Broadscale dampening of uncertainty adjustment in the aging brain

Abstract

The ability to prioritize task-relevant inputs enables efficient behavior across the human lifespan. However, contexts in which feature relevance is ambiguous require dynamic exploration rather than stable selectivity. Although both cognitive flexibility and stability generally decline with ageing, it is unknown whether the aging brain differentially adjusts to changing uncertainty. Here, we comprehensively assess the dynamic range of uncertainty adjustments across the adult lifespan (N = 100) via behavioral modelling and a theoretically informed set of human neuroimaging signatures (EEG-, fMRI-, and pupil-based). As a group, older adults show a broadscale dampening of neuro-computational uncertainty adjustments. In support of a “maintenance” account of brain aging, older individuals with more young-like neural recruitment were better able to select task-relevant features, also in a Stroop task with low perceptual demands. Our results highlight neural mechanisms whose maintenance plausibly enables flexible task set, perception, and decision computations across the adult lifespan.

1. Introduction

The ability to prioritize goal-related signals in perceptual and decision processes is fundamental for adaptive behaviors. Some contexts facilitate this process by designating features to which we should selectively attend ¹. Many contexts do not convey feature relevance, however. Such elevated uncertainty plausibly shifts demands from an emphasis on focused feature selection to a broad, but less precise sensitivity to diverse candidate features ^{2, 3}. An adaptive system should track the degree of such contextual uncertainty, and leverage it to tune perception, guide decisions, and select actions. Conversely, failure to do so may result in maladaptive cognition and behaviors ^{4, 5}. Here, we examine whether a potential failure to adapt to varying uncertainty is a key characteristic of healthy human aging.

Behavioral observations support aging-related deficits in uncertainty-guided processing. In contexts that cue task-relevant dimensions of compound stimuli, older adults remain sensitive also to irrelevant dimensions ^{6, 7}, indicating challenges in stable feature-selection ^8–11. Conversely, older adults show inflexibility when contexts require dynamic feature switches ^12–14, and incur substantial “fade-out” costs when transitioning from dynamic to stable single-feature contexts ¹⁵. Such observations suggest that older adults may be stuck in a suboptimal ‘middle ground’ that neither affords stable task selectivity when uncertainty is low, nor flexible task sensitivity in dynamic or uncertain contexts. Although age-related deficits in using uncertainty variations to guide behavior have been observed to impair computational (learning rate) adjustments ¹⁶, it remains unclear whether such underutilization arises from challenges in estimating latent uncertainty, or from leveraging adequate estimates to adjust computations. Crucially, for uncertainty to provide a principled and comprehensive lens on aging-related adaptivity constraints, first evidence is required to establish whether and/or how neural responses to uncertainty differ in the older adult brain.

Although the neural mechanisms of uncertainty resolution remain vague ¹⁷, emerging models point to interacting systems that define task sets, alter perception, and guide decision formation ^18–20. Task set management has been commonly tied to fronto-parietal cortex ^{20, 21}, although more recent evidence also suggests underappreciated thalamic deep brain contributions especially in uncertain contexts ^{22, 23}. When task sets are limited to specific sensory features, perceptual networks in turn appear to specifically tune to relevant information by combining distractor inhibition ²⁴ with target enhancement ²⁵. In contrast, high uncertainty may facilitate sensitivity to multiple features via broad excitability increases ²⁶. Shifts between such regimes may be orchestrated by diffuse neurotransmitter systems that adjust computational precision to changing demands ². In young adults, we observed such an integrated response to rising uncertainty ²⁷, encompassing increased fronto-thalamic BOLD activation, increased pupil diameter as an index of neuromodulation ²⁸, and upregulated EEG-based cortical excitability. These results indicate that multiple systems interact to enable a large dynamic response range to contextual uncertainty variations. Whether and how these systems change in their response to uncertainty across the adult lifespan has not been tested, however.

It is plausible that joint declines of these systems are a feature of brain aging, constraining the dynamic range of uncertainty adjustments. Senescence is characterized by various systemic alterations including diminished prefrontal cortex function ²⁹, metabolic decreases in fronto-thalamic control networks ^30–32, progressive deterioration of subcortical neurotransmitter systems ^33–35, reduced cortical inhibition ^{36, 37}, as well as structural declines of coordinating nodes such as the thalamus ^{38, 39}. However, beyond findings that older adults’ brain activity changes less alongside varying demands in general ^40–42, whether older brains also adjust less to contextual uncertainty is unknown. Beyond the group-level, the “maintenance account of aging” further posits that cognitive deficits with senescence emerge when neural resources become insufficient to meet demands, and that older adults with more “young-like” signatures should be most likely to maintain function. We test this account by examining whether a reduced engagement of neural mechanisms expressed in younger adults constrains the range of uncertainty adaptation in older age.

Here, we use decision modelling and multimodal neuroimaging (EEG, fMRI, pupillometry) in 47 younger and 53 older adults to investigate how contextual uncertainty impacts neural and behavioral computations across the adult lifespan. Participants performed a decision task involving a compound stimulus, for which we overtly manipulated uncertainty regarding the stimulus feature(s) that would be relevant for decisions. By assessing multiple a priori signatures that were observed in younger adults’ response to contextual uncertainty ²⁷, we observed that older adults exhibited a relatively dampened modulation of decision processes and neural responses to varying contextual uncertainty. Older adults expressing more flexible feature selection were marked by more “young-like” modulation of neural signatures, providing first evidence for a brain maintenance account in the context of uncertainty processing.

2. Results

2.1 Older adults express constrained uncertainty modulation of evidence integration

During EEG and fMRI acquisition, participants performed a Multi-Attribute Attention Task (“MAAT” ²⁷; Figure 1a). In the task, participants had to visually sample a moving display of squares that were characterized by four feature dimensions, with two exemplars each: color (red/green), movement direction (left/right), size (large/small), and color saturation (high/low). Stimuli were presented for three seconds, after which participants were probed as to which of the two exemplars of a single feature was most prevalent. Probe uncertainty was parametrically manipulated using valid pre-stimulus cues that indicated the feature set from which a probe would be selected with equal probability. Higher uncertainty necessitated extra-dimensional attention shifts ^{43, 44} between up to four features (“target load”) to optimally inform probe-related decisions. Younger and older adults performed the task above chance level for all visual features (Figure S1-1).

Figure 1. Older adults show constrained decision-related adjustments to rising uncertainty.

(a) Participants performed a Multi-Attribute Attention Task (“MAAT”) during which they had to sample up to four visual features of a compound stimulus for a subsequent perceptual decision. On each trial, participants were first validly cued to a target probe set (here motion direction and color). The compound stimulus (which always included all four features) was then presented for 3 s and was followed by a probe of one of the cued features (here, whether red or green color was more prevalent in the stimulus). The number of pre-stimulus cues manipulated the level of uncertainty. Behavioral data were modelled with a drift diffusion model, in which evidence is presumed to be successively accumulated with a ‘drift rate’ towards either of two bounds, here representing the options of a single feature. (b) Drift rate estimates from behavioral modelling. Older adults exhibited reduced accumulation rates for single targes (top) and were marked by more limited drift reductions under elevated uncertainty (bottom). Individual data points represent averages across EEG and fMRI sessions. Table S1 reports within-group statistics. (c) The Centro-parietal positivity (CPP) provides an a priori neural signature of evidence accumulation. Older adults exhibited reduced integration slops for single targes (top) and were marked by constrained load-related slope shallowing under elevated uncertainty (bottom). To illustrate age- and condition-differences in integration slope, responses have been rescaled to the [0, 1] range for visualization. Fig. S1-3 shows original traces. ***a p=0e-10 ***b p = 5.1e-10 ***c p = 4.5e-05 ***d p = 2.8e-05.

To characterize probe-related decision processes, we fitted a hierarchical drift-diffusion model ⁴⁵ (HDDM) to participants’ responses. The model estimates (a) the drift rate at which evidence is integrated towards a decision bound, (b) the distance between (correct and incorrect) decision bounds, and (a) the non-decision time of visual probe processing and response execution. Across sessions and age groups the best fitting models (see Figure S1-2) consistently included uncertainty-based variations in all three parameters. Here, we focused on the drift rate of evidence integration based on its close association to stimulus processing ²⁷. Text S1-2 reports the remaining parameters. With rising uncertainty, evidence drift rates decreased for both age groups, indicating that uncertainty constrained decision evidence for the probed feature also in older adults. Crucially, relative to younger adults, older participants’ drift rates were reduced following single-attribute cues, and decreased less under increasing uncertainty (Figure 1b). These drift rate effects remained present when only features with age-matched single-target accuracies were included in the model (see Text S1-3). However, we also observed that for features matched in single-target accuracy, older adults suffered stronger accuracy decreases under uncertainty than younger adults, in line with a larger behavioral cost of transitioning into more uncertain task contexts (see Text S1-4).

We assessed the convergence of behavioral results with an a priori neural proxy signature of evidence integration, the slope of the EEG’s centroparietal positive potential (CPP ⁴⁶; Figure 1c, see also Figure S1-4) prior to decision responses. Consistent with behavioral modeling, CPP slopes were flatter for older relative to younger participants in single-target contexts, and older adults’ uncertainty-related modulation of CPP slopes was minimal (Figure 1c). In line with both indices capturing latent evidence integration, CPP and drift estimates were inter-individually related (Fig. S1-4), both for single targets (r(93) = 0.51, 95%CI = [0.34,0.64], p = 1.4e-07; age-partial: r(92) = 0.34, 95%CI = [0.14,0.5] p = 9.3e-04), and their uncertainty modulation (r(93) = 0.45, 95%CI = [0.27,0.59], p = 6.1e-06; age-partial: r(92) = 0.27, 95%CI = [0.08,0.45], p = 0.01; Fig S1-4c). We also probed contralateral beta power as a signature of motor response preparation ⁴⁷ (Figure S1-5) but did not observe clear relations to drift rate or CPP estimates (Text S1-5), suggesting that it may be a less suitable evidence integration index here. Taken together, older adults’ decisions were marked by reduced evidence integration rates for single targets, and more constrained drift rate reductions under uncertainty.

