Dissociable Neural Mechanisms Underlie the Effects of Attention on Visual Appearance and Response Bias

Attention induces changes in contrast appearance and response bias as measured with behavior

To examine the effects of attention on changes in perceived contrast and response bias, we performed two complementary analyses that compared the probabilities of reporting either the cued or the uncued stimulus as having a higher contrast. First, we computed the probability of reporting a cued stimulus at each contrast level as having a higher perceived contrast than the uncued stimulus at all other contrast levels. For example, we computed the probability of reporting a 5% contrast cued stimulus as having a higher contrast than a 0%, 5%, 10%, . . ., 100% uncued stimulus, and we repeated this exhaustive analysis for each possible contrast level of the cued stimulus. We then performed an analogous analysis quantifying the probability of reporting an uncued stimulus at each contrast level as having a higher perceived contrast when it was paired with a cued stimulus of all possible contrast values. For purposes of data exposition, we always refer to the stimulus being held constant in a plot as the standard stimulus and the stimulus being varied in a plot as the test stimulus. Thus, in the example above, the 5% contrast cued stimulus would be the standard stimulus that is compared with uncued test stimuli that ranged in contrast systematically from 0% to 100%. Importantly, both the cued and uncued stimuli served as standard and test stimuli depending on the nature of the analysis being performed. This allowed us to plot summary data for cued and uncued stimuli on the same axes as shown in Figure 3, with data from the cued stimulus plotted in green and data associated with the uncued stimulus in purple.

To better quantify these behavioral data, we fit each psychophysical function using an NR equation (Eq. 1) to estimate the baseline-offset (B), contrast gain (G_c), and response gain (G_r), which control the baseline, the horizontal position (e.g., leftward shift), and the slope of the behavioral CRFs, respectively (see Fig. 2a-c; Materials and Methods) (Heeger, 1992; Geisler and Albrecht, 1997; Kim et al., 2007; Itthipuripat et al., 2019a). Here, we used changes in the baseline-offset (B) to track the magnitude of cue-induced response biases because reporting the contrast of the cued stimulus as higher than the contrast of the uncued stimulus, when the cued stimulus was rendered at 0% contrast, likely reflects bias (because no stimulus was actually presented) (Prinzmetal et al., 2008; Itthipuripat et al., 2019a). On the other hand, we used changes in contrast gain (G_c) to index changes in contrast appearance because this parameter controls left/right translations along the x axis, which would be expected if the actual physical contrast of the test stimuli increased or decreased, respectively. We focused on G_c instead of the point of subjective equality (PSE) because others have suggested that PSE overestimates changes in contrast appearance when there are significant amounts of cue-induced response bias (Schneider, 2006, 2011; Prinzmetal et al., 2008; Schneider and Komlos, 2008; Itthipuripat et al., 2019a; Schneider and Malik, 2021b).

Consistent with a recent report from our group, we found that the effects of attention on contrast appearance and response bias depend on the overall level of stimulus contrast and stimulus uncertainty (i.e., whether the test and standard stimuli were rendered at the similar contrasts) (Itthipuripat et al., 2019a). When there was a 0% contrast stimulus presented at both the cued and the uncued locations, subjects were more likely to report that the cued location had a higher contrast than the uncued location (although no stimulus was presented on either side in this condition, see Fig. 3a, leftmost panel). This response bias resulted in an increase in the baseline-offset parameter (B) of the psychometric function. Importantly, this response-bias-induced baseline-offset became smaller as the contrast of the uncued standard stimulus increased. To statistically evaluate these effects, we performed a two-way repeated-measures ANOVA with attention (cued vs uncued) and the contrast of the standard as factors. There were a significant main effect of attention on B: F_(1,19) = 60.82, p < 0.001, a significant main effect of the contrast of the standard stimulus on B: F_(6,114) = 159.93, p < 0.001, and a significant interaction between attention and the contrast of the standard stimulus on B: F_(6,114) = 114.94, p < 0.001. An analogous ANOVA was also performed on only the correct trials (i.e., we only counted the responses on trials where subjects correctly discriminated the orientation offset of the chosen visual stimulus). This analysis revealed the same pattern of results: there were a significant main effect of attention on B: F_(1,19) = 61.73, p < 0.001, a significant main effect of the contrast level of the standard stimulus on B: F_(6,114) = 128.96, p values < 0.001, and a significant interaction between attention and the contrast of the standard stimulus on B: F_(6,114) = 100.88, p < 0.001. Post hoc paired t tests revealed that attention effects on B were significant for standard contrast levels of 0%, 5%, 10% and 20% (t₍₁₉₎ values = 3.26-24.00 and 2.88-24.57 for all trials and correct-only trials, respectively, all p values ≤ 0.0042, Holm–Bonferroni-corrected). However, differences in B were not significant for standard contrast levels of 40%, 80% and 100% (t₍₁₉₎ values = 1.36-1.80 and 1.38-1.88 for all trials and correct-only trials, respectively, all p values ≥ 0.0751). Overall, these results suggest that attention induced significant response bias, especially for low-contrast standard stimuli.

