Peak selection and latency jitter correction in developmental event-related potentials

Abstract

Event-related potentials (ERPs) provide great insight into neural responses, yet developmental ERP work is plagued with inconsistent approaches to identifying and quantifying component latency. In this analytical review, we describe popular conventions for the selection of time windows for ERP analysis and assert that a data-driven strategy should be applied to the identification of component latency within individual participants’ data. This may overcome weaknesses of more general approaches to peak selection; however, it does not account for trial-by-trial variability within a participant. This issue, known as ERP latency jitter, may blur the average ERP, misleading the interpretation of neural mechanisms. Recently, the ReSync MATLAB toolbox has been made available for correction of latency jitter. Although not created specifically for pediatric ERP data, this approach can be adapted for developmental researchers. We have demonstrated the use of the ReSync toolbox with individual infant and child datasets to illustrate its utility. Details about our peak detection script and the ReSync toolbox are provided. The adoption of data processing procedures that allow for accurate, study-specific component selection and reduce trial-by-trial asynchrony strengthens developmental ERP research by decreasing noise included in ERP analyses and improving the representation of the neural response.

1 |. INTRODUCTION

Event-related potentials (ERPs) are popular in infant and child neuroscience research for their insight into the timing and nature of neural responses to discrete stimuli. ERPs can be utilized in the investigation of cognitive maturation and are ideal for the examination of neural responses in pediatric populations, based on their excellent temporal resolution, noninvasive application, and minimal task demands. Unfortunately, developmental ERP work is plagued with inconsistent methodological approaches, which contribute to variable results. This is especially true for methods adopted in the identification and quantification of ERP component latency. In the current analytical review, we assert that a data-driven analysis strategy should be applied to the selection of component latency within individual participants’ data to provide an accurate picture of one’s specific dataset and to contribute to the production of more reliable results. We will begin by highlighting three methods for determining appropriate time windows, including (1) a priori time window selection based on previous research, (2) time window selection based on visual review of grand average data, and (3) automated peak latency identification for selection of time windows in individuals’ grand average data. Variable latency ERP responses may be observed across and within individuals’ data, which is known as latency jitter. We will discuss the impact of latency jitter on appropriate time window selection and introduce a possible solution, the ReSync toolbox (Ouyang, 2020). A summary of these three approaches and the ReSync procedure, including advantages, disadvantages, and recommendations for each approach, is included in Table 1. Finally, we will demonstrate the use of the ReSync toolbox to illustrate its utility in the processing of developmental ERP data featuring the P1, N290, and P400 ERP components.

TABLE 1.

Summary of time window selection methods for ERP analysis

	Advantages	Disadvantages	Recommendations
A priori time window selection based on previous literature	Considered the gold standard approach in the adult literature Guided by previous research Not biased by the researchers’ perception of the data	Component latency changes across infancy and childhood and may be poorly defined ERP components may be misidentified, if latency is not clearly defined	Appropriate for use with well-defined components not greatly impacted by latency jitter Unique time windows should be used to compare different age groups in infancy and childhood
Time window selection based on visual inspection	Data-driven approach Ensures that the component of interest is captured in the analyses Can be utilized with novel or ill-defined components	May result in inappropriate selection of data if safeguards are not employed	Appropriate for use with novel or ill-defined components not greatly impacted by latency jitter and when trial numbers are equal across conditions Time windows should be selected by an unbiased experimenter using the ERP grand average
Semiautomated and individualized time window selection	Data-driven approach May prevent researcher bias Ensures that the component of interest is captured in the analyses Addresses latency variability across participants	May require monitoring to ensure appropriate data is identified Does not account for trial-by-trial variability within participants	Appropriate for components with distinct peak that are not greatly impacted by latency jitter
ReSync latency jitter correction	Data-driven approach May prevent researcher bias Addresses latency variability across and within individual participants Recreates ERP with latency jitter correction	Not appropriate for use with all components Time windows must be selected by researchers	Appropriate for components with distinct peak that are impacted by latency jitter Should be combined with semiautomated approaches to define ERP time windows Appropriate for use with individual participants’ data

1.1 |. Time window selection based on past research

Popular practices for ERP component time window selection include a priori selection informed by previous research studies and post hoc selection based on visual review of average waveforms. A priori time window selection is generally regarded as the gold standard approach and has been adopted in several infant and child ERP studies. We conducted a brief review of recent developmental ERP research (i.e., pediatric ERP studies published in Developmental Psychobiology and Developmental Cognitive Neuroscience journals in 2020) and found that a priori time selection was adopted in five of the 21 studies identified. Table 2 presents the studies examined and section (a) features studies that utilized an a priori approach to time window selection.

TABLE 2.

