Claims 46
From the paper Validation of Decision-Making Models and Analysis of Decision Variables in the Rat Basal Ganglia · claims file · checked 2026-10-21: ready
- trial-counts
The analysis covers 60,221 trials in 70 recording sessions from six rats, of which 39,175 are choice trials, 12,707 no-choice trials and 8329 error trials.
Here, we report the results of all 60,221 trials in 70 recording sessions performed by six rats, consisting of 39,175 choice trials (65.1%), 12,707 no-choice trials (21.1%), and 8329 error trials (13.8%)
Results, Behavioral data, paragraph 1How it will be judged
The rat count must match exactly. Choice trials are successful tone-A trials, no-choice trials successful tone-B trials, and error trials those ended by the error tone (Behavioral task). Readings of the total are (a) all trials started; (b) trials in which a cue tone was presented. Reproduced when, under one reading, every count is within tolerance.
- block-structure
A block took 105.6 trials on average, and a session held 8.1 blocks on average, from 5 to 12.
On average, 105.6 trials were required in one block, and 8.1 blocks (from 5 to 12 blocks) were performed in one session
Results, Behavioral data, paragraph 1How it will be judged
The paper does not say which trials a block counts or whether a session's last, possibly unfinished, block counts. Readings are (a) all trials, including no-choice and error trials, every block; (b) as (a), completed blocks only; (c) choice trials only, every block; (d) as (c), completed blocks only. Reproduced when, under one reading, all four values are within tolerance (the range bounds are whole numbers, so the fewest blocks must be 5 and the most 11 to 13).
- reaction-time
In choice trials the time from center-hole exit to L or R hole entry was under 1 s in 72% of trials, with a median of 0.75 s.
The reaction time in the choice trials, the time period from the exit of the center hole to the entry of the left or right hole, was <1 s in 72% of the trials and had a median of 0.75 s
Results, Reaction time and tone discrimination, paragraph 2How it will be judged
Over all choice trials of the 70 sessions, pooled. Reproduced when both values are within tolerance.
- tone-discrimination-errors
Rats rarely failed to poke after tone A (0.39% of choice trials) or poked wrongly after tone B (0.73% of no-choice trials), so they discriminated the tones well.
The errors in the choice trials, in which rats did not perform any nose pokes within 1 min for tone A presentation, were only 0.39% of the choice trials. The errors in no-choice trials, in which rats performed a nose poke at the left or right hole within 1 s for tone B presentation, were 0.73% of the no-choice trials
Results, Reaction time and tone discrimination, paragraph 2How it will be judged
The denominators are not fixed. Readings are (a) the error count over the successful trials of that tone (39,175 and 12,707); (b) the error count over the successful trials of that tone plus those errors. Errors from leaving the center hole during the cue are not counted in either. Reproduced when, under one reading, both values are within tolerance.
- fig2b-reward-stronger-than-noreward
Both L1 and R0 in the previous trial raise the left-choice probability (0.68 and 0.62), more after L1; both R1 and L0 lower it (0.30 and 0.39), more after R1. So rewarded experience modulates choice more than unrewarded experience.
Both the experience L1 and R0 in the previous trial increased the probability of a left choice in the current trial [Fig. 2B, green arrow, P(L|L1) = 0.68; green arrowhead, P(L|R0) = 0.62]. The left choice probability after L1 was significantly higher than after R0 (Mann–Whitney U test, p < 0.0001) ... The left probability after R1 was significantly lower than after L0 [Fig. 2B, brown arrow, P(L|R1) = 0.30; brown arrowhead, P(L|L0) = 0.39] (Mann–Whitney U test, p < 0.0001)
Results, History dependence of action choice, paragraph 1; Figure 2BHow it will be judged
Conditional probabilities as the Methods define them (opening comment), pooled over all sessions; histories as O7; tests as O8. Reproduced when, under every reading, the four values are within tolerance, P(L|L1) > P(L|R0) > 0.5 > P(L|L0) > P(L|R1), and both tests give p < 0.0001.
- fig2b-two-back-history
The experience two trials back still affects choice; P(L|R0 L1) = 0.66 differs from P(L|L1 L1) = 0.72.
the conditional probability P(L|R0 L1) = 0.66 (Fig. 2B,C, blue arrowhead), which is the probability of a left choice (at trial t) after R0 (at trial t − 2) and L1 (at trial t − 1), significantly differs from P(L|L1 L1) = 0.72 (Fig. 2B,C, blue arrow) ... (Mann– Whitney U test, p < 0.0001)
Results, History dependence of action choice, paragraph 2; Figure 2B and 2CHow it will be judged
As fig2b-reward-stronger-than-noreward (O7, O8). Reproduced when, under every reading, both values are within tolerance, P(L|L1 L1) > P(L|R0 L1), and the test gives p < 0.0001.
