matrix. Two of their observations motivate this question:
- dimensionality varies between cortical populations;
- dimensionality collapses on error trials, which links it to behaviour.
Later theory treated dimensionality as a trade-off. High-dimensional codes support flexible linear readout; low-dimensional, abstract codes generalise across conditions (Fusi, Miller & Rigotti 2016; Bernardi et al. 2020). Cross-validated spectral methods made it possible to estimate dimensionality without the variance contributed by trial noise (Stringer et al. 2019).
The IBL Brain Wide Map (IBL et al. 2025) recorded Neuropixels probes in 139 mice performing one standardised visual decision task. The task has these variables:
- stimulus side;
- contrast;
- the previous choice;
- a block prior;
- choice and feedback.
Earlier BWM work found that choice- and movement-related signals are spread across many brain regions, while stimulus encoding is more restricted. The BWM therefore makes it possible to ask Rigotti’s questions across the whole brain, and to test the link with behaviour on thousands of error trials.
3 Refinement of the question
3.1 Data, conditions and split
Data are the bwm_ephys 1.2.1 release: good units only, plus the trials table.
Conditions. Each condition is a cell of stimulus side × contrast group × previous choice, using correct trials only. This gives 12 conditions, as in Rigotti’s design. Contrast groups are low (≤ 6.25 %), mid (12.5–25 %) and high (100 %).
Why block and choice are not condition variables.
- Block: the block prior agrees with stimulus side on about 80 % of trials, so cells crossing side with block are almost empty.
- Choice: on correct trials it equals stimulus side.
Population. A population is the good units of one session in one Beryl region.
Data split. Subjects were split 30 / 70 into exploration and confirmation, stratified by lab. The split covered 139 subjects in total. All exploration figures use exploration subjects only.
3.2 From mixed selectivity and dichotomies to per-factor variance
Per-unit models. On example sessions, per-unit ANOVA-style models found few units with a significant nonlinear interaction (< 8 % of units in a window, against a shuffle rate of about 1 %). Most selective units were “pure” side-selective units around movement.
Rigotti’s dichotomy metric. It saturated on one strong axis. Every dichotomy aligned with stimulus side is decodable. An additive surrogate with no interactions at all reproduced the data’s dichotomy curve and its cvPCA spectrum. The fraction of decodable dichotomies therefore measures side-signal strength, not mixed selectivity. It was dropped.
cvPCA participation ratio (PR). We next used the PR of the cvPCA spectrum at a fixed 20 units, averaged over random unit subsets.
- A single random subset is unreliable (intraclass correlation 0.43).
- With 100 subsets, population rankings are stable (rank ρ = 0.98).
The parameter grid then showed that clipped PR is anti-correlated with signal strength (ρ ≈ −0.6 to −0.75): weaker populations look higher-dimensional. Synthetic populations of known PR confirmed the cause. Noise inflates clipped spectra at the signal-to-noise ratio of these data, and counting components above a shuffle null is biased the other way.
The moment PR was the only estimator that was unbiased in both synthetic and real data. It is the ratio of unbiased estimates of \((\mathrm{tr}\,C)^2\) and \(\mathrm{tr}(C^2)\), built from four independent trial quarters. On real data it is about 1.5, i.e. the 12-condition representation is close to one-dimensional.
Per-factor variance. Because the representation is about one-dimensional, the question was recast in terms of how much variance each task variable carries. Condition means are split by orthogonal projectors into four parts: side, contrast group, previous choice, and their interaction. The variance of each part is estimated without bias from independent trial halves, \(\mathrm{tr}((PA)^\top PB)/(Kn)\). It is expressed per unit as a fraction of single-trial variance.
Reliability (split-half, across populations):
| Term | Split-half reliability |
|---|---|
| side | 0.87–0.93 |
| contrast | 0.61–0.77 |
| previous choice | 0.25–0.34 |
| interaction | 0.15–0.22 |
Control window. The pre-stimulus window (−400–0 ms) carries almost no contrast variance, as expected.
Windows. Two windows were locked:
stim= stimulus onset + 50–200 ms;move= first movement −100 to +100 ms.
Windows after movement were excluded because feedback arrives a median 172 ms after movement onset.
3.3 Performance: from dimensionality collapse to a weaker choice axis
Test design. Decoders were trained on correct trials only and then scored on error trials at matched contrast, following the logic of Rigotti Fig. 5.
