Using Novel Neural Measures to Explore the Development of Infant Attention Bias to Threat

Abstract

Attention bias to threat is considered an adaptive cognitive phenomenon that is associated with developmental and psychopathological outcomes across the lifespan. However, investigations into the development of attention bias to threat in infancy have produced mixed results. Steady-state visual evoked potentials (ssVEPs) provide a robust measure of visual cortex processing and attention by capturing brain entrainment to the rhythmic flicker of visual stimuli. This investigation leveraged a novel ssVEP task to examine attention bias to threat via affective expressions and its changes with age within the first two years of life. Infants (N = 118, Mean_age = 9.21 months; Range_age = 3–22 months; 57.61% Female) viewed a series of affective face pairs (neutral with happy, fearful, or angry) in which one face flickered at 6 Hz and the other at 7.5 Hz, while their brain activity was measured with EEG. Infants’ frequency-tagged brain responses were larger to fearful faces, above all other expressions, consistent with the presence of an attention bias to threat in infancy. Affect-biased attention did not change with age. Furthermore, the presence of an attention bias toward fear was found prior to the literature-suggested age of seven months. This study demonstrated the utility of using a robust and novel measure of attention, ssVEPs, to examine attention bias to threat and its development during infancy.

Evolutionary theory suggests that attention bias to threat is advantageous for the survival and proliferation of humans, both as an individual and as a species (Bar-Haim et al., 2007; Montag & Davis, 2018; Panksepp, 2004; van Rooijen et al., 2017). Psychological investigations have demonstrated that individuals preferentially attend to threatening stimuli, such as threatening facial expressions, snakes, spiders, and syringes (LoBue, 2010; LoBue & DeLoache, 2008, 2010; Öhman & Mineka, 2001; Soares et al., 2009; Soares & Esteves, 2013). Bias in the processing of threat-related information is thought to be normative and potentially advantageous, as evidenced by a study showing higher fear bias at 8 months to be associated with higher socioemotional competence at 2 years (Eskola et al., 2023). However, individual differences in threat perception, especially at the extremes, have also been implicated in the development of psychopathology (Abend et al., 2018; Bar-Haim, 2010; Bar-Haim et al., 2007; Dudeney et al., 2015; Hakamata et al., 2010; Morales et al., 2017; Van Bockstaele et al., 2014). Thus, understanding the development of attention bias to threat is crucial not only for elucidating the ontogeny of emotion processing but also for its potential clinical implications. Attention bias to threat is observed in early development, even within the first year of life (LoBue & DeLoache, 2010). However, it is unclear exactly when attention bias to threat arises and how it changes in infancy. The current study uses a novel paradigm to robustly measure infant attention using brain activity and examine the emergence of attention bias to threat in infancy.

In a world characterized by a perpetual stream of information it is necessary to prioritize limited attentional resources. To rapidly detect and respond to danger, it is evolutionarily advantageous for indicators of threat to take precedence, making attention bias to threats an adaptive cognitive phenomenon. As social animals, humans often rely on facial expressions to convey and detect these threat cues (Jessen & Grossmann, 2020; Leppänen, 2011; van Rooijen et al., 2017). For instance, angry expressions may indicate a direct threat from the expressor while fearful expressions signal the presence of a potential threat in the environment (Adams & Kleck, 2003). Thus, an attention bias toward threat may naturally extend to facial expressions. Facial affective cues may be particularly salient for infants, who rely on their caregivers for safety and to learn about the world, especially around the onset of locomotion (Leppänen & Nelson, 2012). Furthermore, infancy may be a sensitive period in the development of face and emotion processing, making infancy an ideal period to examine the development of attention bias to threat (Jessen & Grossmann, 2020; Yrttiaho et al., 2014).

How and when attention bias to threat develops is an area of consideration in both neuroscience and developmental literature. What is typically conceived of as attention and biases thereof are a series of processes requiring the development of multiple visual, neural, and cognitive processes to function. On a fundamental level, the eyes and necessary cortical visual structures need to be sufficiently functional in their ability to identify and distinguish between facial expressions, which is thought to be possible by four months of age (Barrera & Maurer, 1981; T. M. Field et al., 1982; LaBarbera et al., 1976). In adults, the automatic (alerting and orienting) aspects of attention are associated with the amygdala, orbitofrontal cortex and the visual cortex, while volitional control of attention is thought to be mediated by the executive system in areas such as the medial prefrontal cortex (mPFC), ventrolateral PFC, dorsolateral PFC, and anterior cingulate cortex (Morales, Fu, et al., 2016; Posner & Rothbart, 2023). The latter executive system has a slow developmental course, not reaching maturity until later in life, implicating automatic processes in attention bias presentation in infancy. However, the neural systems underlying automatic processes, orienting in particular, are functionally mature between the ages of three and six months (Colombo, 2001; Courage et al., 2006).