ERP studies published in developmental psychobiology and developmental cognitive neuroscience journals in 2020

	Study	Component(s) examined	Age range	Method for time window selection
(a)	Bayet et al. (2020)	N290, Nc	12- to 15-month-olds; adults	Past literature (Balas & Koldewyn, 2013; de Haan, 2013)
	Canada et al. (2020)	Nc, late slow wave (LSW)	4- to 8-year-olds	Past literature (Geng et al., 2018; Marshall et al., 2002)
	Di Lorenzo et al. (2020)	P1, N290, P400, Nc	5-month-olds	Past literature
	Jessen (2020)	Nc	7-month-olds	Past literature (Jessen & Grossmann, 2014; 2016; 2019)
	Zaadnoordijk et al. (2020)	Mismatch negativity (MMN)	3- to 4.5-month-olds	Past literature (Basirat et al., 2014; Trainor et al., 2003)
(b)	Arslan et al. (2020)	Frontal and parietal central areas within a time-window of 200–600 ms	10- and 14-month-olds	Past literature (Parise et al., 2010) and visual inspection
	Deveney et al. (2020)	N1a, N1p, P1, P2	48.2- to 85.2-month-olds	Past literature (Eldar et al., 2010; Mueller et al., 2009; Perez-Edgar et al., 2006; Rossignol et al., 2013) and visual inspection
	Forgács et al. (2020)	N400	14-month-olds	Past literature and visual inspection
	Shen et al. (2020)	Somatosensory MMN, late discriminative negativity (LDN)	6- and 7-month-olds	Past literature (Shen et al., 2018) and visual inspection
(c)	Adibpour et al. (2020)	P2	5- to 21.4-week-olds and 6.7- to 28.7-week-olds	Visual inspection
	Chevalier et al. (2020)	Posterior positivity (PP), contingent negative variation (CNV), pre-response negativity (PRN)	6- and 9-year-olds; adults	Visual inspection
	Grenier et al. (2020)	P1, P2, N1, N400	8- to 12-year-olds; adults	Visual inspection
	Jin et al. (2020)	Initial posterior peak, positive-going slow wave	5- to 6-year-olds; 9- to 10-year-olds; adults	Visual inspection
	Kailaheimo-Lönnqvist et al. (2020)	P1, N2, mismatch response (MMR)	6-month-olds	Visual inspection
	Marin et al. (2020)	P1, N700, late frontal positivity (LFP)	3-month-olds	Visual inspection
	Safar and Moulson (2020)	N290, P400, Nc	3-month-olds	Visual inspection
(d)	Bonmassar et al. (2020)	P2, early and late P3a, LDN	7- to 10-year-olds; adults	Principal component analysis (PCA)
	Chong et al. (2020)	Error-related negativity (ERN), correct-related negativity (CRN)	10- to 17-year-olds; adults	Automated peak selection
	Cosper et al. (2020)	N400-like effect	10- to 12-month-olds	Examination of 100 ms intervals
	Nayak et al. (2020)	Lateralized readiness potentials (LRP)	6- to 8-year-olds	Examination of 150 ms intervals
	Steber and Rossi (2020)	N400	18-month-olds	Examination of 50 ms intervals

Note: The table is divided into four sections based on method used to select time windows for ERP component analyses: (a) studies that utilized past literature to select time windows a priori; (b) studies using past literature and visual review of the current data; (c) studies using only visual review of the current data; and (d) studies utilizing more automated, data-driven approaches.

Although a priori time window selection based on past literature has been recommended in much of the adult literature (e.g., Luck & Gaspelin, 2017; Picton et al., 2000), it may be problematic for use in developmental ERP research. Latencies of ERP components change as the brain matures, especially in infants and young children. For example, Brooker and colleagues (2019) report that latencies of both the error-related negativity and P3 ERP components decrease with development, making a priori time window selection more difficult. Developmental changes in ERP component latency reflect increasing neural efficiency and are likely to be observed across other components, as well. Indeed, it is recommended that researchers consider utilizing unique time windows for comparison of different infant age groups because of maturation-related changes in component latency (Hoehl & Wahl, 2012). This approach has not been widely adopted; however, these changes are at the root of guidelines to tightly restrict participant age ranges in the first years of life. Recommendations include adoption of 1-month age windows for infants and children less than 24 months and 1-year age ranges in children under 8 years of age (Picton et al., 2000; Taylor & Baldweg, 2002). Developmental changes in the ERP may include decreasing component latency and increasing or decreasing component amplitude; these changes may mute results if developmental variability is not considered in analyses (Taylor & Baldeweg, 2002).