What the pilot found.
- On errors, the side axis follows choice, not stimulus. Median side AUC on errors, scored by stimulus side, was 0.39 (stim) and 0.29 (move); 0.5 is chance.
- Previous-choice decoding does not collapse: AUC 0.59 on correct vs 0.60 on error trials.
- Session performance is unrelated to the metrics (|ρ| ≤ 0.14).
What did differ is that the side decision value was about 10 % smaller in magnitude on errors.
Timing control. Error trials have longer reaction times (+7 %) and movement durations (+23 %). We therefore reweighted correct trials to the errors’ distribution over contrast × choice × RT tercile × movement-duration tercile. After matching:
- in the
stimwindow the weakening disappeared (ratio 0.95, p = 0.22); - in the
movewindow it remained (ratio 0.88, p < 10⁻⁴).
3.4 Regions: a hierarchical anatomical model
Regional differences were modelled with a linear mixed model.
- Response: \(\mathrm{asinh}(v/0.002)\).
- Fixed covariates: log mean firing rate and log trials per condition.
- Anatomical variance components (crossed, nested through unique labels): EphysCosmos (12 groups), Ephys100 (≈ 97 groups) and Beryl. The Ephys100 / EphysCosmos groupings come from data-driven electrophysiological mappings in a development version of
iblatlas. - Other variance components: lab, subject and session.
In exploration, anatomy explained about 30 % of the variance in side, contrast and total variance, mostly at the EphysCosmos / Ephys100 scale. Subject accounted for a third of the variance in stimulus-window contrast.
Ephys100 re-pooling. Populations were rebuilt per Ephys100 group, giving 20 % more units per population. Region effects agreed closely with the Beryl model (ρ 0.92–0.97). Mixing several Beryl regions within one population did not change the moment PR.
3.5 Revised question
The BWM 12-condition representation is about 1.5-dimensional and dominated by the side/choice axis. The question was therefore restated as two hypotheses:
- H2 (regions): the per-unit variance carried by stimulus side (around movement) and by stimulus contrast (after stimulus onset) differs across brain regions.
- H3 (performance): on error trials the choice axis is weaker than on correct trials, at matched contrast, reaction time and movement duration.
Two hypotheses were set aside:
- Nonlinear mixed selectivity (H1) is rare, and its estimates were unreliable.
- Spike vs LFP dimensionality (H5) was deferred.
4 Locked analysis plan
The plan was locked on 2026-10-11, before any confirmation data were loaded. The exploration pipeline was frozen and applied unchanged.
Eligibility.
- Units fire at ≥ 1 Hz in the window.
- Populations have ≥ 20 such units.
- Sessions have ≥ 5 correct trials in every one of the 12 conditions.
Primary family (3 tests, Holm–Bonferroni, α = 0.05):
| Test | Hypothesis | Window | Statistic |
|---|---|---|---|
| C1 | side variance differs across regions | move | LRT of the 3 anatomical components vs none; parametric bootstrap, 1000 simulations from the null fit |
| C2 | contrast variance differs across regions | stim | same |
| C3 | choice axis weaker on errors | move | per-subject mean log₂(|dv| error / RT×MT-matched correct); one-sided Wilcoxon signed-rank < 0 |
Choice-axis measure (C3). An L2 logistic decoder (C = 1) is trained on correct trials with 5-fold cross-validation, at 6.25 % and 12.5 % contrast. Error trials are scored by the mean of the fold models. Populations need ≥ 10 error trials. Subjects are the replicates.
Pre-registered fallbacks.
- Use the χ̄² p-value if more than 10 % of bootstrap fits fail.
- Drop the session component if the observed fit fails.
Secondary analyses (reported, not part of the family):
- variance components and level-wise LRTs;
- Ephys100 re-pooling;
- pooled moment PR per region;
- C4: regional differences in side AUC on errors;
- C3 in the stim window, and without matching;
- previous-choice decoding on errors, session performance, and interaction variance;
- no-firing-rate sensitivity.
5 Results of confirmatory analysis
The confirmation set has 696 candidate populations from 250 sessions and 85 subjects. Of these, 591 (move) and 590 (stim) population-windows met the eligibility rules.