- fig2c-sequence-frequencies
Of the 128 mirror-paired four-trial sequences, L1 L1 L1 L1 (with R1 R1 R1 R1) is the most frequent (6.87%), R0 L1 L1 L1 (with L0 R1 R1 R1) the second (2.87%), R0 L0 R0 L0 (with L0 R0 L0 R0) the third (2.55%), and L0 L0 L0 L0 with R0 R0 R0 R0 make 1.28%.
the sequence L1 L1 L1 L1 and R1 R1 R1 R1 was observed most frequently (6.87% of all observed sequences of four trials). The sequence R0 L1 L1 L1 and L0 R1 R1 R1 was observed second most frequently (2.87%) ... The sequences R0 L0 R0 L0 and L0 R0 L0 R0 were observed third most frequently (2.55%) ... L0 L0 L0 L0 and R0 R0 R0 R0, which consisted of 1.28% of all observed sequences across the four trials
Results, History dependence of action choice, paragraphs 2 and 3How it will be judged
All windows of four consecutive choice trials, histories as O7. Whether a pair's percentage is the sum of the two mirror sequences or each one's share is not fixed; readings are (a) the sum; (b) the mean of the two. Reproduced when, under one reading, all four values are within tolerance and the first three are the three most frequent pairs in that order.
- fig2c-reward-effects-fade
After repeated L1 the left-choice probability rises and converges near 0.75, and the effect of an earlier R0 fades under following L1s, so P(L|L1 L1 L1 L1) and P(L|R0 L1 L1 L1) do not differ (both 0.74).
increased with the number of L1 experiences and converged around 0.75 ... There was no significant difference between P(L|L1 L1 L1 L1) = 0.74 and P(L|R0 L1 L1 L1) = 0.74 (Mann–Whitney U test, p = 0.94)
Results, History dependence of action choice, paragraph 2; Figure 2CHow it will be judged
Prefix probabilities as the opening comment, histories as O7. The absence is judged on the paper's test, a Mann-Whitney U test at p < 0.05 (two-sided), under reading O8 (a), the 0/1 choices that follow each history; reading O8 (b) is shown as a variant. Reproduced when the test does not find a difference, both values are within tolerance, and P(L|L1) < P(L|L1 L1) < P(L|L1 L1 L1 L1), as the paper's own values are (0.68, 0.72, 0.74), with P(L|L1 L1 L1 L1) within 0.075 (10%) of 0.75; P(L|L1 L1 L1) is reported. The difference is reported either way.
- fig2d-switching
After an unrewarded choice rats tended to choose the other action (switching), as seen in the alternating unrewarded sequences R0 L0 R0 L0 and L0 R0 L0 R0.
These results illustrate that there was a “switching” behavior in which rats tended to choose the other action after one action was not rewarded
Results, History dependence of action choice, paragraph 3; Figure 2DHow it will be judged
Prefix probabilities for R0 L0 R0 L0 and L0 R0 L0 R0 (O7), with 99% intervals. Reproduced when, under every reading, after each prefix ending in L0 the left-choice probability is below 0.5 and after each prefix ending in R0 it is above 0.5. The paper gives no test; the 99% intervals are reported.
- fig2e-persevering
After two or more unrewarded left choices the left-choice probability gradually rises, and P(L|L0 L0 L0) and P(L|L0 L0 L0 L0) are both above 0.5, so rats sometimes persevere after repeated non-reward.
The probability of a left choice gradually increased after two or more left choices with no reward. In particular, P(L|L0 L0 L0) and P(L|L0 L0 L0 L0) were both >0.5
Results, History dependence of action choice, paragraph 3; Figure 2EHow it will be judged
Prefix probabilities for L0 L0 L0 L0 (and mirrored for R0 R0 R0 R0, as 1 - P(L)), O7. The paper gives no test. Reproduced when, under every reading, P(L|L0 L0) < P(L|L0 L0 L0) < P(L|L0 L0 L0 L0) ("gradually increased after two or more") and both P(L|L0 L0 L0) and P(L|L0 L0 L0 L0) exceed 0.5 for the left sequence; partially reproduced when this holds for the probabilities above 0.5 but not for the rise.
- fig3-lsq-switching
Fitted by least squares to the six four-trial sequences of Fig. 2C and 2D, the local matching law and standard Q-learning cannot represent switching, while F-Q and DF-Q can; the averaged residual errors are 0.0102, 0.0073, 0.0010 and 0.0002.
