Pupil size is a cheap, continuous, non-invasive signal that changes with how much a person expects to gain. This makes it an attractive candidate for tracking how dopamine shapes valuation in the brain, something that otherwise requires positron emission tomography (PET) or functional MRI. There is a hidden problem with the standard experiment, which gives a dopamine-boosting drug such as L-DOPA and checks whether reward-related pupil dilation changes. Dopamine also acts in the retina, where dopaminergic amacrine cells (local retinal interneurons) set light adaptation. If the drug changes how strongly cones or melanopsin-containing ganglion cells drive the pupil, a "dopamine changed valuation" result could simply mean "dopamine changed the eye's light reflex." This proposal uses chromatic pupillometry to measure that retinal gain directly in the same session and remove it. Chromatic pupillometry means using colored light pulses that each stimulate one class of photoreceptor.
The core claim is testable in two sentences. In a placebo-controlled L-DOPA gambling task, the anticipatory pupil response to risky gain options should scale with each person's drug-induced increase in approach bias. That relationship should survive after the drug's effect on cone- and melanopsin-driven light reflexes, measured with receptor-isolating flashes, is regressed out. If it does not survive, the published pupil–dopamine link is partly retinal. If it does survive, the pupil becomes a validated, low-cost readout of central dopaminergic valuation for aging and Parkinson's research.
What the two literatures give us, and the gap between them
On the valuation side, Dopaminergic Modulation of Decision Making and Subjective Well-Being shows that 150 mg L-DOPA selectively increases gambling on gain trials. Modelling attributes the effect to a Pavlovian approach parameter $\beta_{gain}$, not to utility curvature or loss aversion. That paper supplies our task and our behavioral target, but it recorded no physiology during deliberation. An opponent striatal circuit for distributional reinforcement learning shows that in mice, D1 and D2 striatal neurons encode the optimistic and pessimistic tails of a reward distribution with matched mean and different variance. Mikhael (2016) models tonic dopamine as a dial between those channels. Together they motivate including mean-matched fixed-versus-variable options, because a D1-favoring drug should shift valuation toward the optimistic tail. Attenuation of dopamine-modulated prefrontal value signals underlies probabilistic reward learning deficits in old age shows that anticipatory value signals in ventromedial prefrontal cortex (vmPFC), rather than outcome-phase prediction errors, track learning and depend on accumbens D1 receptors. This tells us to look at the anticipation window, not only at outcomes.
On the pupil side, Reimer et al. (2016) show that fast pupil dilations follow noradrenergic locus coeruleus bursts and slow dilations follow cholinergic tone. They explicitly leave dopamine untested. This constrains the proposal: any dopaminergic pupil signal must be dissociated from noradrenergic surprise. A unified model of the task-evoked pupil response supplies a method. It fits pupil traces with a shared midbrain drive scaled by an arousal gain, but it uses neutral gratings and has no reward terms. Chromatic work supplies the stimulus logic. Analysis for Extracted Features of Pupil Light Reflex to Chromatic Stimuli in Alzheimer's Patients shows that blue pulses isolate melanopsin ganglion-cell function while white-light responses mostly track age, using 1 s red, blue and white pulses recorded at 60 Hz. Finally, Retinal alpha-synuclein accumulation correlates with retinal dysfunction and structural thinning in the A53T mouse model of Parkinson's disease documents cone-first retinal dysfunction and compensatory changes in retinal tyrosine hydroxylase. Retinal dopamine is therefore a real, separately varying quantity, which is exactly the confound we target.
None of the connecting work in the file combines these pieces. White et al. used isoluminant cues in superior colliculus recordings without reward (Competitive Integration of Visual and Goal-related Signals on Neuronal Accumulation Rate). Burlingham et al. modelled pupils without reward. The dopamine papers recorded no pupils. The obvious combination would be "use isoluminant colored cues in a reward task." That idea alone is weak, because spatial-pattern cues of identical spectrum control luminance more simply. Our use of chromatic stimulation is different: the colors are not the cues. They are a separate assay of the light-reflex machinery that the drug may alter.