To measure changes in contrast appearance, we next examined attentional modulations of the contrast gain parameter (G_c) that controls the horizontal shift of the psychometric functions. We found that attention reduced the G_c parameter, which led to a leftward shift of the psychometric functions. However, these leftward shifts were most pronounced at low-to-middle standard contrast levels and then became smaller as the standard contrast approached 100%. A two-way repeated-measures ANOVA on G_c with attention and standard contrast as factors revealed a significant main effect of attention: F_(1,19) values = 48.08 and 35.42, p values < 0.001, a significant main effect of standard contrast: F_(6,114) values = 111.89 and 103.21, p values < 0.001, and a significant interaction between the two factors: F_(6,114) values = 9.64 and 9.34 for all trials and correct-only trials respectively, with p values < 0.001. Post hoc paired t tests showed that, in the all-trial analysis, attention effects on G_c were significant for standard contrast levels of 5%, 20%, 40%, and 80% (t₍₁₉₎ values = −2.77 to −9.15 p values ≤ 0.0122, Holm–Bonferroni-corrected), but were not significant for standard contrast levels of 0%, 10%, and 100% contrasts (t₍₁₉₎ values  = 1.36-1.80, p values ≥ 0.0153, not passing the corrected threshold of 0.0125). For the correct-only trials, attentional modulations of G_c were significant for standard contrast levels of 10%-80% contrast (t₍₁₉₎ values = −9.15 to −2.77, p values ≤ 0.0122, Holm–Bonferroni-corrected) but were not significant for standard contrast levels of 0% and 100% contrast (t₍₁₉₎ values = −0.96 and −2.14, p values ≥ 0.0457, not passing the corrected threshold of 0.025). These results suggest that attention alters contrast appearance but does so primarily at low-to-mid-level contrasts.

The lack of a significant contrast gain modulation (G_c) at the highest standard contrast was in part because the psychometric functions in this condition did not reach the maximum possible value of 1 (i.e., p(stimulus of interest > paired stimulus) was <1). Thus, attentional modulations of the psychophysical functions at the highest standard contrast (100%) manifested as an increase in the response gain parameter (G_r), corresponding to a steeper slope (t₍₁₉₎ values = 10.93 and 10.23 for all trials and correct-only trials, respectively, p values < 0.001, two-tailed, Holm–Bonferroni-corrected). A recent study demonstrated that attention-induced gain changes measured at high standard contrasts might not exclusively reflect changes in appearance per se. Instead, these gain changes might be influenced by response bias related to high stimulus uncertainty as subjects were unsure which of the two already high contrast stimuli had a slightly higher contrast (thus; they were biased to follow the attentional cue) (Itthipuripat et al., 2019a).

Possible neural mechanisms driving attention-related changes in stimulus appearance

Several neural mechanisms have been proposed to explain attention-related modulations in information processing and attention-related changes in perceptual performance and stimulus appearance (Fig. 2d-f) (Reynolds et al., 2000; di Russo et al., 2001; Martínez-Trujillo and Treue, 2002; Carrasco et al., 2004; Williford and Maunsell, 2006; Buracas and Boynton, 2007; Kim et al., 2007; Murray, 2008; Lee and Maunsell, 2009; Reynolds and Heeger, 2009; Herrmann et al., 2010; Lauritzen et al., 2010; Pestilli et al., 2011; Wang and Wade, 2011; Cutrone et al., 2014; Hara and Gardner, 2014; Itthipuripat et al., 2014b, 2015, 2017, 2018, 2019a,b; Sprague et al., 2018). The contrast gain account posits that attention shifts the neural CRFs horizontally to the left, consistent with attention increasing the contrast sensitivity of neural responses that respond to cued stimuli (Fig. 2d) (Reynolds et al., 2000; Martínez-Trujillo and Treue, 2002; Carrasco et al., 2004; Itthipuripat et al., 2014b). In addition, multiplicative response gain models posit that attention can amplify neural activity to increase the slope of neural CRFs, thereby increasing sensitivity to detect small differences in contrast (Fig. 2e) (di Russo et al., 2001; Kim et al., 2007; Lee and Maunsell, 2009; Itthipuripat et al., 2014b, 2017, 2018, 2019b; Sawetsuttipan et al., 2023). Last, an additive baseline shift would suggest that attention can lead to increases in the baseline activity of neural CRFs in a manner that is invariant to stimulus contrast (Fig. 2f) (Williford and Maunsell, 2006; Buracas and Boynton, 2007; Murray, 2008; Pestilli et al., 2011; Cutrone et al., 2014; Hara and Gardner, 2014; Sprague et al., 2018; Itthipuripat et al., 2019b).