The results showed that the local matching law (Fig. 3A) and the standard Q-learning model (Fig. 3B) were not able to represent switching behavior, whereas the F-Q- and DF-Q-learning models were (Fig. 3C,D). The averaged residual errors of the local matching law, the standard Q, F-Q, and DF-Q models were 0.0102, 0.0073, 0.0010, and 0.0002, respectively
Results, Modeling rats' strategies with a reinforcement learning model, paragraph 3; Figure 3; Table 1How it will be judged
Data are the prefix probabilities of L1 L1 L1 L1, R1 R1 R1 R1, R0 L1 L1 L1, L0 R1 R1 R1, R0 L0 R0 L0 and L0 R0 L0 R0 (24 points). Each model starts from QL = QR = Q0 (a free parameter), applies the sequence and predicts P(L) after each prefix; parameters minimize the squared error (Results, paragraph 3; Table 1). Model equations as O1. The "averaged residual error" is not defined; readings are (a) the mean squared error over the 24 points; (b) the mean squared error over the 12 points of the three left-ending sequences with each mirror pair averaged. "Represents switching" is read from Fig. 3 as the model's P(L) after each L0 of R0 L0 R0 L0 below 0.5 and after each R0 above 0.5. Reproduced when, under every reading, the residuals are in the order local matching > standard Q > F-Q > DF-Q, each within tolerance, and switching is represented by F-Q and DF-Q but not by the local matching law and standard Q. Partially reproduced when the order and the switching hold but a value is out of tolerance.
- table1-lsq-parameters
The least-squares fits give the parameters of Table 1 (alpha1, alpha2, kappa1, kappa2, Q0) for each model.
How it will be judged
As fig3-lsq-switching. A reported 0.00 cannot take 10% of itself, so it is given 0.01, one step of the table's printed precision (two decimals). Readings as O1 (the parameter values depend on the form of the update). Reproduced when, under one reading of O1, all values are within tolerance; partially reproduced when at least 10 of the 14 are. The O1 reading that gives the match is reported.
- fig5-markov-accuracy
The first- to fourth-order Markov models reach normalized likelihoods of 0.528, 0.532, 0.534 and 0.532 on test data; the third order is best and the fourth is lower, presumably by overfitting its 256 parameters.
The normalized likelihood of the first-order Markov model was 0.528 ... The normalized likelihood for the second- and third-order Markov models increased to 0.532 and 0.534, respectively ... The performance of the fourth-order model was 0.532, lower than that of the third-order model
Results, Prediction of choices by models, paragraph beginning "Figure 5 shows"; Figure 5How it will be judged
Markov prediction as the Methods give it, counts from the 35 training sessions, scored on the 35 test sessions with the normalized likelihood (O2); split O3; first trials O6; histories over choice trials within a session. Reproduced when, under every reading, all four values are within tolerance and order 3 is the highest. Partially reproduced when order 3 is highest but a value is out of tolerance.
- fig5-constant-models
With constant parameters, standard Q, F-Q and DF-Q (0.527, 0.531, 0.531) predict much better than the local matching law (0.508), but worse than the first- or second-order Markov models.
Predictions by the standard Q-, F-Q-, and DF-Q-learning models with constant parameters (α and κ) were much better (0.527, 0.531, and 0.531, respectively) than the predictions of the local matching law (0.508), but poorer than that of the first- or second-order Markov models
Results, Prediction of choices by models, paragraph beginning "Predictions by the standard Q-"; Figure 5; Table 2How it will be judged
Parameters constant across sessions, fitted by maximum likelihood on the training sessions (O4), Qi(1) = 0 in each session, scored on the test sessions (O2, O3); equations O1; local matching law with P(L) = QL / (QL + QR). Reproduced when, under every reading, all four values are within tolerance and the three Q models exceed the local matching law. The comparison with the Markov models is reported; the paper's own numbers place F-Q and DF-Q (constant) above the first-order Markov model (opening comment), so it is not part of the outcome.
- fig5-variable-models-best
With time-varying parameters standard Q, F-Q and DF-Q improve to 0.5304, 0.5375 and 0.5368, and F-Q with time-varying parameters predicts best of all models.
With time-varying parameters, the performance of the standard Q, F-Q, and DF-Q models improved (0.5304, 0.5375, and 0.5368, respectively). The F-Q-learning model with time-varying parameters produced the best prediction performance
Results, Prediction of choices by models, paragraph beginning "With time-varying parameters"; Figure 5How it will be judged
Particle filter as the supplemental Methods and the opening comment give it, 5000 particles, sigma_alpha and sigma_kappa fitted on the training sessions (O4), drift O5, equations O1, scored on the test sessions (O2, O3). Each variable model is reported against its constant version. Reproduced when, under every reading, the three values are within tolerance, each exceeds its constant version, and F-Q (variable) has the highest normalized likelihood of all models of Fig. 5; partially reproduced when F-Q or DF-Q (variable) is highest and the values are in tolerance but F-Q is not first.
- fig5-fq-dfq-beat-all-models
Per test session, the F-Q and DF-Q models with time-varying parameters predict significantly better than every other model, including the third-order Markov model.
the prediction accuracies of both the F-Q- and DF-Q-learning models with time-varying parameters were significantly higher than those of all other models (paired-sample Wilcoxon's signed rank test, p < 0.05 for third Markov model, p < 0.01 for other models)
Results, Prediction of choices by models, paragraph beginning "A statistical test"; Figure 5 legendHow it will be judged
Per-session normalized likelihoods of the 35 test sessions (O2), two-sided paired Wilcoxon signed-rank tests of F-Q (variable) and of DF-Q (variable) against each of the other nine models of Fig. 5 (standard Q variable, three constant Q models, local matching law, Markov orders 1 to 4). The paper does not say whether the tests are one- or two-sided; both are readings. Readings O1, O3 to O6. Reproduced when, under every reading, both models exceed every other model with p < 0.05 against the third-order Markov model and p < 0.01 against the rest; partially reproduced when all comparisons give p < 0.05 with the F-Q and DF-Q medians higher.
