When a brain region "encodes" a decision variable, the region may be computing that variable itself, or it may simply be receiving a copy computed somewhere else. Neuroscientists call this the problem of signal origin. Intertemporal choice is the decision between a smaller reward now and a larger one later. In that setting the question has a concrete form. Does the hippocampus build an estimate of how costly a delay is and send it to orbitofrontal cortex, or does orbitofrontal cortex compute subjective value and broadcast it back? The best human data on this come from rare single-neuron recordings in epilepsy patients, typically fewer than ten people. Small samples make two separate problems worse. Origin claims are rarely tested, and claims about individual differences ("patient people show stronger hippocampal coding") rest on correlations across a handful of subjects.
This note proposes a specific analysis method plus a ground-truth test bed. The method calls a region's signal "locally computed" only if it carries information about the decision variable that cannot be predicted from the recent past of the other recorded regions, after removing the shared time-tracking activity that dominates prefrontal firing. Every claim is placed at the level where data are plentiful: within-subject trials for origin, and population behavioral cohorts for calibrating any across-subject claim. Before touching real data, we check on simulated networks whose wiring we know whether the method, and its simpler competitors, can tell "computed here" from "inherited" from "both driven by an unrecorded third area" at realistic unit counts.
Core Claim
The motivating dataset is Multiregional representations of intertemporal decision making in human single neurons. It recorded 193 units (50 orbitofrontal, 68 amygdala, 75 hippocampal) in nine patients choosing between immediate cash and $10 delayed by 2–365 days. Orbitofrontal cortex decoded choice at 78% accuracy, hippocampus at 73.8% and amygdala at 72.6%, and hippocampal coding strength tracked individual discount rates. The authors state that connectivity analysis is needed to establish direction. Region-wise decoding accuracy cannot do this. A copied signal can be decoded as well as the original, and a region with more or noisier units can appear "earlier" for reasons unrelated to routing.
Our claim is that origin becomes testable in such data only after three conditions are met. Time-locked activity common to all trials must be separated out first. Functional, But Not Anatomical, Separation of "What" and "When" in Prefrontal Cortex showed that about 80% of persistent prefrontal variance tracks elapsed time and is driven by external input, while under 20% carries task content. That 80% is a ready-made source of spurious shared signal. The test must also be conditional and lagged. And its behavior under an unobserved common driver must be measured, not assumed. We define a hippocampus-to-orbitofrontal origin score as the cross-validated gain in predicting a decision variable $y_t$ (value difference, delay, or choice) from hippocampal activity $H_t$ once the orbitofrontal past $O_{t-\tau:t}$ is already known:
$$ \Delta_{H \mid O}(\tau) = \mathrm{CVLL}\left(y_t \mid H_t, O_{t-\tau:t}\right) - \mathrm{CVLL}\left(y_t \mid O_{t-\tau:t}\right), $$
where CVLL is held-out log-likelihood of a regularized decoder, $\tau$ is the lag window, and both $H_t$ and $O$ are projected onto task-content subspaces after removing the condition-independent time component. The reverse score $\Delta_{O \mid H}$ is defined the same way. Asymmetry between the two scores, compared against a trial-shuffled null, is the origin evidence. Each score is computed separately within each subject. Across subjects, we report prevalence: in how many of the nine the asymmetry holds. We do not report a pooled p-value.
Research Program
The first component is a ground-truth benchmark. We train two- and three-area recurrent neural networks, models made of interconnected rate units, on the exact task of Gill et al. using hyperbolic discounting $V = A/(1+kD)$. Each network is trained with an individual $k$ drawn from a realistic distribution, and wiring is constrained into one of three regimes. In the first, a "hippocampal" area computes delay cost and projects to a "orbitofrontal" area. In the second, the orbitofrontal area computes value and feeds it back. In the third, both areas receive value from a third, unrecorded area. The third regime is the hard case. It mirrors the unidentified tonic external drive found in Flexible sensorimotor computations through rapid reconfiguration of cortical dynamics. We then subsample units to match Gill's counts per region, add Poisson spiking, match trial counts, and generate cohorts of nine "subjects." The deliverable is a confusion matrix of origin attribution for each method. Of greatest interest is the rate at which each method wrongly declares direct routing in the common-input regime. If no method avoids that error at these unit counts, that negative result is itself the main finding, because it bounds what any nine-patient study can claim.
The second component applies the methods to real multiregional data. Gill's dataset is the target and requires collaboration with its authors; the file does not indicate public release. The public development set is A stable, distributed code for cue value in mouse cortex during reward learning. It recorded 5,332 simultaneous units across 11 regions in 5 mice and found cue value everywhere, but explicitly left open how value gets there. That study is Pavlovian, not intertemporal, so it serves to validate the method rather than to answer the discounting question.
The third component calibrates across-subject claims with population behavior. A brain–behavior correlation across nine people is limited by how reliably $k$ is measured from the trials each patient completed. Under the classical attenuation relation, the observed correlation is bounded by $r_{\text{true}}\sqrt{\rho_{\text{brain}}\,\rho_k}$, where $\rho$ denotes reliability. In addition, with $N = 9$ the smallest correlation reaching two-tailed $p < .05$ is about 0.67, which follows from the $t$ distribution with 7 degrees of freedom. We estimate $\rho_k$ for Gill-length task designs by resampling trial subsets in large behavioral cohorts. A Computer Adaptive Measure of Delay Discounting provides calibration data on 4,190 participants. The globalizability of temporal discounting provides 13,629 people and shows that within-country variance dwarfs between-country variance, so the spread of $k$ can be estimated well. From $\rho_k$ and the spread of $k$, we simulate the expected inflation of any significant nine-subject correlation (the winner's curse) and report the hippocampal–patience result alongside that expected inflation.