All three primary hypotheses were confirmed. Neither fallback was needed: 2.8 % (C1) and 0.7 % (C2) of bootstrap fits failed, and every observed fit converged with all components included.
| Test | n | Statistic | p | Holm p | Effect |
|---|---|---|---|---|---|
| C1 side variance, regions | 591 populations, 85 subjects, 138 regions | LR = 117.2 | 0.001 (bootstrap floor) | 0.002 | anatomy 29 % of variance (exploration LR 26.8) |
| C2 contrast variance, regions | 590 populations, 85 subjects, 139 regions | LR = 70.6 | 0.001 (bootstrap floor) | 0.002 | anatomy 20 % of variance (exploration LR 18.8) |
| C3 choice axis on errors | 567 populations, 83 subjects | W = 566 | 4.6 × 10⁻⁸ | 1.4 × 10⁻⁷ | median ratio 0.88; 76 % of subjects < 1; rank-biserial r = −0.68 |
C1 and C2. The likelihood ratios are 2.6–4.4 times their exploration values, in line with the roughly 2.8-fold larger sample. Every bootstrap simulation fell far below the observed values. Region effects estimated independently in the two subject sets agree (ρ = 0.71 for side, 0.67 for contrast, across 77 shared Beryl regions).
- Side variance around movement is highest in MOs, MOp, CP, SCm, PL, GRN and IRN. It is lowest in hippocampus (CA1, CA3), IC and cerebellar cortex.
- Contrast variance after the stimulus is highest in visual cortex (VISp, VISa), SUB, PL and CP. It is lowest in cerebellar cortex (CENT2/3, ANcr), PO and VPM.
C3. On error trials the side/choice axis is about 12 % weaker than on correct trials with the same contrast, choice, reaction time and movement duration. This replicates exploration (ratio 0.84 on 36 subjects).
6 Post-hoc analyses
The pre-specified secondary analyses come first, followed by one post-hoc analysis that was not in the locked plan (cortico-thalamic hierarchy).
Scale of regional variation.
- All levels contribute. Each anatomical level adds variance on its own (level-wise boundary-χ² LRTs, mostly p ≤ 0.03), with the coarsest level (EphysCosmos) usually the strongest.
- Exception. For contrast variance in the move window, Ephys100 adds nothing (p = 0.25), while the Beryl level carries most of the effect (LR = 21).
- Subject effects. Subject is the largest single non-residual component for stimulus-window contrast (22 % of total variance).
- Firing-rate covariate. Removing it strengthens the regional effect for side (LR 117 → 158) and weakens contrast slightly (71 → 63). The rate covariate is therefore conservative, as anticipated.
Dimensionality. Signal-gated moment PR, pooled per coarse region, is 1.3–1.8 in 10 of 11 regions in both windows. The exception is hypothalamus, at about 2.5, which rests on only 5 populations. OLF and RHP intervals are wide. In the Ephys100 model, moment PR also differs by anatomy, more clearly around movement (LR 22, p = 5 × 10⁻⁶) than after the stimulus (LR 5.9, p = 0.02). However, its range is narrow.
Error trials.
- Choice-following. The side axis follows choice on errors (session median AUC by stimulus 0.33 in move and 0.41 in stim; both p < 10⁻¹⁴).
- C4 (regional differences in choice-following). Choice-following is strongest in medulla, pons, midbrain and cerebellum and weakest in hippocampus, olfactory areas and cortical subplate (LR = 20.9, χ̄² p = 2.7 × 10⁻⁵). Exploration had not supported this (p = 0.18), so it is a new result that needs independent replication.
- Stimulus window. The error-trial weakening is also present there (matched ratio 0.91, p = 1.7 × 10⁻⁷), unlike in exploration, where it was explained by timing.
- Matching barely changes C3. The unmatched ratios are 0.87 (move) and 0.91 (stim).
No Rigotti-style collapse. Previous-choice AUC does not drop on errors (median difference 0.006–0.009; p = 0.65 move, p = 0.11 stim).
Session performance. Within region, it correlates weakly with stimulus-window contrast variance, total variance and moment PR (Spearman ρ = 0.11–0.15, p = 0.002–0.016, uncorrected). It does not correlate with any movement-window metric.
Interaction variance is negligible: a median 0.4–0.5 % of single-trial variance. It exceeds the shuffle null in 29–35 % of populations.
