DANDI:000678

There was no discernible change in task performance, and response rates to the distractors were <10%

STAR Methods, Training and inclusion criteria, Stage 6

Distractor trials have outcome 4 (false alarm, a lick) or 8 (correct reject). Under reading (c), 2,446 of 47,037 stage 6 distractor trials had a lick, 5.2%. Under readings (a) and (b), 7 of the 8 mice with stage 6 sessions are below 10%; BW053 is at 16.2% pooled and 15.3% over sessions. "No discriminable change in performance" is not tested; the fraction correct of responded target trials differs between stages by at most 5.4 points.

we fit GLM-HMMs separately to each of the 13 mice that reached proficiency on the sensory discrimination task (8 auditory and 5 visual task performers). While there was diversity in the best number of states across subjects (3–5 states

Results, Mice switch between several performance states, paragraph 6; Figure S2

With 10 random initialisations per fold, as the paper, both readings of O1 give 3 states in five mice and 4 in eight, inside the claimed 3 to 5 as the criterion requires; no mouse has 5 states, so the largest count is 4 against the paper's 5. The MAP plateau is never larger than the MLE plateau. 38 of the 9,100 cross-validation fits stopped at 1,000 iterations.

How the test differed from the paper’s method

• The GLM-HMM of the paper's ssm fork is reimplemented (hulsey.py) with the same model, priors, initialisation scheme, EM updates and stopping rule; its random initial values differ from ssm's. The test log probability includes the log prior, as ssm's log_probability does.
• The visual stimulus values follow loadbehavior.get_stim_vals of the paper's code (uobrainflex at ef48985), which for a file with a visual_gabor_angle column maps the difficulty codes 0, 1 and 2 to the session's distinct values of abs(angle - 45) / 45, largest first, and otherwise to 1, 0.8 and 0.6. 118 of the 141 visual-task sessions get 1, 0.6 and 0.2 and the other 23 get 1, 0.8 and 0.6. The first version of this attempt used 1, 0.8 and 0.6 for every session.
• Fit on the 376 sessions that pass the inclusion criteria (see cohort-inclusion).

A one-state GLM-HMM, equivalent to assuming stationary performance during training, correctly predicted the choice of the example mouse on 70% of trials, while the best-fit 4-state GLM-HMM correctly predicted 78% of choices

Results, Mice switch between several performance states, paragraph 5; Figure 1G

BW046 gets 4 states under both readings of O1; one-state accuracy 70.2%, 4-state accuracy 78.1%, computed as in fig2d-prediction-accuracy.

How the test differed from the paper’s method

• As for figs2-n-states and fig2d-prediction-accuracy.

Each subject exhibited both optimal (state 1) and disengaged (state 2) states, along with at least one additional sub-optimal performance state

Results, Mice switch between several performance states, paragraph 6; Figure 2A

All 13 mice have an optimal, a disengaged and at least one other state, and each of the 21 other states is matched to a sub-optimal template. Mice per type are bias left 6, avoid right 5, bias right 6 and avoid left 4 (Figure 2A gives 6, 4, 5 and 5; not deciding). Under M3 the other states are matched only to the four sub-optimal templates, so that part of the criterion cannot fail; the distances to all six templates, printed for BW061 and BW048, put each of their states nearest the template of its label. In BW059 one state has no trial with a posterior of at least 0.8 in the sessions where the optimal state occurs, so its psychometric is all missing; Figure2.py's nansum would score it 0 against every template and mislabel the mouse's optimal state. That state (432 trials over all sessions) is typed on all sessions, as avoid right. The question is asked of the paper's authors (findings).

How the test differed from the paper’s method

• Assumption. A state with no discrete-state trial in the sessions where the optimal state occurs is typed on all of the mouse's sessions; Figure2.py has no rule for it (it would score the state as equal to every template). Affects one state of BW059.
• As for figs2-n-states.