Although few studies have directly linked these attention-related changes in neural activity with subjective visual appearance, prior psychophysical studies have used ideal-observer models to determine which modulations of hypothetical neural CRFs might best explain attention-induced changes in the perceived contrast of visual stimuli. One of the earliest studies adopting this approach demonstrated that response gain (i.e., Fig. 2e) and contrast gain (i.e., Fig. 2d) failed to explain attention-induced changes in perceived contrast compared with a model that only incorporated a baseline-offset at low contrasts (and compression at higher contrasts, referred to as activity gain) (Cutrone et al., 2014). However, this study did not take into account the potential for response bias driven by exogenous cues, especially when the visual stimulus was absent at the cued location (i.e., effects of attention on 0% contrast stimuli). Recently, we systematically tracked attentional effects on both perceived contrast and response bias and found that cue-induced response biases for 0% contrast stimuli were largely explained by shifts in the baseline-offset of hypothetical neural CRFs (Itthipuripat et al., 2019a). In contrast, attention-induced changes in appearance for low-to-mid contrast stimuli were better explained by changes in response gain (Itthipuripat et al., 2019a). However, there is still a lack of neural evidence that can inform linking models to relate neural data with perceptual experience. Here we tested whether (1) dissociable patterns of attentional modulations in different neural markers of visual processing, such as the P1 component and α band activity, could be observed; and (2) if dissociable neural components explain attention-induced changes in contrast appearance and response bias measured with behavior. Our modeling approach was based on SDT and adopted from previous psychophysical studies (Cutrone et al., 2014; Itthipuripat et al., 2019a). The primary difference was that we used EEG data in the present study — the P1 and α band activity — instead of using hypothetical neural CRFs. Thus, we could directly link changes in behavioral responses with concurrent modulations in simultaneously collected EEG data.

Motivation for linking the amplitude of P1 and α band responses with behavior

Here, we targeted two EEG indices that are thought to track different aspects of visual information processing: (1) the P1 component, which is an early visually evoked potential that peaks ∼60-90 ms ms after stimulus onset; and (2) the amplitude of posterior occipital α band oscillations (i.e., EEG oscillations in the ∼9-12 Hz band). We used these two EEG markers because they have been previously linked to bias in subjective contrast perception (Störmer et al., 2009, 2019; Balestrieri and Busch, 2022). That said, we hypothesized that the attentional modulations of the visual P1 component and α band activity would differentially relate to the effects of attention on visual contrast appearance and response bias for several reasons. First, attention enhances the amplitude of the P1 component (van Voorhis and Hillyard, 1977; Mangun and Hillyard, 1990; Woldorff et al., 1997; Hillyard and Anllo-Vento, 1998), and attentional gain of P1 amplitude has been linked to improved detection and discrimination for low-level visual features, such as contrast and object shape (Mangun and Hillyard, 1990; Itthipuripat et al., 2014a, 2017). Recent studies have also found that selective attention induces a multiplicative response gain of neural CRFs based on P1 amplitude (see Fig. 2e), and quantitative models suggest that these gain modulations predict attention-related changes in perceptual contrast discrimination thresholds (Itthipuripat et al., 2014a, 2017). Importantly, especially for the present experiment, attentional gain of P1 amplitude has been previously related to an increase in the perceived contrast of cued compared with uncued visual stimuli (Störmer et al., 2009, 2019). Although changes in subjective experience were not quantitatively linked to changes in P1 amplitude using a formal linking model, this finding (Störmer et al., 2009, 2019), coupled with suggestive earlier work (Itthipuripat et al., 2014b, 2017; Störmer et al., 2019), is consistent with the hypothesis that multiplicative response gain of the P1 is tightly coupled with attention-induced changes in perceived contrast (Itthipuripat et al., 2019a).

