The temporal dynamics of the Stroop effect from childhood to young and older adulthood

It is well admitted that children and older adults tend to show longer response latencies at the Stroop task than young adults. The present study aims at clarifying the rational of such changes from childhood to adulthood and in ageing by comparing the impacted cognitive processes across age groups. More precisely, the aim was to clarify if all processes take more time to be executed, hence implying that longer latencies rely mainly on processing speed or if an additional process lengthens the resolution of the conflict in children and/or older adults. To this aim we recorded brain electrical activity using EEG in school-age children, young and older adults while they performed a classic verbal Stroop task. To decompose the signal in the underlying brain networks, we used microstates analyses and compared congruent, incongruent and neutral trials across the three age-groups. Behaviorally, children and older adults presented longer latencies and larger Stroop effects relative to young adults. The microstates results showed that children tend to present different brain configurations compared to both adult groups, even though some brain configurations remained identical among the three groups. In particular, additional brain networks were involved in children to perform the Stroop task, which party account for the longer latencies in this group. By contrast, in aging the results favor the general slowing hypothesis rather than a decline in a specific process since all involved brain networks were similar in the two adult groups but slowed down in the older one.

implying that longer latencies rely mainly on processing speed or if an additional process lengthens the 23 resolution of the conflict in children and/or older adults. To this aim we recorded brain electrical 24 activity using EEG in school-age children, young and older adults while they performed a classic verbal 25 Stroop task. To decompose the signal in the underlying brain networks, we used microstates analyses 26 and compared congruent, incongruent and neutral trials across the three age-groups. Behaviorally, 27 children and older adults presented longer latencies and larger Stroop effects relative to young adults. 28 The microstates results showed that children tend to present different brain configurations compared 29 to both adult groups, even though some brain configurations remained identical among the three 30 groups. In particular, additional brain networks were involved in children to perform the Stroop task,31 which party account for the longer latencies in this group. By contrast, in aging the results favor the 32 general slowing hypothesis rather than a decline in a specific process since all involved brain networks 33 were similar in the two adult groups but slowed down in the older one.

35
Attentional control is probably one of the most studied topics in cognitive psychology. Historically, this 36 process has been investigated with many paradigms among which the Stroop task [1], extensively 37 used. Even though the exact design of the task varies across studies, the Stroop paradigm can generally 38 be defined as the presentation of two congruent or incongruent pieces of semantically related 39 information while the subject has to "name" (manually or orally) only one of them, usually the color 40 of the ink in which words are displayed. It is typically observed that subjects are slower to process 41 incongruent items (e.g. naming the ink color of the word "blue" written in red) than congruent ones 42 (e.g. naming the ink color of the word "green" written in green), and the congruent condition seems 43 to lead to less errors than the incongruent one [2]. This difference in reaction times and accuracy is 44 known as the Stroop effect. Since 1935, a large number of studies have tried to understand the core 45 processes underlying the Stroop effect, namely how the conflictual information of the word reading 46 and color naming processes is handled from a cognitive and neural point of view. Authors globally 47 agree on a two levels of conflict model explaining the Stroop interference. The first one appears after 48 the perception of an incongruency in the stimulus (stimulus conflict), while the second one is a 49 response-based conflict, occurring when the response has to be selected among the two possible 50 answers [3][4][5][6]. Moreover, the Stroop task involves some additional processes relative to other 51 attentional control tasks such as the flanker or the Simon tasks. Indeed, it involves components 52 engaged in word production as well as executive processes allowing to select the correct answer [7].

53
To go beyond the conclusions obtained from behavioral studies, researchers used neuroimaging 54 approaches, such as magnetoencephalography -MEG - [8,9] and electroencephalography -EEG -[10-55 21], allowing to infer on the cognitive processes and their dynamics. The results of these electro-and 56 magnetophysiological studies converge on some central conclusions. Indeed, most studies reported a 57 more negative deflection for incongruent relative to congruent trials around 400 to 450ms associated 58 with negative centroparietal topographies (referred to as N400 hereafter). A second component has 59 also been described for the same comparison (congruent versus incongruent trials) at a later time-60 period. The latter sometimes is referred to as the slow potential component (SP), late positive complex 61 (LPC), the P600 or SP600 (SP600 hereafter), and is widely spread around 600ms, showing a 62 centroparietal positivity. Notwithstanding converging results on the components underlying the 63 Stroop effect, the cognitive processes behind these components have been widely debated. A general 64 agreement has however emerged on the idea that the N400 is associated to the detection of the 65 conflict (conflict at the stimulus level) while the second component has been associated with conflict 66 resolution (conflict at the response level).

67
It is also noteworthy to mention that the vast majority of behavioral and neuroimaging studies used 68 manual responses (mapping between the color response and a keyboard key), and very few studies 69 had participants giving a verbal response [8,18,[21][22][23]. These two response modalities do not always 70 lead to the same results. Some studies tried to understand the electrophysiological differences 71 characterizing the verbal and manual versions of the task. Initially, Liotti and colleagues (2000) 72 confirmed the behavioral finding already reported in earlier studies that the Stroop effect was larger 73 in the verbal modality [2] and showed increased N400 and SP600 effects. However, when comparing 74 the response latencies with the timing of the components, it appears that the SP600 effect is located 75 at the same temporality as the response time. It implies that the observed difference on the SP600 76 component could simply be attributable to a shift in the response artifact latencies between the 77 congruent and incongruent conditions. Classically, to overcome this limitation, the ERP signal is aligned 78 first to the stimulus (stimulus-aligned signal) and then to the response (starting the temporal window 79 from the onset of the response and extracting anterior time points, i.e. response-aligned signal) 80 [24,25]. By using this procedure, some authors showed significant differences between congruent and 81 incongruent conditions only for the vocal version of the task close to the production of the response 82 [23]. These results were interpreted as an interference effect appearing not only at a semantical or 83 lexical level, but also at the word form encoding level. In other words, the verbal and manual Stroop 84 tasks are comparable in terms of involved cognitive processes except that the verbal Stroop task 85 involves also a word form encoding conflict component explaining increased interference effect for 86 this modality.