Subjects spent the largest proportion of trials in the optimal performance state ... and only 13.1% ± 1.6% of trials were in an indeterminate state

Results, Mice switch between several performance states, paragraph 6; Figure 2B

What the paper reports

• optimal against disengaged occupancy: Wilcoxon signed-rank test, p = 2.4 × 10⁻⁴ (n = 13 mice)
• optimal against sub-optimal occupancy: Wilcoxon signed-rank test, p = 2.4 × 10⁻⁴ (n = 13 mice)
• optimal against indeterminate occupancy: Wilcoxon signed-rank test, p = 2.4 × 10⁻⁴ (n = 13 mice)

On sessions in which the optimal state occurs, the optimal state holds 50.5% +/- 3.3% of trials, disengaged 12.7%, sub-optimal 23.7% and indeterminate 13.1% +/- 1.6%, within the expected 13.1 +/- 1.6. The optimal share is larger than each other share in 13 of 13 mice (Wilcoxon signed-rank p = 0.00024 for each).

How the test differed from the paper’s method

• As for figs2-n-states.

the best-fit GLM-HMM correctly predicted 75.7% ± 1.7% of choices, a significant 10.8% ± 1.3% improvement over the classical static-performance assumption

Results, Mice switch between several performance states, paragraph 6; Figure 2D

What the paper reports

• Wilcoxon signed-rank tests, p = 2.4 × 10⁻⁴ (n = 13 mice)

Best model 76.2% +/- 1.6%, one state 65.1% +/- 2.2%, gain 11.0 +/- 1.3 points under reading (a) (17.8% +/- 2.5% relative under (b), outside). Better in 13 of 13 mice, Wilcoxon signed-rank p = 0.00024 (rank-sum p = 0.00025).

How the test differed from the paper’s method

• The paper's code computes the filtered posterior of the preceding test trials by re-running expected_states on each prefix; here one forward pass gives the same filtered probabilities.
• As for figs2-n-states.

The inverted-U relationship was found in nearly all subjects, and the pupil diameter that gave the highest probability for being in the optimal state was calculated for each

Results, Arousal and uninstructed movement, paragraph 2; Figure 3C; Figure S3

Reading (i) holds in 13 of 13 mice and reading (ii) in 13 of 13, so "nearly all" holds under (a) and (b). Polynomial degrees 2 to 6 were selected by the elbow rule.

How the test differed from the paper’s method

• As Figure3.py, a state absent from a bin with more than 5 trials leaves that bin missing, so a bin with no optimal trial is not a zero; this follows the code.
• As for figs2-n-states.

Subjects performing the auditory task had significantly smaller optimal pupil sizes than those performing the visual task

Results, Arousal and uninstructed movement, paragraph 2; Figure 3D

What the paper reports

• Wilcoxon rank-sum test, p = 0.008 (n = 8 auditory and 5 visual mice)

Median optimal pupil 0.649 of the session maximum for the 8 auditory mice and 0.739 for the 5 visual mice; Wilcoxon rank-sum p = 0.013 (exact Mann-Whitney p = 0.011; paper 0.008).

we found a prominent U relationship between the probability of being in the disengaged state and pupil diameter ... These relationships were maintained when considering only trials where the mice were stationary

Results, Arousal and uninstructed movement, paragraph 3; Figure 3E

On all trials, over 23 bins with at least 7 mice, the second-order fit to mean P(optimal) has a negative quadratic term (vertex -0.9% from the optimal pupil) and to P(disengaged) a positive one (vertex -1.6%). On stationary trials under reading (b), absolute speed below 0.02 m/s (32,495 trials, 21 bins), both hold (vertices -3.3 and -6.1). Counting the 214 trials at or below -0.02 m/s as stationary, as the first version did, changes the vertices by at most 0.1. Reading (a), speed exactly 0, leaves 1,821 trials and no bin with 7 mice, so it cannot be tested. Exact zeros depend on rounding in the filter (9,869 more trials have a speed below 1e-9 m/s that is not exactly 0), and a run on another machine counted 1,879.

How the test differed from the paper’s method

• Stationary under reading (b) is taken as absolute speed below 0.02 m/s, so that backward movement is not counted as stationary.

Using this combined movement index, a clear inverted-U pattern emerged, showing that optimal performance state occupancy peaked at intermediate levels of uninstructed movement, similar to that of pupil diameter (regression of movement index vs. optimal state, cross-validated r 2 = 0.63 ± 0.06

Results, Arousal and uninstructed movement, paragraph 5; Figure 3F; Figure S5D

What the paper reports

• movement index against pupil r2: Wilcoxon signed-rank test, p = 0.34 (n = 13 mice)

The mean P(optimal) over the 0.25 movement-index bins with at least 7 mice has a negative quadratic term with its vertex at 0.28, inside -4.25 to 3.75. The r2 is 0.574 +/- 0.057 under (a), within 0.63 +/- 0.06 and close to its lower bound, and 0.533 under (b), outside. It does not differ from pupil's (difference -0.12, p = 0.22 under (a); p = 0.17 under (b)). Over 10 shuffled splits per mouse the r2 is 0.573 (a) and 0.522 (b).