2.2 Decoding indicates uncertainty-induced trade-offs between feature specificity and sensitivity

Higher single-target drift rates and larger evidence reductions may reflect an adaptive trade-off between reduced single-feature specificity and elevated sensitivity to multiple features under higher uncertainty. However, as decisions were tied to the probed feature, they cannot elucidate how unprobed feature dimensions were processed. To clarify this question, we performed fMRI decoding analyses. We created pairwise classifiers that targeted the sensory representation of each feature’s prevalent option (e.g., left vs. rightward movement) based on BOLD responses in visual cortex (see Methods: fMRI decoding of prevalent feature options). The prevalent option of individual features could be decoded above chance during the approximate time of stimulus presentation (Fig. 2a). Robust decoding was observed for all cued features except for luminance, for which discrimination was also behaviorally most challenging (Fig. S1-1). Above-chance decoding was not observed for uncued feature options, except for motion discrimination (see Fig. 2b), indicating that participants mainly discriminated task-relevant feature options ¹⁸.

Figure 2. Decoding of prevalent options from visual cortex.

(a) Decoding accuracy for cued and uncued features across age groups (means +/- SEM). Grey shading indicates the approximate timing of stimulus presentation considering the temporal lag in the hemodynamic response. Lines indicate periods of statistically significant differences from chance decoding accuracy (50%) as assessed by cluster-based permutation tests. The inset highlights the visual cortex mask from which signals were extracted for decoding. (b) Same as in a, but for each feature probe. (c) Decoding accuracy for probed (left) and unprobed (center) features as a function of the number of cued targets; and decoding accuracy for al features as a function of age (right). Accuracy was averaged across significant decoding timepoints for cued features. Means +- within-subject SEM for (un)probed features, means +- SEM for age analysis.

Next, we assessed uncertainty and age effects on decoding accuracy. First, we applied classifiers to trials in which target features were probed, which mirrors the participant’s behavioral task. A linear mixed effects model indicated a significant reduction in decoding accuracy with increasing uncertainty (β = -0.18, SE = 0.05, t = -3.56, p = 0.00037; Figure 2c), as well as reduced decoding accuracy for older adults (β = -0.862, SE = 0.31, t = -2.77, p = 0.007), but no significant interaction (p = 0.76). Crucially, such uncertainty-related precision losses may trade-off against sensitivity to other cued, but ultimately unprobed features. We tested this possibility by considering decoding accuracy across all unprobed features in any given trial. This analysis indicated that uncertainty indeed slightly increased decoding accuracy across unprobed features (β = 0.077, SE = 0.026, t = 2.94, p = 0.0033). Decoding accuracy trended to be lower in older compared to younger adults (β = -0.259, SE = 0.134, t = -1.92, p = 0.0574). Again, no significant interaction was observed (p = 0.434). Consistent with opposing uncertainty effects on probed and unprobed features, no significant uncertainty effect was indicated when all trials were considered (β = 0.012, SE = 0.024, t = 0.53, p = 0.5927), but decoding accuracy was overall reduced in older adults (β = -.41, SE = 0.144, t = -2.84, p = 0.0056). Decoding analyses thus suggest that rising uncertainty in both age groups increased sensitivity to more diverse features, albeit at the cost of reduced precision for single features.

2.3 MAAT performance generalizes to feature selection in the context of low perceptual demands

Relative to younger adults, older adults appear to have encoded less single-target evidence for downstream decisions. However, the multifaceted demands of the MAAT do not resolve whether such differences arise from task idiosyncrasies such as the necessity to resolve high perceptual uncertainty for each feature, or whether they capture differences related to flexible feature selection. To adjudicate between these accounts, participants also performed a Stroop task, which probes the capacity to inhibit prepotent responses to one of two feature dimensions (the color vs. semantics) of a presented word ⁴⁸. We recorded voice responses as a more naturalistic modality for older adults ⁴⁹. To estimate speech onset times (SOTs ∼ reaction times), we labeled the onset of voiced responses in each trial’s recording (see methods). Labeled SOTs showed high validity as the neural CPP peaked immediately prior to SOTs (Fig. 3a). In line with the Stroop literature ⁴⁹, older adults incurred larger behavioral interference costs (Fig. 3b) than younger adults. These behavioral results were mirrored by neural CPP slopes: interference shallowed pre-response CPP slopes in both age groups, but to a larger extent in older adults, and the CPP shallowing tracked behavioral interference costs across subjects (Fig. S3-1). Crucially, participants with higher MAAT drift rates were also faster responders in the incongruent condition (Fig. 3c), pointing to a better capacity to inhibit prepotent responses. Notably, relations between MAAT drift rates and SOTs in the Stroop interference condition (r(93) = -0.65, 95%CI = [-0.75,-0.51], p = 1.2e-12) held after controlling for age and SOTs in the congruent condition (r(91) = -0.29, 95%CI = [-0.46,-0.09], p = 0.01), whereas the opposite was not observed (congruent SOTs-drift: r(93) = -0.4, 95%CI = [-0.56,-0.22], p = 4.7e-05, age- and incongruent SOT-partial: r(91) = 0.13, 95%CI = [-0.07,0.33], p = 0.2). As such, selective inhibition of interfering features, as opposed to processing speed, appears to be a key contributor to individual MAAT drift rates. Taken together, these findings suggest that individual and age differences in MAAT drift rates generalize to flexible feature selection also in perceptually unambiguous contexts.

Figure 3. MAAT evidence integration relates to prepotent response inhibition.

(a) Centro-Parietal Positivity (CPP) traces and speech signal power suggest high validity for the semi-automatically labeled speech onset times (SOTs). The CPP trace has been averaged across age and congruency conditions and displays means +/- SEM. The inset shows the mean EEG topography during the final 300 ms prior to speech onset. (b) The voiced Stroop task indicated robust interference costs whose magnitude was larger in older adults. Table S1 reports within-group statistics. (c) Participants with larger MAAT drift rates showed faster responses to incongruent trials (e.g., responding blue to the inset stimulus), also after accounting for categorical age (squares: younger; diamonds: older) and covariation with congruent SOTs (see main text).

2.4 Theta power and pupil diameter upregulation with elevated uncertainty dampens in older age

Our results indicate age-related constraints in adjusting perceptual and decision processes to varying uncertainty. To test whether such constraints are rooted in a reduced neural uncertainty response as expected under a maintenance account of cognitive & brain aging, we assessed several a priori signatures (see ²⁷) during MAAT stimulus presentation by means of two-group task partial-least-squares analyses (PLS, see methods). First, we assessed the effect of uncertainty on frontocentral theta power, an index of cognitive control ⁵⁰ and exploration under uncertainty ⁵¹. Uncertainty increased theta power in both age groups (Figure 4a), but to a lesser extent in older adults (Figure 4a). Next, we assessed phasic changes in pupil diameter, a signature that covaries with neuromodulation and arousal ^{52, 53}, has been related to frontal control ^{2, 27, 54–56}, and is sensitive to rising demands ⁵⁷ such as dynamically changing and uncertain contexts ^{58, 59}. Once again, we observed that uncertainty increased pupil diameter in both age groups, with more constrained upregulation in older adults (Fig. 4b). The extent of pupil modulation was related to individual theta power increases (r(98) = .28, 95%CI = [0.09, 0.46], p = 0.005; age-partial: r(97) = .19, 95%CI = [0, 0.38], p = 0.05), indicating a joint uncertainty modulation. These results indicate that both age groups were sensitive to rising uncertainty, albeit older adults to a dampened extent.

Figure 4. Uncertainty increases theta power (a) and pupil diameter (b) across the adult lifespan.

(Center) Age comparison of linear uncertainty effects (∼age x target load interaction). Red bars indicate significant within-group differences from zero, as assessed via one-sample t-tests (see Table S1). Both signatures exhibited significant uncertainty modulation in younger, as well as older adults, with constrained modulation in older adults. For condition-wise plots, see Fig. S4-1. Statistics refer to unpaired t-tests. (Right) Time series data are presented as means +- within-subject S.E.Ms. Orange shading in a indicates the timepoints across which data have been averaged for the task PLS. Black lines in b indicate time points exceeding a BSR of 3 (∼99% threshold). The uncertainty modulation of pupil diameter occurred on top of a general pupil constriction due to stimulus-evoked changes in luminance upon task onset (see inset), that by stimulus design did not systematically differ across load levels. YA = Younger adults. OA = Older adults.

2.5 Only younger adults adjust posterior cortical excitability to varying uncertainty

Elevated contextual uncertainty may impact perception by altering sensory excitability. To test this, we focused on three indices related to cortical excitability: alpha power, sample entropy, and aperiodic 1/f slopes ^{27, 60}. We constrained analyses to posterior sensors as we targeted perceptual changes in visual-parietal cortices. Text S5-3 reports whole-channel analyses. In younger adults, we observed uncertainty effects on all three signatures (Fig. 5 a-c), akin to those we previously reported ²⁷. In line with putative excitability increases, posterior alpha power decreased alongside uncertainty, while sample entropy increased and the aperiodic spectral slope shallowed. However, we found no evidence of a similar modulation in older adults for any of the probed signatures (Fig. 5, see also Fig. S4-1), indicating a failure of the aged system to adjust to changing uncertainty demands. Such failure may be rooted in a less precise estimation of environmental uncertainty in the aged neural system ¹⁶. However, we reduced inference demands in our design by providing overt cues on each trial, and keeping the cue set identical for eight consecutive trials. In line with age-invariant sensitivity to uncertainty cues, we observed comparable increases in pre-stimulus alpha power alongside uncertainty in both age groups (Fig. S5-1, see also Text S5-1). However, these increases were not associated with subsequent behavioral drift rate adjustments (Fig. S5-1 and Text S5-1), arguing against a direct role of pre-stimulus alpha power in adjudicating uncertainty. We additionally considered the steady-state visual evoked potential (SSVEP) as a proxy of bottom-up processing. Despite robust and comparable SSVEPs in both age groups, we found no evidence of uncertainty modulation in either group (Fig. S5-2, see also Text S5-2). Given that the 30 Hz flicker frequency was shared between all stimulus features, this suggests that sensory processing of the compound stimulus was similar between uncertainty conditions and age groups. Taken together, our results suggest that older adults may have suffered from a relative failure to adjust perceptual excitability to changing feature relevance, rather than insensitivity to the level of contextual uncertainty or an inability to encode the undifferentiated stimulus.

Figure 5. Only younger adults upregulate cortical excitability under increased uncertainty.

(a-c) Results of task partial least squares (PLS) models, assessing relations of alpha power (a), sample entropy (b) and aperiodic 1/f slope (c) to uncertainty. (Left) Topographies indicate mean bootstrap ratios (BSR). Orange dots indicate the sensors across which data were averaged for data visualization. (Center) Age comparison of linear uncertainty effects (∼age x uncertainty interaction). All three signatures exhibited a significant uncertainty modulation in younger, but not in older adults. For condition-wise plots, see Fig. S4-1. Statistics refer to unpaired t-tests. Table S1 reports within-group statistics. (Right) Time series data are presented as means +- within-subject S.E.Ms. Orange shading in a indicates the timepoints across which data have been averaged for the respective task-PLS. Black lines in b indicate time points exceeding a BSR of 3 (∼99% threshold). YA = Younger adults. OA = Older adults.