1.2 |. Time window selection based on visual inspection

A priori methods are recommended when possible; however, the adoption of alternative techniques is endorsed when a novel ERP paradigm is employed (e.g., Luck & Gaspelin, 2017). Much of the ERP literature is newer in pediatric than adult populations, which increases study novelty. A priori time window selection may lead to undesirable variability in results across studies unless the component is very well understood, and its latency is very clearly defined in the age group of interest. A popular solution is to select ERP component time windows based on visual review of grand average waveforms, often in conjunction with information from previous research. As shown in Table 2, visual inspection for time window selection was included in 11 of the 21 studies reviewed. Seven studies, presented in Table 2, section (c), utilized visual inspection of grand average data solely to determine appropriate time windows for ERP analysis. An additional four studies, presented in Table 2, section b, utilized visual inspection of grand average data combined with guidance from past research. If a visual inspection approach is adopted, it is strongly recommended that a grand average including all data collected and an unbiased experimenter, blind to the study’s conditions or hypotheses, identify time windows for component analysis (Luck, 2014; Luck & Gaspelin, 2017). When these safeguards are not employed, this approach has been associated with inappropriate rejections of the null hypothesis in adult research (Luck & Gaspelin, 2017). This may be because an informed researcher is more likely to unintentionally direct their analyses toward time windows characterized by unique responses across study variables. Additionally, it is important to consider the number of participants and trials submitted to the grand average across participant groups and stimulus conditions. An imbalance in the number of participants belonging to each group or the number of trials completed for each stimulus condition could bias the grand average data (Keil et al., 2014).

1.3 |. Semiautomated and individualized time window selection

Well-justified conceptual and data-driven methods for ERP latency selection can help to prevent bias and increase power (Keil et al., 2014). There are a variety of automated and semiautomated data-driven approaches to ERP component time window selection, as demonstrated in Table 2, section (d). It can be seen that these approaches were utilized in five of the 21 recent studies identified and that specific methods for time window selection varied across each study. In our infant ERP work (Conte et al., 2020; Guy et al., 2016, 2018; Xie & Richards, 2016, 2017), we have applied a semiautomated procedure to define peak component latency in individual ERP averages. This type of unbiased, algorithm-based approach has been recommended when a priori time window selection is not feasible (Picton et al., 2000). Additionally, individual differences in neural responses can contribute to variability in results observed within an experimental group. Unique trajectories in neural development may lead to different ERP component latencies in children belonging to the same age group (Brooker et al., 2019). We have created a custom MATLAB script to identify the point of greatest amplitude for each stimulus condition at specific electrodes within a predetermined time window. Each peak is inspected after detection and misidentified peaks are adjusted. Unique latencies can be selected for each stimulus condition within each participant (i.e., reviewers are not viewing the conditions in the same plot or averaged across participants), which may reduce bias. Peak amplitude may be identified and analyzed following identification of the peak latency. This could be defined as the amplitude value at the peak latency or as a measurement of mean around the peak (e.g., mean amplitude ± 20 ms of the peak latency).

Much of our recent work has focused on the early development of specialized face processing (e.g., Conte et al., 2020; Guy et al., 2016, 2018). In this line of research, we have examined developmental change in the ERP in response to face and nonface stimuli, focusing on the P1, N290/N170, P400, and Nc components. The P1 is a positive peak occurring approximately 100 ms after stimulus onset, the N290 is a negative peak at approximately 290 ms after stimulus onset, and the P400 is a positive peak occurring about 400 ms after stimulus onset (e.g., Conte et al., 2020). These components are observed at medial and lateral posterior electrode sites. Amplitude and latency of the P1 are believed to reflect low-level processing of visual stimuli. The N290 is believed to be the precursor to the adult and child N170 and, like the N170, is greater in amplitude to face than nonface visual stimuli (Conte et al., 2020; Di Lorenzo et al., 2020; Guy et al., 2018). Although the P400 has been investigated in several studies of visual information processing, the functional role of the P400 is still unclear. Discrimination of face and nonface stimuli has been observed in a small number of studies based on latency (de Haan & Nelson, 1999) and amplitude (Guy et al., 2016). The Negative central (Nc) is a broad negative deflection in the ERP that occurs at frontal and central midline electrodes between 300 and 800 ms after stimulus onset (Courchesne et al., 1981; de Haan et al., 2003; Reynolds & Richards, 2005, 2009). The Nc is not face specific, but is believed to reflect attentional engagement, as it is often greater in amplitude to novel and salient stimuli (e.g., Carver et al., 2003; de Haan & Nelson, 1999; Guy et al., 2013; Reynolds & Richards, 2005) and during heart rate-defined periods of attention (Guy et al., 2016; Reynolds et al., 2010; Reynolds & Richards, 2005, 2009; Richards, 2003).

Developmental changes and individual differences in amplitude and/or latency have been observed in all of these components, highlighting the need for more individualized analysis strategies. In a recent study examining development of cortical face specialization from 4.5 to 12 months of age, we report increasing N290, P400, and Nc amplitude with age (Conte et al., 2020). Additionally, Di Lorenzo and colleagues (2020) tested participants longitudinally at 5 and 10 months of age and found significant differences in N290 latency across sessions. They most frequently observed earlier N290 responses at 5 months compared with 10 months; however, they regularly observed the opposite pattern, as well. These findings support the use of an individualized approach to ERP component time window selection.