How the test differed from the paper’s method

• As for figs2-n-states.

There is a marked U-shaped relationship between both pupil diameter and movement index and the probability of being in the disengaged state

Figure 3F legend

The mean P(disengaged) over 34 movement-index bins with at least 7 mice has a positive quadratic term (0.0225) with its vertex at 0.08, inside -4.25 to 4.0.

we combined these two measures into a single continuous measure of uninstructed movement levels termed the “movement index,” which is equally correlated with each measure

Results, Arousal and uninstructed movement, paragraph 4; Figure S5B and S5C

The two correlations agree to better than 0.0001 in every mouse (as they must by construction), with r from 0.674 to 0.855; reading (b), above 0.5, holds, and reading (a), above 0.7, fails in BW071 (0.674).

a trend toward larger raw face motion energy during optimal state occupancy (optimal: 0.48 ± 0.02 a.u., disengaged: 0.43 ± 0.3 a.u., sub-optimal: 0.44 ± 0.03 a.u.; opt. vs. dis. p = 0.06, opt. vs. sub. p = 0.04

Results, Arousal and uninstructed movement, paragraph 5

What the paper reports

• optimal against disengaged: Wilcoxon signed-rank tests, p = 0.06 (n = 13 mice)
• optimal against sub-optimal: Wilcoxon signed-rank tests, p = 0.04 (n = 13 mice)

Reading (a) of O3 gives 0.482, 0.441 and 0.448 a.u., all within tolerance; optimal is the largest, above sub-optimal with p = 0.017 and above disengaged with p = 0.19 (paper 0.06). Reading (b) gives 0.485, 0.435 and 0.453 (p = 0.048 and 0.08).

but no significant differences in raw locomotion speed (optimal: 0.14 ± 0.02 m/s, disengaged: 0.14 ± 0.2 m/s, suboptimal: 0.14 ± 0.02 m/s; opt. vs. dis. p = 0.84, opt. vs. sub. p = 1

Results, Arousal and uninstructed movement, paragraph 5

What the paper reports

• optimal against disengaged: Wilcoxon signed-rank tests, p = 0.84 (n = 13 mice)
• optimal against sub-optimal: Wilcoxon signed-rank tests, p = 1 (n = 13 mice)

Optimal 0.136, disengaged 0.137, sub-optimal 0.136 m/s under reading (a), differences of at most 0.002 m/s, p = 0.84 and 0.89; reading (b) 0.137, 0.132 and 0.139 m/s, p = 0.68 and 0.95. Speeds are read as m/s although the file labels them cm/s (findings).

During optimal state epochs, the correlation between pupil diameter and the movement index is significantly reduced across subjects

Results, Arousal and uninstructed movement, paragraph 7; Figures 4C and 4D

What the paper reports

• optimal against disengaged (stars drawn by Figure4.py): Wilcoxon signed-rank tests, p < 0.001 (n = 13 mice)
• optimal against sub-optimal: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)
• disengaged against sub-optimal: Wilcoxon signed-rank tests, p < 0.05 (n = 13 mice)

Mean r 0.077 (optimal), 0.439 (disengaged), 0.267 (sub-optimal); optimal lower than disengaged in 13 of 13 (p = 0.00024) and than sub-optimal in 12 of 13 (p = 0.00049). Disengaged is higher than sub-optimal (p = 0.0081), as the figure marks.

This decrease in correlation is also seen when considering individual measures of face motion energy and locomotion speed

Results, Arousal and uninstructed movement, paragraph 7; Figure S6B and S6D

What the paper reports

• pupil and face motion energy correlation, optimal against other states: Wilcoxon signed rank tests, p < 0.05 (n = 13 mice)

Pupil with face, optimal lower than disengaged (p = 0.0024) and sub-optimal (p = 0.00049). Pupil with locomotion, lower than disengaged (p = 0.0017), not significantly lower than sub-optimal (p = 0.068). The face with locomotion correlation is also lower in the optimal state (p = 0.00049 and 0.01).