Ephys100 re-pooling. Populations were rebuilt per Ephys100 group (565 populations), and the anatomical effects persist (all LRT p < 10⁻¹⁵). Region effects match the Beryl model (ρ 0.94–0.98, 66 groups). Pooling several Beryl regions into one population has no measurable effect on any metric (mixing-covariate p > 0.2).
6.1 Post-hoc: region effects vs cortico-thalamic hierarchy
This analysis was requested after unblinding, so it is descriptive.
- Hierarchy score. We used the anatomical hierarchy score of Harris et al. (2019): the Cre-confidence-corrected, full cortico-cortical + thalamo-cortical + cortico-thalamic model. It covers 61 cortical and thalamic areas.
- Mapping regions to scores. Beryl regions map directly onto Harris areas. Ephys100 groups that span several areas (e.g. VIS, DORpm, DORsm, AUD) get a score averaged over their units in the confirmation populations. Groups with fewer than 50 % of units in scored areas were excluded.
- Regression. We regressed the covariate-adjusted region effects on the score, separately per mapping, by weighted least squares (weights = populations per region).
Side variance around movement does not track hierarchy (slope 0.22, p = 0.17 for Beryl; 0.15, p = 0.52 for Ephys100). It peaks in mid-hierarchy sensorimotor cortex (MOs, MOp, SSs, SSp-bfd).
Contrast variance after the stimulus has a positive weighted slope (Beryl 0.52, p = 6 × 10⁻⁴; Ephys100 0.57, p = 5 × 10⁻⁴). However, the rank correlation is weak (ρ = 0.22 and 0.10, both not significant), and the slope disappears within cortex (cortex-only p = 0.98 and 0.88). It is produced by somatosensory and auditory thalamus (VPM, VPL, MG, PO), which sits low on the hierarchy and carries little visual-contrast variance. Visual cortex carries high contrast variance at every hierarchy level, from VISp to VISam. The two mappings give nearly parallel fits.
7 Discussion
Regions. Task-variable variance is organised anatomically at several scales, and the organisation replicates across independent sets of mice.
- Choice/side variance around movement is strongest in frontal motor cortex, dorsal striatum, superior colliculus and the medullary reticular formation. This matches the brain-wide but uneven choice and movement signals reported by the IBL Brain Wide Map and by Steinmetz et al. (2019).
- Contrast variance after the stimulus is concentrated in visual cortex and a few associative/striatal regions. This matches the more restricted stimulus encoding in the same studies.
Most anatomical variance sits at the coarse EphysCosmos / Ephys100 scale. Beryl subdivisions add measurable but smaller effects.
Performance. Errors are not explained by a collapse of dimensionality.
- The representation is already about 1.5-dimensional, so there is little to collapse.
- Variables other than side/choice keep their information on errors: previous-choice coding is intact.
- What changes is the choice axis itself. It follows the animal’s (wrong) choice, and does so about 12 % more weakly than on correct trials with the same timing. The decision is therefore committed with less population drive, which is consistent with low-confidence choices rather than a scrambled representation.
Session-level performance is at most weakly related to the representation metrics.
Relation to Rigotti et al. The high dimensionality and nonlinear mixed selectivity found in monkey PFC during a complex rule-based task did not appear here. This is not a contradiction: the IBL task has one dominant decision variable and short reaction times, so we expect low-dimensional, abstract-like geometry (Bernardi et al. 2020).
Hierarchy. Neither variable follows the cortico-thalamic connectivity hierarchy within cortex (post-hoc). Regional differences reflect modality and function (visual, motor) more than hierarchical position.
7.1 Caveats
- Low dimensionality limits the original question. PR differences between regions are small and partly carried by populations with low signal-to-noise. The confirmed effects are on per-factor variance, not on “dimensionality” in the strict sense.
- Side and choice are confounded on correct trials, and movement is fast. “Side variance” around movement includes motor and choice signals. It cannot be read as a pure stimulus code.
- Anatomy vs recording strategy. Labs targeted different regions. The lab component absorbs some of this (lab explains 2–10 % of variance), but residual confounding with probe trajectories cannot be excluded.
- Unreleased mapping. Ephys100 / EphysCosmos come from an unreleased version of
iblatlas, pinned to a specific commit. They enter only as nested variance components, and the results agree with the Beryl-only analyses. - Bootstrap floor. p-values for C1 and C2 are at the 1000-simulation floor; the true p-values are smaller (χ̄² p < 10⁻¹⁵).