Development of Attention Bias

The development of attention bias has been theoretically described by Field and Lester (2010) in three different potential developmental models, called the Integral, Moderation, and Acquisition models. The Integral model posits that individual factors, such as genetics and temperament, determine the degree to which an attention bias exists and that it is evident early and remains relatively unchanged across the lifespan (A. P. Field & Lester, 2010). In this model, attention bias to threat is innate and infants with the associated individual factors would show a similar bias at different ages when assessed with developmentally appropriate measurement (i.e., attention bias to threat would not change with development; A. P. Field & Lester, 2010). The Moderation model suggests an existing attention bias to threat is present in all infants, decreases over the course of normative development, and may persist/increase in the case of non-normative development due to individual-level factors (A. P. Field & Lester, 2010). Finally, the Acquisition model posits that attention bias emerges gradually through the shaping of developmental experiences (A. P. Field & Lester, 2010). In this model, infants would not initially show an attention bias to threat, but it would develop with age in those with the associated experiences (i.e., learning), akin to a symptom of a disorder (Burris et al., 2019). Most research currently supports the Moderation and Acquisition models over the Integral model (A. P. Field & Lester, 2010; Morales, Fu, et al., 2016). However, attention bias to threat may be better represented by a hybrid model that combines two models, such as that proposed by Morales and colleagues, which suggests that an early bias exists which is moderated by both individual and environmental factors in a potentially cyclical manner (Morales, Fu, et al., 2016).

A general threat bias is seen as early as the first year of life, as evidenced by findings that infants as young as eight months orient more quickly to snakes than to flowers and frogs (LoBue & DeLoache, 2010). Indeed, evidence suggests that this bias also exists with facial expressions. For example, eight- to fourteen-month olds will orient faster towards angry faces than happy faces when presented side-by-side (LoBue & DeLoache, 2010). A large body of research suggests that by seven months, infants will look longer at fearful facial expressions and are less likely to disengage when presented with a distractor than with happy and neutral expressions (Leppänen et al., 2010; Nelson & Dolgin, 1985; Peltola et al., 2008; Peltola, Leppänen, Mäki, et al., 2009). Studies additionally provide neurophysiological evidence for a threat bias including heightened brain responses, as measured by event-related potentials (ERPs), for fearful faces compared to happy and neutral expressions (Leppänen, Kauppinen, et al., 2007; Nelson & De Haan, 1996; Peltola, Leppänen, Mäki, et al., 2009).

Most of these studies suggest that these threat biases emerge or are present by around seven months of age (Aktar et al., 2021; Hoehl & Striano, 2010; Kotsoni et al., 2001; Leppänen et al., 2010; Leppänen, Kauppinen, et al., 2007; Leppänen et al., 2018; Nelson & De Haan, 1996; Nelson & Dolgin, 1985; Peltola et al., 2008; Peltola, Leppänen, Mäki, et al., 2009; Peltola, Leppänen, Vogel-Farley, et al., 2009; Peltola et al., 2011, 2018; Reilly et al., 2022), and are not present before 7 months (Hoehl & Striano, 2010; Jessen & Grossmann, 2016; Leppänen et al., 2018; Peltola et al., 2013; Peltola, Leppänen, Mäki, et al., 2009). However, a smaller subset of studies suggests these biases may emerge earlier in development, with some studies finding a threat bias before 7 months (Bayet et al., 2017, 2021; Heck et al., 2016, 2016; Miguel et al., 2019; Safar & Moulson, 2020; Yrttiaho et al., 2014). For example, Heck and colleagues found that infants paid more attention to fearful faces versus happy or neutral faces when using dynamic stimuli (Heck et al., 2016) – although these findings have been debated (Grossmann & Jessen, 2017; Heck et al., 2017). While this same study did not support bias existing at three and a half months (Heck et al., 2016), another examination by Bayet and colleagues suggests that infants at this age show a detection advantage for fearful faces compared to happy faces (Bayet et al., 2017; Heck et al., 2016). This implies that the foundations for attention bias to threat may exist and be detectable earlier than is currently reported in the literature (i.e., seven months). Studies that examine attention bias to angry facial expressions using age as a continuous variable also find mixed results (Burris et al., 2017; Fu et al., 2019; Morales et al., 2017; Pérez-Edgar et al., 2017). For instance, one eye-tracking investigation by Pérez-Edgar and colleagues indicated that with age infants spent greater time attending to affective faces, especially angry faces (Pérez-Edgar et al., 2017). Other studies found no change in attention bias across ages (Burris et al., 2017; Fu et al., 2019; Morales et al., 2017), while longitudinal studies support an attention bias to angry faces emerging with age (Reider et al., 2022). Thus, questions still exist about the age at which infants begin to demonstrate an attention bias to threat via affective expressions.

Measuring Attention Bias in Infancy

Inconsistencies in results across studies may arise from methodological limitations, as a primary challenge of studying the development of attention bias during infancy is accurate and reliable measurement. Given that infants cannot report on their attentional states nor participate in tasks requiring a response (e.g., pressing a button as in the Dot-Probe Task; MacLeod et al., 1986), researchers use multiple other methods to assess attention in infancy. Some studies rely on peripheral measures, such as heart-rate (Leppänen et al., 2010; Peltola et al., 2011, 2013), which may indicate but not directly capture attentional processes.