2.6 BOLD modulation links neuro-behavioral responses to uncertainty across the adult lifespan

Finally, we investigated uncertainty-related changes in whole-brain fMRI BOLD activation during stimulus presentation, extending sensitivity also to subcortical areas like the thalamus that are considered critical for managing contextual uncertainty ^{27, 61, 62}. We targeted associations between uncertainty-related BOLD modulation and the a priori neurobehavioral signatures (i.e., uncertainty-induced changes in drift rate, theta power, pupil diameter, alpha power, 1/f slopes, and sample entropy) using a multivariate behavioral PLS analysis (see Methods; Text S4-1 reports a task PLS targeting the main effect of uncertainty). We identified a single latent variable (LV; permuted p < 1e-3) with positive frontoparietal and thalamic loadings, and most pronounced negative loadings in medial PFC and hippocampus (Fig. 6a, Table S5). Older adults expressed this LV to a lesser extent than younger adults as indicated by lower “Brainscores” (Fig. 6b), indicating dampened BOLD modulation in the face of changing uncertainty. Brainscores were associated with the latent score of neurobehavioral input signatures (Fig. 6c), such that less dampened, more “young-like” BOLD modulation tracked a larger modulation of decision, EEG, and pupil signatures. Fig. 6d depicts relations to the individual signatures of the model: across age groups, greater BOLD modulation corresponded to larger drift rate reductions, more pronounced theta power and pupil diameter increases, and larger excitability modulation (see Fig. S6-2 for more signatures). As the PLS model leveraged variance both from within and across age groups, we used linear-mixed-effects models to assess the age-dependency of these relations. These models indicated that all a priori signatures, except sample entropy and 1/f modulation, predicted Brainscores also after accounting for the shared main effects of age (Table 1). This indicates a robust coupling of uncertainty effects between most signatures, while aligning with unobserved posterior excitability modulation in older adults. Control analyses indicate that within- and between-group differences in BOLD uncertainty sensitivity are robust to matched feature accuracy (see Fig S6-3).

Figure 6: Multivariate relation of EEG/pupil/behavioral signatures to fMRI BOLD uncertainty modulation.

(a) Results of a behavioral partial least squares (PLS) analysis linking linear changes in BOLD activation to interindividual EEG, pupil, and behavioral differences. Table S4 lists peak coordinates. (b) The multivariate expression of BOLD changes alongside rising uncertainty was reduced in older compared with younger adults. Table S1 reports within-group statistics. (c) Individual Brainscore differences were related to a behavioral composite score, also after accounting for age covariation. Squares = younger individuals; diamonds = older individuals. (d) Signature-specific Brainscore relations. All signature estimates refer to linear changes as a function of uncertainty. Error bars represent bootstrapped 95% confidence values. (e) Major nuclei and projection zones in which behavioral relations are maximally reliable according to average Bootstrap ratios (red) and the percentage of voxels in each subregion exceeding a BSR of 3. See Methods for abbreviations. Strongest expression is observed in nuclei that project to fronto-parietal cortical targets. (f) Visualization of uncertainty modulation for the mediodorsal nucleus, a “higher order” nucleus, and the LGN, a visual relay nucleus. Traces display mean +/- SEM. The green shading indicates the approximate stimulus presentation period after accounting for the delay in the hemodynamic response function.

Table 1: Summary of Brainscore predictors, while controlling for categorical age.

Separate linear-mixed-effects models assessed effects of target signature, categorical age, and age x signature interactions on Brainscores. We observed no significant interaction in any of the models (all p > 0.05), pointing to consistent relations across age groups; therefore, all reported models only include main effects of signature and age. Fig. S6-2 reports similar results using partial regressions. Degrees of freedom: 92 in all models.

Behavioral relations were closely tracked by BOLD activation in the thalamus. To obtain insights within this differentiated structure, we assessed regional loadings based on projection zones and nucleus segmentations (Fig, 6e). Loadings were highest in subregions with frontoparietal projections, including the mediodorsal nucleus (Fig. 6f). In contrast, a traditional visual “relay” nucleus of the thalamus, the lateral geniculate nucleus, did not show sensitivity to our uncertainty manipulation (Fig. 6f). This indicates a specificity of thalamic effects that coheres with functional subdivisions and alludes to uncertainty-invariant sensory processing of the compound stimulus. These results indicate that the mediodorsal thalamus contributes to maintained uncertainty adjustments across the adult lifespan.

3. Discussion

Managing uncertainty is vital for navigating the flux of life. While some environments prioritize specific inputs over others, many contexts provide few, contrasting, or ambiguous cues. Here, we show that healthy older adults exhibit markedly dampened adaptations to such varying uncertainty across coupled EEG/fMRI/pupil signatures. Our results extend observations that older adults rely less on uncertainty representations to guide internal computations ¹⁶ by characterizing several plausible neural mechanisms for this shortfall. Our results suggest that such computational constraints do not exclusively stem from an inadequate sensitivity to latent uncertainty, as the current task provides overt uncertainty cues that are similarly processed by both age groups. Rather, our findings support the “maintenance” account of cognitive/brain aging ⁶³ in the context of uncertainty processing, wherein individuals with a more “young-like” neural recruitment are better able to leverage comparable uncertainty estimates to adjust ongoing computations.

3.1 Fronto-thalamic circuits may enable stable and flexible feature selection across the adult lifespan

As part of the neural uncertainty response, we observed a behaviorally relevant upregulation of anterior cingulate cortex (ACC) BOLD activation and (presumably ACC-based ^{50, 64}) mediofrontal theta power. By charting the progression through multiple task contexts ^65–67, the ACC can estimate contextual volatility ⁶⁸ and uncertainty ^{16, 69} to guide exploration of alternative goals, strategies, and attentional targets ^{51, 70–72}. Non-human animal studies suggest that high contextual uncertainty switches ACC dynamics to a state of increased excitability ^{60, 73} and stochastic activity ⁷⁴, which benefits concurrent sensitivity to alternate task rules ⁷⁵. Also in humans, the ACC is sensitive to stimulus features before they behaviorally guide task strategies ^{74, 76}, suggesting that the ACC contributes to the exploration of alternate features whose significance remains contextually unclear ^{77, 78}. While our results align with such contribution, we also localize high uncertainty sensitivity in the mediodorsal (MD) thalamus, which aligns with the MD being a key partner for selecting, switching, and maintaining cortical task representations ^{23, 79, 80} especially in uncertain contexts ^{27, 61, 62} . Extrapolating from this emerging perspective, the MD-ACC circuit may regulate task set stability vs. flexibility ^81–83 according to contextual demands (Fig. 7a). Partial evidence for such a notion is provided by models that link task stability in low-uncertainty contexts to thalamic engagement ⁸⁴. The current observations suggest a complementary thalamic role in task flexibility. While maintained across the adult lifespan, BOLD and theta power signals indicated that such MD-ACC upregulation was dampened in older adults ^{85, 86}. Indeed, the ACC network is particularly susceptible to age-related metabolic declines ^30–32 as well as structural atrophy ³⁸. Retained ACC function on the other hand is a hallmark of cognitive reserve ⁸⁷, relates to maintained executive function ³², and is a fruitful target of cognitive interventions in older adults ⁸⁶. Given evidence of a key role of the MD thalamus in the coordination of ACC engagement and our observations of reduced MD-ACC sensitivity to uncertainty in older age, the thalamus may be an underappreciated site for cascading age-related dysfunctions in cognitive stability and flexibility.

Figure 7. Schematic model summary.

(a) In static contexts, prefrontal-hippocampal networks may signal high confidence in the current task state, which enables stable task sets, and a targeted processing of specific sensory representations with high acuity. Such selective processing of specific task-relevant features benefits their efficient evidence integration. Such selectivity would be suboptimal in contexts with uncertain or changing task sets, however. An MD-ACC circuit may track such uncertainty and enhance stochastic task set flexibility in changing or ambiguous contexts. In coordination with posterior-parietal cortex, this feasibly enables more diverse albeit less precise perceptual representations. (b) The neural system of younger adults adjusts system dynamics to the degree of environmental uncertainty. Observed effects align with a switch between a specific processing of individual features with high acuity, as exemplified by a single, deep attractor (blue), and a more diverse, if less precise processing of multiple features, as indicated by a more unstable attractor landscape (red; see also Thiele & Bellgrove, 2018). In contrast, the aged neural system may be stuck in a suboptimal middle ground that affords neither stable precision, nor flexible imprecision. mPFC = medial prefrontal cortex; HC = hippocampus; ACC = anterior cingulate cortex; MD = mediodorsal thalamus.

3.2 Neuromodulation may sculpt the dynamic range of uncertainty adjustments

Neurotransmitter systems provide a candidate substrate for computational adjustments under uncertainty. In response to rising uncertainty, phasic norepinephrine release can sensitize the system to incoming signals ^{88, 89} by increasing neuro-behavioral activation ^{52, 90}. Pupil diameter, an index that is partially sensitive to noradrenergic drive ⁵⁷, robustly increases alongside uncertainty during learning ⁵⁸ and attention ⁹¹, environmental exploration ⁹², and change points in dynamic environments ^{58, 59, 93}. Here, we show that such pupil sensitivity to rising uncertainty is retained across the adult lifespan, but dampens in older age. Such dampening hints at declining noradrenergic responsiveness in older age ^{94, 95}, arising from reduced LC integrity ⁹⁶, and/or decreased LC recruitment. Notably, pupil sensitivity to volatility has been related to the ACC as a primary source of cortical LC input ^{28, 97}, and joint modulation of ACC and pupil diameter in uncertain, or dynamic contexts has consistently been observed in studies that record both signals ^{2, 27, 54–56}. While future studies need to clarify the origin of constrained pupil adjustments in older age, our results affirm the relevance of the extended LC system for attentional function across the lifespan ⁹⁵. In contrast to noradrenaline’s potential role in sensitizing, cholinergic innervation from the basal forebrain may foster selectivity via cortical gain increases ^{98, 99}. Notably, basal forebrain BOLD activation decreased under uncertainty alongside regions such as the medial prefrontal cortex and hippocampus, that are sensitive to subjective confidence ¹⁰⁰, suggesting that it may support stable task beliefs when contextual uncertainty is low ¹⁰¹ (Fig. 7a). The constrained BOLD modulation observed in older adults may thus point to reduced task set stability in low-uncertainty contexts (Fig. 7b) ¹¹, plausibly as a consequence of limited cholinergic gain control. Similar ideas have been captured in the cortical gain theory of aging ¹⁰², but in the context of the dopamine system ^{34, 103}. Computational models and pharmacological studies indeed support a role of dopamine availability in task set stability and flexibility ^{104, 105}. For instance, amphetamines (operating via the DA system) can in- and decrease task set stability in ACC ^{106, 107} depending on baseline dopamine levels in frontoparietal cortex and thalamus ¹⁰⁸. Given that our results align with the fronto-thalamic system being a primary neural substrate of cognitive aging ^{34, 39, 109}, the potential contribution of age-related dopamine depletion to constrained uncertainty adjustments deserves future clarification.