Baselines
The naive reference point is the analysis already published: region-wise choice decoding accuracy plus peak decoding latency, taking the region that decodes earliest or best as the origin. Its reported numbers are 78% orbitofrontal, 73.8% hippocampal and 72.6% amygdala choice decoding Multiregional representations of intertemporal decision making in human single neurons. We rerun it on our simulations with identical folds and decoder families. Beating it means lower false attribution in the common-input regime at equal or better accuracy in the two direct regimes.
The strongest existing competitor for direction is pairwise lagged prediction on raw population activity, essentially Granger-style influence estimation. For directed connectivity in discounting, no reported result was found on Lacuna. We give it the same regularization grid, lag range and cross-validation as our method. It is the real threat, because our score is a refinement of it.
Two ablations isolate the proposal's ingredients. The first drops the removal of time-locked activity, which tests whether the "what versus when" separation matters. The second drops the trial-shuffled common-input null and the prevalence reporting, which tests whether our calibration, not the predictor, does the work. For the across-subject analysis, the baseline is the uncorrected Pearson correlation across nine subjects as reported. We beat it by showing calibrated intervals that predict how the effect behaves when subjects are left out.
Why not just…?
Why not just compare decoding latencies across regions? Latency is how origin has been inferred in vision, for example the delayed onset of contextual modulation in Figure-ground perception in the awake mouse and neuronal activity elicited by figure-ground stimuli in primary visual cortex and the recurrence-dependent delays in Evidence that recurrent circuits are critical to the ventral stream's execution of core object recognition behavior. Those comparisons have stimulus onset as an anchor and large timing gaps. In discounting, value can be present before the choice period, and decodability onset depends on unit count and firing rate, which differ across the three regions. The simulation benchmark tests this directly. If latency alone attributes origin as well as our score, we will adopt it and the proposal reduces to the calibration contribution.
Why not just run Granger causality on the existing recordings? This is the most plausible competitor, and it may work. Our concern is specific. Shared time-tracking drive, roughly 80% of prefrontal variance per the "what/when" study, can create directed-looking prediction when one region follows a common input with a slight lag. The common-input regime in the benchmark settles the question. If plain Granger matches our false-attribution rate there, the demixing step is unnecessary and we will say so.
Why not just recruit more patients or pool sites? Pooling is the right long-term answer. Federated analysis in COINSTAC reveals functional network connectivity and spectral links to smoking and alcohol consumption in nearly 2,000 adolescent brains shows that decentralized analysis can match centralized precision. But intracranial patients accrue slowly, and more subjects do not fix an origin test that is biased by common input. Our benchmark tells a consortium which analysis to federate and how many subjects a given prevalence claim needs.
Why not just use large structural MRI cohorts? Diminished Cortical Thickness Is Associated with Impulsive Choice in Adolescence robustly linked orbitofrontal thinning to discounting in 427 youths. Static thickness, however, cannot say whether a signal is computed or relayed. Are neuroanatomical phenotypes for psychiatric disorders robust? An assessment of the reproducibility of grey matter differences in mental illness also warns that structural phenotypes often fail to replicate. Structural cohorts answer a different question, and we use population data only where it applies: calibrating behavioral reliability.
Decisive Checks
The method succeeds only if three things hold on the benchmark. First, it attributes origin correctly in both direct regimes at Gill's unit counts. Second, it keeps false direct-routing claims in the common-input regime near the shuffled-null rate, where both baselines exceed it. Third, the time-removal ablation measurably degrades that performance. On the mouse data, it should produce stable asymmetries that replicate across animals rather than in one. On the human data, the deliverable is a per-subject origin map for value and delay, plus the hippocampal–patience correlation reported with its population-calibrated expected inflation.
The outcome is falsifiable either way. If the asymmetry appears in most of the nine subjects and survives the common-input null, that is real evidence for the direction of hippocampal–orbitofrontal communication. If the benchmark shows that nine-patient, 50–75-unit designs cannot separate inheritance from common drive, that bound is equally worth publishing.
References
- Multiregional representations of intertemporal decision making in human single neurons
- Functional, But Not Anatomical, Separation of “What” and “When” in Prefrontal Cortex
- Flexible Sensorimotor Computations through Rapid Reconfiguration of Cortical Dynamics
- A stable, distributed code for cue value in mouse cortex during reward learning
- A Computer Adaptive Measure of Delay Discounting
- The globalizability of temporal discounting
- Figure-ground perception in the awake mouse and neuronal activity elicited by figure-ground stimuli in primary visual cortex
- Evidence that recurrent circuits are critical to the ventral stream’s execution of core object recognition behavior
- Federated Analysis in COINSTAC Reveals Functional Network Connectivity and Spectral Links to Smoking and Alcohol Consumption in Nearly 2,000 Adolescent Brains
- Diminished Cortical Thickness Is Associated with Impulsive Choice in Adolescence
- Are neuroanatomical phenotypes for psychiatric disorders robust? An assessment of the reproducibility of grey matter differences in mental illness