- C4 and the stimulus-window weakening reversed between exploration and confirmation. They are hypotheses for future work, not established results.
- C3 depends on the decoder. It uses a single linear decoder trained on correct trials, so the “choice axis” is the correct-trial axis. Errors could also involve a rotation that this measure does not capture.
- Spike vs LFP comparison (H5) was not carried out.
7.2 Lessons for AI-assisted analyses of these data
- Test estimators on synthetic data at the data’s signal-to-noise ratio before using them. The standard clipped PR looked plausible but was dominated by noise.
- When one variable dominates, decompose the variance per factor. Summary dimensionality measures (dichotomy fraction, PR) mostly measured side-signal strength.
- Control for reaction and movement time in error-trial analyses. In exploration, an apparent error-trial effect in the stimulus window disappeared after matching.
- Use a hierarchical anatomical model. It answers “at which scale do regions differ” directly, and it handles the many sparsely sampled Beryl regions by partial pooling.
- Splitting by subject was essential. Two exploration-stage conclusions (C4 and the stimulus-window weakening) did not hold up on new subjects.
7.3 References
- Rigotti M, Barak O, Warden MR, Wang X-J, Daw ND, Miller EK, Fusi S (2013). The importance of mixed selectivity in complex cognitive tasks. Nature 497:585–590.
- Fusi S, Miller EK, Rigotti M (2016). Why neurons mix: high dimensionality for higher cognition. Curr Opin Neurobiol 37:66–74.
- Bernardi S, Benna MK, Rigotti M, Munuera J, Fusi S, Salzman CD (2020). The geometry of abstraction in the hippocampus and prefrontal cortex. Cell 183:954–967.
- Stringer C, Pachitariu M, Steinmetz N, Carandini M, Harris KD (2019). High-dimensional geometry of population responses in visual cortex. Nature 571:361–365.
- Harris JA, Mihalas S, Hirokawa KE, et al. (2019). Hierarchical organization of cortical and thalamic connectivity. Nature 575:195–202.
- Steinmetz NA, Zatka-Haas P, Carandini M, Harris KD (2019). Distributed coding of choice, action and engagement across the mouse brain. Nature 576:266–273.
- International Brain Laboratory et al. (2025). A brain-wide map of neural activity during complex behaviour. Nature.
8 Statistical information
All tests were run on the confirmation set. C1–C3 are the primary family (Holm correction); all other tests are uncorrected secondary analyses. For the LRTs, df is the number of variance components tested; the null distribution is a χ̄² mixture because variances are tested on their boundary at 0.
| Test | Figure | Comparison | Statistic | df / n | p | Effect size |
|---|---|---|---|---|---|---|
| C1 LRT, parametric bootstrap (1000 sims) | Primary tests, left | side variance (move): anatomy components vs none | LR = 117.2 | 3 components; n = 591 | 0.0010 (Holm 0.0020) | anatomy 28.5 % of variance |
| C2 LRT, parametric bootstrap (1000 sims) | Primary tests, middle | contrast variance (stim): anatomy vs none | LR = 70.6 | 3; n = 590 | 0.0010 (Holm 0.0020) | anatomy 20.1 % of variance |
| C3 one-sided Wilcoxon signed-rank | Primary tests, right | subject mean log₂(|dv| error / matched correct) < 0 (move) | W = 566 | n = 83 subjects | 4.6 × 10⁻⁸ (Holm 1.4 × 10⁻⁷) | median ratio 0.88; r = −0.68 |
| C1 / C2 χ̄² approximation | — | same as C1 / C2 | LR = 117.2 / 70.6 | 3 | 5 × 10⁻²⁶ / 6 × 10⁻¹⁶ | — |