In addition to changes in perceived contrast, attention is also thought to influence other aspects of behavioral performance that may be more closely tracked by other neural markers, such as α band oscillations. For example, prior work suggests that attention cues, particularly the peripheral cues used in most comparative judgment tasks, can induce a bias such that subjects are more likely to select the cued stimulus as having a higher contrast, independent of the perceptual experience of the subject (Schneider, 2006, 2011; Schneider and Komlos, 2008; Beck and Schneider, 2017; Itthipuripat et al., 2019a; Schneider and Malik, 2021b). A large corpus of prior studies demonstrates that endogenous attention cues decrease α amplitude in posterior visual areas contralateral to the attended visual field (Foxe et al., 1998; Fries et al., 2001; Sauseng et al., 2005; Rihs et al., 2007; Yamagishi et al., 2008; Kelly et al., 2009; Händel et al., 2011; Bosman et al., 2012; Keefe and Störmer, 2021). Recent studies using exogenous cues presented at the peripheral locations also report similar cue-induced decreases in contralateral α band activity. Together, these findings suggest that lateralized α activity reflects visuocortical biasing across both exogenous and endogenous attention (Song et al., 2014; Keefe and Störmer, 2021).

In addition to these well-documented topographic modulations related to the locus of spatial attention, we hypothesized that α might also track attention-induced changes in response bias. First, past studies have found that α amplitude at the time of stimulus onset predicts shifts in response bias (i.e., response criterion) but not shifts in perceptual sensitivity in some visual detection and discrimination tasks (Limbach and Corballis, 2016; Benwell et al., 2017, 2018, 2020; Iemi et al., 2017; Foster and Awh, 2019). Second, attention has been shown to reduce the amplitude of contralateral posterior occipital α band oscillations, even in the absence of visual stimuli, suggesting that it may simply reflect top-down inputs from downstream areas in visual cortex onto early sensory areas and does not tract the interaction between attention and sensory inputs (Itthipuripat et al., 2019b; Foster et al., 2021). Consistent with this idea, recent studies have found that attention shifts the baseline-offset of CRFs based on the amplitude of α oscillations, reflecting a shift in general arousal or responsiveness that does not interact with the actual intensity of the stimulus (Fig. 2f) (Itthipuripat et al., 2019b; Foster et al., 2021). Based on these observations, we predicted that a shift in the baseline-offset of α-based CRFs would be systematically linked with attention-induced response bias (Itthipuripat et al., 2019a).

Attention amplifies response gain of the early visually evoked P1 component

We first sought to isolate stimulus-specific P1 activity from activity evoked by the presentation of the attention cue. Thus, we computed ERPs from trials where the cue was followed by 0% contrast cued and uncued stimuli (termed cue-only trials). We then subtracted this cue-only ERP from the ERPs evoked on trials that contained a cue plus a stimulus. This subtraction was performed separately for stimulus-present trials from each attention condition and contrast level (Fig. 4a; see Materials and Methods) (see similar methods in Greenwood and Goff, 1987; Iragui et al., 1993; Kiss et al., 1998; Chica et al., 2010; Itthipuripat et al., 2014a, 2017, 2019b; Störmer et al., 2019). This procedure yielded a clear P1 component that peaked ∼60-90 ms after stimulus over posterior occipital electrodes that were contralateral to the stimulus of interest. Next, we plotted the mean amplitude of the isolated P1 component as a function of stimulus contrast (i.e., test contrast) to obtain the P1-based CRFs for each attention condition and contrast level of the standard stimulus (Fig. 5a). Then, we fit these P1-based CRFs with an NR equation (Eq. 1) to examine changes in the baseline-offset (B), response gain (G_r) and contrast gain (G_c) of the neural CRFs. Since the estimated G_c parameters can potentially go beyond the realistic range of stimulus contrast (≫100%), the contrast gain parameter was allowed to vary only between 0% and 100% contrast (i.e., the realistic range of the physical contrast value). Moreover, we reparameterized the response gain and contrast gain of neural CRFs as the maximal response (R_max or the response at 100% contrast minus the baseline-offset) and the semisaturation contrast (C₅₀ or the contrast at which the response reached half-maximum), respectively (see Early sensory-evoked response).