87
To sum up, the ERP components characterizing the cognitive processes involved in the Stroop task are 88 globally well defined and their underlying mental processes are consistently described. However, this 89 is only the case for young adults: other populations such as typical children and healthy older adults 90 might show differences in mental processes and brain activations when performing the Stroop task. 91 As further detailed below, the literature shows that the timing of the ERP components elicited by the 92 task is modulated by the age of the participant. It is nonetheless still unclear (1) how the Stroop effect 93 evolves across the lifespan both behaviorally and neurophysiologically and (2) if the involved processes 94 are identical at all ages. Hereafter we will review the available evidence on these two issues and the 95 questions that are still debated or unresolved. 96 97 The large majority of what we know about the Stroop task comes from studies based on young adult 98 participants. A few studies have investigated changes in executive control over the entire lifespan with 99 the Stroop task or some other tasks [17,[26][27][28][29][30], but these studies are not numerous in the wide 100 literature of the evolution of executive control abilities. That is why in the following, we will first 101 examine the evolution of the Stroop effect from childhood to adulthood, and at a second stage, focus 102 on the mechanisms explaining the decrease of performances in aging. 103 104 The literature on the evolution of performance from childhood to adulthood suggests that even after 105 a short exposure to reading, children show a strong and reliable Stroop effect despite longer latencies 106

From childhood to adulthood
and an increased error rate [2,31]. In school-age children, the Stroop effect evolves in a non-linear 107 pattern, which seems entirely related to reading acquisition as when accounting for reading level, a 108 linear trend appears [32]. This observation involves that the development of attentional control and 109 reading abilities interact in the evolution of the performances on the Stroop task across childhood. 110 Using an ERP approach with children aged from 6 to 12 years old performing a semantic Stroop task 111 (e.g. naming the font of the word "sky" written in blue), Jongen and Jonkman (2008) claimed that 112 stimulus conflict (reflected by the N400 component) appears early in the development, which 113 translates by an early setup of the inhibition mechanisms, while the response conflict (reflected by the 114 SP600) continues to mature until 12 years old. This result is in accordance with a literature review, 115 stressing that the processes at play when handling interference are not fully developed before the age 116 of 12 years or later [34], as well as a study investigating the evolution of the conflict adaptation 117 (adaptation of the attentional control system from one trial to the next) [35]. The literature on other 118 attentional control tasks such as the flanker task suggests that conflict detection processes reach their 119 maturation at the age of 10 years old while, surprisingly, conflict resolution does not seem to evolve. 120 The absence of evolution of conflict resolution processes could be related to the non-verbal character 121 of the task [36]. Overall, authors conclude that since the ERP components are already present in 122 childhood, processes are equivalent but differ in their maturational status. However, even though 123 processing speed evolves during development, it cannot explain the entire changes observed between 124 childhood and adulthood [37]. 125

126
At the other end of the lifespan, it is well admitted that cognitive performances decline, and this is 127 particularly well described regarding conflict processing abilities. Different explanations have been 128 proposed regarding the general alteration of cognitive performances with aging . Two hypotheses have  129  been suggested: the general slowing hypothesis and a specific alteration of executive functions with  130 aging. These hypotheses were derived partly from the decline observed in the Stroop task.

131
The general slowing hypothesis explains the longer latencies in aging by a slowdown of all processes. 132 The decline could then simply be explained by a decrease in attentional resources but a stability of the 133 inhibition abilities [38][39][40][41][42] results were interpreted as a specific decline of these mechanisms in aging, although the authors did 155 not conclude on the origin of such decline. It seems that both stimulus and response conflict are 156 impacted since the N400 as well as the SP600 are modulated, nevertheless further clarifications are 157 needed to better characterize this phenomenon and to disentangle the general slowing from specific 158 deficit theories.

159
At a first glance the two theories are contradictory, but it is possible that they are two faces of the 160 same coin. All processes can be slowed down in aging due to attentional decline but some may be 161 affected more than others, reflecting a decline in specific executive processes. Moreover, a slowdown 162 in latencies does not exclude that compensatory mechanisms are involved. The complementarity of 163 the two theories has been investigated using behavioral approaches [50]. Applying a hierarchical 164 regression model, the authors observed that processing speed (measured by a simple reaction time 165 task) accounted for a large part of the variance but age explained a specific part of the residual 166 variance, supposing a specific deficit in executive functions, although this result may be further clarified 167 using electrophysiological approaches.

168
To sum up, state of the art literature agrees that brain activations change from childhood to aging. 169 Indeed, components measured on the EEG signal vary across ages in amplitude and timing of the peaks. 170 The changes in latencies and error rates in the Stroop task are generally interpreted as the 171 consequences of less developed processes in childhood compared to adulthood and a decline in these 172 same processes with aging. Differences in timing could translate a linear development of attentional 173 control as well as a linear decline of the performances of the latter, but no arguments were highlighted 174 in the neurophysiological literature regarding a specific decline of attentional control, as argued partly 175 in the behavioral literature. 176

A single mechanism underlying congruent and incongruent trials,
177 at least in adulthood 178 One of the core questions tackled in the first neuroimaging literature dedicated to the Stroop paradigm 179 was to find how the conflict is processed and whether it requires an additional mechanism as 180 compared to congruent or neutral trials. EEG results usually indicate similar components across 181 conditions, except for a difference in amplitude of the N400 component and the SP600 (but see  182 concerns about the SP600 in Stroop tasks with manual -button press -responses reported by Zahedi 183 and colleagues (2019) and summarized above). Since latencies vary between the congruent and 184 incongruent conditions, the amplitude difference observed between the components might be 185 attributed to a shift in the signal rather than to different processes (i.e. differences in amplitude) 186 underlying the two conditions. The issue of whether there is a pure shift of the same components or 187 different brain processes across conditions is best investigated using microstates analyses. This 188 method allows to investigate the presence of different global electric fields in the signal, i.e. of different 189 underlying mental processes [51-54], but also their relative onset and duration. The analysis would 190 allow to further describe whether the same processes underpin the two conditions in the Stroop task 191 and whether this is the case across age-groups. Several studies investigated the Stroop effect using 192 microstates [16,[55][56][57], but to our knowledge, only two of them have investigated the brain networks 193 involved in both congruent and incongruent conditions. Khateb and colleagues proposed a modified 194 Stroop task to young adult participants and segmented the ERP signal in periods of quasi-stable global 195 electric fields [16]. Their results confirmed that the brain processes are common between congruent 196 and incongruent trials, but that one microstate was extended in the incongruent trials compared to 197 the other condition, reflecting differences in response latencies. The differences in latencies could thus 198 be explained by the increased duration of a particular topography in the signal of incongruent trials. 199 The Stroop interference would then be recruiting a large network and the results speak against a 200 specific module at play to detect and resolve conflicts, in accordance with the rest of the literature. loci of the effect remain unclear, but might be due to the design of the tasks, since Khateb and 214 colleagues used a passive pseudo-Stroop task.