3.3 Excitability modulation as a mechanism for acuity/sensitivity trade-offs

Uncertain contexts motivate perceptual exploration over a selective encoding of individual features. Our decoding results indeed indicate that higher uncertainty benefitted sensitivity to multiple features at the cost of feature-specific precision (or “acuity”) ³. Perceptual representations thus depend on whether a feature is included in the active task set ¹⁸, but also on the degree of competition with other task set elements for neuro-computational resources ¹¹⁰. Excitability changes in parietal/sensory cortices provide a candidate mechanism that may implement such trade-off. One index of (decreased) cortical excitability is alpha power. Models suggest that broad alpha power increases reflect active inhibition of irrelevant information ^111–115, while alpha desynchronization in target regions can selectively disinhibit relevant information ³⁸. With advancing adult age, alpha power decreases, which has been linked to inhibitory deficits in older age ^{95, 116–119} . Such filtering deficits manifest in maladaptive sensitivity also to irrelevant ⁷ and non-salient features ¹²⁰ of compound stimuli ⁶ that impairs selective feature discrimination as required in the MAAT. Decoding and decision analyses indeed indicate that older adults’ task performance suffered from reduced single-feature information, in line with filtering deficits ^{121, 122}. Alpha desynchronization, in turn, is thought to reflect increased sensitivity to multiple input features ²⁶. In line with such a notion, stronger alpha suppression is observed when multiple features must be jointly tracked ^{123, 124} and retained in working memory ^125–128. In addition to alpha power, aperiodic dynamics such as the spectral slope of the EEG potential ¹²⁹ and signal entropy ¹³⁰ may also index levels of neural excitability ^{60, 129}. Here, we reproduce the observation that uncertainty increases excitability as assessed by all three signatures in younger adults ²⁷, but find no evidence for a comparable modulation in older adults. Such deficits in excitability modulation may be rooted in age-related declines of GABAergic inhibition ^{36, 37}. Aperiodic dynamics at rest suggest increased excitatory tone with increased adult age ^131–133, including in the current sample ¹³⁰. Our results suggest that such imbalances ¹³⁴ may constrain the dynamic range of excitability modulation in older age, both on- and off-task ^{42, 135}. Ultimately, this may point to dual challenges in implementing selective attention, as well as diverse feature coding under uncertainty (Fig. 7b). It is also possible that the consistently high level of perceptual uncertainty, i.e., the difficulty of arbitrating between the two options of each feature, was overly taxing especially for older participants. Based on behavioral and decoding results, younger adults were indeed better able to arbitrate feature-specific options at all levels of contextual uncertainty, relative to older adults. In this scenario, preserved excitability modulation may be observed if individual features were perceptually less uncertain. However, performance on the Stroop task suggests that age-related deficits (and individual differences) in feature selection generalize to contexts of low perceptual uncertainty. As perceptual uncertainty resolution relies on partially dissociable circuits from those implicated in feature selection ^136–138, future work needs to chart the ability to resolve either type across the lifespan.

3.4 Conclusion

Changes in uncertainty provide an important signal that adaptive systems can use to adjust their internal computations. We highlight that such uncertainty-related adjustments present a principled challenge for the aged brain. Our results thus argue that uncertainty provides a useful lens on healthy cognitive aging and underline the need to better understand the integrated neural basis of estimating and computationally leveraging uncertainty signals across the lifespan.

Online Methods

Sample

47 healthy young adults (mean age = 25.8 years, SD = 4.6, range 18 to 35 years; 25 women) and 53 healthy older adults (mean age = 68.7 years, SD = 4.2, range 59 to 78 years; 28 women) performed a perceptual decision task during 64-channel active scalp EEG acquisition. 42 younger adults and all older adults returned for a subsequent 3T fMRI session. Participants were recruited from the participant database of the Max Planck Institute for Human Development, Berlin, Germany (MPIB). Participants were right-handed, as assessed with a modified version of the Edinburgh Handedness Inventory ¹³⁹, and had normal or corrected-to-normal vision. Participants reported to be in good health with no known history of neurological or psychiatric incidences, and were paid for their participation (10 € per hour). All older adults had Mini Mental State Examination (MMSE) ^{140, 141} scores above 25. All participants gave written informed consent according to the institutional guidelines of the Deutsche Gesellschaft für Psychologie (DGPS) ethics board, which approved the study.

Procedure: EEG Session

Participants were seated 60 cm in front of a monitor in an acoustically and electrically shielded chamber with their heads placed on a chin rest. Following electrode placement, participants were instructed to rest with their eyes open and closed, each for 3 minutes. Afterwards, participants performed a Stroop task (see below), followed by the visual attention task instruction & practice (see below), the performance of the task and a second Stroop assessment. Stimuli were presented on a 60 Hz 1920×1080p LCD screen (AG Neovo X24) using PsychToolbox 3.0.11 ^142–144. The session lasted ∼3 hours. EEG was continuously recorded from 60 active (Ag/AgCl) electrodes using BrainAmp amplifiers (Brain Products GmbH, Gilching, Germany). Scalp electrodes were arranged within an elastic cap (EASYCAP GmbH, Herrsching, Germany) according to the 10% system ¹⁴⁵, with the ground placed at AFz. To monitor eye movements, two additional electrodes were placed on the outer canthi (horizontal EOG) and one electrode below the left eye (vertical EOG). During recording, all electrodes were referenced to the right mastoid electrode, while the left mastoid electrode was recorded as an additional channel. Online, signals were digitized at a sampling rate of 1 kHz. In addition to EEG, we simultaneously tracked eye movements and assessed pupil diameter using EyeLink 1000+ hardware (SR Research, v.4.594) with a sampling rate of 1kHz.

Procedure: MRI session

A second testing session included structural and functional MRI assessments. First, participants took part in a short refresh of the visual attention task (“MAAT”, see below) instructions and practiced the task outside the scanner. Then, participants were placed in the TimTrio 3T scanner and were instructed in the button mapping. We collected the following sequences: T1w, task (4 runs), T2w, resting state, DTI, with a 15 min out-of-scanner break following the task acquisition. The session lasted ∼3 hours. Whole-brain task fMRI data (4 runs á ∼11,5 mins, 1066 volumes per run) were collected via a 3T Siemens TrioTim MRI system (Erlangen, Germany) using a multi-band EPI sequence (factor 4; TR = 645 ms; TE = 30 ms; flip angle 60°; FoV = 222 mm; voxel size 3×3×3 mm; 40 transverse slices. The first 12 volumes (12 × 645 ms = 7.7 sec) were removed to ensure a steady state of tissue magnetization (total remaining volumes = 1054 per run). A T1-weighted structural scan (MPRAGE: TR = 2500 ms; TE = 4.77 ms; flip angle 7°; FoV = 256 mm; voxel size 1×1×1 mm; 192 sagittal slices) and a T2-weighted structural scan were also acquired (GRAPPA: TR = 3200 ms; TE = 347 ms; FoV = 256 mm; voxel size 1×1×1 mm; 176 sagittal slices).

The multi-attribute attention task (“MAAT”)

The MAAT requires participants to sample up to four visual features in a compound stimulus, in the absence of systematic variation in bottom-up visual stimulation (see Figure 1). Participants were shown a dynamic square display that jointly consisted of four attributes: color (red/green), movement direction (left, right), size (small, large) and saturation (low, high). The task incorporates features from random dot motion tasks which have been extensively studied in both animal models ^146–148 and humans ^{46, 149}. Following the presentation of these displays, a probe queried the prevalence of one of the four attributes in the display (e.g., whether the display comprised a greater proportion of either smaller or larger squares) via 2-AFC (alternative forced choices). Prior to stimulus onset, a varying number of valid cues informed participants about the active feature set, out of which one feature would be chosen as the probe. We parametrically manipulated uncertainty regarding the upcoming probe by systematically varying the number of cues between one and four.

The perceptual difficulty of each feature was determined by (a) the fundamental feature difference between the two alternatives and (b) the sensory evidence for each alternative in the display. For (a) the following values were used: high (RGB: 128, 255, 0) and low saturation green (RGB: 192, 255, 128) and high (RGB: 255, 0, 43) and low saturated red (RGB: 255, 128, 149) for color and saturation, 5 and 8 pixels for size differences and a coherence of .2 for directions. For (b) the proportion of winning to losing option (i.e., sensory evidence) was chosen as follows: color: 60/40; direction: 80/20; size: 65/35; luminance: 60/40. Parameter difficulty was established in a pilot population, with the aim to produce above-chance accuracy for individual features. Parameters were held constant across age groups to retain identical bottom-up inputs.

The experiment consisted of four runs of ∼10 min, each including eight blocks of eight trials (i.e., a total of 32 trial blocks; 256 trials). The size and constellation of the cue set was held constant within eight-trial blocks to reduce set switching and working memory demands. At the onset of each block, the valid cue set, composed of one to four target features, was presented for 5 s. Each trial was structured as follows: recuing phase (1 s), fixation phase (2 s), dynamic stimulus phase (3 s), probe phase (incl. response; 2 s); ITI (un-jittered; 1.5 s). At the offset of each block, participants received performance feedback for 3 s. The four attributes spanned a constellation of 16 feature combinations (4×4), of which presentation frequency was matched within participants. The size and type of the cue set was pseudo-randomized: Within each run, every set size was presented once, but never directly following a block of the same set size. In every block, each feature in the active set acted as a probe in at least one trial. Moreover, any attribute served as a probe equally often across blocks. The dominant options for each feature were counterbalanced across all trials of the experiment. To retain high motivation during the task and encourage fast and accurate responses, we instructed participants that one response would randomly be drawn at the end of each block; if this response was correct and faster than the mean RT during the preceding block, they would earn a reward of 20 cents. However, we pseudo-randomized feedback such that all participants received an additional fixed payout of 10 € per session. This bonus was paid at the end of the second session, at which point participants were debriefed.