Most recent studies use eye-tracking measures, such as first fixations and dwell times, to assess infant attention bias and its development by inferring attention from looking behavior (e.g., Aktar et al., 2022; Bayet et al., 2017; Bierstedt et al., 2022; Burris et al., 2017, 2022; Eskola et al., 2021, 2023, 2024; Fu et al., 2019; Heck et al., 2016; Kataja et al., 2022; Leppänen et al., 2018; LoBue et al., 2017; LoBue & DeLoache, 2010; Miguel et al., 2019; Morales et al., 2017; Peltola et al., 2015; Pérez-Edgar et al., 2017; Reider et al., 2022; Reilly et al., 2022). However, eye-tracking methodologies can be difficult to implement with infants, who are often uncooperative during eye-tracking calibration. In addition, the immature muscle control of the neck as well as the facial and eye structure of young infants can be a challenge for eye-tracking (Gharib & Thompson, 2022). Because of this, eye-tracking tends to be more successful with older infants and there are relatively few eye-tracking studies with young infants (e.g., younger than 5 months of age; (Gharib & Thompson, 2022; Hessels & Hooge, 2019). Even when eye-tracking has been successful in capturing looking behavior, from which researchers infer overt attention, covert aspects of attention are not directly measured, potentially reducing its comprehensiveness (Christodoulou et al., 2018; Vettori et al., 2020).

Neuroscience approaches have also been used to examine visual processing and attention processes at different levels attention in infancy and may enable examination of covert attention. For example, studies have employed functional near-infrared spectroscopy (fNIRS; e.g., Bayet et al., 2021; Porto et al., 2020) and EEG-based ERPs (e.g., Bowman et al., 2022; Hoehl & Striano, 2008; Kane-Grade et al., 2024; Leppänen, Kauppinen, et al., 2007; Leppänen, Moulson, et al., 2007; Moulson et al., 2009; Nelson & De Haan, 1996; Peltola et al., 2018; Peltola, Leppänen, Mäki, et al., 2009; Yrttiaho et al., 2014). However, fNIRS captures a slow hemodynamic response that is not sensitive to millisecond-level changes in attention. ERP tasks share the characteristics of long run times and considerable data loss with eye-tracking paradigms, commonly losing 50–60% of data or more and limiting potential analyses (Bowman et al., 2022; Burris et al., 2022; Leppänen et al., 2018; Morales et al., 2017; Pérez-Edgar et al., 2017; Vallorani et al., 2021). To better understand attention bias and its development in infancy, it is crucial to use a direct measure of attention that is robust to the challenges of infant data collection.

The Potential of Steady-State Visual Evoked Potentials as a Measure of Affect-Biased Attention

Steady-state visual evoked potentials (ssVEPs) show promise in their ability to assess attention bias in infancy while overcoming some limitations of current methodologies. ssVEPs are a neurophysiological response of attention measurable via electroencephalography and produced by the synchronization of neural oscillations to the frequency of flickering stimuli (Kabdebon et al., 2022; Norcia et al., 2015; Notbohm et al., 2016; Wieser et al., 2016). For example, if a stimulus is flickered at 6 Hz, a corresponding response at 6 Hz is observed in the brain activity of areas processing that stimulus (e.g., visual cortex; Norcia et al., 2015). Adult validation studies suggest that not only are ssVEPs a robust and direct measure of visuocortical engagement, but they have been found to be strongly enhanced for attended stimuli, relative to unattended stimuli (Morgan et al., 1996; for a review, see Norcia et al., 2015). Because of this, ssVEPs are frequently used in studies of visual attention and have several advantages that enable their use for studying attention bias in infants (Figueira et al., 2022; Kabdebon et al., 2022; Norcia et al., 2015).

One strength of ssVEPs is their utility in measuring attention in multi-stimulus displays. For example, if multiple stimuli are flickered at different temporal rates (e.g., 6 Hz and 7.5 Hz), activity at those frequencies can be readily separated using frequency-domain analysis. The resulting spectral power values provide measures of visuocortical engagement that specifically reflect the region tagged at one frequency, compared to visuocortical engagement underlying the region tagged at the second frequency, providing stimulus specificity even when stimuli are competing for attention or overlapping in space (Deweese et al., 2016; Müller et al., 2008; Wieser et al., 2011).

Additionally, ssVEPs may be a more comprehensive and inclusive measure of attention than traditionally used methods, including eye-tracking. Studies with adults and infants suggest that ssVEPs reflect both overt and covert attention, distinguishing between active sustained attention and inattentive staring (Christodoulou et al., 2018; Robertson et al., 2012; Vettori et al., 2020). Christodoulou and colleagues, for example, validated the use of ssVEPs with infants in a frequency-tagging approach (e.g., Figueira et al., 2022) and extended paradigms used in adult work to show increased SNR for more attractive stimuli in infant’s peripheral vision while the infants simultaneously fixated on a central stimulus (Christodoulou et al., 2018). This illustrates ssVEPs’ ability to detect covert attention processes in infants that are not evident with eye-movement-based measures (e.g., looking time). Furthermore, one study detected differences in attention to faces between children with and without autism in ssVEP measures when eye-tracking was unable to do so, demonstrating the feasibility of using ssVEPs in populations where measures of covert attention may enable a more thorough examination of attention (Vettori et al., 2020). As infants cannot report their attention, this more comprehensive and direct approach to measuring attention may be particularly useful.