215
In the present study, we went a step further by investigating whether the same underlying processes 216 for congruent and incongruent trials are also observed during development and in aging. As 217 summarized above, despite behavioral and ERP changes observed across the lifespan on the Stroop 218 task, the underlying processes seem to be the same, although less efficient during maturation and in 219 aging. This issue has however never been confirmed directly by considering both children and older 220 adults in a same study. To address this, we designed a classical serial Stroop task including congruent, 221 incongruent and neutral items (rows of symbols displayed in different colors) while recording the 222 participants brain activity with high density EEG in school-age children (10 to 12 years old), young 223 adults (20 to 30 years old) and older adults (from 58 to 70 years old). EEG signal was first analyzed with 224 peak and mass univariate tests on the waveform amplitudes, and in a second step, using topographical 225 or microstates analysis. Finally, source localization analysis was performed on the time-window of the 226 microstates carrying the Stroop effect to clarify the underlying brain structure and associated cognitive 227 processes. Even though children, young and older adults were included in the same analysis to 228 distinguish better the different trend of development and aging, the results were interpreted 229 separately, as distinct research questions.

230
According to the literature on the development of the Stroop effect reported above, we expected that 231 children would have the same brain networks at play when performing a congruent or an incongruent 232 trial. However, it is still unclear if children are slower and display larger Stroop effects relative to young 233 adults because of the immature attentional resources impacting all processes or if only some processes 234 need more time to become fully efficient. In other words, if microstates are identical in children and 235 in young adults but are only different in duration, then it might suggest that the executive processes 236 are already mature in school-age children but attentional resources are not developed enough to allow 237 reaching the performances of young adults. Another possible explanation would be that one specific 238 mechanism is still immature in children and explains the slower reaction times. This hypothesis would 239 be confirmed if additional microstates were identified in children relative to young adults or if a specific 240 microstate was significantly more impacted in children.

241
The same rationale may be applied regarding aging. A lengthening of all microstates would favor the 242 general slowing hypothesis while a supplementary process (additional microstate) in older adults 243 relative to young adults would favor a compensatory mechanism. Finally, a disproportionately more 244 present process in the older adults group would favor the decline of a specific brain network in aging.

247
Seventy-four participants were recruited, divided into three age groups (children, young adults and 248 older adults). Among them, only participants with a minimum of 25 artifact-free ERP epochs per 249 condition (see pre-analyses) were retained. The final sample thus included 53 participants: 17 children 250 from 10 to 12 years old (mean age: 11; SD = 0.79; 7 females), 18 young adults from 20 to 30 years old 251 (mean age: 23.9; SD = 2.97; 13 females) and 18 older adults from 58 to 70 years old (mean age: 64.9 ; 252 SD = 3.6; 14 females). This sample size seemed reasonable given that previous studies included from 8 253 to 25 participants in average [14,18,59]. Older adults aged less than 70 years old were preferred to 254 elderly to avoid cognitive decline and selection bias related to the fact that only the more cognitively 255 performant older persons agreed to participate to neuroscientific studies. All participants were right 256 handed [60] native French-speakers, and did not report any language, neurological, psychiatric or color 257 vision impairment. All participants or their legal representative gave their written consent before the 258 beginning of the procedure and received a financial compensation for their participation. The entire 259 procedure was approved by the local Ethics Committee.

260
The participants were part of a larger group from which only the behavioral data has been analyzed in 261 previous study [40].

263
A 180 trials, classical four colors Stroop task requiring verbal responses was used. The verbal stimuli 264 were four color names in French ("bleu" ; "jaune"; "rouge"; "vert", respectively blue, yellow, red and 265 green) as well as symbols displayed in different colors, were those of Fagot and colleagues (2008). The 266 task set encompassed 60 congruent (the color font and the color words match, e.g. the word "bleu" 267 written in blue), 60 incongruent (there is a discrepancy between the color font and the color word, e.g. 268 the word "vert" written in red) and 60 neutral items ("++++"; "^^^^"; " '''' "; "****") displayed in one 269 of the four different colors. All verbal stimuli were presented in lower case at the center of the screen.

271
Participants sat at approximately 80 cm of a 17 inches computer screen (refreshment rate: 50Hz). 272 Stimuli were presented using the E-Prime software (E-studio). Oral responses were recorded by a 273 microphone and sent to the E-Prime software to be recorded and labelled. To estimate reaction times, 274 the onset of the production was manually retriggered offline using the CheckVocal software [62].
275 Participants received the instruction to name in which color the stimuli were displayed. They were also 276 asked to produce their response orally as fast and accurately as possible. Regarding the instructions 277 relative to the EEG recording, in order to avoid artifacts in the signal, the participants were asked to 278 remain still and blink only after the end of the production of their responses.

279
The trial structure was identical for all the age groups. First, a white fixation cross was displayed on a 280 black background for 500ms. To avoid visual responses' contamination on the temporal window of 281 interest due to the fixation cross, a black screen was presented for 200ms. The stimulus was then 282 presented for 1500ms, followed by a variable interstimuli black screen for which the duration was 283 ranging randomly from 1000 to 1200ms.

284
Before the beginning of the task, a training phase of 32 items including all the possible stimuli was 285 proposed to the participant. The purpose was first to familiarize the participants with the task, but also 286 to avoid, or at least attenuate, an eventual novelty effect on the first trials.

294
A trial was excluded if the participant gave an incorrect color name, corrected himself/herself 295 immediately after the production of an error, or did not give any response. Minor hesitations such as 296 phonological transformation (example: "r-rouge") were considered as correct. Regarding latencies, all 297 response times under 200ms were excluded, as well as values outranging the threshold of 2.5 standard 298 deviations above and under the mean of the age group. The respective rejected trial rates for 299 congruent, incongruent and neutral condition was: 0.85%, 3.85% and 0.35%.

300
The data were analyzed using linear mixed models with the lmer function (lme4 package) for latencies 301 and generalized linear mixed models fitting a binomial distribution with the glmer function (from the 302 same package) regarding accuracy. Post-hoc analyses were estimated using the emmeans function 303 performing a Tukey test. It is noteworthy to mention that since the model initially did not converge, 304 the optimization algorithm was changed for "nlminb". For both models, post-hocs were estimated by 305 a Tukey test targeting relevant comparisons. Moreover, to obtain main effects on generalized linear 306 mixed models, the Anova function from the car package was used instead of the classical anova 307 function from base R used for lmer models. 308 309 The electroencephalogram (EEG) was recorded continuously during the task by 128 electrodes placed 310 on a nylon cap following the 10-5 system. The signal was acquired by a Biosemi amplifier including two 311 active electrodes (ActiveTwo system, Biosemi V.O.F. Amsterdam, Netherlands) at a sampling rate of 312 512Hz including an online DC filter ranging from 0.01 to 104Hz and a 3dB/octave slope.