| No-rate sensitivity, χ̄² LRT | — | C1 / C2 without firing-rate covariate | LR = 158.2 / 63.0 | 3 | 7 × 10⁻³⁵ / 2 × 10⁻¹⁴ | — |
| Level-wise LRT (boundary χ²) | Secondary, top left | one level dropped, side (move): EphysCosmos / Ephys100 / Beryl | LR = 9.1 / 3.6 / 5.3 | 1 | 0.001 / 0.029 / 0.011 | variance 0.14 / 0.14 / 0.10 |
| Level-wise LRT (boundary χ²) | Secondary, top left | one level dropped, contrast (stim) | LR = 16.6 / 2.6 / 15.4 | 1 | 2 × 10⁻⁵ / 0.053 / 4 × 10⁻⁵ | variance 0.17 / 0.07 / 0.08 |
| Spearman correlation | Secondary, bottom | region effects, exploration vs confirmation (side move / contrast stim) | ρ = 0.71 / 0.67 | 77 regions | — | — |
| C4 χ̄² LRT | Errors, left | side AUC on errors (move): anatomy vs none | LR = 20.9 | 3; n = 557 | 2.7 × 10⁻⁵ | median AUC 0.33 |
| One-sided Wilcoxon signed-rank | — | C3-type, stim window, matched / unmatched | — | 83 subjects | 1.7 × 10⁻⁷ / 3.5 × 10⁻⁸ | ratio 0.91 / 0.91 |
| One-sided Wilcoxon signed-rank | — | C3-type, move window, unmatched | — | 83 subjects | 1.3 × 10⁻¹¹ | ratio 0.87 |
| Wilcoxon signed-rank vs 0.5 | Errors, left | session side AUC on errors (move / stim) | — | 220 / 219 sessions | 4 × 10⁻³⁰ / 4 × 10⁻¹⁵ | median 0.33 / 0.41 |
| Wilcoxon signed-rank | Errors, right | previous-choice AUC, correct − error (move / stim) | — | 237 / 236 sessions | 0.65 / 0.11 | median 0.006 / 0.009 |
| Spearman correlation | — | session performance vs within-region deviation of contrast / total / moment PR (stim) | ρ = 0.13 / 0.11 / 0.15 | 489 populations | 0.005 / 0.016 / 0.002 | — |
| Spearman correlation | — | same, move window (side / contrast / total / PR) | ρ = 0.07 / −0.01 / 0.05 / 0.05 | 492 | 0.14 / 0.87 / 0.28 / 0.28 | — |
| Ephys100 re-pooling, χ̄² LRT | Ephys100 | side (move) / contrast (stim), Ephys100 + EphysCosmos vs none | LR = 124.3 / 67.1 | 2; n = 559 / 560 | 3 × 10⁻²⁸ / 8 × 10⁻¹⁶ | ρ with Beryl model 0.98 / 0.98 |
| Ephys100 re-pooling, χ̄² LRT | Ephys100 | log moment PR, move / stim | LR = 22.3 / 5.9 | 2; n = 454 / 440 | 5 × 10⁻⁶ / 0.02 | — |
| Post-hoc WLS regression on hierarchy score | Hierarchy | side (move) region effect, Beryl / Ephys100 | slope = 0.22 ± 0.16 / 0.15 ± 0.24 | 45 / 26 regions | 0.17 / 0.52 | R² = 0.04 / 0.02; ρ = 0.15 / 0.01 |
| Post-hoc WLS regression on hierarchy score | Hierarchy | contrast (stim) region effect, Beryl / Ephys100 | slope = 0.52 ± 0.14 / 0.57 ± 0.14 | 47 / 26 regions | 6 × 10⁻⁴ / 5 × 10⁻⁴ | R² = 0.23 / 0.40; ρ = 0.22 / 0.10 |
| Post-hoc WLS, cortex only | — | side (move) / contrast (stim), Beryl | slope = 0.43 / 0.00 | 32 / 33 regions | 0.21 / 0.98 | R² = 0.05 / 0.00 |
| Post-hoc WLS, cortex only | — | side (move) / contrast (stim), Ephys100 | slope = 0.47 / −0.03 | 24 / 24 regions | 0.30 / 0.88 | R² = 0.05 / 0.00 |
| Wald test | Ephys100, bottom left | mixing covariate (log effective number of Beryl regions), 8 fits | β ∈ [−0.10, 0.10] | — | all > 0.2 | — |
8.1 Methods summary
- Data. IBL Brain Wide Map,
bwm_ephys1.2.1 (good units, trials table). Anatomy uses Beryl (iblatlas); Ephys100 and EphysCosmos use the iblatlas development mapping (pinned commit), with MOs/MOp layer groups collapsed to Beryl. - Software. statsmodels
MixedLM(ML for LRTs, REML for variance components), scikit-learn logistic regression, SciPy. - Code. The code is numbered by step in
exploratory-analyses/andconfirmatory-analyses/of the project directory.