ssVEPs also may address procedural challenges of examining attention in infancy. Unlike eye-tracking, ssVEP paradigms do not require calibration, reducing the likelihood of data loss in individuals with limited ability to have stable fixations and track moving objects (e.g., young infants; Aslin & Salapatek, 1975), or remain still for an extended time. Furthermore, the high signal-to-noise (SNR) ratio of ssVEPs and fast precision enables short recordings more suitable for infants and their limited attention span, compared to longer ERP and eye-tracking tasks (Figueira et al., 2022; Kabdebon et al., 2022). Additionally, high SNR makes ssVEPs robust to data loss typical in infant populations due to relatively low data quality (Figueira et al., 2022). These strengths are evidenced in the successful use of ssVEPs in research with infants (de Heering & Rossion, 2015), and even newborns (Buiatti et al., 2019).

The Current Investigation

The current investigation seeks to leverage the advantages of ssVEPs to examine the emergence of attention bias toward threat in infants aged three to twenty-four months. For this, we utilize a task that presents two faces next to each other (similar to the Dot-Probe Task; (MacLeod et al., 1986) flickering at different rates (6 and 7.5 Hz), involving angry, fearful, happy, and neutral facial expressions. The addition of both fearful and angry facial expression is notable given that it allowed us to test the specificity of the threat bias. We hypothesize that generally infants will show an attention bias as measured via the SNR of ssVEPs toward threat cues (fearful and angry faces), compared to other affective expressions, such as happy and neutral facial expressions. Developmentally, based on existing literature, we anticipate increased SNR for affective threat cues versus other affective expressions to emerge at or before seven months of age and to increase with age. By using this novel measure of attention, we aim to better understand the emergence of attention bias to threat and how it changes with development across infancy. In addition to using these innovative methods to characterize attention bias to threat and its development, we hope to demonstrate the utility of ssVEP paradigms for examining attention in infancy.

Method

Participants

Participants consisted of 120 infants from a larger study examining the development of infant attention. 118 infants (57.61% female; 3–22 months, mean age = 9.21 months) had useable EEG data following pre-processing (described below). In terms of ethnicity, 56.78% identified as Hispanic/Latino. For race, 33.05% identified as White, 28.81% as Identity Not Listed, 9.32% as Black, 7.37% as Multi-Racial, 7.63% as Asian, and 0.85% as Native Hawaiian/Pacific Islander. Infants were included in the study if they met the following criteria: a) be between three and twenty-four months of age, b) were born within three weeks of their due date, c) be without serious medical complications, and d) have no family or personal history of seizures due to the flickering nature of the EEG task. Participants were recruited from the community in and around Los Angeles, CA via flyers, Facebook advertisements, in-person events, and BuildClinical’s research recruitment services. All recruitment materials were available in both English and Spanish. Caregivers gave informed consent prior to participation and all aspects of the protocol were approved by the IRB of the University of Southern California (UP-21-01082).

Procedures

After providing informed consent and prior to beginning EEG data collection, a trained research assistant asked the infant’s caregiver for information on their demographic characteristics, including infant birthdate, race, ethnicity, and sex, as well as estimated gross income across primary and any secondary caregivers.

EEG was collected using a 128-channel EGI HydroCel Geodesic Sensor net (adapted to 124 channels by removing eye channels E125-E128), NetAmps 400, and Net Station software (Magstim EGI, 2021). The timing presentation of the stimuli were measured using a photocell and a Cedrus StimTracker. Infants sat on their primary caregiver’s lap at approximately 60 cm from the monitor throughout the duration of the task. Before the task began, impedances were measured to ensure they fell below 50 kΩ. The ssVEP task was modeled off the Dot-Probe Task to capture attentional engagement while competing for attentional resources with other social stimuli. Infant participants viewed four blocks of 18 trials (total 24 trials per condition) in which two flickering faces appeared side-by-side on the screen for 4000 ms in the following affective pairs: Neutral-Happy, Neutral-Fearful, and Neutral-Angry. One face flickered at 6 Hz and the other at 7.5 Hz in accordance with a frequency-tagging approach that would both minimize overlap in frequency bins and their harmonics for analysis (Figueira et al., 2022). Moreover, these slower frequencies, compared to adult work, were selected to minimize the risk of inducing a photosensitive epileptic seizure. Six female faces of differing racial and ethnic backgrounds were used from the NimStim stimuli set, with a single model appearing as both faces in each trial (Tottenham et al., 2009). Faces were presented on a gray background and matched for luminance using Adobe Photoshop and the SHINE toolbox in MATLAB (Adobe, 2024; Willenbockel et al., 2010). The images were presented on a 19.5 in Lenovo ThinkVision monitor (1440 × 900 resolution) and measured 6 cm (width, approximately 5.7 degrees of horizontal visual angle) and 8.5 cm (height) with 6 cm in between the two faces, resulting in the inner edge of each face being 2.9 degrees excentric relative to fixation. The order of the conditions, side of the screen that each expression appeared on, and rate of flickering were counterbalanced across trials. Trials started with an attention-grabber fixation at the center of the screen and an associated sound (e.g., an illustration of a duck with a quacking noise) to get the infant’s attention.