Correlations between all measures at the start of a trial and prior inter trial intervals are weak and centered around zero across mice. (D) Average evoked pupil responses of mice are not significantly different across performance states

Figure S1C and S1D

What the paper reports

• evoked pupil response across performance states: Kruskal-Wallis test, p = 0.54

Median r with the preceding ITI, pupil -0.044, locomotion -0.007, face -0.044, none different from zero (p = 0.13, 0.89, 0.27), so the ITI part holds under readings (a) and (b). The Kruskal-Wallis test of the per-mouse median evoked pupil response over the four groups gives H = 4.38, p = 0.22 (three discrete states, p = 0.15); group means -1.19, 1.57, 0.2 and -0.51% of the maximum.

How the test differed from the paper’s method

• The pupil at onset is the filtered value at the stimulus sample; FigureS1.py takes it from the trial's pupil trace, 1 ms later. A standard choice that cannot change the result.

weights for the quadratic terms fit to predict optimal state probability were negative for pupil diameter across mice ... and the weights were positive when fit to disengaged state probabilities ... This was also true for quadratic terms of locomotion speed, while face motion energy had weights centered around zero across subjects

Results, Arousal measures predict optimal performance state, paragraph 3; Figure S8A

Optimal fits, pupil squared, locomotion squared, pupil SD and locomotion SD weights negative (medians -1.11, -1.10, -0.26, -0.33; Wilcoxon against zero p = 0.00024, 0.0012, 0.0081, 0.027). Disengaged fits, the same weights positive (0.56, 0.64, 0.19, 0.20; p = 0.00049, 0.00024, 0.033, 0.0081). Face squared and face SD weights do not differ from zero (p 0.19 to 0.84). Reading (a) holds, and so does (b).

the only consistently weighted feature across fits to sub-optimal state probabilities was the interaction between pupil diameter and the movement index ( Figure S8B ; 0.42 ± 0.11), which was also significantly different than weights fit to optimal state probability (−0.95 ± 0.21

Results, Arousal measures predict optimal performance state, paragraph 3; Figure S8B

What the paper reports

• Wilcoxon rank-sum test, p = 3.4 × 10⁻⁶ (n = sub-optimal state fits against 13 optimal state fits)

Pupil-by-movement weight 0.41 +/- 0.10 over 21 sub-optimal fits and -0.95 +/- 0.21 over 13 optimal fits, rank-sum p = 3.8e-06; quadratic movement weight of the optimal fits -1.30. Among the sub-optimal weights only the interaction differs from zero (p below 0.001).

We found that for all subjects, performance states could be significantly predicted using the recorded arousal measures with high accuracy (opt. vs. dis. = 88.2% ± 1.8%; opt. vs. sub. = 77.7% ± 2%

Results, Arousal measures predict optimal performance state, paragraph 1; Figures 5B and 5C

Eight features unshuffled (condition 13), 88.9% +/- 1.7% over 13 optimal against disengaged classifications and 78.4% +/- 1.8% over 20 optimal against sub-optimal classifications; the "All measures" condition 0 gives 88.5% and 77.7%. All 33 classifications are significant under readings (a), (b) and (c). The paper has 19 sub-optimal classifications; this attempt's fits give 21 sub-optimal states, 20 with enough trials.

How the test differed from the paper’s method

• 3 runs per shuffled condition and 20 null runs, in place of 1000, to fit the time limit. The unshuffled condition does not vary between runs. With 20 null runs the null SD and its 95th percentile are estimated less precisely.
• Rests on this attempt's state fits (figs2-n-states).

Pupil diameter (value and trial-to-trial variability) was the best individual raw measure for state classification

Results, Arousal measures predict optimal performance state, paragraph 2; Figure S7

What the paper reports

• pupil against face motion energy (Results; the S7 legend prints p = 0.004): Wilcoxon signed-rank tests over optimal state classifications, p = 0.0039 (n = 32 classifications)
• pupil against locomotion speed (Results; the S7 legend prints p = 0.14): Wilcoxon signed-rank tests over optimal state classifications, p = 0.106 (n = 32 classifications)

Reading (a), accuracy, pupil 71.5%, face 66.1%, locomotion 68.1%; pupil above face in 24 of 33 (p = 0.0011) and above locomotion on average (p = 0.053; 0.051 in a run on another machine). The z-scores and reading (b) agree for face (p from 0.00029 to 0.0013); under (b) pupil is also above locomotion (p = 0.029 and 0.022).