313
Regarding data cleaning, all the manipulations were performed with the Cartool software (V. 3.91) [74] 314 and with the R software (V.3.8) [63]. The signal was first filtered by using a notch filter at 50Hz and a 315 zero-phase shift order 2 bidirectional butterworth bandpass filter including a highpass filter at 0.3Hz 316 and a lowpass filter at 30Hz. The signal was then downsampled at 256Hz to reduce the computing time 317 and the number of multiple comparisons of the statistical analyses. A first interpolation was performed 318 on the entire signal for noisy electrodes following a 3-D spline method [75]. The mean number of 319 interpolated electrodes was 4.79 (SD = 1.84) (with at maximum 10 interpolated electrodes, or 7.81% 320 of the 128 electrodes). Regarding the event related potentials (ERP) extraction, two time windows 321 were selected: a stimulus-aligned period and a response-aligned period. The rationale of two 322 alignment points has been presented in the Introduction [24,25]. The data aligned to the stimulus for 323 a shorter time period than the behavioral latencies allow to track the N400 effect; the data aligned to 324 the response onset (backwards) allows the investigation of later stages of word form encoding. Finally, 325 this approach is also adopted to avoid the presence of artifacts due to articulation in the EEG signal. 326 The first period included the first 25 time frames (~100ms) before the stimulus presentation, which 327 was taken as baseline, and the 130 time frames (~520ms) following the stimulus presentation. This 328 stimulus-locked ERPs lasting ~520ms ensured the entire signal to be at least 100ms shorter than the 329 fastest RTs (629ms in the C condition in young adults, see Table 1). The second window was selected 330 backward from the response to the stimulus and lasted 100 time frames (~400ms). For the purpose of 331 the EEG analyses, the response latencies used as onset of the response-aligned period were reduced 332 by 100ms to avoid measuring the premises of articulation [25]. Since no automatic algorithm of 333 artifacts correction were used, all the epochs were visually inspected and only the artifact-free epochs 334 were retained for the averaging, and grouped according to the Stroop conditions (congruent, 335 incongruent and neutral). Finally, a second interpolation was performed on the averaged data of each 336 subject (mean number of interpolated electrodes: 3.38 (SD = 2.75); maximum 9 electrodes or 7.03% 337 of the total number of electrodes) using the same method as described above but avoiding 338 interpolation of previously interpolated electrodes or their direct neighbors. When combining the two 339 interpolations, on average, 8.17 (SD = 3.3) electrodes were interpolated per participant and the 340 maximum of interpolated electrodes was 16, or 12.5% of the total number of electrodes. On average, 341 17.48% (SD = 10.87), 26.82% (SD = 13.94) and 19.78% (SD = 13.08) of the trials were rejected for the 342 congruent, incongruent, and neutral conditions, respectively, including the ones already rejected 343 during the behavioral analyses. 345 To ensure comparison with the rest of the literature, first, the waveforms of the signal was analyzed. 346

EEG analyses
This was achieved by a peak analysis and by a mass-univariate statistic. In a second step, the 347 topography of the data was considered. It was verified that the signal was topographically consistent 348 using a topographical consistency test before running an analysis of variance between the three 349 conditions and age groups on the signal (tANOVA). This analysis will guide the interpretation of the 350 microstates and will allow to relate topographical differences with those observed on the waveform 351 analyses. The signal was then segmented in microstates and the characteristics of the microstates 352 (presence, duration and onset) were extracted. Since these microstates tend to be difficult to relate to 353 cognitive processes, source localization analyses were performed to identify the generators behind the 354 most relevant brain networks identified by the microstates analyses.  382 The microstates analysis allows to decompose the EEG/ERP signal in periods of topographical stability 383 (or quasi-stability). It has been proposed that when a cognitive process is taking place, the topography 384 of the scalp EEG remains stable, reflecting the activation of a specific brain network. During this 385 stability, different brain areas are connected according to one or several firing frequency/ies [77][78][79][80]. 386

Microstates analysis
Each of these periods, or microstates, can be segmented from the evoked potentials and are known to 387 reflect the activation of a network, generally underlying one or several cognitive processes.

388
Microstates analysis was performed separately on stimulus aligned and response aligned ERP signals 389 across groups using the Ragu software [53,54]. The software performed a segmentation on the 390 averaged ERP of each subject using a cross-validation method based on the topographical atomize and 391 agglomerate hierarchical clustering (TAAHC) algorithm, for a user pre-defined user range of 392 topographies (here from one to 30). At each fold (data are segmented in subsamples named folds and 393 models are trained on a majority of the folds and tested on the rest), the segmentation is operated on 394 the training data and the global explained variance is reported on both training and test data.

440
As shown in Table 1, children presented the slowest latencies in all conditions (mean = 851ms; SD = 441 207), and young adults (mean = 682ms; SD = 157) were faster than the older adults' group (mean = 442 788ms ; SD = 164). Children also displayed the highest standard deviations regarding errors and 443 reaction times, while young and older adults showed comparable variances despite differences in 444 reaction times, as highlighted by Figure 1. were less prone to errors than incongruent trials (Z = 6.68; p<0.001), but none of the other comparisons 468 reached significance. The age groups effect shows that accuracy was significantly lower in children 469 compared to young adults (Z = -4.52; p < 0.001), however no significant differences were observed 470 between children and older adults (Z = -0.02; p ≈ 1) nor between young and older adults (Z = -0.02; p 471 ≈ 1).

473
[ Fig 2 about 480 The results of the mass univariate tests on the waveform amplitudes run separately on the stimulus-481

Mass univariate test and peak analysis
and response-aligned ERPs are presented in Figure 3. Since neutral trials were very different from 482 congruent and incongruent trials, it was suspected that most of the effect would be carried by the 483 neutral condition compared to the other two conditions. That is what was observed since significant 484 difference in amplitudes appeared on virtually the entire time window for both main effects of 485 conditions and age groups as well as for the interaction as presented in Appendix S1. Since the present 486 study focused mainly on the Stroop effect, contrasts including neutral conditions were discarded, and 487 in Figure 3, only the contrasts between the incongruent and congruent condition are displayed for 488 each age group.

489
The results of the mass univariate statistic highlighted different time windows of significant variations 490 in amplitudes between congruent and incongruent trials in the three age groups (displayed in gray on 491 Figure 3). First, the older adult group is the only one yielding significant differences between the two 492 conditions in the stimulus-aligned analysis in the time-window from 260 to 520ms and on large clusters 493 of electrodes. On the response-aligned analysis, children and young adults show significant effects of 494 conditions in a similar time window (from -500 to -360ms), although not on the same electrodes, and 495 with an additional time window (from -350 to -180ms) in children. Older adults presented a time 496 window of significance closer to production (from -210 to -100ms).