- fig5-fq-dfq-no-difference
The F-Q and DF-Q models with time-varying parameters do not differ in prediction accuracy (p = 0.067), so F-Q, with fewer parameters, is regarded as the better model.
There was no significant difference between these two models (paired-sample Wilcoxon's signed rank test, p = 0.067)
Results, Prediction of choices by models, paragraph beginning "A statistical test"How it will be judged
A claim of absence, judged on the paper's test, a two-sided paired Wilcoxon signed-rank test of per-session normalized likelihoods of F-Q (variable) against DF-Q (variable) over the 35 test sessions at p < 0.05, under every reading of O1, O3 to O5. The paper does not give the side; the test is run two-sided, as a test of no difference with no direction named, and the one-sided p is shown as a variant. Reproduced when the test does not find a difference; the median difference is reported either way.
- binary-prediction-accuracy
Taking left when the predicted P(L) exceeds 0.5 and right otherwise, the best model predicts 68% of choices correctly, averaged over the 35 test sessions.
If we take binary predictions from the choice probability [left for PL(t) > 0.5, right for PL(t) < 0.5], the average percentage of correct predictions over 35 test sessions is 68%
Results, Prediction of choices by models, paragraph beginning "With time-varying parameters"How it will be judged
The sentence follows the statement that F-Q (variable) is best and does not name the model; read as F-Q (variable). Percentage correct per test session, averaged over the 35 sessions; trials with P(L) = 0.5 exactly count as half correct. Readings O1, O3 to O5. Reproduced when, under every reading, the value is within tolerance.
- table2-ml-parameters
The maximum-likelihood parameters on the training data are those of Table 2 (for example F-Q constant alpha 0.55, kappa1 1.25, kappa2 0.55; F-Q variable sigma_alpha 0.005, sigma_kappa 0.07).
How it will be judged
As fig5-constant-models and fig5-variable-models-best. The parameter values depend on O1 and O3 to O5, and a grid step (O4 a) is coarser than 10% for the smallest values. O1 is a reading of what the reported values are (the parameterization), O3 to O5 readings of method. Reproduced when, under one reading of O1 and every reading of O3 to O5, all 16 values are within tolerance; partially reproduced when at least 8 are. The O1 reading that matches is reported.
- table3-parameter-averages
Across all data the time-varying estimates average, for F-Q, alpha1 0.47 ± 0.25, kappa1 1.61 ± 0.77, kappa2 1.10 ± 0.87, and for DF-Q, alpha1 0.45 ± 0.19, alpha2 0.46 ± 0.19, kappa1 1.69 ± 0.77, kappa2 1.24 ± 0.92; standard Q alpha1 0.45 ± 0.23, kappa1 1.86 ± 0.68.
How it will be judged
Per-trial estimates (O13) from the time-varying models with the Table 2 sigmas, over all 70 sessions. The paper does not say over what the average and SD are taken; readings are (a) over all choice trials pooled; (b) over sessions, of per-session means. The tolerance of each mean is its SD (protocol); each SD is also reported against the table. Readings O1, O5, O13. Reproduced when, under every reading, all nine means are within tolerance.
- fig4-strategy-tracking
With time-varying parameters, F-Q and DF-Q capture win-stay-lose-switch behaviour by raising kappa2 close to kappa1, and persevering behaviour by a negative kappa2, which the standard Q model cannot.
This behavior was captured by the F-Q- and DF-Q-learning models, with both models showing an increase in the nonreward aversion parameter κ2 close to the reward reinforcement parameter κ1 ... The F-Q and DF-Q models captured this behavior by exhibiting a negative value for the aversion parameter κ2 (Fig. 4H)
Results, Prediction of choices by models, paragraphs beginning "Figure 4A–F" and "Figure 4, B, D, and F"; Figure 4How it will be judged
The paper shows representative sessions and does not identify them, so the claim is tested as existence of such episodes. Win-stay-lose-switch episode, a run of at least 10 choice trials in which the rat repeated every rewarded choice and switched after every unrewarded one, as Results describe it. Persevering episode, a run of consecutive choices of the less advantageous side, rewarded or not, in a (10, 50) or (50, 10) block, as the Results describe Fig. 4B, D, F ("continued to choose the less advantageous left action"). The paper gives no run length; 5, 10 and 15 trials are readings for both kinds of episode, and the outcome is the one they all give. Reproduced when, under every reading of O1, O5 and O13, some win-stay-lose-switch episode has F-Q (variable) kappa2 within 25% of kappa1 at its end, some persevering episode has F-Q (variable) kappa2 below 0, and over all such episodes the F-Q (variable) mean per-trial likelihood exceeds that of standard Q (variable). The 25% reads "close to"; the paper gives no bound.