Measures

SNR

Continuous EEG data was first aligned with the stimulus onset markers and converted from a .mff file format to a .set format. These labeled .set files were run through the HAPPE+ER pipeline (version 4), of which a detailed description of the pre-processing steps can be found in Monachino et al. (2022). In short, the pipeline first corrects artifacts present in the data using line noise reduction via the Cleanline EEGLAB plug-in (Delorme & Makeig, 2004; Mullen, 2012) and wavelet thresholding. The pipeline then bandpass filters with EEGLAB’s FIR filter (Delorme & Makeig, 2004) to the frequencies of interest (0.1 – 45 Hz) and segments the data into epochs ranging 4100 milliseconds that start 100 ms prior to stimulus presentation and end four seconds post-presentation. Remaining artifact-laden segments are rejected if they contain amplitudes of ±200 μV in occipital electrodes E70, E71, E75, E76, or E83. Finally, data is re-referenced to the average. The full parameters used to process the data are included in the Supplement as Table S1.

Following pre-processing, data was baseline corrected by subtracting the time window −100 to 0 ms from the time series and averaged for each condition. Averaged data was then transformed into the frequency domain through a fast Fourier transform (FFT) on the full 4000 ms segment, resulting in a frequency resolution of 0.25 Hz. SNR was calculated by dividing each frequency value by the average of the nearest ten frequency values (five below, five above; Vettori et al., 2020). Finally, SNR was averaged separately for the frequencies of interest (6 Hz and 7.5 Hz) and their harmonics (12 and 18, 15 and 22.5, respectively) for electrode Oz (E75) in a frequency-tagging approach (Figueira et al., 2022; Vettori et al., 2020). The use of Oz was determined a priori (e.g., McTeague et al., 2017) and confirmed via visual inspection of the topographical maps of activation across the scalp (Figure 1). Furthermore, sensitivity analyses utilizing a cluster approach around Oz (E70, E75, E80) showed reduced SNR, suggesting a localized scalp distribution. Importantly, the sensitivity analyses with the cluster showed the same pattern of results as the ones presented in this manuscript (see Supplement, Figure S1 and Figure S2).

Figure 1. Topographical Map of SNR Across the Scalp.

Note. Topographical maps of SNR across the scalp for each affective condition: Neutral, Happy, Fearful, and Angry. Areas with higher SNR are indicated by warmer color values.

Internal consistency of the SNR measures were assessed using trial-level SNR estimates via split-half reliability as implemented in the READIE toolbox (Xu et al., 2024). The internal consistency reliability estimates of the SNR measures for each condition were found to be acceptable and good (Angry = .61, Fear = .64, Happy = .69, and Neutral = .87). Although it is more common to use baseline-corrected amplitude rather than SNR in ssVEP studies, we decided to use SNR as our main measure given it demonstrated higher reliability than baseline-corrected amplitude. However, in order to provide a comparison with other studies, we also present results with baseline-corrected amplitude, finding the same pattern of results (see Supplement, Figure S3 and Figure S4).

Infant Age

Infants’ caregivers reported their child’s date of birth, which was used in conjunction with the date of participation to calculate the infant participant’s age in months.

Caregiver Income

Infants’ caregivers indicated the total estimated gross income for both primary and secondary caregivers (if applicable) by selecting from the following options: Less than $15,000; $15,000–$29,999; $30,000–$44,999; $45,000–$59,999; $60,000–$74,999; 75,000–$89,999; 90,000–$104,999; 105,000–$119,999; 120,000–$134,999; 135,000–$149,444; and $150,000 or more. These bins were assigned a number from one to eleven in ascending order to create a continuous scale of income.

Analyses

To examine the first aim of the study, examining differences between affective facial expressions, differences in SNR across affective conditions were tested using a multi-level model (MLM), which is the equivalent to a one-way repeated-measures ANOVA with Affect as a within-subjects factor. A significant effect of Affect was followed with paired t-tests to determine specific differences between conditions. Specifically, we examined affective expressions (i.e., Angry, Fear, and Happy) compared to Neutral and differences among the affective expressions. We focused on comparing to Neutral a priori given that this is what is commonly done in the Dot-Probe Task with infants (Burris et al., 2022; Pérez-Edgar et al., 2017) and children (Morales et al., 2015, 2020; Morales, Pérez-Edgar, et al., 2016). However, as a follow-up test, we also examined differences between the three affective expressions.

To test the second aim of this study, evaluating age effects in attention to different affective facial expressions, we ran a mixed effects model equivalent to a repeated-measures ANCOVA with Affect as a within-subject factor, Age as a between-subject factor, and their interaction. A significant interaction would indicate that the attention to affective facial expressions changed with age as a continuous measure. These models were conducted using the nlme package in R (Pinheiro et al., 2024), and included total number of trials (Mean = 55.41, SD = 20.13), infant sex, and combined caregiver income as covariates.

To further rule out differences being due to different numbers of trials (i.e., the Neutral condition had more trials as it was present all trials), as an additional sensitivity analysis we also modeled the data at the trial level (i.e., estimated SNR values for each trial). This sensitivity analysis allowed us to better model the different number of trials and estimate effects within trial (i.e., the difference between affective expressions [Angry, Fear, and Happy] to the accompanying Neutral face), showing similar results. Furthermore, to examine potential habituation effects and reduced salience of neutral faces, we tested whether the difference between conditions changed significantly over the course of the task by evaluating the interaction between Affect and trial number.