As illustrated in Figure 5a, the best fit curves from the NR equation (Eq. 1) explain the observed baseline-subtracted P1-based CRFs reasonably well (goodness-of-fit R² = 0.7977). Importantly, we found that attention increased the maximum response (i.e., R_max) of the P1-based CRF (main effect of attention: F_(1,19) = 4.46, p= 0.0482) (Fig. 5b). In addition, there were larger attentional modulations at the low-to-mid-level standard contrasts (5%-20%) compared with when the standard was absent (0% contrast) and when the standard had a higher contrast (40%-100%). This gave rise to a significant interaction between attention and the contrast of the standard stimulus (F_(6,114) = 2.24, p = 0.0441). Importantly, attention had a selective effect on the R_max parameter of the P1 response, as there were no changes in any other parameters (Fig. 5b; main effect of attention: F_(1,19) = 2.20, p = 0.1546 for B, F_(1,19) = 1.51, p = 0.2973 for C₅₀; main effect of standard contrast level: F_(6,114) = 1.05, p = 0.3961 for B, F_(6,114) = 1.91, p = 0.0854 for C₅₀; interaction between the attention and standard contrast: F_(6,114) = 0.24, p = 0.9614 for B; F_(6,114) = 1.03, p = 0.4079 for C₅₀). The elevated response gain of the P1-based CRFs (i.e., R_max) at low-to-mid-level contrast levels was consistent with the fact that changes in appearance, as indexed by changes in contrast gain (G_c), were observed most prominently at these contrast levels (Fig. 3b). The best fit curves illustrated in Figure 5a were not obtained by fitting the P1-based CRF data after averaging across all subjects (i.e., Fig. 5a, diamonds and squares). Instead, the best fit curves were reconstructed based on the averaged fit parameters computed at the individual-subject level (Fig. 5b, diamonds and squares). We adopted this approach so that the best fit curves better represent the statistics we performed on the parameters obtained from fitting individual subjects’ data separately.

Since the ERP subtraction method is based on the assumption that the stimulus- and cue-evoked responses linearly sum (Greenwood and Goff, 1987; Iragui et al., 1993; Kiss et al., 1998; Chica et al., 2010; Itthipuripat et al., 2014a, 2017, 2019b; Störmer et al., 2019), we did additional analyses to determine whether a similar pattern of response gain modulations was observed without performing the ERP subtraction (Fig. 5c,d). Overall, we found that the NR equation (Eq. 1) fit the nonsubtracted CRF data slightly better than the subtracted CRF data with R² = 0.8305. In addition, the overall pattern of attentional modulations were consistent across the subtracted and nonsubtracted data: there were still a significant main effect of attention (F_(1,19) = 5.03, p = 0.0371) and a significant interaction between attention and standard contrast on R_max (F_(6,114) = 2.57, p = 0.0225). However, attention and the contrast of the standard had no significant impact on the C₅₀ parameter and the two factors did not interact (all F values ≤ 1.91, p values = 0.1030-0.5503).

For the baseline-offset parameter (B), we found that the attention cue reduced the baseline value of the nonsubtracted P1-based CRF (it became more negative), resulting in a significant main effect of attention on the baseline-offset (F_(1,19) = 5.88, p= 0.0255). This was expected because the stimulus-elicited P1 component overlaps with the early negative N1 component evoked by the cue that was presented just 100 ms before stimulus onset (see Fig. 4a) (see also similar results in Itthipuripat et al., 2019b). The baseline-offset value also increased as a function of standard contrast (F_(6,114) = 2.87, p = 0.0121). This is also expected because the nonsubtracted responses contained positive evoked potentials elicited by both test and standard stimuli. The increased baseline-offset values as a function of contrast observed at the contralateral posterior-occipital sites could be a result of the spread of activity elicited by the standard stimulus from the opposite hemisphere (see topographic maps showing more spread of activity for the nonsubtracted P1 data in Fig. 4b), and this is why we subtracted out the baseline activity to isolate brain activity related to the stimulus of interest (i.e., the test stimulus). We also found a significant interaction between attention and standard contrast on the baseline values of the nonsubtracted P1 CRFs (F_(6,114) = 2.82, p = 0.0136). This interaction was driven by a higher degree of attention-induced reduction in baseline values on trials with low compared with high standard contrasts and compared with 0% standard contrast.

Overall, there was a consistent pattern of response gain modulations in the baseline-subtracted and non–baseline-subtracted P1 data, suggesting that the baseline subtraction method did not change the pattern of modulations of the response gain in early sensory responses.

Attention induces an additive shift in the amplitude of posterior occipital α oscillations

Next, we examined the effects of attention on the contralateral posterior occipital α band activity (i.e., EEG oscillations at ∼9-12 Hz), another commonly used neural index of visuospatial attention (Foxe et al., 1998; Fries et al., 2001, 2008; Sauseng et al., 2005; Kelly et al., 2006, 2009; Klimesch et al., 2007; Rihs et al., 2007; Bosman et al., 2012; Foster et al., 2016, 2017; Samaha et al., 2016; Voytek et al., 2017; Hakim et al., 2019; Itthipuripat et al., 2019b; Wang et al., 2021). Consistent with previous observations, we found a significant cue-related reduction in the amplitude of α oscillations, compared with a precue baseline period, that grew more pronounced as the contrast of the test stimulus increased. These α amplitude modulations were most prominent over the posterior occipital electrodes that were contralateral to the stimulus of interest (Figs. 6 and 7).