Stroop performance

Participants performed a voiced Stroop task before and after the main MAAT task in the EEG session. EEG signals were acquired during task performance. One subject did not complete the second Stroop acquisition. In the Stroop task, we presented three words (RED, GREEN, BLUE) either in the congruent or incongruent display color. Each of the two runs consisted of 81 trials, with fully matched combinations, i.e., 1/3^rd congruent trials. Stimuli were presented for two seconds, followed by a one-second ITI with a centrally presented fixation cross. Participants were instructed to indicate the displayed color as fast and accurately as possible following stimulus onset by speaking into a microphone. During analysis, speech on- and offsets were pre-labeled automatically using a custom tool (Computer-Assisted Response Labeler (CARL); doi: 10.5281/zenodo.7505622), and manually inspected and refined by one of two trained labelers. Voiced responses were manually labeled using the CARL GUI. Speech onset times (SOTs) were highly reliable across two Stroop sessions preceding and following the MAAT (r = .83, p =5e-26), as were individual interference costs (r = .64, p =5e-13). We therefore averaged SOTs estimates across both runs, where available. For EEG analyses, single-trial time series were aligned to SOTs, and averaged according to coherence conditions. The centroparietal positive potential was extracted from channel POz, at which we observed a maximum potential during the average 300 ms prior to SOT (see inset in Fig. 3a).

Behavioral estimates of probe-related decision processes

Sequential sampling models, such as the drift-diffusion model, have been used to characterize evolving perceptual decisions in 2-alternative forced choice (2AFC) random dot motion tasks ⁴⁶, memory retrieval ¹⁵⁰, and probabilistic decision making ¹⁵¹. We estimated individual evidence integration parameters within the HDDM 0.6.0 toolbox ⁴⁵ to regularize relatively sparse within-subject data with group priors based on a large number of participants. Premature responses faster than 250 ms were excluded prior to modeling, and the probability of outliers was set to 5%. 7000 Markov-Chain Monte Carlo samples were sampled to estimate parameters, with the first 5000 samples being discarded as burn- in to achieve convergence. We judged convergence for each model by visually assessing both Markov chain convergence and posterior predictive fits. Individual estimates were averaged across the remaining 2000 samples for follow-up analyses. We fitted data to correct and incorrect RTs (termed ‘accuracy coding‘ in Wiecki, et al. ⁴⁵). To explain differences in decision components, we compared four separate models. In the ‘full model’, we allowed the following parameters to vary between conditions: (i) the mean drift rate across trials, (ii) the threshold separation between the two decision bounds, (iii) the non-decision time, which represents the summed duration of sensory encoding and response execution. In the remaining models, we reduced model complexity, by only varying (a) drift, (b) drift + threshold, or (c) drift + NDT, with a null model fixing all three parameters. For model comparison, we first used the Deviance Information Criterion (DIC) to select the model which provided the best fit to our data. The DIC compares models based on the maximal log-likelihood value, while penalizing model complexity. The full model provided the best fit to the empirical data based on the DIC index (Figure S1c) in both the EEG and the fMRI session, and in either age group. Posterior predictive checks indicated a suitable recovery of behavioral effects using this full solution. Given the observation of high reliability between sessions ²⁷ (see also Figure S1-2), we averaged parameter estimates across the EEG and fMRI sessions for the main analysis. In contrast with previous work ²⁷, we did not constrain boundary separation estimates ¹⁵² here given our observation of CPP threshold differences in older adults (see Figure S1-3a). See also Text 1-2 for a brief discussion of NDT and boundary separation.

EEG preprocessing

Preprocessing and analysis of EEG data were conducted with the FieldTrip toolbox (v.20170904) ¹⁵³ and using custom-written MATLAB (The MathWorks Inc., Natick, MA, USA) code. Offline, EEG data were filtered using a 4^th order Butterworth filter with a passband of 0.5 to 100 Hz. Subsequently, data were downsampled to 500 Hz and all channels were re-referenced to mathematically averaged mastoids. Blink, movement and heart-beat artifacts were identified using Independent Component Analysis (ICA; ¹⁵⁴) and removed from the signal. Artifact-contaminated channels (determined across epochs) were automatically detected using (a) the FASTER algorithm ¹⁵⁵, and by (b) detecting outliers exceeding three standard deviations of the kurtosis of the distribution of power values in each epoch within low (0.2-2 Hz) or high (30-100 Hz) frequency bands, respectively. Rejected channels were interpolated using spherical splines ¹⁵⁶. Subsequently, noisy epochs were likewise excluded based on a custom implementation of FASTER and on recursive outlier detection. Finally, recordings were segmented to stimulus onsets and were epoched into separate trials. To enhance spatial specificity, scalp current density estimates were derived via 4^th order spherical splines ¹⁵⁶ using a standard 1005 channel layout (conductivity: 0.33 S/m; regularization: 1^-05; 14^th degree polynomials).

Electrophysiological estimates of probe-related decision processes

Centro-Parietal Positivity (CPP)

The Centro-Parietal Positivity (CPP) is an electrophysiological signature of internal evidence-to-bound accumulation ^{46, 152, 157}. We probed the task modulation of this established signature and assessed its convergence with behavioral parameter estimates. To derive the CPP, preprocessed EEG data were low-pass filtered at 8 Hz with a 6^th order Butterworth filter to exclude low-frequency oscillations, epoched relative to response and averaged across trials within each condition. In accordance with the literature, this revealed a dipolar scalp potential that exhibited a positive peak over parietal channel POz (Fig. 1c). We temporally normalized individual CPP estimates to a condition-specific baseline during the final 250 ms preceding probe onset. As a proxy of evidence drift rate, CPP slopes were estimates via linear regression from -250 ms to -100 ms surrounding response execution, while the average CPP amplitude from -50 ms to 50 ms served as an indicator of decision thresholds (i.e., boundary separation; e.g., ¹⁵²).

Contralateral mu-beta

Decreases in contralateral mu-beta power provide a complementary, effector-specific signature of evidence integration ^{47, 152}. We estimated mu-beta power using 7-cycle wavelets for the 8-25 Hz range with a step size of 50 ms. Spectral power was time-locked to probe presentation and response execution. We re-mapped channels to describe data recorded contra- and ipsi-lateral to the executed motor response in each trial, and averaged data from those channels to derive grand average mu-beta time courses. Individual average mu-beta time series were baseline-corrected using the -400 to -200 ms prior to probe onset, separately for each condition. For contralateral motor responses, remapped sites C3/5 and CP3/CP5 were selected based on the grand average topography for lateralized response executions (see inset in Figure S2a). Mu-beta slopes were estimated via linear regression from - 250 ms to -50 ms prior to response execution, while the average power from -50 ms to 50 ms indexed decision thresholds (e.g., ¹⁵²).

Electrophysiological indices of top-down modulation during sensation

Low-frequency alpha and theta power

We estimated low-frequency power via a 7-cycle wavelet transform (linearly spaced center frequencies; 1 Hz steps; 2 to 15 Hz). The step size of estimates was 50 ms, ranging from -1.5 s prior to cue onset to 3.5 s following stimulus offset. Estimates were log10-transformed at the single trial level ¹⁵⁸, with no explicit baseline-correction.

Steady State Visual Evoked Potential (SSVEP)

The SSVEP characterizes the phase-locked, entrained visual activity (here 30 Hz) during dynamic stimulus updates (e.g., ¹⁵⁹). These features differentiate it from induced broadband activity or muscle artefacts in similar frequency bands. We used these properties to normalize individual single-trial SSVEP responses prior to averaging: (a) we calculated an FFT for overlapping one second epochs with a step size of 100 ms (Hanning-based multitaper) and averaged them within each uncertainty condition; (b) spectrally normalized 30 Hz estimates by subtracting the average of estimates at 28 and 32 Hz, effectively removing broadband effects (i.e., aperiodic slopes), and; (c) we subtracted a temporal baseline -700 to -100 ms prior to stimulus onset. Linear uncertainty effects on SSVEPs were assessed by paired t-tests on linear uncertainty slope estimates across posterior channel averages.

Time-resolved sample entropy

Sample entropy ¹⁶⁰ quantifies the irregularity of a time series of length N by assessing the conditional probability that two sequences of m consecutive data points will remain similar when another sample (m+1) is included in the sequence (for a visual example see Figure 1A in ¹³⁰). Sample entropy is defined as the inverse natural logarithm of this conditional similarity: The similarity criterion (r) defines the tolerance within which two points are considered similar and is defined relative to the standard deviation (∼variance) of the signal (here set to r = .5). We set the sequence length m to 2, in line with previous applications ¹³⁰. An adapted version of sample entropy calculations implemented in the mMSE toolbox (available from https://github.com/LNDG/mMSE) was used ^{130, 161, 162}, wherein entropy is estimated across discontinuous data segments to provide time-resolved estimates. The estimation of scale-wise entropy across trials allows for an estimation of coarse scale entropy also for short time-bins (i.e., without requiring long, continuous signals), while quickly converging with entropy estimates from continuous recordings ¹⁶¹. To remove the influence of posterior-occipital low-frequency rhythms on entropy estimates, we notch-filtered the 8-15 Hz alpha band using 6^th order Butterworth filter prior to the entropy calculation ¹³⁰. Time-resolved entropy estimates were calculated for 500 ms windows from -1 s pre-stimulus to 1.25 s post-probe with a step size of 150 ms. As entropy values are implicitly normalized by the variance in each time bin via the similarity criterion, no temporal baseline correction was applied.

Aperiodic (1/f) slopes

The aperiodic 1/f slope of neural recordings is closely related to the sample entropy of broadband signals ¹³⁰ and has been suggested as a proxy for cortical excitation-inhibition balance ¹²⁹. Spectral estimates were computed by means of a Fast Fourier Transform (FFT) over the final 2.5 s of the presentation period (to exclude onset transients) for linearly spaced frequencies between 2 and 80 Hz (step size of 0.5 Hz; Hanning-tapered segments zero-padded to 20 s) and subsequently averaged. Spectral power was log10-transformed to render power values more normally distributed across participants. Power spectral density (PSD) slopes were estimated using the fooof toolbox (v1.0.0-dev) using default parameters ¹⁶³.