Automated peak selection within an individual participant’s ERP is most appropriate for components with a distinct and visible peak, such as the P1 and N290. Mid-latency components, including the P400 and Nc, are characterized by broad peaks, rendering peak-based adjustment more difficult. The broader peak is likely due to the increased presence of slow wave activity during this time window (DeBoer et al., 2007). We do not apply our peak identification program to the Nc component for this reason. However, we do attempt to identify a peak P400. This is because the P400 includes an “early” peak that occurs approximately 400–500 ms after stimulus onset, and although increased activation is often prolonged beyond this peak, we believe that the early-occurring, distinct positive peak is representative of the P400, and that the continued positivity beyond this peak reflects neural activation more closely associated with the broad Nc component.

1.4 |. Latency jitter and the ReSync procedure

Individual identification of peak amplitude latency has helped to overcome weaknesses of a “one-size-fits-all” approach to peak selection; however, it does not account for trial-by-trial variability within a participant. This issue, known as ERP latency jitter, may blur the average ERP, misleading the interpretation of neural mechanisms and reducing peak amplitude. All ERP components are not identically time-locked with stimulus onset, increasing peak latency variability even within a participant. Latency jitter may occur more frequently when the component is produced as the result of a cognitive process that occurs with variable timing (Picton et al., 2000). Furthermore, time window selection for analysis of mean amplitude may require special consideration when latencies are variable between different experimental conditions or participant groups.

A further step in the data preprocessing pipeline can be taken to estimate and correct latency jitter. Woody (1967) first identified and approached this problem by estimating latencies on a trial-by-trial basis and applying a cross-correlational technique to create a reconstructed ERP. In the reconstructed ERP, the component of interest is realigned based on identified latencies rather than stimulus onset. This method is rarely applied in developmental ERP research, despite the presence of ERP component latency variability in infant data (Hoehl & Wahl, 2012). In part, this may be because there are weaknesses to the use of the “Woody filter.” First, the ERP cannot be viewed as a single component, because it contains multiple independent components, and doing so may weaken the value of realigning the entire ERP segment. Second, approaches to resynchronizing ERP data may be outside of the technical skillset of many ERP researchers. To address these concerns, the ReSync MATLAB toolbox (Ouyang, 2020) has been designed as an accessible solution made freely available for correction of ERP component latency jitter across individual trials.

The ReSync procedure is applied to a single participant’s data to determine whether an ERP component is impacted by latency jitter, to correct for jitter, and to create a new average including jitter-corrected trials (Ouyang, 2020). Researchers first identify a component that they believe may be influenced by latency jitter. They then select electrodes and a time window that include the component and run the ReSync program, which calculates estimated component latencies for individual trials and creates a visual representation of the trials in order of increasing component latency. One single electrode, multiple electrodes, or all electrodes may be selected. Data are averaged across the selected electrodes when multiple electrodes are selected. The decomposition-and-reconstruction method is used to isolate the component of interest, realign it across trials based on peak latency, and to create a new individual average. The ERP is then recreated with the jitter-corrected component and plotted in comparison to the original ERP. The jitter-corrected component is integrated into the ERP in a manner that prevents distortion of other components (i.e., components that were not ReSynced should not be changed or jitter-corrected through the ReSyncing of another component). The procedure is carried out separately for each ERP component of interest, allowing researchers to determine whether latency jitter is a concern in one component but not others (e.g., jitter may be found in the infant N290, but not the P1), so that only component-specific jitter is corrected.

Not all components are equally benefitted from the use of ReSync. For example, a low signal-to-noise ratio may interfere with appropriate realignment due to arbitrary latency identification (Ouyang, 2020). This issue may be reduced somewhat in infants due to their thinner skull relative to adults, which is associated with a stronger signal. However, infant and child ERP data are frequently impacted by artifact originating from participant movement. The component’s characteristics also influence its appropriateness for the ReSync program. A component that demonstrates variable latency and a sharp peak is more appropriate for ReSyncing than a blunt component equal in latency variability. Ouyang (2020) recommends continued data analysis with the standard average ERP in these instances and when ERP component latency is not very variable. Decisions about the execution of the ReSync program can be made at the individual level. For example, if one participant contributes suitable data, components influenced by latency variability could be ReSynced, whereas if they contribute a minimum amount of data or have a low signal-to-noise ratio, their data may be averaged following traditional methods. Ouyang (2020) created a measure of “relative latency variability” (RLV) to assist in decisions about the appropriateness of data for ReSyncing. When the RLV or the signal-to-noise ratio is low, it is recommended that the standard average ERP is used.

1.5 |. Current demonstration

Developmental ERP researchers have the capability of decreasing noise included in ERP analyses and improving the representation of the neural response by adopting data processing procedures that allow for accurate, study-specific peak selection and reduce trial-by-trial asynchrony. In the current analytical review, we illustrate the use of the ReSync program with data collected from one infant and one child participant during a single ERP experiment. Individual participants were selected for this demonstration, as the ReSync program is intended to be applied to individual participant’s data. We also present data from a group of 6-month-old infants to depict how use of the ReSync procedure may influence group averages. Although the ReSync toolbox was not created specifically for developmental ERP data, this approach can be adapted for developmental ERP researchers. We used the ReSync procedure to realign the P1, N290/N170, and P400 components within each participant. We hypothesized that greater variability in latency would be observed for the N290/N170 and P400 than for the P1. The P1 was expected to demonstrate less latency variability as it is typically observed within a narrow time window and likely reflects automatic sensory processing. Alternatively, the N290/N170 and P400 components are associated with cognitive processing and face recognition and may be observed across a broader time window (particularly the P400). Recent work providing evidence of N290 latency variability across testing sessions (Di Lorenzo et al., 2020) provides further support for the potential of the ReSync program in increasing the accuracy of measurement of this component.