497
The temporal window of interest of the N400 component did not show any significant effect for the 498 children and young adult groups on the mass univariate tests. To ensure that our results were 499 comparable to the rest of the literature, we conducted a peak analysis on a temporal window 500 compatible with this component and showing the maximum difference based on a visual inspection. 501 Repeated measures ANOVAs were conducted on an averaged time window ranging from 300 to 420ms 502 (as shown on Figure 2) for the FPz, Fz, Cz, Pz and Oz electrodes. The models from all five electrodes are 503 detailed in the Appendix (Appendix S2) but only the significant results on FPz are presented here. First, 504 the model highlighted a significant main effect of conditions (F(2,100) = 6.94; p = 0.001), a significant 505 main effect of age groups (F(2,50) = 4.45; p = 0.017) and a significant interaction between conditions 506 and age groups (F(4,100) = 4.75; p = 0.001). Post-hoc decomposition achieved by a Tukey test showed 507 that young adults present a significantly more negative deflection in the congruent condition 508 compared to the incongruent one (t(100) = -2.45; p = 0.042), while this results was marginally 509 significant in the older adults group (t(100) = -2.26; p = 0.066). Complete decomposition of the 510 differences between conditions among the three age groups are presented un Appendix S3. 511 512 513 The results of the topographic analysis of variance (tANOVA) displayed in Figure 3 showed different 514 patterns across groups but were globally concordant with the mass univariate test. Regarding the 515 stimulus-aligned signal, in children, a small time window appeared significant during the baseline, close 516 to stimulus onset. In young adults, two small time periods suggested a change in topography, occurring 517 at about 400 and 520ms. In older adults, several periods of significance appeared around 400ms post-518 stimulus, but also briefly during the baseline. In the response aligned signal, the same pattern as the 519 mass univariate results were highlighted for children and young adults, namely two periods of 520 significance for children and one, far from the response onset, for the young adults. However, older 521 adults showed a discrepancy between the mass univariate statistic and the waveforms results. Indeed, 522 the tANOVA shows a significant time window far from the response onset, while the mass univariate 523 reveals an effect close to the onset of the response.   534 The TAAHC algorithm identified five microstates explaining the signal on the stimulus-aligned (Maps 1 535

Microstates analyses: segmentation and characteristics of the maps
to 5) and four microstates for the response-aligned signal (Maps 6 to 9). Descriptive statistics about 536 duration of the maps can be found in Table 2.  The first analyses were performed on the presence of each map in the individual averaged signal per 542

Maps
condition. For detailed presence information of each microstate map, see Appendix S4. Since the aim 543 was to assess whether the maps explained the signal among the different age groups or conditions, 544 the model tested for an interaction between conditions and maps as well as between maps and age 545 groups. Results showed a significant main effect of maps (stimulus-aligned: χ 2 = 39.2; p <0.001; 546 response-aligned: χ 2 = 13.08; p = 0.004), and a significant interaction between the maps and the age 547 groups (stimulus-aligned: χ 2 = 79.69; p <0.001; response-aligned: χ 2 = 84.15; p <0.001). The other 548 effects did not reach significance. Detailed decomposition was achieved by a Tukey test. The most 549 relevant comparisons showed that some significant differences were observed on the stimulus-aligned 550 signal among the three groups. Maps 1 and 2 were more present in the signal of older adults than in 551 children (respectively: z = -2.91; p = 0.01 and z = -2.37; p = 0.047). Map 3 was more specific to children 552 (older adults -children: z = 4.91; p <0.001; young adults -children: z = 4.89; p <0.001), while map 5 553 was more present in young adults, compared to older adults (z = -3.87; p <0.001). Regarding the 554 response-aligned signal, results consistently show that map presence differs drastically among children 555 and adults. Indeed, only comparisons involving children and adults returned significant while none of 556 the comparisons between the two adult groups did. As presented in Figure 4, maps 6 (Children -Older 557 adults: z = 2.73; p = 0.017; Children -Young adults: z = 2.81; p = 0.014) and 8 (Children -Older adults: 558 z = 3.28; p = 0.003; Children -Young adults: z = 4.44; p <0.001) were significantly more characteristic 559 of children while maps 7 (Children -Older adults: z = -2.86; p = 0.012; Children -Young adults: z = -560 4.06; p <0.001) and 9 (Children -Older adults: z = -3.01; p = 0.007; Children -Young adults: z = -3.5; p 561 = 0.001) were more representative of adults, young and older. As a consequence of these results on 562 the presence of maps across groups, only maps present on a sufficient number of subjects within a 563 group were compared across conditions. Even though all data were still included in the statistical 564 model at all time, for each comparison among the groups and/or conditions, only the comparisons 565 among the maps present in Figure 4 were interpreted. 566

567
The statistical model on microstates duration (Hurdle model) in the individual ERPs included as fixed 568 factor the maps labels, the age-groups, as well as the conditions of the Stroop task. All interactions 569 were entered in the model. The random structure included the subjects' ID as random intercept, and 570 post-hoc decomposition of the effects was achieved by Tukey tests. Median durations of the 571 microstates maps per age group are presented in Appendix S5

572
As shown in Table 3, results of the conditional model (the zero-inflated part will not be interpreted) on 573 both stimulus-and response-aligned signal suggest a significant main effect of maps and conditions, 574 no main effects of age groups, but significant interactions between the maps and conditions, a 575 significant interaction between maps and age groups and a significant interaction between conditions 576 and age groups. Finally, a triple interaction between conditions, age groups and maps appeared 577 significant as well.  It is noteworthy to mention that the differences observed on the most distant map from the alignment 589 point (i.e. microstates at the upper and lower borders of stimulus-and response-aligned data 590 respectively: for children, Maps 5 and 6, for young adults, Maps 2 and 7, and for older adults, Maps 1 591 and 5 for stimulus-aligned and Map 7 for response-aligned signals) were discarded. Indeed, these maps 592 are susceptible to carry a bias due to differences in latencies among conditions and age groups whereas 593 the analyzes were carried out on identical fixed periods. In other words, the duration of the most 594 distant maps from the alignment point might be modulated by the fixed length of the epoch across 595 conditions and age groups.

596
First the comparison between age groups was considered (disregarding the differences among the 597 conditions) and in second time, the differences among the conditions will be reported.

598
As shown in the previous analysis on the presence/absence, larger differences are observed between 599 children and adults than between the two adult groups for the stimulus-aligned signal. According to 600 Figure 4, Map 1 can be compared between the two adult groups and only Maps 2 and 5 can be 601 compared between the children and the two adult groups. Results suggest that Map 1 had a 602 significantly longer duration in the older adults group compared to the young adults group (t(704) = 603 3.26; p = 0.003). Maps 2 and 5 did not show any significant difference between any of the groups. 604 Furthermore, some maps differentiate the congruent and incongruent conditions. For the children's 605 group, only Map 4 was marginally significantly more present in the incongruent condition compared 606 to the congruent one (t(704) = -2.07; p = 0.097). For the young adults group, no effect was reported, 607 and concerning the older adults group, Map1 (t(704) = -3.95; p <0.001) and Map 5 (t(704) = -3.12; p = 608 0.005) were significantly more present in the incongruent condition. Nevertheless, it is worthy to 609 mention that on the segmentation depicted on Figure 4, Map 5 seems much more present in the 610 congruent condition. This might be related to the fact that, as presented in Figure 4 and above in the 611 absence/presence analysis, only 50% of older adults showed this map in the incongruent condition.  621 To confirm that differences in duration explained differences among the groups, the onset time of each 622 microstate was analyzed for each subject. The model included all main effects and interactions among 623 maps, conditions and age groups. For this analysis, maps at the extremity of the temporal window have 624 been considered.