Research Program
Each participant completes two sessions, L-DOPA and placebo, in double-blind crossover order. Each session has three parts.
The first part is a retinal assay before and after dosing. It extends the Nowak protocol with silent substitution, meaning multi-primary light mixtures that change the excitation of one photoreceptor class while holding the others constant. The battery contains L+M-cone-isolating pulses, S-cone-isolating pulses, and melanopsin-isolating pulses. It also includes standard red and blue pulses so the results stay comparable to the Alzheimer's study. From each pulse we extract constriction amplitude, latency, maximum re-constriction velocity and post-illumination sustained constriction.
The second part is the gambling task. It follows the Rutledge structure of gain, loss and mixed trials, with an added block of mean-matched fixed-versus-variable gain options modelled on the Lowet design. All option cues share an identical spectrum and mean luminance and differ only in spatial pattern, so the displays produce no cue-locked light-reflex difference. The background is held at a fixed adapting spectrum.
The third part is a short block of receptor-isolating probe flashes placed in inter-trial intervals throughout the task. These track any drift in retinal gain as plasma L-DOPA rises, without contaminating the anticipation window.
The analysis model treats the pupil trace $p(t)$ on each trial as a sum of kernels convolved with event regressors:
$$p(t) = \sum_k h_k * x_k(t) \;+\; g_{ret}\,\big(h_{L} * \ell(t)\big) \;+\; \epsilon(t)$$
Here the $x_k$ are cognitive regressors: expected value, outcome variance, chosen-gamble indicator, and unsigned outcome surprise. The $h_k$ are their fitted response kernels. $\ell(t)$ is any residual light input, and $h_L$ is the light-reflex kernel. $g_{ret}$ is a per-session retinal gain fixed in advance from the chromatic probes rather than fitted to task data. The key quantity is the drug effect on the anticipation-window kernel for risky gain choices, $\Delta h_{gain}$. We test whether it correlates across people with the drug effect on behavioral $\beta_{gain}$, after partialling out the drug effect on each probe metric.
The noradrenergic alternative is handled by timing and sign. Locus coeruleus–driven dilation should follow unsigned surprise at outcome. An approach signal should appear before the outcome and be signed by gain domain, present for gains but not losses, mirroring the behavioral selectivity Rutledge reported.
A second stage applies the validated pipeline to older adults. Here age-related miosis and retinal senescence are severe (Analysis for Extracted Features of Pupil Light Reflex to Chromatic Stimuli in Alzheimer's Patients), and attenuated vmPFC anticipation is expected (Attenuation of dopamine-modulated prefrontal value signals underlies probabilistic reward learning deficits in old age). The question is whether age differences in reward pupillometry survive retinal correction. The two-armed bandit paradigm from that paper is added so pupil anticipation can be related to learning.
Baselines
The naive reference point is standard pupillometry. It uses luminance-matched achromatic cues, an event-related average of pupil dilation by condition, and no retinal assay. We will compute the drug-by-gain-domain effect on anticipatory dilation exactly this way from the same recordings. Because both analyses use identical data, any difference in conclusions comes from the correction alone.
The strongest existing pupil model is the linear-nonlinear linking model of A unified model of the task-evoked pupil response, which reports a median $R^2$ of 81% for predicting task-evoked pupil area. We will fit it to our task data with its arousal-gain term free per condition and the same cross-validation folds and fitting budget as our model. Beating it means two things. First, our model must match or exceed its held-out $R^2$ on our data. We do not expect to beat 81% on a different task. Second, only our model must yield drug effects that track behavior. Burlingham-style arousal gain could absorb value effects as generic gain, which is the main rival account.
The strongest behavioral baseline is the approach-avoidance choice model of Dopaminergic Modulation of Decision Making and Subjective Well-Being. That paper reports a gain-trial gambling increase at $p = 0.04$ and a dose correlation at $p < 0.01$; the file gives no BIC values. The pupil earns its place only if it adds information choices lack. Concretely, adding anticipatory pupil kernels should improve cross-validated prediction of the per-person drug effect on $\beta_{gain}$, or of later bandit performance, beyond what choices alone provide.