2 |. METHOD

In this section, we describe our procedure for identification of the peak amplitude of the P1, N290, and P400 ERP components and application of the ReSync latency jitter correction MATLAB toolbox (Ouyang, 2020) to infant and child EEG data. A graphic representation of the pipeline is depicted in Figure 1. The pipeline has been applied to individual data recorded from one infant and one child EEG during a face-processing ERP task in which participants were presented with upright and inverted faces and houses. The EEG recordings were processed with the EEGLAB (version 14.1.1b) and ERPLAB toolboxes (Delorme & Makeig, 2004; Lopez-Calderon & Luck, 2014). Details about the paradigm, recording procedure, and preprocessing of the data can be found in our previous publications (Conte et al., 2020; Guy et al., 2016, 2018). The MATLAB code and data described in the methods and results sections and in the Supporting Information are provided at https://osf.io/4dc3p/.

FIGURE 1.

Latency peak detection and jitter correction pipeline

The components under investigation and their respective regions of interest are depicted in Figure 2. Electrodes in the occipital cluster were utilized for the C1 and P1 components; the lateral parietal cluster was the area in which the lateral P1 and N290 were more prominent; the occipital and lateral parietal cluster was utilized for the P400 component. The C1 was only identified in the current study for calculation of the peak-to-peak amplitude of the P1. It is the earliest occurring visual ERP component and is observed approximately 50–90 ms after stimulus onset at occipital and parietal electrodes. It is not shown to vary across stimulus types or with attention (e.g., Clark & Hillyard, 1996; Di Russo et al., 2003), and was not investigated in the current study.

FIGURE 2.

Infant and child prototypical face-processing ERPs and electrode clusters

Note: Top panels depict the ERP waveforms elicited during a face-processing task in an infant and a child participant. Note that the y-axis scale maximizes the data distribution in each panel. The topographic maps with electrode clusters are represented in the bottom panels. Empty dots represent the EGI 128-channel montage, whereas filled gray dots represent the location of the electrodes in the 10–10 system. Colored channels identify the electrode clusters utilized for the analyses in the current paper.

2.1 |. Participants

Sample data and results of the pipeline are provided for one 6-month-old infant (M = 6.20 months) and one 6-year-old child (M = 81.72 months). Both participants were full term (i.e., at least 38 weeks gestational age, birth weight at least 2500 g) and healthy at birth. Participants were randomly selected from larger datasets of ongoing studies based upon successful completion of the ERP study. Informed parental consent was obtained in accordance with ethics approval from the Institutional Review Board of University of South Carolina. The infant participant contributed 184 good trials (upright face: 43; inverted face: 43; upright house: 48; inverted house: 50) to the analyses, whereas the child participant contributed 258 good trials (upright face: 61; inverted face: 73; upright house: 65; inverted house: 59) to the analyses. This amount of data is representative of average performance in our research examining face processing in typically developing infants and children. The procedure was also applied to a representative group of 6-month-old infants (n = 7; M_age = 6.42 months), randomly selected from an ongoing study. On average participants had 103 good trials after preprocessing (upright face: 25; inverted face: 26; upright house: 25; inverted house: 27).

2.2 |. Pipeline description

The pipeline begins with the selection of time windows around each ERP component under investigation within an individual participant’s data, which will be used for the latency jitter correction. Our PeakDetection function automatically identifies the latency of peak values for each ERP component and electrode of interest and plots the median latency values on ERP line graphs for visual inspection. The participant’s ERP dataset is loaded into ERPLAB and the function pop_geterpvalues is utilized to measure the latency to the peak of the components of interest. An initial latency identification is performed on the grand average ERP. The PeakDetection procedure starts with an “anchor” component (e.g., P1), which is defined within an a priori user-defined time window and subset of electrodes (e.g., occipital cluster; see Figure 2). Therefore, we recommend that a relatively stable component with a visible peak be used as the “anchor.” ERP components reflecting early sensory processing appropriately serve this purpose. A median latency value across electrodes of the cluster of interest is calculated and utilized as a reference for the time window identification of the remaining components. We identified the C1 peak between the stimulus onset and the median P1 latency, the N290 peak between the median P1 latency and 400 ms, and the P400 peak between the median N290 latency and 900 ms. The same strategy was utilized to detect the peak latency value for all of the experimental conditions. Detected latency values for each electrode and component are saved in a separate file per each participant (see Supporting Information for details) and the median latency values are plotted on the ERP line graphs. We recommend visual inspection of the ERP plots to evaluate the accuracy of the peak selection procedure. Misidentified peaks can be reported in the LatencyAdjustemt.txt file before performing the final peak detection (see Supporting Information for details).