Pupil diameter

Pupil diameter was recorded during the EEG session using EyeLink 1000 at a sampling rate of 1000 Hz and was analyzed using FieldTrip and custom-written MATLAB scripts. Blinks were automatically indicated by the EyeLink software (version 4.40). To increase the sensitivity to periods of partially occluded pupils or eye movements, the first derivative of eye-tracker-based vertical eye movements was calculated, z-standardized, and outliers >= 3 STD were removed. We additionally removed data within 150 ms preceding or following indicated outliers. Finally, missing data were linearly interpolated, and data were epoched to 3.5 s prior to stimulus onset to 1 s following stimulus offset. We quantified phasic arousal responses via the rate of change of pupil diameter traces as this measure (i) has higher temporal precision and (ii) has been more strongly associated with noradrenergic responses than the overall response ¹⁶⁴. We downsampled pupil time series to 100 Hz. For visualization, but not statistics, we smoothed pupil traces using a moving average median of 300 ms.

fMRI-based analyses

Preprocessing of functional MRI data

fMRI data were preprocessed with FSL 5 (RRID:SCR_002823) ^{165, 166}. Pre-processing included motion correction using McFLIRT, smoothing (7mm) and high-pass filtering (.01 Hz) using an 8^th order zero-phase Butterworth filter applied using MATLAB’s filtfilt function. We registered individual functional runs to the individual, ANTs brain-extracted T2w images (6 DOF), to T1w images (6 DOF) and finally to 3mm standard space (ICBM 2009c MNI152 nonlinear symmetric) ¹⁶⁷ using nonlinear transformations in ANTs 2.1.0 ¹⁶⁸ (for one participant, no T2w image was acquired and 6 DOF transformation of BOLD data was preformed directly to the T1w structural scan). We then masked the functional data with the ICBM 2009c GM tissue prior (thresholded at a probability of 0.25), and detrended the functional images (up to a cubic trend) using SPM12’s spm_detrend. We also used a series of extended preprocessing steps to further reduce potential non-neural artifacts ^{135, 169}. Specifically, we examined data within-subject, within-run via spatial independent component analysis (ICA) as implemented in FSL-MELODIC ¹⁷⁰. Due to the high multiband data dimensionality in the absence of low-pass filtering, we constrained the solution to 30 components per participant. Noise components were identified according to several key criteria: a) Spiking (components dominated by abrupt time series spikes); b) Motion (prominent edge or “ringing” effects, sometimes [but not always] accompanied by large time series spikes); c) Susceptibility and flow artifacts (prominent air-tissue boundary or sinus activation; typically represents cardio/respiratory effects); d) White matter (WM) and ventricle activation ¹⁷¹; e) Low-frequency signal drift ¹⁷²; f) High power in high-frequency ranges unlikely to represent neural activity (≥ 75% of total spectral power present above .10 Hz;); and g) Spatial distribution (“spotty” or “speckled” spatial pattern that appears scattered randomly across ≥ 25% of the brain, with few if any clusters with ≥ 80 contiguous voxels). Examples of these various components we typically deem to be noise can be found in ¹⁷³. By default, we utilized a conservative set of rejection criteria; if manual classification decisions were challenging due to mixing of “signal” and “noise” in a single component, we generally elected to keep such components. Three independent raters of noise components were utilized; > 90% inter-rater reliability was required on separate data before denoising decisions were made on the current data. Components identified as artifacts were then regressed from corresponding fMRI runs using the regfilt command in FSL. To reduce the influence of motion and physiological fluctuations, we regressed FSL’s 6 DOF motion parameters from the data, in addition to average signal within white matter and CSF masks. Masks were created using 95% tissue probability thresholds to create conservative masks. Data and regressors were demeaned and linearly detrended prior to multiple linear regression for each run. To further reduce the impact of potential motion outliers, we censored significant DVARS outliers during the regression as described by ¹⁷⁴. We calculated the ‘practical significance’ of DVARS estimates and applied a threshold of 5 ¹⁷⁵. The regression-based residuals were subsequently spectrally interpolated during DVARS outliers as described in ¹⁷⁴ and ¹⁷⁶. BOLD analyses were restricted to participants with both EEG and MRI data available (N = 42 YA, N = 53 OA).

fMRI decoding of prevalent feature options

We performed a decoding analysis to probe the extent to which participants’ visual cortices contained information about the prevalent option of each feature. N = 2 older adults with two missing runs each were not included in this analysis due to the limited number of eligible trials. We trained a decoder based on BOLD signals from within a visual cortex mask that included Jülich parcellations ranging from V1 to area MT. We resliced the mask to 3mm and created an intersection mask with the cortical grey matter mask used throughout the remaining analyses. For classification analyses, we used linear support-vector machines (SVM) ¹⁷⁷ implemented with libsvm (www.csie.ntu.edu.tw/~cjlin/libsvm). As no separate session was recorded, we trained classifiers based on all trials (across uncertainty conditions) for which the target feature was probed, therefore necessitating but not exhaustively capturing trials on which the respective feature was also cued. By experimental design, the number of trials during which a target feature was probed was matched across uncertainty levels. We used a bootstrap classification approach in the context of leave-one-out cross-validation to derive single-trial estimates of decoding accuracy. To increase the signal-to-noise ratio for the decoders, we averaged randomly selected trials into three folds (excluding any trial used for testing) and concatenated two pseudo-trials from each condition to create the training set. Trained decoders were then applied to the left-out trial. This train- and-test procedure was randomly repeated 100 times to create bootstrapped single-trial estimates. Finally, decoding accuracy was averaged across trials based on condition assignment (e.g., whether a given feature was cued or uncued). To assess above-chance decoding accuracy in time, we used univariate cluster-based permutation analyses (CBPAs). These univariate tests were performed by means of dependent samples t-tests, and cluster-based permutation tests ¹⁷⁸ were performed to control for multiple comparisons. Initially, a clustering algorithm formed clusters based on significant t-tests of individual data points (p <.05, two-sided; cluster entry threshold) with the spatial constraint of a cluster covering a minimum of three neighboring channels. Then, the significance of the observed cluster-level statistic (based on the summed t-values within the cluster) was assessed by comparison to the distribution of all permutation-based cluster-level statistics. The final cluster p-value was assessed as the proportion of 1000 Monte Carlo iterations in which the cluster-level statistic was exceeded. Cluster significance was indicated by p-values below .025 (two-sided cluster significance threshold). To test uncertainty and age effects, we initially fitted linear mixed effects models with random intercepts and fixed effects of uncertainty, age, and an uncertainty x age interaction. As no significant interaction was indicated for any of the models (probed: p = 0.760; unprobed: p = 0.434; all: p = 0.625), we removed the interaction term for the main effect estimation. We constrained analysis to timepoints for which the cluster-based permutation analysis indicated above-chance decoding for cued features. We focused on probed and unprobed feature trials, as they are matched in trial number at each uncertainty level.

BOLD modulation by uncertainty and relation to external variables

We conducted a 1^st level analysis using SPM12 to identify beta weights for each condition separately. Design variables included stimulus presentation (4 volumes; separate regressors for each uncertainty condition; parametrically modulated by sequence position), onset cue (no mod.), and probe (2 volumes, parametric modulation by RT). Design variables were convolved with a canonical HRF, including its temporal derivative as a nuisance term. Nuisance regressors included 24 motion parameters ¹⁷⁹, as well as continuous DVARS estimates. Autoregressive modelling was implemented via FAST. Output beta images for each uncertainty condition were finally averaged across runs. We investigated the multivariate modulation of the BOLD response at the 2^nd level using PLS analyses (see Multivariate partial least squares analyses). Specifically, we probed the relationship between voxel-wise 1^st level beta weights and uncertainty within a task PLS. Next, we assessed the relationship between task-related BOLD signal changes and interindividual differences in the joint modulation of decision processes, cortical excitability, and pupil modulation by means of a behavioral PLS. For this, we first calculated linear slope coefficients for voxel-wise beta estimates. Then, we included the behavioral variables reported on the left of Figure 6c. For visualization, spatial clusters were defined based on a minimum distance of 10 mm, and by exceeding a size of 25 voxels. We identified regions associated with peak activity based on cytoarchitectonic probabilistic maps implemented in the SPM Anatomy Toolbox (Version 2.2c) ¹⁸⁰. If no assignment was found, the most proximal assignment to the peak coordinates was reported.

Temporal dynamics of thalamic engagement

To visualize the uncertainty modulation of thalamic activity, we extracted signals within a binary mask of thalamic divisions extracted from the Morel atlas ¹⁸¹. Preprocessed BOLD timeseries were segmented into trials, spanning the period from the stimulus onset to the onset of the feedback phase. Given a time-to-peak of a canonical hemodynamic response function (HRF) between 5-6 seconds, we designated the 3 second interval from 5-8 seconds following the stimulus onset trigger as the stimulus presentation interval, and the 2 second interval from 3-5 s as the fixation interval, respectively. Single-trial time series were then temporally normalized to the temporal average during the approximate fixation interval.

Thalamic loci of behavioral PLS

To assess the thalamic loci of most reliable behavioral relations, we assessed bootstrap ratios within two thalamic masks. First, for nucleic subdivisions, we used the Morel parcellation scheme as consolidated and kindly provided by Hwang et al. ¹⁸² for 3 mm data at 3T field strength. The abbreviations are as follows: AN: anterior nucleus; VM: ventromedial; VL: ventrolateral; MGN: medial geniculate nucleus; LGN: lateral geniculate nucleus; MD: mediodorsal; PuA: anterior pulvinar; LP: lateral-posterior; IL: intra-laminar; VA: ventral-anterior; PuM: medial pulvinar; Pul: pulvinar proper; PuL: lateral pulvinar. Second, to assess cortical white-matter projections we considered the overlap with seven structurally derived cortical projection zones suggested by Horn & Blankenburg ¹⁸³, which were derived from a large adult sample (N = 169). We binarized continuous probability maps at a relative 75% threshold of the respective maximum probability, and re-sliced masks to 3mm (ICBM 2009c MNI152).

Statistical analyses

Outlier handling

For each signature, we defined outliers at the subject-level as individuals within their respective age group whose values (e.g., estimates of linear modulation) exceeded three scaled median absolute deviations (MAD) as implemented in MATLAB. Such individual data points were winsorized prior to statistical analysis. For repeated measures analyses, such individuals were removed prior to statistical assessment.

Linear uncertainty effect estimates

To estimate the linear uncertainty modulation of dependent variables, we calculated 1^st level beta estimates (y = intercept+β*target load+e) and assessed the slope difference from zero at the within-group level (see Table S1) using two-sided paired t-tests. Similarly, we compared linear uncertainty effect estimates between groups using two-sides unpaired t-tests. We assessed the relation of individual linear load effects between measures of interest via Pearson correlations.