The key ablation removes $g_{ret}$, that is, it fits the same model without the chromatic probes. A second ablation replaces the spectrum-identical cues with photopically isoluminant colored cues. Such cues match in brightness but differ in melanopsin drive. This ablation shows how much spurious cue-locked "value" signal ordinary isoluminant designs would create.
Why not just…?
Why not just use luminance-matched grayscale cues and standard pupillometry? For the cues themselves we largely agree, and our cues are spectrum-identical patterns rather than colors. The threat lies elsewhere. Matching cues does nothing about a drug that changes retinal light adaptation, because the background, fixation light and screen still drive the light reflex, and L-DOPA can alter its gain. Grayscale pupillometry cannot detect this, since white light mixes all receptor classes and mostly reflects age (Analysis for Extracted Features of Pupil Light Reflex to Chromatic Stimuli in Alzheimer's Patients). If the probes show no drug effect on any receptor-isolated reflex, this alternative wins on parsimony. We will report that outcome as a reassuring null for prior pupil–dopamine studies.
Why not just model choices? Choices give one number per trial. The aging work shows that the informative signal can be anticipatory rather than outcome-locked (Attenuation of dopamine-modulated prefrontal value signals underlies probabilistic reward learning deficits in old age). The incremental-prediction test above settles this. If pupils add nothing beyond $\beta_{gain}$, the alternative is right.
Why not just use saccades? In the Burlingham model, saccades are threshold crossings of the same drive that the pupil follows continuously, so they miss sub-threshold gain changes. We record saccades anyway and include saccade-rate regressors, so this can be checked directly.
Why not just call it noradrenaline? This is the most serious threat. Reimer et al. (2016) tie pupil dilation to locus coeruleus and cholinergic activity, and L-DOPA is a precursor of noradrenaline as well as dopamine. Timing and sign help: a gain-selective anticipatory effect does not look like unsigned surprise. Pharmacology alone cannot fully separate the two transmitters. A follow-up with a dopamine-receptor antagonist arm would be needed for a clean claim, and we will state the conclusion at the level of "L-DOPA-sensitive central valuation," not "dopamine specifically."
Decisive Checks
The program succeeds if three results hold. First, the receptor-isolated probe battery detects whether L-DOPA shifts cone or melanopsin reflex gain, a question that is itself unreported in the file. Second, after retinal correction, the drug-induced anticipatory pupil change on gain trials, but not loss trials, correlates with the drug-induced change in $\beta_{gain}$. Third, in the mean-matched block, L-DOPA enlarges the pupil difference between variable and fixed options, as an optimistic, D1-weighted shift (An opponent striatal circuit for distributional reinforcement learning) predicts.
The proposal fails informatively in either of two ways. The correlation may appear only without retinal correction, which would show that previous pupil–dopamine links are partly ocular. Or the effect may track outcome surprise rather than anticipation, which would favor the noradrenergic account.
References
- Dopaminergic Modulation of Decision Making and Subjective Well-Being
- An opponent striatal circuit for distributional reinforcement learning
- Learning Reward Uncertainty in the Basal Ganglia
- Attenuation of dopamine-modulated prefrontal value signals underlies probabilistic reward learning deficits in old age
- Pupil fluctuations track rapid changes in adrenergic and cholinergic activity in cortex
- A unified model of the task-evoked pupil response
- Analysis for Extracted Features of Pupil Light Reflex to Chromatic Stimuli in Alzheimer’s Patients
- Retinal alpha-synuclein accumulation correlates with retinal dysfunction and structural thinning in the A53T mouse model of Parkinson’s disease
- Competitive Integration of Visual and Goal-related Signals on Neuronal Accumulation Rate: A Correlate of Oculomotor Capture in the Superior Colliculus