- suppdata1-choice-vs-q-difference
The actual left-choice probability, plotted against QL - QR from the standard Q, F-Q and DF-Q models, confirms that the models' predictions are meaningful.
the meaningful performances of the standard Q, F-Q, and DF-Q models were also conformed by plotting the actual left-choice probability for the difference between QL and QR (supplemental Data 1
Results, Prediction of choices by models, paragraph beginning "With time-varying parameters"; supplemental Data 1 (not available)How it will be judged
Supplemental Data 1 is not in the text read, so its bins and models (constant or variable) are not known. Readings are the variable and the constant versions of each model, with QL - QR (O13 a) in ten equal-count bins over all choice trials. Reproduced when, under every reading, the left-choice fraction rises with QL - QR (Spearman correlation of bin centres and fractions above 0.9) for all three models.
- neuron-counts
Analysis used 148 NAc neurons and 66 VP neurons from six rats in 70 sessions.
Subsequent analysis was performed for a total of 148 NAc neurons and 66 VP neurons from six rats during 70 sessions
Results, Neuronal recording data, paragraph 1How it will be judged
Units whose estimated position lies in the NAc or VP (Histology). Reproduced when both counts are within tolerance.
- firing-rates
VP neurons fire faster than NAc neurons during the task (10.2 ± 9.1 against 2.6 ± 2.6 Hz, mean ± SD).
VP neurons showed a higher firing rate than NAc neurons overall. The average and SD of the firing rate of VP neurons and NAc neurons during the task were 10.2±9.1 and 2.6±2.6 Hz, respectively
Results, Neuronal activity in the NAc and VP, paragraph 1How it will be judged
Per neuron, spikes over the task period divided by its duration (O15); mean and SD over neurons per region. The paper gives no test. Reproduced when, under every reading, the VP mean exceeds the NAc mean and the two means are within tolerance; the SDs are reported and partially reproduced is given when a mean holds but an SD is outside 10%.
- fig8b-state-coding
After cue onset (phase 2), 54% of NAc neurons (80 of 148) and 50% of VP neurons (33 of 66) fire differently for the choice and the no-choice tone (state-coding neurons).
After the onset of the cue tone (phase 2), 54% (80 of 148) of NAc neurons and 50% (33 of 66) of VP neurons showed different firing patterns in response to the choice tone and the no-choice tone
Results, Neuronal activity in the NAc and VP, paragraph 4; Figure 8BHow it will be judged
Mann-Whitney U test at p < 0.01 of phase-2 spike counts, choice against no-choice trials (opening comment), trials O14. Reproduced when, under every reading, both counts are within tolerance.
- fig8b-action-value-psth
By block comparison of firing, few neurons code an action value; QL-coding neurons are 6, 8 and 7% of NAc neurons and 11, 18 and 21% of VP neurons in phases 1 to 3, and QR-coding neurons 7, 8 and 5% (NAc) and 11, 9 and 11% (VP).
Before the onset of the nose poke action at the center hole (phase 1), the percentage of QL-coding neurons in the NAc was 6% (9 of 148), and 11% (7 of 66) in the VP ... The percentage of QR-coding neurons in the NAc was 7% (11 of 148), and 11% (7 of 66) in the VP ... The percentage of QL-coding neurons in the NAc was 8% (12 of 148) and 18% (12 of 66) in the VP. The percentage of QR-coding neurons in the NAc was 8% (12 of 148) and 9% (6 of 66) in the VP. Before the initiation of action (phase 3), QL-coding neurons represented 7% (11 of 148) of the neurons in the NAc and 21% (14 of 66) of those in the VP. QR-coding neurons represented 5% (8 of 148) of the neurons in the NAc and 11% (7 of 66) of those in the VP
Results, Neuronal activity in the NAc and VP, paragraph 4; Figure 8B; Figure 7E and 7FHow it will be judged
Block comparisons by Mann-Whitney U test at p < 0.01 on the last 20 choice trials of each block (opening comment). Several blocks of one type in a session are pooled. The 10% tolerance is under one neuron for most counts, so each count is also judged within ±2 neurons. Reproduced when, under that rule, all 12 counts match; partially reproduced when at least 9 do and every count is below 25% of its region.
- fig8b-action-command-coding
Before action initiation (phase 3), 11% of NAc neurons (17 of 148) and 18% of VP neurons (12 of 66) code the action to be chosen (action command-coding neurons).
Neurons encoding the action to be chosen before execution represented 11% (17 of 148) of the neurons in the NAc and 18% (12 of 66) of those in the VP (action command-coding neurons)
Results, Neuronal activity in the NAc and VP, paragraph 4; Figure 8BHow it will be judged
Mann-Whitney U test at p < 0.01 of phase-3 spike counts, left against right choice trials. Reproduced when both counts are within tolerance.
- fig8b-action-coding
After action onset (phase 4), 41% of NAc neurons (61 of 148) and 39% of VP neurons (26 of 66) code the selected action.