625
Regarding the stimulus-aligned data, the results completed the ones observed on the duration of the 626 maps and highlighted a significant main effect of maps (χ 2 = 230.25; p <0.001), a significant main effect 627 of age group (χ 2 = 11.24; p = 0.004), and a significant interaction between maps and age groups (χ 2 = 628 175.58; p <0.001). As stated above, only relevant post-hoc comparisons were interpreted, namely 629 differences among all age groups for Maps 2 and 5, and also Map 1 for the adult groups. Map 2 had a 630 significantly earlier onset for children (z = -5.57; p <0.001) and older adults (z = -5.26; p <0.001) 631 compared to young adults. Regarding Map 5, as presented in Figure 4, this map had a significantly 632 delayed onset compared to both young (z = 6.04; p <0.001) and older (z = 4.77; p <0.001) adult groups. 633 Regarding Map 1, older adults have a significantly later onset compared to young adults (z = 2.86; p = 634 0.012).

635
Regarding the response-aligned model, the results point out a significant effect of maps (χ 2 = 101.22; 636 p <0.001), a significant interaction between maps and age groups (χ 2 = 45.64; p <0.001), as well as a 637 marginally significant interaction between maps and conditions (χ 2 = 11.27; p = 0.08). Since the triple 638 interaction including age groups was not significant, this interaction was not decomposed. Post-hoc 639 decomposition of the interaction between Maps, age groups and conditions did not highlight any 640 relevant significant effect. Indeed, the only relevant comparison among the groups was between 641 young and older adults on Map 9 according to the segmentation presented on Figure 4, which failed 642 to reach significance (z = -1.29; p = 0.403). 644 To better understand the brain networks involved in the microstate maps, a source localization 645 procedure was performed on the grand averages, based on the timing of the microstates maps bearing 646 the differences among groups or conditions ( Figure 5). In the stimulus-aligned signal, Map 3 appeared 647

Source localization of the microstates
to be specific to the children group but did not show a significant difference between the congruent 648 and incongruent conditions. This map reflects mainly activations in the bilateral (but mainly right) 649 inferior temporo-parietal regions. In the adult groups, Map 5, which encompasses the time window of 650 the N400, seems to carry the Stroop effect in the older adults group but not in the young adults group. 651 In young adults, congruent items elicit mostly parieto-occipital regions, while the maximum activation 652 located in the ACC and fronto-mesial regions are reported in the incongruent condition. In the older 653 adults group, the condition does not seem to impact the brain regions engaged. This age group displays 654 a globally similar pattern as the one observed in young adults in the congruent condition.

655
Regarding the response aligned signal, in the last 300ms before the response onset, the three age 656 groups show different patterns. Children and young adults show a strong fronto-mesial activation in 657 the incongruent condition. In the congruent condition, young adults show a distributed left temporal 658 network while children consistently show an activation of the fronto-mesial areas. Conversely to the 659 other two age groups, older adults show a temporo-insular activation in both congruent and 660 incongruent conditions.

666
The present study aimed at understanding how the Stroop effect with its two levels of conflict 667 (perception-and resolution-based conflicts) evolves with age, since latencies are longer for children 668 and older adults. The present results confirmed the larger latencies in both children and older adults 669 reported in previous studies, but the ERP results indicate different underlying processes in 670 development and aging. In the following, we will first briefly comment the behavioral results before 671 digging separately the issues related to development and aging. Finally, we will discuss additional 672 findings such as the robustness of the N400 and the reliability of the SP600 components before 673 concluding.

675
The behavioral results supported the expectations, showing a strong Stroop effect on the latencies for 676 each group while the measurement of accuracy did not highlight clear effects. Indeed, even though 677 the congruent condition was less prone to errors than the incongruent one, no modulation across the 678 age groups was emphasized. Across the three conditions, children were the group showing the largest 679 error rate. Regarding older adults, even though the difference was not significant, it is worthy to note 680 that this group was more accurate compared to their young counterparts, which could translate a 681 change in the speed-accuracy tradeoff between the two groups. Young adults might be more focused 682 on rapidity to provide a swift response while older adults might rather favor accuracy. This might be 683 the first argument claiming that aging is not maturation in reverse [27]. To explore more precisely how 684 maturation and aging differ, microstates analyses were performed on the task to help going beyond 685 the limitations of behavioral and classical ERP methods.

688
The signal of the three age groups includes the same electrophysiological components, namely the 689 P100, the N1/N170, the P300 and a long lasting positivity usually labeled as the SP600. However, as 690 shown in Figures 2 and 4 Stroop tasks) showed different components between age groups [33,35,86]. Since this particular 695 pattern has been shown only in picture naming and by the results of the present study, it is most 696 probably attributable to language processing.

697
Microstates allowed to segment the ERP signal in periods of quasi-stable electrical configurations 698 reflecting activated brain networks at certain temporalities. We will first describe here the within-699 groups comparisons for children and young adults, and in a second time discuss the differences in the 700 brain networks involved between the two groups on the stimulus-aligned signal, and then discuss 701 response-aligned results. First, our data confirmed that in children and young adults groups, the same 702 processes were involved in congruent and incongruent trials and only differed in duration. Such a result 703 was already reported in the literature [16,57] with young adults and favors the idea of a single conflict 704 detection mechanism activated continuously for all types of trials.

705
The Stroop effect (incongruent vs. congruent) was not significant on the duration of any of the 706 microstates found on the children stimulus-aligned signal, however, a trend was highlighted regarding 707 Map 4, which appeared on the N400 time window (367 -470ms). It is worth to mention that there 708 were no significant differences across conditions in this time window on the waveforms analyses nor 709 the tANOVA. In accordance with this tendential result, none of the maps showed significant differences 710 between conditions in the young adults group. This absence of results is not in line with those by 711 Khateb and colleagues but rather replicate those by Ruggeri et al [57], who also showed an elongation 712 of a microstate map close to the onset of the production.