Participant-specific peak latencies were utilized to identify the time window(s) in which the latency jitter correction was performed. In our LatencyJitterCorrection function, a range around the median peak latency of a given component is calculated from the latency values identified in the previous step. Latency jitter correction is estimated and corrected in the original EEG dataset using the ReSync toolbox (Ouyang, 2020). The ReSync toolbox was designed to use continuous, nonepoched datasets. A critical preprocessing step in visual ERP studies with pediatric populations consists of removing epochs in which the participant was not attending to the screen. This step occurs early in the preprocessing pipeline and requires segmentation of the continuous recording, so that unattended stimuli can be eliminated from further processing. Use of the ReSync toolbox with segmented datasets produces artifactual amplitudes at each epoch’s boundary. Therefore, we zero-padded each epoch before initiating the ReSync toolbox, and removed all paddings after completing the latency jitter correction. The procedure allowed us to automatically define a time window that was centered to the peak of each component. The automatically generated time window can be adjusted by the user if necessary (see details in Supporting Information). Estimated latency windows and amplitude values computed in the ReSync.m function are plotted for each participant (Figure S4). These values could be saved and separately analyzed. We reported the average latency of peak across trials for each component and participant in Table 3. Single-trial latency values are plotted in Figure 3. A jitter-corrected ERP is saved for each component under investigation at the end of the procedure. At this point, the latency peak detection can be performed anew on the jitter-corrected datasets (LJC_PeakDetection). The final peak latency values are saved in a separate file per each participant and can be utilized for further analyses.

TABLE 3.

Peak latency values (ms) across all trials per each component and participant

	P1		N290/170		P400
	M	SD	M	SD	M	SD
Infant	96.61	11.72	266.07	40.66	373.30	67.59
Child	119.61	24.28	181.05	34.12	294.14	95.47

Note: M = mean latency of peak across experimental conditions; SD = standard deviation value across experimental conditions.

FIGURE 3.

Individual and group ERP waveforms before and after latency jitter correction

Note: Panel (a) shows both original and latency jitter-corrected ERPs across electrodes belonging to the respective component cluster. The gray boxes indicate the component-specific time window calculated using data-driven latency peak values. In panel (b), the grand average ERPs before and after latency jitter correction of the N290 component are plotted for two representative channels of the lateral parietal cluster in a group of 6-month-old infants (n = 7).

3 |. RESULTS

The average ERP from the original and jitter-corrected dataset is plotted along with the utilized time window. In Figure 4, the original and latency jitter-corrected ERP are depicted for each component and participant. An adjustment was made to the P400 time window for the infant participant to better fit the ERP component. Details on the procedure can be found in the Supporting Information. The procedure adequately created subject-specific time windows for all of the remaining components. The latency jitter correction enhanced the peak of all ERPs under investigation. Table 4 reports the average peak amplitude across the experimental conditions before and after the jitter correction.

Figure 4.

Peak latency at each trail for the three ERP components under investigation

TABLE 4.

Amplitude values (μV) before and after latency jitter correction for each ERP component

		P1		N290/170		P400
		M	SD	M	SD	M	SD
Infant	Original	7.09	3.43	−1.09	2.67	13.84	6.84
	Corrected	7.70	3.39	−1.99	2.76	13.83	6.87
Child	Original	33.58	19.09	2.25	8.15	22.74	10.58
	Corrected	37.12	20.10	0.20	7.78	23.20	11.11

Note: M = mean amplitude value across experimental conditions; SD = standard deviation value across experimental conditions.

In Figure 5, we reported the ERP waveforms in response to upright faces before and after the latency correction, along with the latency peak of each component. Most of the components showed a change in the amplitude and latency of the peak after the correction for both sample participants. Similar changes were visible after the procedure was applied to additional participants in the 6-month-old group (n = 7). Figure 4 presents the original and ReSynced ERP data in grand averages created from the original and ReSynced data of seven 6-month-olds.

FIGURE 5.

ERP response to faces and component-specific latencies of peak before and after latency jitter correction

Note: ERP line graphs in response to upright faces before (solid line) and after (dotted line) the latency jitter correction. The P1 ERP is plotted over one representative channel of the occipital cluster (i.e., Oz). The N290 and P400 ERPs are plotted over a representative channel of the lateral parietal cluster (i.e., P8). Peak latencies pre- and postcorrection are reported for each component.