How the test differed from the paper’s method

• As for fig5c-state-classification.

13 of 31 mice met inclusion criteria, with a total of 381 of their 500 stage 5 and 6 training sessions included for further analysis

STAR Methods, Training and inclusion criteria, last paragraph; Results, Mice switch between several performance states, paragraph 6

The dandiset holds 13 mice, 8 on the auditory and 5 on the visual task, each with at least 10 passing sessions. Of its 391 stage 5 and 6 sessions, 376 pass the STAR Methods criteria counted from the reward series and 375 counted from the hits, against 381 in the paper. The sessions failing either count are listed in the notebook; 14 fail both (13 with fewer than 100 rewards, 1 with less than 20% at one port) and 3 fail one count only. The paper's 500 stage 5 and 6 sessions and its 18 excluded mice are not in the dandiset.

How the test differed from the paper’s method

• The sessions of every later analysis are the 376 that pass the criteria by the reward count; the paper analysed 381, and the dandiset does not say which.

Additionally, epochs of the optimal state lasted longer than those of sub-optimal states ... and the median transition time between discrete states (dwell time in indeterminate states) was 3.4 ± 0.3 trials

Results, Mice switch between several performance states, paragraph 6; Discussion, paragraph beginning "Here, we used recently implemented GLM-HMMs"; Figure 2C

What the paper reports

• optimal against disengaged dwell times: Wilcoxon signed-rank test, p = 2.4 × 10⁻⁴ (n = 13 mice)
• optimal against sub-optimal dwell times: Wilcoxon signed-rank test, p = 2.4 × 10⁻⁴ (n = 13 mice)

Optimal epochs are longer than disengaged and sub-optimal epochs in 13 of 13 mice under every reading (p = 0.00024). The median indeterminate dwell is 3.31 +/- 0.35 trials, within 3.4 +/- 0.3. The 37 is not matched under any reading. Under reading (a) the optimal dwell is 43.8 trials with the code's dwell (y) and 42.8 as run length (x), above 40.7; the disengaged dwell is 14.0 (13.8), within 15 +/- 1.5, and the sub-optimal 19.1 (18.1), above it. Readings (b) and (c) give optimal dwells of 73.2 to 74.1 and 49.3 to 50.3. The measured values are the indeterminate dwell and the optimal and disengaged dwells under (a) and (y).

How the test differed from the paper’s method

• flex_hmm.get_dwell_times adds a last epoch of dwell 0 when a session's last trial differs from the one before it, which also adds a zero indeterminate dwell. The code's output is used, as the claims file reads it; without the 3 such epochs the indeterminate dwell and the medians are unchanged (shown in the notebook).
• As for figs2-n-states.

pupil diameter: r 2 = 0.72 ± 0.03; face motion energy: r 2 = 0.49 ± 0.08, Wilcoxon signed-rank test vs. pupil: p = 0.04, locomotion speed: r 2 = 0.44 ± 0.07, Wilcoxon signed-rank test vs. pupil: p = 0.002

Results, Arousal and uninstructed movement, paragraph 4; Figure S4

What the paper reports

• face motion energy against pupil r2: Wilcoxon signed-rank test, p = 0.04 (n = 13 mice)
• locomotion speed against pupil r2: Wilcoxon signed-rank test, p = 0.002 (n = 13 mice)

Pupil r2 0.690 +/- 0.039 (within 0.72 +/- 0.03). Face 0.428 +/- 0.092 with plotting bins (a), within, and 0.409 with 50 bins (b), just outside. Locomotion 0.567 (a), outside (0.570 in a run on another machine), and 0.413 (b), within. No bin reading puts all three in tolerance. Both are lower than pupil on average (by 0.12 to 0.28), but by Wilcoxon only locomotion under (b) is lower at p < 0.05 (p = 0.011); face p = 0.11 and 0.057, locomotion (a) p = 0.41. The criterion's r2 comes from one shuffled split per mouse, as 4_optimal_pupils.py; the mean over 10 splits gives pupil 0.682 +/- 0.043 (BW051 from 0.001 to 0.562 over the splits) and the same test results (face p = 0.15 and 0.068, locomotion p = 0.34 and 0.013). The measured values are pupil, face (a) and locomotion (a).

How the test differed from the paper’s method

• As for figs2-n-states.