Because solely testing the interaction with age does not perfectly test the emergence (i.e., presence vs absence) of an affective bias at a given age, we conducted additional sensitivity analyses to more explicitly test this. Furthermore, a continuous measure of age may fail to characterize a discrete presence of bias between ages as in stage-based development. Given that seven months is the most supported age for the emergence of attention bias to threat, we conducted additional analyses with infants younger than seven months. The presence of attention bias to threat in infants younger than seven months would contradict the current view that attention bias to threat emerges at seven months (Bayet & Nelson, 2020; Hoehl & Striano, 2010; Hoemann et al., 2020; Leppänen et al., 2010; Leppänen et al., 2018; Leppänen, Moulson, et al., 2007; Leppänen & Nelson, 2009; Nelson & De Haan, 1996; Nelson & Dolgin, 1985; Peltola et al., 2013; Peltola, Leppänen, Mäki, et al., 2009). This analysis was identical to the one described above for the first aim, except limiting the sample to infants younger than seven months. Using this approach, we aimed to identify a potential window of emergence for attention bias. All analyses were run in RStudio using R version 4.4.1 (Posit Team, 2024; R Core Team, 2024).

Transparency and Openness

We report all data exclusions and all measures in the study, and we follow JARS (Appelbaum et al., 2018). HAPPE software is freely available at https://github.com/PINE-Lab/HAPPE. Further research materials can be made available at reasonable request. This study’s design and its analysis were not pre-registered.

Results

Attention to Affective Expressions

The MLM revealed a significant difference in attention, as measured via SNR, to the different affective expressions in infants, F(3, 338) = 17.90, p < .001, as shown in Figures 2 and 3. Comparisons between affective conditions (i.e., Angry, Happy, and Fear) and the neutral condition showed that SNR was greater for fearful faces (t(117) = 7.08, p < .001, d_z = .65) and for happy faces (t(116) = 4.00, p < .001, d_z = .37) than for neutral faces. However, there was no significant difference in SNR for angry and neutral faces (t(117) = 1.91, p = .059, d_z = .18). Pairwise comparisons between affective conditions showed that SNR was greater for fearful faces versus both happy faces (t(116) = 3.19, p = .002, d_z = .30) and angry faces (t(117) = 4.88, p < .001, d_z = .45), while SNR was greater for happy faces than for angry faces (t(116) = 2.07, p = .041, d_z = .19). All pairwise tests survived correction for multiple comparisons using the false discovery rate.

Figure 2. SNR by Affective Condition.

Note. A bar plot showing the SNR for each affective condition, including error bars representing the within-subject 95% confidence intervals (Loftus & Masson, 1994). Angry is shown in red, Fear in yellow, Happy in green, and Neutral in blue. The y-axis begins at 1 to better illustrate differences across conditions.

Figure 3. Comparisons of SNR Across Frequency for Each Affective Condition Versus Neutral.

Note. Comparisons of SNR across frequencies for each trial type: A) Angry-Neutral, B) Happy-Neutral, and C) Fear-Neutral. The blue line in each graph represents SNR when the non-neutral expression (e.g., angry) is flickering at 6 Hz. The red line in each graph represents SNR when the non-neutral expression is flickering at 7.5 Hz. The vertical dotted lines in each plot indicate 6 Hz, 7.5 Hz, 12 Hz, 15 Hz, 18 Hz, and 22.5 Hz. The labels above each peak indicate hypothesized results such that the top condition should have higher SNR (e.g., Fear over Neutral indicates Fear should have higher SNR) with the text color indicating which line matches the rate of flickering at that frequency.

Age Effects

When examining the effects of age, we observed a significant effect of age on overall SNR for viewing faces, regardless of affect, such that SNR decreased with age, F(1, 109) = 13.51, p < .001. The model did not indicate a significant Affect by Age interaction for SNR, F(3, 335) = 1.56, p = .200. This was true when examining changes relative to Neutral or changes relative to Happy (without Neutral in the model).

Because the above interaction examines changes with age as a continuous measure, rather than the presence (vs absence) of a bias, we conducted a further analysis focused on identifying whether a bias was present prior to seven months. This analysis including only infants younger than 7 months included 53 infants. The MLM revealed a significant difference in attention, as measured via SNR, to the different affective expressions in infants younger than seven months, F(3, 147) = 9.17, p < .001, as illustrated in Figure 4. Comparisons between affective conditions (i.e., Angry, Happy, and Fear) and the neutral condition showed that SNR was greater for fearful faces (t(52) = 5.01, p < .001, d_z = .69) and for happy faces (t(52) = 3.53, p < .001, d_z = .48) than for neutral faces. However, there was no significant difference in SNR for angry and neutral faces (t(52) = 0.66, p = .514, d_z = .09). Pairwise comparisons between affective conditions showed that SNR was greater for both fearful faces (t(52) = 3.97, p < .001, d_z = .54) and happy faces (t(52) = 2.53, p = .014, d_z = .35) compared to angry faces. There was no significant difference in SNR between fearful and happy faces prior to seven months of age (t(52) = 1.59, p = .117, d_z = .22). All pairwise associations survived correction for multiple comparisons using the false discovery rate.