713
Despite similar processes across conditions in children, as already observed in young adults, it remains 714 to be discussed if children are recruiting the same brain networks as young adults and if the timing of 715 the shared microstates is different. When comparing the topographic maps in children and young adult 716 groups, it was observed that Map 3 (263 -362ms) was specific to children. This Map 3 appears 717 following the N170 component, which seems common across groups (Map 2). However, the time 718 period of Map 3 is not yielding the interference effect since the map was not significantly different 719 between the congruent and incongruent conditions. Seemingly, the process involved in the P300 720 component is still immature in children. Here, it may be related to reading [87] or to lexical selection 721 [88] processes, or both. The spatial localization of signal in this time window showed an activation in 722 the parieto-occipital regions, and especially in the right hemisphere, which have been related to the 723 processing of a word compared to a fixation cross [89,90], as well as words reading compared non-724 words [90], rather suggesting that Map 3 represents word reading processes in children. Alternatively, 725 a change in the same temporality has been proposed in the aging literature, stating that the P300 was 726 modulated between young and older adults [12,17,48]. As described below, the authors attributed this 727 modulation to a preparation component related to the attentional processes aiming at facilitating the 728 processing of the conflict.

729
In other words, on the stimulus-aligned signal, children and adults showed globally the same maps 730 except for an additional one in an early time window. Now, regarding the duration of each of these 731 common processes, Map 5 was delayed in children (475 -520ms) compared to young adults (265 -732 477ms). This difference is compatible with the results described above suggesting the involvement of 733 an additional map in the children's group. The delay is probably the consequence of this additional 734 activation (see Figure 4). Interestingly, this map is also the one present in the time window of the N400 735 component for the young adults, supporting the hypothesis that children show immature specific brain 736 networks involved in conflict detection instead of a lack of maturation of the attentional network, 737 slowing down each process. 738

739
The results on the response-aligned signal are much clearer than on the stimulus-aligned data. First, 740 once again, the results suggest that the brain processes involved within groups in the congruent and 741 incongruent conditions are identical and only differ in duration between the two conditions. Indeed, 742 for the children's group a Stroop effect was highlighted on Map 8, which appeared in the signal close 743 to the response onset (-279 --100ms) and a similar finding was reported in young adults on Map 9 (-744 224 --100ms). In other words, both groups showed a Stroop effect globally at the same temporality 745 but on different topographic maps. According to the source localization analysis, children show a much 746 more reliable activation of the ACC compared to the other two groups who tend to use more strongly 747 their left temporal lobe to process Stroop items. Nevertheless, it appears that the maximum activation 748 for young adults is located in the ACC for the incongruent condition (as for children), but in the 749 temporal lobe for the congruent condition despite the absence of significant differences in the 750 topographies. This discrepancy might be explained by the fact that during this time window, left-751 temporal structures as well as the ACC were activated but only the maximum activation switched from 752 one region to the other. This finding implies that the conflict resolution processes in children are not 753 yet identical to those of young adults (different brain networks involved) but since the timing of the 754 maps that encompass the interference effect is comparable, these mechanisms are probably the 755 immature version of the adult ones and no additional process is needed to resolve the conflict in 756 children. This interpretation is also in line with developmental studies showing that the SP600 is still 757 maturing after the age of 12 [33,34], since different brain networks were involved in this time period.

787
The last point to be discussed is the difference in involvement of the brain networks among the 788 different maps between young and older adults. The results showed that globally the same 789 topographies were involved in both adult groups, except for Map 5 (young adults: 265 -477ms; older 790 adults: 388 -520ms) which tended to be more present in the young adults group. These results could 791 be explained by two different hypotheses and have to be related with the extreme duration of Map 1 792 in the older adults group. It is possible that, the duration of Map 1 pushed Map 5 outside the temporal 793 window of interest. An alternative explanation would be that Map 5 was suppressed to compensate 794 the wide duration of Map 1. In other words, the processes involved in older adults take longer, at least 795 for the conflict detection processes without the involvement of compensatory mechanisms, which 796 supports the general slowing hypothesis. To claim that this hypothesis explains best the results, we 797 need to confirm that microstates on the response-aligned signal follow the same trend.

799
In the response-aligned signal the data also showed the same microstates involved in congruent and 800 incongruent conditions. Even though the effect reached significance in the young adults group while 801 being marginally significant in the older adults group, Map 9 (young adults C: -204 --104; young adults 802 I: -252 --104; older adults C: -296 --104; older adults I: -224 --104), observed around 150ms before 803 the vocal onset, seems to play a role in the response-based conflict processing of both adult groups. 804 The weaker effect in older adults compared to young adults might be attributed to a higher cognitive 805 load level already for congruent items, hence reducing the difference between congruent and 806 incongruent trials. This could imply that simply dealing with two stimuli dimensions (color font and 807 color word) is interferent without needing the two dimensions to be contradictory. When considering 808 results of the source localization analysis, it appears that older adults activate strongly the left 809 temporal areas to process congruent and incongruent items, while young adults seem to rely more on 810 the frontal regions, despite similar topographies observed in the two groups for this time window. 811 Here again, the same brain network is activated but only the maximum activation changes from one 812 condition to another.

813
To conclude on aging, both stimulus-and response-aligned signals showed that the two adult groups 814 displayed globally similar maps however differing in duration, which fits best the general slowing 815 hypothesis. As a consequence, the less efficient performances observed in aging are rather to be 816 attributed to a decline in attentional resources than in a specific decline in executive functions. 817 818 To sum up, the present results led us to conclude that during development, the brain mechanisms 819

Integration of results on development and aging
involved are different than those observed in young adults. An additional network is required to 820 process conflict detection and the other mechanisms seem to be involved differently. In aging, the 821 data told a different story, favoring the general slowing hypothesis. Even though the two groups show 822 radically different patterns, some elements can be relevant for both maturation and decline. On the 823 response-aligned signal, all three groups showed an interference effect (incongruent vs. congruent 824 condition) on the last brain network activated before the onset of production. This microstates 825 however underlies different sources for the three groups. It appears that children rely mostly on the 826 ACC to solve the conflict and even process congruent items, while young adults activate strongly this 827 structure only in incongruent situations along with older adults using mostly their left temporal lobe. 828 When considering the three groups as a continuum, it might be possible that with age, the interference 829 processing uses more semantic features than executive functions, as suggested by Spreng and Turner 830 [91]. They hypothesized that since at first, children can rely only on fluid intelligence and since in aging, 831 adults tend to show a higher crystallized intelligence but lower fluid intelligence abilities, architectural 832 changes along the lifespan experience a transition from fluid to crystallized intelligence engaged to 833 perform cognitive tasks. This theory could explain the changes observed in the source localization 834 analysis, however interference effect (incongruent vs. congruent trials) within each age group remains 835 to be discussed. The interference effect present in all age groups in the response-aligned ERPs can be 836 discussed in the framework of two different cognitive models, namely general models of language 837 production [88,92] as well as the horse race model [93]. In neuropsycholinguistic models, the time 838 window close to production has been associated with the phonetic encoding or motor speech 839 planning, i.e. the preparation of the articulatory scores, encoding the motor pattern to produce the 840 response or with self-monitoring processes [94]. From this point of view, the Stroop effect is 841 compatible with a response-based conflict model but the linguistic stage impacted by this interference 842 effect is rather speech encoding than language. According to these same models, this timing of 843 activation might be related to the self-monitoring of the response [95,96]. Since models on language 844 production and of the Stroop task evolved in parallel with very few interactions, the monitoring 845 processes engaged close to the onset of the production and described in the general model on 846 language production could reflect the same mechanisms as those involved in the conflict resolution of 847 the Stroop. This interpretation is also in line with the response exclusion hypothesis [97,98]. According 848 to the authors, the closer the distractor from the target semantically, the smaller the interference 849 effect. These results were interpreted as an argument against the concept of competition in the lexical 850 selection process, implying that the speaker suppresses the distractor (thus favoring the production of 851 the target) during the post-lexical pre-articulatory stages of the word production model. The horse 852 race model of Dunbar and MacLeod (1984) is now often used to describe the history of the cognitive 853 models explaining the Stroop effect rather than to build new theoretical concepts on it. This model 854 describes the cognitive processes involved in the Stroop task as two different processes, specifically 855 color naming and word reading, triggered at the same time and competing only when reaching the 856 response production stage. Nowadays, it has been shown that both processes interact more than 857 expected by the model in early processing stages, although this model describes our data quite well.