Within-subject centering

To visually emphasize effects within participants, we use within-subject centering across repeated measures conditions by subtracting individual cross-condition means and adding global group means. For these visualizations, only the mean of the dependent values directly reflects the original units of measurement, as individual data points by construction do not reflect between-subject variation averaged across conditions. This procedure equals the creation of within-subject standard errors ¹⁸⁴. Within-subject centering is exclusively used for display and explicitly noted in the respective legends.

Multivariate partial least squares analyses

For data with a high-dimensional structure, we performed multivariate partial least squares analyses ^{185, 186}. To assess main effect of probe uncertainty, we performed Task PLS analyses. Task PLS begins by calculating a between-subject covariance matrix (COV) between conditions and each neural value (e.g., time-space-frequency power), which is then decomposed using singular value decomposition (SVD). This yields a left singular vector of experimental condition weights (U), a right singular vector of brain weights (V), and a diagonal matrix of singular values (S). Task PLS produces orthogonal latent variables (LVs) that reflect optimal relations between experimental conditions and the neural data. We ran a version of task PLS in which group means were removed from condition means to highlight how conditions were modulated by group membership, i.e., condition and condition-by-group effects. To examine multivariate relations between neural data and other variables of interest, we performed behavioral PLS analyses. This analysis initially calculates a between-subject correlation matrix (CORR) between (1) each brain index of interest (e.g., 1^st level BOLD beta values) and (2) a second ‘behavioral’ variable of interest (note that although called behavioral, this variable can reflect any variable of interest, e.g., behavior, pupil diameter, spectral power). CORR is then decomposed using singular value decomposition (SVD): SVDCORR = USV’, which produces a matrix of left singular vectors of cognition weights (U), a matrix of right singular vectors of brain weights (V), and a diagonal matrix of singular values (S). For each LV (ordered strongest to weakest in S), a data pattern results which depicts the strongest available relation between brain data and other variables of interest. Significance of detected relations of both PLS model types was assessed using 1000 permutation tests of the singular value corresponding to the LV. A subsequent bootstrapping procedure indicated the robustness of within-LV neural saliences across 1000 resamples of the data ¹⁸⁷. By dividing each brain weight (from V) by its bootstrapped standard error, we obtained “bootstrap ratios” (BSRs) as normalized robustness estimates. We generally thresholded BSRs at values of ±3.00 (∼99.9% confidence interval). We also obtained a summary measure of each participant’s robust expression of a particular LV’s pattern (a within-person “brain score”) by multiplying the vector of brain weights (V) from each LV by each participant’s vector of neural values (P), producing a single within-subject value: Brain score = VP’.

Data and code availability

Experiment code is available from https://git.mpib-berlin.mpg.de/LNDG/multi-attribute-task. Analysis code, primary EEG, fMRI, and behavioral data will be made available upon publication (for younger adults see https://osf.io/ug4b8/). Structural MRI data are exempt from public sharing according to obtained informed consent. All data are available from the corresponding authors upon reasonable request.

Author contributions

JQK: Conceptualization, Methodology, Investigation, Software, Formal analysis, Visualization, Writing – original draft, Writing – review and editing, Validation, Data Curation; UM: Conceptualization, Writing – review and editing, UL: Conceptualization, Resources, Writing – review and editing, Supervision, Funding acquisition; DDG: Conceptualization, Methodology, Software, Resources, Writing—review and editing, Supervision, Project administration, Funding acquisition.

Competing interests

The authors declare no competing interests.

Supplementary Information for

Figure S1-1. Average accuracy across target load conditions.

Younger (red) and older adults (grey) performed the task above chance for all attributes that were probed. Statistics are based on one sample t-tests against chance level (.5 in this 2AFC task). *** p < .001.

Figure S1-2. Age-related uncertainty adjustments to decision processes.

(a) DIC-based model comparison indicates that a model, including uncertainty modulation of drift rates, non-decision times, and boundary separation provides the best group fit to the behavioral data. (b) Posterior predictive checks for the full model (shown for the EEG session). Negative RTs indicate incorrect responses. Model-based (“posterior predictive”) values were sampled 50 times within each subject and condition (as implemented in the HDDM package), and probability density (100 RT bins) was estimated first within-subject across all samples, and then averaged across participants. In empirical data, probability densities were estimated across all participants due to the sparse within-subject RT counts. (c) Uncertainty modulation of HDD parameter estimates, averaged across sessions. Statistics refer to paired t-tests of linear slopes against zero. Data are within-subject centered. (d) Age comparison of single-target parameter estimates (left) and linear uncertainty effects (∼age x target load interaction). Statistics refer to unpaired t-tests.

Text 1-2. Uncertainty and age effects on non-decision time and boundary separation. The main analyses targeted drift rate as the main parameter of interest. Given that the best-fitting model (Figure S1-2ab) included uncertainty variation also for non-decision times as well as boundary separation, we explored the potential variation of the latter two parameters with age and uncertainty (Figure S1-2c). In contrast with younger adults, older adults had significantly longer non-decision times, and larger boundary separation, suggesting that more evidence was collected prior to committing to a choice. There is some evidence from 2AFC tasks that older adults adopt decision boundaries that are wider than the boundaries of younger adults (Starns & Ratcliff, 2010, 2012) [but see (McGovern et al., 2018)], which may signify increased response caution . In both age groups, we observed uncertainty-related increases in non-decision times, albeit more constrained in older adults, as well as similar increases in boundary separation as a function of rising uncertainty (see Figure S1-2d). Notably, the uncertainty effect on boundary separation was not consistently reproduced by either the integration threshold of the domain-general CPP (Figure S1-4b), or the effector-specific contralateral beta power threshold (Figure S1-5d), highlighting uncertainty regarding the true effect on behavioral response caution, or neural proxy signatures thereof. These discrepancies deserve further attention in future work and may suggest that a model with alternative parameter constellations could provide a more coherent description. Convergence of the current model with our previous results in younger adults (Kosciessa et al., 2021) ultimately argues for robust drift rate inferences that were independent from the specific model choice.

Figure S1-3. Drift rate differences do not arise from accuracy ceiling or floor effects.

(a) Difficulty with selectively distinguishing individual features is a major component of the task, which may contribute to age differences in single-target drift rates. While on average, younger adults’ single-target responses were more accurate than those of older adults, this differed between features (see Fig. S1-1). Excluding the most accurate (‘best’) feature for younger adults and the least accurate (‘worst’) feature for older adults matched groups regarding their accuracy in the single-target condition. In this scenario, older adults showed more pronounced accuracy decreases under uncertainty, relative to younger adults. Data are means +/- SEs and include data from EEG and fMRI sessions. n.s.: p = 0.13; ***1: p = 2.6e-05; ***2: p = 6.1e-04; ***3: p = 8.2e-04. (b) Drift rate estimates for an HDDM model that only included age-matched features (“match” in a). This model indicated retained age differences in single-target drift rate and drift changes under uncertainty (right). (c) Older (vs. younger) adults showed stronger relative drift rate reductions from the single-target baseline.

Text S1-3. Drift rate effects for accuracy-matched features. Our analysis indicated that older adults on average showed reduced behavioral uncertainty costs. However, these uncertainty costs are thought to arise from attending to a varied feature set, whose discrimination also varies between age groups when only a single feature is relevant. To examine whether potential ceiling or floor effects in feature-specific accuracy (e.g., due to varying perceptual uncertainty) acts as a between-group confound, we sorted features according to their single-target accuracy in each participant, and averaged accuracy according to such ” preference” within each age group. This revealed that three out of the four features elicited comparable single-target accuracy between age groups, whereas only the best feature of younger adults, and the worst feature of older adults could not be matched (Figure S1-3a). To test the robustness of unmatched drift rate estimates (Fig. 1b), we created HDDM models that excluded the most preferred feature of younger adults, and the individually least preferred feature in older adults (i.e., only including “matched” features). Results from this control analysis are shown in Figure S1-3b. We observed retained age differences in single-target drift rates, as well as uncertainty-related drift rate changes that mirrored our main results. These results indicate that baseline feature differences are likely not the principal origin of age and uncertainty drift rate differences.

Text S1-4. More pronounced relative performance decreases in older adults. Compared with younger adults, older adults’ drift rates were lower across levels of target load (Fig. 1b). To probe whether drift rates across all set sizes show similar proportional age changes, we calculated relative drift rate changes. Arguing against uncertainty-independent age differences in drift rate, we observed larger relative drift rate decreases under uncertainty in older as compared with younger adults (see Fig. S1-3b right for feature-matched HDDM; similar results were obtained in the main model). This indicates that despite being smaller in absolute terms, older as compared to younger adults suffered stronger relative drift rate losses once uncertainty was introduced. This mirrored larger accuracy decreases in matched features once uncertainty was introduced (Fig. S1-3a). Taken together, this indicates that uncertain contexts present an outsized challenge to older adults’ performance, over and above challenges in single-target specificity. For our main analyses that target inter-individual relations, we focus on absolute uncertainty-related drift rate changes due to their relation to neural uncertainty adjustment in prior work (Kosciessa et al., 2021), and the computational interpretability of absolute drift rates at each target load.

Figure S1-4. Centroparietal Positive Potential (CPP) as a signature of domain-general evidence integration.

(a) Modulation of CPP as a neural signature of evidence accumulation (mean ± within-subject SEM). The integration slope of the response-locked CPP decreases with increasing probe uncertainty. Traces are mean ± within-subject SEM. Insets show CPP slope estimates from −250 to −50 ms relative to response execution. (b) Age comparison of CPP integration slopes (yellow background) and CPP integration thresholds (grey backgrounds). (c) CPP estimates of evidence integration converge with behavioral drift rate estimates at the interindividual level, both w.r.t the single-target condition (r(93) = 0.45, 95%CI = [0.27,0.59], p = 6.1e-06; age-partialed: r(92) = 0.27, 95%CI = [0.08,0.45] p = 0.01) and linear effects of target number (r(93) = 0.51, 95%CI = [0.34,0.64], p = 1.4e-07; age-partialed: r(92) = 0.34, 95%CI = [0.14,0.5] p = 9.3e-04).

Figure S1-5. Contralateral beta power as a signature of motor-specific response preparation.

(a) Pre-response desynchronization of contralateral mu-beta power shallow with increasing number of targets. Traces show means +/- within-subject SEM. (b) Linear slope estimates, estimated via linear regression from -250 ms to -50 ms, relative to response. Data are within-subject centered for visualization. Statistics refer to paired t-tests of linear slopes against zero. (c, d) Age comparison of linear modulation of beta slopes (c) and integration thresholds (d) by target load. Statistics refer to unpaired t-tests.