Figure 4. SNR by Affective Condition for Infants Younger than 7 Months of Age.

Note. A bar plot showing the SNR for each affective condition, including error bars representing the within-subject 95% confidence intervals (Loftus & Masson, 1994). Angry is shown in red, Fear in yellow, Happy in green, and Neutral in blue. The y-axis begins at 1 to better illustrate differences across conditions. This analysis includes 53 infants younger than 7 months.

Sensitivity Analyses

To examine the robustness of these results, we conducted several sensitivity analyses. First, as shown in the Supplement, these results were replicated when using trial-level data (see Supplement, Table S2, Table S3, Figure S5 and Figure S6) and the interaction between Affect and Trial Number was not significant (see Supplement). These results ameliorate concerns about the observed differences being due to different number of trials or potential habituation effects to neutral faces. In another sensitivity analysis, we also replicated the results when examining baseline-corrected amplitude rather than SNR (see Supplement, Figure S3 and Figure S4), suggesting our results are robust to several methodological decisions.

Discussion

Attention bias to threat is an adaptive cognitive phenomenon that has theoretically deep evolutionary origins, potentially shapes how infants process and interpret affective information, and at the extreme, has been associated with clinical outcomes later in life (Abend et al., 2018; Bar-Haim, 2010; Bar-Haim et al., 2007; Dudeney et al., 2015; Grossmann, 2023; Hakamata et al., 2010; Montag & Davis, 2018; Morales et al., 2017; Panksepp, 2004; Van Bockstaele et al., 2014; van Rooijen et al., 2017). However, research to date varies on how and when attention bias to threat emerges, a discrepancy that may, in part, be due to the methodological limitations of existing studies. This investigation sought to characterize attention bias to threat in infancy and clarify the age at which it emerges using a novel measure of attention that directly captures visual cortex processing, ssVEPs.

In the first aim of the study, we used a novel EEG measure, ssVEPs, to examine infant attention to different affective expressions; namely neutral, fearful, happy, and angry faces. Across all infants, we found an attention bias to fearful faces over and above all other affective expressions, as indexed by a higher SNR. We also found a bias for happy faces compared to neutral and angry faces. While we hypothesized that angry faces would additionally serve as a threat cue and therefore be subject to more attention, infant attention to angry expressions was not significantly different than for neutral faces. Increased attention to fearful expressions over all other facial expressions is in line with previous studies that also find increased attention to fearful expressions over other expressions (Aktar et al., 2022; Bayet et al., 2021; Heck et al., 2016; Hoehl & Striano, 2008; Leppänen et al., 2010; Leppänen et al., 2018; Leppänen, Kauppinen, et al., 2007; Leppänen, Moulson, et al., 2007; Miguel et al., 2019; Morales et al., 2017; Moulson et al., 2009; Peltola et al., 2011; Peltola, Leppänen, Mäki, et al., 2009; Peltola, Leppänen, Vogel-Farley, et al., 2009), including angry expressions (Hoehl & Striano, 2008; Miguel et al., 2019). These findings suggest that there may be important differences in both indicators of threat and information conveyed by different affective expressions, warranting further examination.

In the second aim of the study, we examined attention bias in relation to age. Overall, infants paid less attention to faces as age increased. This could be the result of maturation of facial and emotion-processing circuits leading to faster processing or from increased familiarity with screens and affective stimuli. We did not find a significant emotion-by-age interaction, contrary to our hypothesis that attention to affective expressions changes differentially over time.

Furthermore, to try to more directly examine the presence of affect-biased attention in young infants, we also examined differences in affect-biased attention for infants younger than seven months. A relative bias to fearful faces compared to neutral and angry faces was found, contradicting the literature suggesting that attention bias emerges at approximately seven months of age. Together these findings lend support to recent studies finding an earlier age of emergence for attention bias to threat (Heck et al., 2016). Interestingly, infants demonstrated a bias for happy faces compared to neutral and angry faces prior to seven months. In conjunction, comparisons against happy faces suggests the presence of an initial favorability by infants to pay attention to happy faces, consistent with literature that young infants demonstrate preferences for happy faces (Farroni et al., 2007; Johnson et al., 2015).