After the action onset (phase 4), the proportion of action-coding neurons increased to 41% (61 of 148) in the NAc and 39% (26 of 66) in the VP
Results, Neuronal activity in the NAc and VP, paragraph 4; Figure 8BHow it will be judged
Mann-Whitney U test at p < 0.01 of phase-4 spike counts, left against right choice trials. Reproduced when both counts are within tolerance.
- fig8b-reward-coding
After the reward or no-reward tone (phase 5), 63% of NAc neurons (93 of 148) and 68% of VP neurons (45 of 66) code reward availability.
After the onset of the reward tone or no-reward tone (phase 5), 63% (93 of 148) of NAc neurons and 68% (45 of 66) of VP neurons coded reward availability
Results, Neuronal activity in the NAc and VP, paragraph 4; Figure 8BHow it will be judged
Mann-Whitney U test at p < 0.01 of phase-5 spike counts, rewarded against unrewarded choice trials. Reproduced when both counts are within tolerance.
- fig8b-nac-vp-same-event-coding
The NAc and VP do not differ in the percentages of neurons coding trial type, action command, chosen action and reward outcome.
Comparison of the percentages of coding neurons in the NAc and VP revealed no differences with trial type, action command, chosen action, and reward outcome (Mann–Whitney U test, p = 0.59, p = 0.19, p = 0.80, and p = 0.45, respectively)
Results, Neuronal activity in the NAc and VP, paragraph 5; Figure 8BHow it will be judged
A claim of absence, judged on the paper's test, a Mann-Whitney U test as O9 at p < 0.05, for each of the four types. Reproduced when none of the four finds a difference; not reproduced when any does. The four differences of percentages are reported either way.
- fig8b-ql-vp-higher
QL-coding neurons are a larger share of VP than of NAc neurons after cue onset (phase 2) and before action initiation (phase 3).
Surprisingly, the percentages of QL-coding neurons both after cue onset (phase 2) and before action initiation (phase 3) were significantly higher in the VP than in the NAc (Mann–Whitney U test, p = 0.03 < 0.05, p = 0.004 < 0.01, respectively)
Results, Neuronal activity in the NAc and VP, paragraph 5; Figure 8BHow it will be judged
Mann-Whitney U test as O9 on QL-coding indicators, NAc against VP, per phase. Reproduced when, in both phases, the VP percentage is higher with p < 0.05; partially reproduced when this holds in one phase.
- fig8b-qr-no-region-difference
QR-coding neurons do not differ in share between the VP and NAc in phases 2 and 3.
whereas those of QR-coding neurons were not significantly different between areas at the two different times (Mann–Whitney U test, p = 0.81, and 0.17)
Results, Neuronal activity in the NAc and VP, paragraph 5; Figure 8BHow it will be judged
A claim of absence, judged on the paper's test, a Mann-Whitney U test as O9 at p < 0.05 on QR-coding indicators, NAc against VP, in phases 2 and 3. Reproduced when neither finds a difference; the differences are reported either way.
- preference-balance
Among state-coding neurons (48% NAc, 36% VP preferring tone A), action command-coding neurons (53%, 50% preferring left) and action-coding neurons (57%, 50% preferring left), the share with each preference does not differ from 50%.
Of the state-coding neurons, 48% (38 of 80) in the NAc and 36% (12 of 33) in the VP showed a greater response to tone A than to tone B. Of the action command-coding neurons, 53% (9 of 17) in the NAc and 50% (6 of 12) in the VP showed greater activity in left-selected trials than in right-selected trials. In the action-coding neurons, 57% (35 of 61) in the NAc and 50% (13 of 26) in the VP showed greater activity in left-selected than in right-selected trials. All of these populations were not significantly different from 50% (Wilcoxon's signed rank test, p > 0.05)
Results, Neuronal activity in the NAc and VP, paragraph 6How it will be judged
A claim of absence, judged on the paper's test, a Wilcoxon signed-rank test at p < 0.05 within each of the six populations. The test as written is reading O11 (a); the sign test, O11 (b), is shown as a variant. Preference is the sign of the difference of mean spike counts in the coding bin. Reproduced when none of the six tests finds a departure from 50%; the six percentages are reported against the expected values either way.
- nac-noreward-preference
73% of NAc reward-coding neurons (68 of 93) respond more to the no-reward than to the reward tone, more than the 50% of chance, while the VP share (56%, 25 of 45) does not differ from 50%.
However, 73% (68 of 93) of the reward-coding neurons in the NAc showed a larger response to the no-reward than to the reward tone. This percentage of cells was significantly larger than the 50% predicted by chance (Wilcoxon's signed rank test, p < 0.0001). The percentage of reward-coding neurons in the VP (56%; 25 of 45), however, was not significantly different from 50% (Wilcoxon's signed rank test, p > 0.05)
Results, Neuronal activity in the NAc and VP, paragraph 6How it will be judged
Preference by the sign of the phase-5 difference of mean spike counts. For the NAc, under both readings of O11, the NAc percentage is above 50% and the test gives p < 0.0001. The VP part is a claim of absence judged on O11 (a) at p < 0.05, with O11 (b) as a variant. Reproduced when the NAc part holds under every reading and the VP test finds no departure; partially reproduced when one part holds.