These correlations translate to a significant decrease in the absolute difference from the calculated optimal pupil diameter during optimal state occupancy as compared to sub-optimal or disengaged states (optimal: 7% ± 0.3%, disengaged: 12.6% ± 1%, sub-optimal: 9% ± 0.5%

Results, Arousal and uninstructed movement, paragraph 5

What the paper reports

• optimal against disengaged: Wilcoxon signed-rank tests, p = 2.4 × 10⁻⁴ (n = 13 mice)
• optimal against sub-optimal: Wilcoxon signed-rank tests, p = 4.8 × 10⁻⁴ (n = 13 mice)

Under reading (a) of O3 the absolute distance is 8.8% +/- 1.0% in the optimal state, 16.0% in the disengaged and 11.8% in the sub-optimal states, lower in the optimal state than in the disengaged (13 of 13, p = 0.00024) and sub-optimal states (13 of 13, p = 0.00024); reading (b) gives 8.8, 15.8 and 11.5 (p = 0.00024 and 0.0012). The order and the tests hold; all three values are larger than the paper's 7, 12.6 and 9.

How the test differed from the paper’s method

• Rests on the optimal pupils of fig3c-inverted-u-pupil, from this attempt's state fits.

For all arousal measures recorded−pupil diameter, face motion energy, and locomotion speed−we found significantly reduced variability during optimal performance state occupancy as compared to sub-optimal or disengaged states

Results, Arousal and uninstructed movement, paragraph 6; Figure 4A

What the paper reports

• pupil, optimal against disengaged (stars drawn by Figure4.py): Wilcoxon signed-rank tests, p < 0.001 (n = 13 mice)
• pupil, optimal against sub-optimal: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)
• face motion energy, optimal against disengaged: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)
• face motion energy, optimal against sub-optimal: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)
• locomotion speed, optimal against disengaged: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)
• locomotion speed, optimal against sub-optimal: Wilcoxon signed-rank tests, p < 0.05 (n = 13 mice)

All six differences have the stated sign. Pupil, optimal against disengaged p = 0.00024 and against sub-optimal p = 0.0012; face p = 0.0024 and 0.017; locomotion p = 0.0012 and 0.057. Five of six tests pass.

Strikingly, this decrease in variability predicts transitions into and out of the optimal state

Results, Arousal and uninstructed movement, paragraph 6; Figure 4B

What the paper reports

• pupil, entry (as drawn by Figure4.py): Wilcoxon signed-rank tests, p = 0.07 (n = 13 mice)
• pupil, exit: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)
• face motion energy, entry: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)
• face motion energy, exit: Wilcoxon signed-rank tests, p = 0.12 (n = 13 mice)
• locomotion speed, entry: Wilcoxon signed-rank tests, p < 0.05 (n = 13 mice)
• locomotion speed, exit: Wilcoxon signed-rank tests, p < 0.01 (n = 13 mice)

All six changes have the stated sign on average over mice (entry, pupil 0.0319 to 0.0283, face 0.0750 to 0.0680, locomotion 0.0620 to 0.0545; exit, 0.0301 to 0.0346, 0.0713 to 0.0780, 0.0553 to 0.0640). Of the four comparisons the figure marks, pupil exit (p = 0.0012), locomotion entry (p = 0.04) and locomotion exit (p = 0.0012) pass, face entry (p = 0.11) does not. 9 to 85 entries and 18 to 125 exits per mouse.

Interestingly, we found that the best predictor of performance state was the computed movement index incorporating face motion energy and locomotion speed, which encoded more information about states compared to pupil diameter

Results, Arousal measures predict optimal performance state, paragraph 2; Figure 5C

What the paper reports

• Wilcoxon signed-rank tests over optimal state classifications, p = 0.024 (n = 32 classifications)

Main reading (movement index value and SD kept, all else shuffled) gives 68.6% against 71.5% for pupil alone, movement higher in 14 of 33 (p = 0.086); z 4.51 against 5.21 (p = 0.068). The movement index is not better than pupil. The code's conditions shown as variants give higher accuracy for the movement-index classifier, condition 5 (movement index kept with the pupil value) 75.7% (p = 1.3e-05) and condition 1 (face and locomotion shuffled) 77.9% (p = 2.3e-10), but both keep pupil information, so they compare pupil and movement together against pupil alone and do not test the claim.

How the test differed from the paper’s method

• As for fig5c-state-classification.