Together, our results suggest the presence of a strong bias to fear and a moderate bias to happy facial expressions that are present early in infancy and do not significantly change with age. These findings contradict existing literature that suggests affect-biased attention emerges in infancy (Bayet & Nelson, 2020; Hoehl & Striano, 2010; Hoemann et al., 2020; Jessen & Grossmann, 2016; Leppänen et al., 2010; Leppänen et al., 2018; Leppänen & Nelson, 2009; Nelson & De Haan, 1996; Nelson & Dolgin, 1985; Reider et al., 2022) and Field and Lester’s Acquisition model (A. P. Field & Lester, 2010). The lack of changes in affect-biased attention with age suggest a consistent attention bias more closely aligned with the Integral model (A. P. Field & Lester, 2010), which suggest an innate attention bias to threat. Importantly, in the current study we did not test for potential moderators of age-related changes like child temperament (e.g., Fu et al., 2019; Pérez-Edgar et al., 2017; Ravicz et al., 2015; Vallorani et al., 2022, 2023), maternal factors (Bowman et al., 2022; Kataja et al., 2020; Morales et al., 2017; Porto et al., 2020; Reilly et al., 2022; Vallorani et al., 2023; Waters et al., 2015), or the infant-caregiver relationship (Eskola et al., 2024; Peltola et al., 2015). As such, it is possible that these results also support a hybrid model of attention bias development and there may be different developmental trajectories that vary with individual-level factors, context, and affective conditions (Morales, Pérez-Edgar, et al., 2016). It is also unknown if these early affective biases in attention are adaptive (Eskola et al., 2021, 2023; Peltola et al., 2015) or maladaptive (Abend et al., 2018; Bar-Haim, 2010; Bar-Haim et al., 2007; Dudeney et al., 2015; Hakamata et al., 2010; Morales et al., 2017; Van Bockstaele et al., 2014); thus future longitudinal studies are needed to answer these open questions.

Another goal of this investigation was to illustrate the utility of ssVEPs in infant attention bias research. Not only were we able to successfully detect attention bias in infancy using this novel measure, but we also kept 98.33% of participants who attempted the task with an average of 81.44% trials retained and adequate to good split-half reliability. These reliability measures in particular are either in-line with or better than existing measures for assessing infant attention, including EEG (Xu et al., 2024), eye-tracking (Leppänen et al., 2018), and some scales of maternal reports of infant socioemotional behavior (e.g., the Infant-Toddler Social and Emotional Assessment [ITSEA; Carter et al., 2003]; the Infant Behavior Questionnaire - Revised [IBQ-R; Putnam et al., 2014]). Furthermore, it is likely that the reliability estimates can be improved by optimizing the task (e.g., increasing the number of trials per condition) or advances in pre-processing. Finally, the paradigm also does not require calibration like eye-tracking measures and takes an average of approximately eight minutes to complete, making it tolerable and feasible for infant participants. Future investigations may seek to leverage ssVEPs as a means for studying the development of attention biases given ssVEPs’ advantages and comprehensiveness compared to other commonly used measures that involve longer paradigms and high data loss (e.g., eye-tracking, ERPs).

Furthermore, ssVEPs may provide unique insight into the neural mechanisms of attention bias to threat. Recent models of threat processing extend beyond the amygdala and emphasize the role of the visual cortex as a key node in the threat network, encoding threat information and sending rapid and specific signals to the rest of the network to guide responses (Li & Keil, 2023). This is supported by evidence that feedback mechanisms between the amygdala and the primary visual cortex show that the visual cortex is modulated by threat signals during the earliest stages of processing (Liu et al., 2022). As ssVEPs directly capture the processing of the visual cortex, they may represent activation relevant to the overall threat detection network and provide a robust measure of the neural mechanisms of attention bias to threat during developmental periods that are difficult to measure with other neuroimaging modalities (e.g., fMRI).

Limitations and Future Directions

The findings of this study should be interpreted in the context of some limitations. First, while ssVEPs are commonly used in other areas of research, they are still a novel paradigm for use in developmental populations so direct comparisons with alternative and traditionally used measures, including eye-tracking and ERPs, are needed to validate them. Furthermore, while the reliability of this measure is equal to or better than other commonly used measures, it can potentially be improved. Secondly, although validation studies in infants and adults demonstrate that ssVEPs can capture covert attention (Christodoulou et al., 2018; Robertson et al., 2012; Vettori et al., 2020), we do not distinguish nor assess differences between covert and overt attention in this investigation. Additionally, this data is cross-sectional so we cannot perform growth-curve analyses to investigate the developmental trajectory of attention bias to threat. We also lack sufficient distribution in ages to compare attention bias in additional age bins (e.g., younger than five months of age), limiting our ability to isolate the age of emergence. Finally, our sample size may be underpowered to detect small effect sizes. Future investigations would benefit from including a greater number of infants at a wider range of ages and employing longitudinal designs to expand upon and refine the trajectories of attention bias to threat with age. Furthermore, research examining relations between attention bias with caregiver and infant characteristics (e.g., temperament, maternal mental health) and nuances of attention (e.g., disengagement) may provide additional insight into underlying mechanisms and developmental patterns.

Conclusion

The present study increases our understanding of attention bias toward threat in infancy. Namely, it demonstrates the capacity of ssVEPs to successfully measure attention bias in infants. Our findings support an attention bias to fearful expressions over other affective expressions that is consistent across infancy. Furthermore, we found the presence of an attention bias toward fearful faces compared to neutral faces in infants seven months and younger, implicating an earlier emergence of attention bias than primarily indicated in existing research. Future research may leverage ssVEPs to further examine the developmental trajectory of attention bias to threat and further elucidate the age and mechanisms of its emergence.

Public Significance Statement:

This study used a novel steady-state visual evoked potentials (ssVEPs) task to examine attention bias to threat and its age-related changes in infancy. Results suggest the presence of an attention bias to fear in infancy detectable in the first half of the first year. Overall, findings illustrate the utility of using ssVEPs to conduct affect-biased attention research in infancy.