- core-shell-no-difference
Splitting NAc neurons into tentative core and shell groups by electrode position, the same analyses find no difference between core and shell neurons.
The same analysis described above found no significant differences between core and shell neurons (supplemental Data 3
Results, Neuronal activity in the NAc and VP, paragraph 7; supplemental Data 3 (not available)How it will be judged
Supplemental Data 3, which holds the tests, is not in the text read. A claim of absence, read as the analysis of fig8b-nac-vp-same-event-coding and fig8b-ql-vp-higher (Mann-Whitney U test as O9 at p < 0.05 on per-neuron coding indicators) with core and shell in place of NAc and VP, for state, action command, action, reward, and QL and QR coding in phases 1 to 3. Core and shell as assigned in the dataset from Fig. 6; if the dataset does not assign them, the claim cannot be tested. Reproduced when no test finds a difference.
- fig8c-regression-action-value
By regression on the block reward probabilities, QL-coding neurons are 9, 10 and 7% of NAc neurons and 15, 24 and 26% of VP neurons in phases 1 to 3, and QR-coding neurons 11, 9 and 6% (NAc) and 9, 17 and 14% (VP).
QL-coding neurons, which were found to have a significant regression coefficient to QL (t test, p < 0.01) but not to QR, were present in percentages of 9, 10, and 7% in the NAc, and 15, 24, and 26% in the VP for different time bins (Fig. 8A, phases 1–3) ... QR-coding neurons, which were found to have a significant regression coefficient to QR (t test, p < 0.01) but not to QL, were present in percentages of 11, 9, and 6% in the NAc, and 9, 17, and 14% in the VP
Results, Neuronal activity in the NAc and VP, paragraph 8; Figure 8CHow it will be judged
Regression as the opening comment, y = b0 + b1 QL + b2 QR with QL and QR the block's reward probabilities as fractions; the equation is missing, and the text names no other regressor, so none is added. The 10% tolerance is under one neuron for most values, so each is also judged within ±2 neurons of the reported percentage. Reproduced when, under that rule, all 12 values match; partially reproduced when at least 9 do.
- fig8c-ql-vp-higher
In the regression analysis, QL-coding neurons are a larger share of VP than of NAc neurons in phases 2 and 3.
In phases 2 and 3, there were significant differences in the size of the population of QL-coding neurons in the NAc, and in the VP (p = 0.007 and 0.0001 < 0.01, Mann–Whitney U test)
Results, Neuronal activity in the NAc and VP, paragraph 8; Figure 8CHow it will be judged
Mann-Whitney U test as O9, NAc against VP, on regression QL-coding indicators. Reproduced when, in both phases, the VP share is higher with p < 0.01; partially reproduced when both give p < 0.05 or one gives p < 0.01.
- fig8c-regression-detects-more
Regression detects slightly more action value-coding neurons (Fig. 8C) than the block comparison (Fig. 8B), and the two are consistent.
These regression analysis results were consistent with those obtained by comparing the perievent time histograms (Fig. 8B). However, the number of the detected action value-coding neurons was slightly larger in Figure 8C than in Figure 8B
Results, Neuronal activity in the NAc and VP, paragraph 8How it will be judged
Total QL- plus QR-coding neurons over phases 1 to 3 and both regions, regression against block comparison. Reproduced when the regression total is larger, and the region with the larger QL share in phases 2 and 3 is the VP in both analyses.
- fig8d-state-value-vp-higher
State value-coding neurons are 12, 8 and 8% of NAc neurons and 18, 20 and 20% of VP neurons in phases 1 to 3, with the VP share larger in phases 2 and 3.
State value-coding neurons, which were found to have a significant regression coefficient to [QL(t) + QR(t)] but not to [QL(t) − QR(t)], were present in percentages of 12, 8, and 8% in the NAc, and 18, 20, and 20% in the VP (Fig. 8D). In phases 2 and 3, these differences between the sizes of the populations in the NAc and VP were significant (p = 0.02 and 0.02 < 0.05, respectively; Mann–Whitney U test)
Results, Neuronal activity in the NAc and VP, paragraph 8; Figure 8DHow it will be judged
Regression y = b0 + b1 (QL + QR) + b2 (QL - QR) on the last 20 choice trials of each block (Methods, Regression analysis); t test at p < 0.01; Mann-Whitney U test as O9. Percentages judged within ±2 neurons where 10% is under one neuron. Reproduced when the six values match and the VP share is higher in phases 2 and 3 with p < 0.05; partially reproduced when the test result holds and at least four values match.
- fig8d-policy
Policy-coding neurons are 10, 9 and 7% of NAc neurons and 11, 15 and 14% of VP neurons in phases 1 to 3.