Lastly, we used bootstrapping to estimate the standardized measurement error (SME) for both peak amplitude and latency on the single-trial EEG epochs before and after the latency jitter correction. The SME is a measure of data quality and quantifies the measurement error for a given score (Luck et al., 2020). Thus, small SME values are an indication of a better measurement. A total of 10,000 iterations were performed for each ERP component and experimental condition and separately for the two sample participants. SME calculations were performed on component-specific time windows and electrode clusters utilized for the jitter correction of each participant. The average SME values across conditions are reported in Table 5. A reduction in the measurement error seemed to occur for the peak latency measurements of all infant components and for the P400 child ERP. Less evident changes were obtained for the peak amplitude measure. In addition, we measured the ERP’s internal consistency by calculating Cronbach’s alphas for each component before and after the latency jitter correction. Peak amplitude and latency values within each component-specific time window were used as measures, the component-specific channels were considered as items, and the participant’s trials were utilized as observations. Excellent reliability is usually indicated by a coefficient alpha exceeding .90; high reliability is indicated by alphas between .70 and .90, moderate between .50 and .70, and low with alpha coefficients below .50. Results are reported in Table 5. Peak amplitude showed excellent internal consistency regardless of the latency jitter correction. A change in internal consistency from high to excellent characterized the latency values of the infant P1 and N290, and the child P400 component.

TABLE 5.

Amplitude and latency standardized measurement errors and Cronbach’s alphas across conditions before and after latency jitter correction

		P1		N290/170		P400
	SME	Amplitude	Latency	Amplitude	Latency	Amplitude	Latency
Infant	Original	2.61	6.50	3.37	14.34	3.35	29.71
	Corrected	2.61	5.29	3.34	12.44	3.33	28.85
Child	Original	2.40	7.79	2.33	12.29	2.95	21.66
	Corrected	2.40	7.76	2.34	12.31	2.95	18.33
	α	Amplitude	Latency	Amplitude	Latency	Amplitude	Latency
Infant	Original	0.96	0.84	0.95	0.80	0.97	0.92
	Corrected	0.96	0.91	0.95	0.90	0.97	0.93
Child	Original	0.95	0.86	0.94	0.84	0.97	0.87
	Corrected	0.95	0.86	0.94	0.84	0.97	0.91

A similar analysis was performed for the group of 6-month-old infants. Cronbach’s alpha coefficients were calculated for the upright face condition using participants as observations and repeated with varying numbers of trials included in the ERP (i.e., 5,10,15, and 20 trials). Figure 6 shows the changes of alpha coefficients before and after latency jitter correction for different trial counts. Overall, the group coefficients indicated high or excellent reliability. The latency jitter correction slightly improved the ERP reliability for the N290 amplitude and latency and the P400 latency.

FIGURE 6.

Cronbach’s alpha coefficients as a function of trial count for the three ERP components before and after latency jitter correction

Note: Y-axis scales maximize the data distribution in each panel.

4 |. DISCUSSION

Thoughtful and appropriate selection of time windows for ERP component analysis is necessary to informative research. In this paper, we have described variety of methods for time window selection, all possessing qualities appropriate for use in some research contexts. The first approach described, that is, a priori time window selection based on a review of the past literature, may be ideal for research examining well-understood and defined ERP components that are not impacted by high levels of latency jitter. The second approach, that is, visual inspection of grand average data, may be valid for the study of a novel or ill-defined component that does not appear to be greatly impacted by latency jitter and when trial numbers are equivalent across experimental conditions. The third approach, that is, a semi-automatic approach utilizing algorithms to identify ERP components’ peaks in individual participants’ data, is useful in many contexts and ideal when latency variability may be present across participants. This approach may be combined with the ReSync procedure to account for variability both across and within an individual. Such approach could also be useful for an automated selection of ERP component peaks inside a single trial. In our demonstration, the ReSync toolbox proved to be effective for the alignment of components possessing latency jitter within an individual infant’s and child’s data. Jitter-corrected outputs led to stronger representation of the ERP components of interest in the averaged data when compared with uncorrected ERPs.

The ReSync toolbox was effective in the identification of component peaks within individual trials for both the infant and child participants. We used our peak detection program to estimate time windows for the P1, N290/N170, and P400 components based on individual participants’ ERP data averaged for each stimulus type. The combination of the time window identified using the peak detection program and the ReSync toolbox led to appropriate time window identification for all components except for the infant P400, which required manual adjustment. Use of the ReSync toolbox led to enhanced ERP component presentation, including more distinct peaks and greater component amplitude. This indicates that all components examined were impacted by some level of latency jitter across trials collected from a single participant. Enhanced components could be observed in grand averages created from multiple participants’ data as well. We tested whether the procedure improved our data quality by performing bootstrap analyses on corrected and uncorrected datasets to estimate the SME scores described in Luck et al. (2020). The comparison between SME values provided evidence of reduced error in the measurement of peak latency following application of the ReSync procedure. This was observed in the infant P1, N290, and P400 components and the child P400 component. Similar results were obtained from the measure of internal consistency obtained through the computation of Cronbach’s alphas. The reliability of the latency measure changed from high to excellent for the infant P1 and N290 components, and the child P400 component. Although we computed psychometric measures of data quality and ERP reliability at the individual level, it is important to verify the subject-level internal consistency in relation to the group-level consistency to assess how well the group characterized individual differences of the single participants. Ultimately, measures of data quality and between-subjects and within-subjects internal consistency should be reported in all studies to increase methodological transparency (for a detailed discussion, see Clayson, Brush, & Hajcak, 2021).