Figure 8 shows the α data plotted as a function of both test and standard contrast levels to form CRFs. Overall, the best fit curves from the NR equation (Eq. 1) explain the observed α data reasonably well (goodness-of-fit R² = 0.8289). In contrast to the P1 data, we found that attention cues modulated the baseline-offset (B) of the neural CRFs based on the postcue reduction of the posterior occipital α activity (main effect of attention; F_(1,19) = 39.35, p < 0.001). These lateralized α changes likely reflect a mixture of low-level sensory-evoked activity elicited by the peripheral visual cues as well as α changes because of attentional biases. This cue-induced reduction in α band activity occurred to a comparable degree across all standard contrast levels (no main effect of standard contrast level: F_(6,114) = 0.71, p = 0.6384; no interaction between attention and standard contrast: F_(6,114) = 0.65, p = 0.6912). Since the degree of cue-induced modulation of α amplitude was relatively more robust at the low compared with the high test contrasts, the R_max parameters describing the α-based CRFs became less negative with attention (i.e., the negative slope of the α-based CRFs became shallower, main effect of attention: F_(1,19) = 18.96, p = 0.0003). On the other hand, R_max became more negative with increasing standard contrast (main effect of standard contrast: F_(6,114) = 3.83, p = 0.0016). However, there was no interaction between attentional cue and contrast on R_max (F_(6,114) = 1.00, p = 0.4314). While there were attention-induced changes in response gain modulations, these modulations occurred in opposition to the effects related to the contrast of the test stimulus, resulting in the CRFs pinching/saturating at the highest test contrast. Thus, unlike the P1-based CRFs where the slope of cued condition was steeper than that of the uncued condition, the slope of the cued α-based CRF was actually shallower than the uncued α-based CRF. Interestingly, past studies have reported a similar pattern of results in fMRI data measured in early visual cortex, suggesting that increases in baseline responses and reductions in response gain at high contrast levels may reflect the influence of top-down attention and response saturation induced by overly strong stimulus inputs into early visual areas, respectively (e.g., Pestilli et al., 2011; Itthipuripat et al., 2019b). Finally, there was no main effect of the attention cue or the contrast of the standard on the semisaturation contrast parameter (C₅₀), and no interaction between the two factors (all F values ≤ 0.98, all p values ≥ 0.4391). Overall, the shift in the baseline-offset of the α-based CRFs was consistent with the robust baseline-offset response bias observed in the behavioral data (Fig. 3).

Based on recent work, it is plausible that the attention effects on the α band activity reflect changes in aperiodic components of EEG signals rather than a change in the amplitude of a true oscillation (Donoghue et al., 2020). For example, a global rotation of the 1/f fall-off of the EEG power spectrum might lead to erroneous conclusions that energy in a specific frequency band increased or decreased. To address this issue, we ran additional analyses to independently examine attention effects on the periodic and aperiodic components of α oscillations following the algorithm described by Donoghue et al. (2020). First, we obtained the spectrogram from the contralateral posterior occipital electrodes during the precue baseline and the poststimulus period (Fig. 7c). Next, we fit the spectrogram in each experimental condition, each time period, and each subject using the FOOOF toolbox developed by Donoghue et al. (2020). This step yielded estimates of the periodic components of α oscillations, including log power, central frequency, bandwidth, as well as the aperiodic offset and exponent. We then computed the differences between these periodic and aperiodic parameters across the stimulus versus precue periods to estimate changes of these parameters with respect to the precue baseline (Fig. 9a-e).

The modulatory pattern of the log power of the α band activity was very similar to the main analysis where we found attention reduced the α band amplitudes more prominently at the baseline-offset (i.e., 0% test contrasts), and this pattern of data was consistent across all standard levels (Fig. 9a). We found a significant main effect of period, showing that α log power decreased significantly during the stimulus period compared with the precue baseline (F_(1,19) = 51.12, p = 0.0004). Importantly, we found significant interactions between attention and period (F_(1,19) = 389.26, p < 0.0001), and between the attention, period, and test contrast (F_(6,114) = 7.86, p < 0.0001) with no main effect of standard contrast and no interactions between standard contrast and any other factors (p values > 0.1038). This set of statistical results could be described by the significant reduction in log α power during the stimulus period with respect to the precue period.