Policy-coding neurons, which were found to have a significant regression coefficient to [QL(t) − QR(t)] but not to [QL(t) + QR(t)], were present in percentages of 10, 9, and 7% in the NAc, and 11, 15, and 14% in the VP (Fig. 8D)
Results, Neuronal activity in the NAc and VP, paragraph 8; Figure 8DHow it will be judged
As fig8d-state-value-vp-higher, for the QL - QR coefficient. Reproduced when the six values match (±2 neurons where 10% is under one neuron); partially reproduced when at least four do.
- fig8-above-chance
Every coding population of Figure 8 is larger than expected by chance.
All percentages shown in Figure 8 were significantly larger than those predicted by chance (p < 0.01, binomial test)
Results, Neuronal activity in the NAc and VP, paragraph 8; Figure 8 legendHow it will be judged
One-sided binomial test of each count of Fig. 8B to 8D against the chance level O10. Reproduced when, under every reading, every population gives p < 0.01; partially reproduced when at least 90% of them do.
- fig9-state-information
Information about the trial type (state) in the NAc and VP rises right after cue onset and again after center-hole exit.
The information coding state in the NAc and VP (green lines) increased immediately after the onset of the cue tone, and again after exiting from the center hole
Results, Neuronal activity in the NAc and VP, paragraph 9; Figure 9A and 9BHow it will be judged
MI as the opening comment, tone A against tone B, choice and no-choice trials, averaged per region, windows O12. "Immediately after" is read as within 0.5 s of cue onset. Reproduced when, under every reading and in both regions, the state MI exceeds the threshold within 0.5 s after cue onset and is higher there than in the 0.5 s before it, and the curve aligned to center-hole exit has a local maximum above threshold after the exit that exceeds its value at the exit.
- fig9-action-information
Information about the chosen action is very weak before center-hole exit, rises rapidly after it, and peaks around entry to the L or R hole.
Information related to action choice was very weak in both the NAc and VP before the onset of the choice (before the exit of the center hole), but increased rapidly after the choice onset, and showed a peak around the time of entry to the left or right hole
Results, Neuronal activity in the NAc and VP, paragraph 9; Figure 9BHow it will be judged
MI between firing and left or right choice, choice trials, aligned to center-hole exit, windows O12. Reproduced when, under every reading and in both regions, the mean MI over windows before the exit is below a fifth of the peak, and the peak lies after the exit within 0.5 s of the median reaction time (L or R entry).
- fig9-reward-information
Information about reward rises when the reward or no-reward tone is presented at L or R entry, then decreases gradually but lasts until the next trial.
Information related to the reward started to increase at the presentation of reward/no-reward tone, which was presented at the time of entry to the left or right hole. The reward information gradually decreased, but was sustained until the subsequent trial
Results, Neuronal activity in the NAc and VP, paragraph 9; Figure 9B; Discussion, Information coding in NAc and VP, paragraph 1How it will be judged
MI between firing and reward outcome, choice trials, aligned to center-hole exit (Fig. 9B) and, for the next trial, to the next cue onset (Fig. 9A, "previous reward"), with the previous trial's outcome as the event; windows O12. Reproduced when, under every reading and in both regions, reward MI is below threshold before the L or R entry, rises above it after, and the previous-trial reward MI is above threshold around the next cue onset.
- fig9-action-value-information-low
Information about the action values QL and QR is above the significance threshold in the NAc and VP but low and without clear peaks, while state, action and reward information are robust.
Information regarding action values QL and QR in the NAc and VP were found to be larger than the threshold values, indicating a significant level of information (p < 0.01) (see Materials and Methods). However, this level of information was still relatively low and without clear peaks, whereas information about state, action, and reward was robustly represented in the NAc and VP
Results, Neuronal activity in the NAc and VP, paragraph 10; Figure 9; Discussion, Information coding in NAc and VP, last paragraph but oneHow it will be judged
QL and QR from F-Q with time-varying parameters (Table 2 sigmas), binarized at the median (O13), MI as the opening comment, windows O12. Reproduced when, under every reading and in both regions, QL and QR MI exceed the threshold in most windows (over half), their peak is below the peaks of state, action and reward MI, and their maximum is less than twice their median over windows (no clear peak); partially reproduced when the comparison with the other events holds but the above-threshold or flatness part does not.
- action-information-outlasts-reward-peak
Information about the chosen action in the NAc and VP lasts beyond the peak of reward information.
In particular, we showed that a representation of information regarding action in the NAc and VP lasted beyond the peak of information regarding reward (Fig. 9)
Discussion, Information coding in NAc and VP, paragraph 3; Figure 9How it will be judged
On the curves of Fig. 9B (O12). Reproduced when, under every reading and in both regions, action MI is above threshold at the time of the reward MI peak and in the window 0.5 s after it.
Pipeline
- Paper
- done (1 of 1)
- Claims
- done
- Claims check
- done (ready)
- Attempt
- todo
- Questions
- none
- Verified
- none