858
Even though children and older adults show an interference effect, it does not seem to be that robust 859 compared to response conflict. Moreover, the data suggest a strong and reliable response conflict 860 which corroborates the predictions made by this model.

861
Even though the present results bring some clarification on the evolution of the Stroop effect, several  862 issues remain open. Undeniably, it is still difficult to attribute a precise process to each of the brain 863 networks highlighted by the microstates analyses. Future studies might need to replicate the result 864 and compare tasks with different involvement of attentional abilities and of cognitive control. It seems 865 also important to better understand the difference between verbal and manual tasks and how it 866 relates to interference. Since differences were observed on the last stages of verbal responses, the 867 divergent results with manual ones could be exploited in a direct comparison to clarify the underlying 868 mental processes. Finally, some unexpected results were observed such as the absence of any Stroop 869 effect on the N400 time window for the young adults group. Future studies need to clarify the reasons 870 why this effect might not be as robust as previously reported, but in the meanwhile, a few of these 871 reasons will be described in the following section.

873
The ERP results on waveforms using a mass univariate test showed that the N400 effect was not robust 874 since it was only observed in the older adults group. When examining the waveforms presented in 875 Figure 2 and given the results of the peak analysis performed on FPz, Fz, Cz, Pz, Oz electrodes, young 876 and older adults show a significantly more negative deflection in the 300-420ms post-stimulus time 877 window. Regarding children, a slightly more negative peak is observed around 420ms for incongruency 878 on both electrodes even though later temporalities show a more negative trend for congruent than 879 incongruent condition. The absence of a robust N400 effect may be due to the presence of neutral 880 trials in the present study. Even though previous studies including neutral trials replicated the N400 881 Stroop effect [14,20,21], these studies usually used rows of "X" and manual responses. By varying the 882 type of neutral trials, hence increasing the stimuli set size as well as the magnitude of the interference, 883 the effect might have been reduced [99]. This, added to the strong correction applied by the mass 884 univariate statistics might explain the disappearance of the effect. Moreover, for young adults, the 885 tANOVA results tended to show a significant effect of condition on the time window of the N400 and 886 children showed an interference effect on Map 4, present in the N400 time window. Another 887 explanation might be related to the structure of the task. Since initially this task was designed to 888 investigate the conflict adaptation effect [40], up to three repetitions of the same condition were 889 allowed. Subjects might have benefit from this feature to anticipate the presence of a conflict in the 890 next item. This point is supported by the results of Larson and colleagues [35] who analyzed the 891 evolution of the conflict adaptation effect in development and did not report any significant N400 892 effect, but only a SP600 effect. Other pieces of evidence such as a small but significant effect in the 893 mass univariate around stimulus presentation in the children signal or an early N400 effect in young 894 adults bring some conviction to this interpretation.

895
Despite some discrepancies in the results relative to classical approaches, here this data-driven 896 method allowed to describe more precisely the data, as stated below. Typically, ERP studies 897 investigating the Stroop effect analyze a time window of about 1000ms following the stimulus onset, 898 while subjects start responding on average around 682ms in our study (earlier with manual responses). 899 For this reason, we preferred using two alignment points, one at the stimulus and the second at the 900 response onset. This method allowed to overcome the response artifact on the signal, and its shift 901 across conditions, which can contribute to the effect. Indeed, even though the SP600 effect was found 902 when participants were asked to respond covertly [18], if the response is executed earlier in the 903 congruent compared to the incongruent condition, the difference observed on the SP600 could be 904 partly attributable to a shift in the motor activity related to the response. Nevertheless, aligning ERPs 905 to the response has the inconvenience that the results cannot be precisely related to a SP600, but 906 since the component is observed close to the response, it is very likely that a time window close to the 907 response will correspond to the SP600 component. Interestingly, the mass univariate test did not 908 highlight one single effect but rather three different time windows allowing to decompose the 909 complexity of the effect.

911
The present study aimed at understanding if longer latencies in children or older adults compared to 912 young adults could be attributable to the suboptimal functioning of a specific process, the addition of 913 a compensatory process or a slowdown of all the involved processes. As discussed above, the results 914 in children showed that the brain networks involved were substantially different from those engaged 915 by the young adults group on both conflict detection and resolution. Children showed an additional 916 microstate around 350ms and a Stroop effect carried by the brain network activated in the N400 time 917 window, showing a more effortful conflict detection process for this group. Regarding conflict 918 resolution, children tended to engage different microstates and an interference effect was again 919 carried by a network activated right before the response onset. These results were interpreted as 920 favoring the idea that the longer latencies compared to young adults could not be interpreted as a 921 slowing of the same mechanisms, but rather the immaturity of specific processes related to conflict 922 detection and resolution. Regarding aging, the results led to the opposite conclusion. Both stimulus-923 and response-aligned signals involve globally the same networks. However, they all seemed to be 924 slowed down compared to young adults. These results favor the general slowing hypothesis, stating 925 that the decrease in performance for an executive function task is not related to a specific decline of 926 executive functions but rather a decrease in the attentional resources availability. It is noteworthy to 927 mention that the results also highlighted that the N400 effect known to be characteristic of the 928 detection of the Stroop conflict is not as robust as described. By using a classical peak analysis the 929 results of princeps studies were replicated but vanished when using a more data-driven method. The 930 SP600 component was however more robust but appeared to be more complex than described so far. 931 Indeed, young adults and children showed a similar time window of the effect, while children 932 presented a second time window of significance closer to the production of the response. Regarding 933 older adults, conflict resolution effect was shown only later, in the last 200ms before production onset.