Next, we fit the neural CRFs based on log α power to compare the response gain, contrast gain, and baseline-offsets across attention and standard contrast conditions. We found results similar to the main α-based CRF analysis (compare Figs. 8a, b and 9f, g). Specifically, we found that attention significantly increased the baseline-offset of the CRF based on the reduced α log power with respect to the precue baseline (F_(1,19) = 38.65, p < 0.0001). Since the reduction of log α power saturated at high test contrast levels, we found no differences in attentional modulations at high test contrasts. This led to a decrease in the response gain parameter that controls the slope of the α-based CRFs (F_(1,19) = 16.27, p = 0.0007). Importantly, there were no significant main effects of standard contrast and no interactions between attention and standard contrast on the baseline-offset and response gain parameters (p values > 0.065), suggesting that the patterns of attention modulations of the α-based CRFs were independent of the level of standard contrast. In addition, there were no changes in contrast gain (p values > 0.1143). Together, the main analysis of the neural CRFs based on α amplitude and this additional analysis of log α power, which separated out the periodic components of α band oscillations from the aperiodic components of the 1/f activity, provides converging evidence that the attentional modulations of α amplitude/power undergo the baseline shift in the neural CRFs.

The other periodic components, including the central frequency and bandwidth of α band oscillations, reduced significantly during the stimulus period with respect to the precued baseline (F_(1,19) values = 51.12 and 8.78 with p values = 0.0004 and 0.0252, respectively) (Fig. 9b,c). For the central frequency, there were no significant interactions between period and any of other factors, including attention, test contrast, and standard contrast, suggesting that the reduction in the central frequency α band oscillations with respect to the precued baseline was not modulated by attention, test contrast, or standard contrast (p values > 0.223). For the bandwidth parameter, however, attention also interacted significantly with period because of a higher degree of bandwidth reduction with respect to the precue baseline in the uncued compared with the cued conditions (F_(1,19) = 6.75, p = 0.0407). That said, there were no significant interactions between period and the other factors, including test and standard contrasts (p values > 0.1616).

The modulatory patterns of aperiodic components of the 1/f were different from the patterns of periodic components of the α band oscillations reported above (see Fig. 9d,e). Both the aperiodic offset and exponent increased during the stimulus period compared with the precue baseline (F_(1,19) values =1600.71 and 3542.80, respectively, p values ≤ 0.0001). The increased aperiodic offset and exponent with respect to the precue baseline increased as a function of test contrast (the interaction between period and test contrast: F_(6,114) values = 10.92 and 5.81, p values < 0.0001 and = 0.0003), but they decreased with attention (the interaction between period and attention: F_(1,19) values = 20.69 and 6.45, p values = 0.0039 and 0.044). Unlike the α log power data, the degree of attentional modulations in the aperiodic offset and exponent were larger at high compared with lower test contrasts. Together, our results suggest that changes in the amplitude/power of α band activity could not be explained by the modulatory patterns of the aperiodic components of the 1/f activity.

Quantitative linking models suggest that different attentional modulations of neural data relate to different modulations of behavioral data

Next, we used a quantitative model to link patterns of attentional modulations of the P1-based and α-based CRFs and the pattern of attentional modulations in the psychometric data (see Materials and Methods) (see Tanner and Swets, 1954; Boynton et al., 1999; Pestilli et al., 2011; Cutrone et al., 2014, 2018; Itthipuripat et al., 2014b, 2015, 2017). The linking model used the patterns of attentional modulations of the observed P1-based and α-based CRFs to predict changes in contrast appearance and response bias in the observed behavioral data. The model is based on the assumption derived from the SDT where observers’ contrast discrimination accuracy relies on the difference in neural responses (ΔR) related to the standard (R_standard(c)) and test stimuli (R_test(c)) divided by the trial-by-trial variability of neural responses (termed as neuronal noise) (Tanner and Swets, 1954; Boynton et al., 1999; Pestilli et al., 2011; Cutrone et al., 2014; Itthipuripat et al., 2014a, 2017, 2019a). For a given pair of standard and test stimuli, the model computed the probability of a test stimulus being perceived as having a higher contrast than a standard stimulus ((p(test > standard)) using a maximum likelihood decision rule with neuronal noise equally distributed across the standard and test stimuli.