Text 1-5. Motor-specific response preparation. In addition to the domain-general CPP, we also investigated motor-specific contralateral beta power (Figure S1-5a). Extending results from behavioral modeling, and CPP integration slopes, we observed a shallowing of pre-response beta power build-up, suggesting decreases in response preparation (Figure S1-5b). However, such shallowing was not statistically different between age groups (Figure S1-5b), thus deviating from the age x load interaction that we observed for the remaining integration signatures. Furthermore, linear changes in beta slope as a function of target load were neither associated with linear drift changes (r(93) = -0.03, 95%CI = [-0.23,0.17], p = 0.77) nor CPP slopes (r(93) = -0.11, 95%CI = [-0.3,0.09], p = 0.29) across age groups. The parameters were also not directly related in the single-target condition (drift rates: r(93) = 0.18, 95%CI = [-0.02,0.37], p = 0.07; CPP slopes: r(93) = -0.06, 95%CI = [-0.26,0.14], p = 0.55). Motor-specific response preparation thus appears to partially dissociate from effector-unspecific evidence integration at the individual level.

Figure S3-1. CPP slope during the Stroop task.

(a) Response-aligned CPP traces split by condition and age group. Time series were smoothed with 60 ms windows for visualization, but not for slope fitting. Linear slopes were estimated during the interval of -600 to -100 ms prior to indicated SOTs, marked by the red line. (a) CPP integration slopes were reduced in magnitude in the mismatch condition in both age groups. (b) Interference effects on CPP slopes were more pronounced in younger ad compared with older adults. The magnitude of individual interference effects was similarly reflected in RTs and CPP slopes with longer RTs being coupled to stronger CPP slope reductions [r(91) = -0.43, 95%CI = [-0.58,-0.25], p = 1.5e-05; partial-correlation accounting for age: r(89) = -0.32, 95%CI = [-0.49,-0.12] p = 3.3e-04].

Figure S4-1. Load modulation for cognitive control (a) and excitability signatures (b).

Statistics refer to paired t-tests of linear slopes against zero. In line with the different excitability indices capturing a shared latent characteristic, the magnitude of uncertainty modulation was inter-individually related among the three parameters (alpha-1/f: r = 0.44, p = 9.8e-06; 1/f-SampEn: r = 0.6, p = 1.1e-10; SampEn-alpha: r = 0.24, p = .02). (c) Sample entropy input spectrum highlighting the exclusion of alpha-range signal content.

Figure S5-1. Pre-stimulus alpha power.

Uncertainty similarly increases pre-stimulus alpha power in younger and older adults but does not relate to individual drift rate adjustments. Light grey squares indicate younger adults, dark grey diamonds correspond to older adults.

Text S5-1 Pre-stimulus alpha power. Evidence on age-related changes in pre-stimulus alpha power are mixed. Early studies suggest that pre-stimulus alpha synchronization (or lateralization) in the context of attentional cueing is observed exclusively for younger, but not older adults (Hong et al., 2015; Vaden et al., 2012). In contrast, (Leenders et al., 2018) indicated similar pre-stimulus lateralization between age groups, whereas they noted age differences in alpha modulation during working memory retention. While our task design does not allow us to assess the lateralization of alpha power, our results indicate that pre-stimulus alpha power increases similarly alongside uncertainty in both age groups, but with no apparent relation to subsequent (delayed) task performance (Figure S3-1).

Figure S5-2. Steady-state visual evoked potential (SSVEP).

(a) Both age groups exhibited a robust SSVEPs. Time-resolved, spectrally normalized, SSVEP power, averaged across occipital channels (O1, Oz, O2), indicates clear SSVEP increases specifically during stimulus presentation. Data are presented as mean values +/- within-subject SEM. Topography insets show stimulus-evoked SSVEP contrast minus baseline. (b) However, estimates from occipital EEG channels (O1, Oz, O2) did not indicate age differences in single-target SSVEP magnitude, a main effect of load in either group, or differences in the strength of linear modulation (∼ age*load interaction).

Text S5-2. SSVEP magnitude. SSVEP magnitude has been suggested as a signature of encoded sensory information that is enhanced by attention (Morgan et al., 1996; Muller et al., 2006; Quigley et al., 2010; Quigley & Muller, 2014) and indicates fluctuations in early visual cortex excitability (Zhigalov et al., 2019). However, despite a clear SSVEP signature of comparable magnitude in both younger and older adults (Fig. S5-2 a, b), we did not observe significant effects of target uncertainty on SSVEP magnitude in either age group (Fig. S5-2). Given that the SSVEP frequency was shared across different features, we could not investigate feature selection via SSVEPs as is commonly the case in attention studies. Studies with feature-specific SSVEPs, suggest that younger adults’ SSVEP magnitude differentiates between attended and unattended features, whereas no robust differentiation is observed in older adults, pointing to deficits in attentional filtering (Quigley et al., 2010; Quigley & Muller, 2014).

Figure S5-3. Task PLS of sample entropy and aperiodic slopes across all channels.

Brainscores for younger adults are shown on the left side, with data for older adults shown to the right. Inset estimates refer to fixed linear effects models. Topographies of bootstrap ratios are unthresholded.

Text S5-3. Exploratory whole-brain task PLS of aperiodic dynamics. In the main analysis, we restricted the PLS to posterior channels with the aim to predominantly characterize signals stemming from parietal and visual cortex. To explore whether this analysis missed uncertainty-related changes in aperiodic dynamics in other regions, we performed an additional task PLS analysis that included all channels. This task PLS averaged sample entropy across the final 2.5s of stimulus presentation. To normalize relative contributions of the two signatures to the PLS, we z-transformed values of each signature across target load levels prior to including them in the model. This joint PLS resulted in two significant latent variables (Figure S5-3). The first latent variable (permuted p = 0.001) indicated uncertainty-related increases in sample entropy and shallowing of aperiodic slopes in younger, but not older adults. Regional contributions were predominantly observed in posterior sensors. This latent variable thus captures the observations in the main analysis. The second latent variable (permuted p = 0.021) was instead marked by quadratic changes (younger adults: p = 1.5e-08; older adults: p = 0.03; linear mixed effects model with fixed and random quadratic effects) as a function of target load. Estimates initially decreased, followed by an increase with load towards higher target load, predominantly at mediofrontal channels.

Figure S6-1. Main effects of target load on BOLD magnitude.

A task partial least squares (PLS) analysis indicated three significant latent variables (loadings shown in panels a-c) that were sensitive to changes in target number. (d) Statistics refer to paired t-tests of linear slopes against zero.

Text S6-1. Main effects of uncertainty on BOLD magnitude across the adult lifespan. We performed a whole-brain task PLS to assess potential main effects of uncertainty on BOLD magnitude. In brief, we observed a similar first latent variable (permuted p < 0.001) to that reported in younger adults (Kosciessa et al., 2021), highlighting uncertainty-related increases dominantly in cortical areas encompassing the frontoparietal and the midcingulo-insular network, as well as in the thalamus (see detailed results of this analysis in Figure S6-1 and Table S2-4.The task PLS indicated two further robust LVs. LV2 (permuted p < 0.001) captured a non-linear pattern in younger adults and linear changes under uncertainty in older adults. Regional contributors partly overlapped with the initial LV (Table S3). Finally, LV3 (permuted p < 0.001) captured nonlinear changes (initial increases in engagement followed by disengagement) in both age groups in a set of regions encompassing positive loadings in frontoparietal components of the executive control network, and negative loadings in temporal-occipital cortex (Table S4).

Figure S6-2. Additional brainscore relations before (left) and after controlling for categorical age (right).

Beyond the a priori signatures included in the behavioral PLS model, post-hoc exploration indicated that individuals with more pronounced BOLD uncertainty modulation also had larger single-target drift rates, and lower single-target boundary separation (“boundary thresholds”), as well as larger increases in the latter as a function of uncertainty, also after controlling for categorical age (see right). In addition to the magnitude of uncertainty-related drift rate modulation, CPP slope modulation was similarly related to Brainscores. Plots indicate Pearson correlation coefficients +- 95%CI after accounting for age covariation.

Figure S6-3. BOLD modulation effects are robust to accuracy differences within (a) and between age groups (b).

(a) Younger and older individuals with larger BOLD uncertainty modulation (LV1, see Fig S6-1) achieve higher drift rates across uncertainty levels at comparable accuracy levels. Data show upper (full lines) and lower (broken lines) groups of a trichotomized split of accuracy and drift rate data based on the magnitude of uncertainty change (234 vs. 1) in the 1^st LV of the task PLS (closely mirroring the change LV1 of the behavioral PLS). Insets illustrate comparable within-group splits (split performed within-group). Data are means +- SEs. (b) Age x uncertainty interaction in mediodorsal thalamus for accuracy-matched features. For this analysis, trials with probes of the best (YA) or worst (OA) features were excluded to achieve group-matched single-target accuracy (see Text S1-3). A linear mixed effects model indicated a retained group x target load interaction for data averaged in the time window of interest (yellow shading; beta=-0.757, SE=0.369, t = -2.0507, dof = 364, p = 0.041, 95%CI = [-1.48, -0.03]). Data are means +- SEs.

Table S1: Statistics for within age-group effects.

Effects were assessed via paired t-tests against zero. YA: Younger adults. OA: Older adults.

Table S2: PLS model peak activations, bootstrap ratios, and cluster sizes for task PLS LV1.

Table S3: PLS model peak activations, bootstrap ratios, and cluster sizes for task PLS LV2.

Table S4: PLS model peak activations, bootstrap ratios, and cluster sizes for task PLS LV3.

Table S5: PLS model peak activations, bootstrap ratios, and cluster sizes for behavioral PLS LV1.

Acknowledgements

This study was conducted within the ‘Lifespan Neural Dynamics Group’ at the Max Planck UCL Centre for Computational Psychiatry and Ageing Research in the Max Planck Institute for Human Development (MPIB) in Berlin, Germany. DDG was supported by an Emmy Noether Programme grant from the German Research Foundation. UL acknowledges financial support from the Intramural Innovation Fund of the Max Planck Society. JQK, DDG and UL were partially supported by the Max Planck UCL Centre for Computational Psychiatry and Ageing Research. The participating institutions are the Max Planck Institute for Human Development, Berlin, Germany, and University College London, London, UK. For more information, see https://www.mps-ucl-centre.mpg.de/en/comp2psych. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. We thank our research assistants and participants for their contributions to the present work.