Ouyang (2020) asserted that the ReSync procedure is not appropriate for use with all EEG components. The Nc is a broad component characterized by an ill-defined peak, and so we chose to exclude it from automated peak detection and the ReSync procedure. Furthermore, our investigation indicates that the beneficial effect of ReSyncing varies based on the component under investigation and the participant’s age. For example, SME of component latency was reduced for the infant P1 and N290 following use of the ReSync procedure, but not for the child P1 and N170 data. This was somewhat surprising, as we predicted that the ReSync procedure would be increasingly effective in reducing jitter with increasing component latency across both ages (i.e., we expected that the N290/N170 would be more benefited by ReSyncing than the P1 in the infant and child data). A larger reduction in measurement error was observed at later ERP components (i.e., the P400) than earlier components at both ages. The SME values for peak amplitude remained constant across all components before and after application of the ReSync procedure. Following Ouyang’s (2020) recommendations, researchers should examine jitter within each component and each participant and thoughtfully consider whether ReSyncing is appropriate for that individual’s data. Decisions should be made for each viable component based on the presence of trial-by-trial variability across all trials that a participant completed (e.g., see Figures 4 and S4) and measures of the degree of latency variability and the signal-to-noise ratio (see Ouyang, 2020). To prevent researcher bias, decisions should not be made based on individual study conditions or grand-averaged data. If a high level of jitter is not observed within an individual’s data, Ouyang (2020) argues that further manipulation of the data may not improve, or may even overcorrect, the results. Additional examination of the change in data quality after the latency jitter correction should be performed on the entire sample of a given study to further assess whether the ReSync procedure should be applied.

Data that have undergone the ReSync procedure may contribute to more accurate measurements of component amplitude and latency. The ReSync procedure diminishes the impact of latency jitter on grand-averaged ERP component amplitudes by aligning the component at each trial. The well-defined peak that results from ReSyncing ERP data may be especially appropriate for analyses of peak amplitude and peak-to-trough measurement approaches, which consider the impact of earlier components. The P400 component, which was characterized by greater latency jitter than other components examined, may especially benefit from the ReSync procedure. Greater jitter of the P400 may have contributed to the variability in results observed in previous studies of face processing. Appropriate data preprocessing might elucidate the role of mid-latency components as neural correlates of face processing. The latency reported in the reconstructed, ReSynced ERP is not appropriate for analyses including latency as a dependent variable. Latency measures are manipulated by the ReSync procedure, as presented in Figure 5. This latency is the result of the realignment individual trials to create a reconstructed grand average. The new grand average contains less noise created by latency jitter, but may also present an artificial peak latency. However, calculation of more accurate component latency may be facilitated through use of data collected during the execution of the ReSync program. Latencies measured in each trial, as presented in Figure 4, may be averaged together to create a measure of mean latency that is less influenced by component amplitude than latencies measured from grand averaged data alone.

Our review of the recent literature of ERPs in pediatric populations revealed that a subset of recent studies utilized additional data-driven and automated approaches to investigate amplitude differences between experimental conditions. In one approach, the entire ERP segment is investigated by comparing the response to different stimuli in subsequent and short (e.g., 50 or 100 ms) time windows (Cosper et al., 2020; Nayak et al., 2020; Steber & Rossi, 2020). This approach has the advantage of selecting time windows independently from previous studies and the experimenters’ visual judgement and is likely more appropriate for broad components without a well-defined peak. The risk of performing comparisons between multiple and successive time windows in components with a sharp visible peak is that the differences between conditions might be underestimated if the single time window fails to cover the component’s peak. Therefore, the decision on the length of each time window would be critical, especially for waveforms with multiple adjacent ERP components. Other studies utilized principal component analysis (Bonmassar et al., 2020) and automated peak detection (Chong et al., 2020).

A goal of the current paper is to increase thoughtfulness surrounding peak selection and analysis in developmental ERP work. Although the convention in the field calls for reliance on previous work and/or examination of grand average data, there are weaknesses to both of these approaches. A priori ERP component time window selection can lead to misidentification of peak activity if the time windows of interest are not already well-defined. Examination of grand average data may allow for researcher bias in time window selection and does not account for individual variability in ERP responses. Although the ReSync procedure may not be appropriate for all datasets, it is a data-driven, automated approach that can address bias and individual variability in selection of the time window of analysis in developmental ERP data. An automated approach to time window selection in individual participant’s data, complemented by use of the ReSync procedure, may contribute to more accurate time window selection and more accurate measurements of component amplitude. As previously recommended by pediatric and adult ERP researchers (Brooker et al., 2019; Keil et al., 2014; Luck & Gaspelin, 2017), it is important that whatever method for ERP component latency selection is applied, researchers practice persistence in reporting and justifying their selection of ERP time windows.

DATA AVAILABILITY STATEMENT

The MATLAB code and data described in Sections 2 and 3 and in the Supporting Information are provided at https://osf.io/4dc3p/.