We first stimulated the behavioral data using the normalized P1-based CRFs (termed here as the P1-based model) and compared the simulated results with those predicted using the normalized α-based CRFs (termed here as the α-based model; see details in Materials and Methods). The amplitude of α band activity generally got smaller with attention and stimulus contrast (see Fig. 8) (see also Itthipuripat et al., 2019b; Foster et al., 2021). Therefore, we flipped the sign of the normalized α-based CRFs before estimating p (test > standard). Last, we compared the results with those predicted by a model that sums the normalized P1-based CRFs and the normalized α-based CRFs (with the flipped sign) to predict the behavioral data based on the sum of the normalized P1 and α responses. For each linking model, neuronal noise was one free parameter shared across all contrast levels and attention conditions. Thus, since each model had the same number of free parameters, we directly compared the goodness-of-fit of individual models (i.e., R² values) and log-likelihood estimates to compare how well they predicted the psychophysical data.

The fivefold cross-validation bootstrapping results demonstrated that the P1-based model accounted for the psychophysical data reasonably well, especially at the low-to-mid-level contrasts (Fig. 10). That said, the P1 model could not capture changes in the baseline-offset in the behavioral data at lower standard contrast levels. This resulted in poorer model fits at these contrast levels. The P1 model also performed poorly at higher standard contrast levels. This poor fit occurred because attentional modulations of the P1-based R_max parameter diminished with increasing standard contrast, although the attentional modulations of the psychophysical data remained robust at these contrast levels (Fig. 3).

Compared with the P1-based model, the α-based model performed significantly worse at predicting the pattern of the behavioral data in general (p values < 0.001 for differences in both log-likelihood estimates and R²). This is because the α-based model could only capture attention-induced changes in the baseline-offset of the psychometric functions at the low standard contrast levels, which only accounted for a small fraction of the variance in the overall behavioral data. That said, when we used the combination of the P1 and α data to predict the psychometric functions (i.e., the combined model), we were able to predict the pattern of the behavioral responses significantly better than using the P1 data or the α alone (p values < 0.001 for differences in log-likelihood estimates and R² between the combined and α models as well as between the combined and P1 models). This improvement in model performance was because the combined model better captured the baseline-offset response bias at the low and high standard contrast levels.

When modeled separately for each standard contrast level (Fig. 10d), we found that the P1 and α models were equally bad at predicting the data at low standard contrast levels and performance did not differ between the models (p values = 0.1000 and 0.3480 for 0% and 5% standard contrasts). Moreover, the combined P1+α model performed significantly better than either the P1 (p values = 0 for both 0% and 5% standard contrasts) or the α model on its own (p values = 0 for both 0% and 5% standard contrasts) (see the first two columns of Fig. 10a). This occurred because the baseline-offset modulations of the α-based CRFs help the combined model better account for changes in the baseline-offset response bias in the psychometric data. On the other hand, for the intermediate standard contrast levels, the P1 model performed very well and significantly better than the α model (p values = 0 for 10% and 20% standard contrasts). The combined model, therefore, did not perform significantly better than the P1 model alone (p values = 0.8160 and 0.1660 for 10% and 20% standard contrasts) (see the third and fourth columns of Fig. 10a). Additionally, no differences between model performance were observed between the P1 and α models at higher standard contrast levels (p values = 0.4060-0.8720). For 40% standard contrast, the combined model was significantly better than the P1 model (p = 0.0140, passing the Holm–Bonferroni-corrected threshold of 0.0167), but it did not perform significantly better than the α model (p = 0.1040). For 80% and 100% standard contrasts, the combined model was slightly better than the P1 model at explaining attentional modulations at of the psychometric functions (see the last two columns of Fig. 10a); however, the two models did not significantly differ (p values = 0.1740-0.6060). Nonetheless, the combined model was marginally and significantly better than the α model at the 80% and 100% contrast levels, respectively (p values = 0.0280 and 0.0140 with a Holm–Bonferroni-corrected threshold of 0.0167).

In addition, we ran an auxiliary analysis where the weights of the P1 and α data were varied systematically before computing the combined response to examine the relative contribution of the P1 and α data at explaining changes in the psychophysical data (Extended Data Fig. 10-1). Overall, the results were consistent with the main analysis where the combined model was significantly better than the P1 or the α model. The R² and log-likelihood estimates were the highest for the model with 6:4 P1/α weight ratio. However, this model was not significantly better than the combined model with equal weighting of the P1 and α data. Together, these results suggest that, under the assumption of linear summation with equal weights given to the P1 and α data, the attentional modulations of these neural markers sufficiently account for attention-induced changes in contrast appearance and response bias. Specifically, the response gain modulations of the P1 component explained attention-induced changes in perceived contrast at the low-mid-level standard contrasts but changes in the baseline-offset of α band activity underlie cued-induced response bias at very low and very high standard contrasts.