Attentional Bias During a Novel Alcohol Cue Virtual Reality Paradigm Authors: Ryan M Sullivan, Benjamin S McKenna, Weichen Liu, Arwynn McKinty, Camarin Laurance, Craig R Rush, Nadir Weibel, Kelly E Courtney Published: 2026 ABSTRACT ABSTRACTBackgroundAttentional bias to alcohol cues is largely investigated using pictorial cues (i.e., two‐dimensional images). The addition of eye‐tracking metrics has improved these investigations yet are limited by a lack of naturalistic cues. Virtual reality (VR) can theoretically improve ecological validity through three‐dimensional alcohol object presentation while utilizing eye‐tracking metrics. This report sought to test the construct validity of the Alcohol Cue VR paradigm in evoking subjective craving and attentional bias through measurement of eye‐gaze characteristics.MethodsParticipants (n = 105, 53% male, Mage = 38), with varying alcohol use histories, underwent the Alcohol Cue VR paradigm, an alcohol visual probe task, and measures of alcohol craving and VR‐related assessments. The Alcohol Cue VR paradigm comprised of three active scenes with alcohol cues present, and three neutral scenes devoid of alcohol cues. Eye‐gaze fixation time (index of “attentional bias”) in response to alcohol and control cues was recorded. Forward‐stepwise models searched for optimal alcohol‐related predictors of attentional bias, then compared against eye‐tracking and response time metrics from the alcohol visual probe task.ResultsAlcohol use history, age of regular use onset, recency of use, and binge episodes were significantly associated with eye‐gaze fixation time to alcohol cues; how present one felt within the VR scenes did not moderate these findings. Alcohol‐related predictors accounted for a larger proportion of variance in VR‐based models of attentional bias (ΔR 2 = 0.10) than in comparable alcohol visual probe task models based on fixation time (ΔR 2 = 0.05) or response time (ΔR 2 = 0.06).ConclusionsThe Alcohol Cue VR paradigm provided commensurate measures of attentional bias in the context of alcohol cues. Unique alcohol‐related characteristics impacted the extent of attentional bias to alcohol‐related cues, exemplifying the utility of VR‐derived measurement in assessing alcohol use. Future research can understand whether VR attentional bias is modifiable through treatment and can serve as clinical endpoints.Keywords: alcohol, attentional bias, craving, cue‐exposure, virtual reality Introduction Substance‐related attentional bias, or the disproportionate focus of attention to drug‐related stimuli relative to neutral stimuli, represents an objectively measurable and clinically relevant phenomenon for the study of addiction (Field and Cox 2008; Topbas et al. 2025). As a behavioral marker of incentive salience (Ostafin and Feyel 2019) and theoretical correlate of substance craving (Franken 2003; Robinson and Berridge 1993; Ryan 2002), substance‐related attentional bias is thought to either cause or index critical processes responsible for drug‐seeking behavior (Field et al. 2009). Attentional bias toward alcohol cues has been demonstrated in social drinkers (Manchery et al. 2017), regular drinkers (Wilcockson et al. 2019), and individuals with problematic alcohol use (Garland et al. 2012; Qureshi et al. 2019); however, a recent review by Bollen et al. (2022) suggests alcohol attentional bias may be more strongly related to drinking intensity at lower severity/subclinical alcohol use levels. Despite this, alcohol attentional bias has been linked to treatment retention in individuals who use both alcohol and cocaine (Diaz‐Batanero et al. 2018) and appears to diminish with prolonged abstinence in individuals with alcohol use disorder (AUD) (Escudero et al. 2024), suggesting the construct may be reflective of a transient, value‐based decision process, as opposed to a stable phenotype, which may prove useful for predicting clinical outcomes. Despite this, the classic attentional bias literature has been fraught with mixed results, low reliability, and small effect sizes. This is due, at least in part, to the frequent use of unreliable indirect attentional bias measurements (e.g., response time from the visual/dot probe and Stroop tasks) (Jones et al. 2018; Pennington et al. 2021). In response, the use of direct eye‐tracking indices of attentional bias toward drug cues is increasingly common (Ghiţă et al. 2019), and mounting evidence supports specific metrics such as eye‐gaze fixation time toward drug cues as providing substantial improvements in bias estimate reliability (Marks et al. 2015; Marks, Pike, et al. 2014; Marks, Roberts, et al. 2014; Soleymani et al. 2020). Through continuous measurement, eye‐tracking provides a more detailed assessment of dynamic attentional engagement toward focal points of interest (e.g., alcohol cues) (Roy‐Charland et al. 2017) and directly measures visual attention by recording how individuals direct their gaze (Godijn and Theeuwes 2003). Assessment within naturalistic settings has also independently improved the reliability (Spanakis et al. 2019) and validity (Monem and Fillmore 2017) of attentional bias measurement, yet the naturalistic constraints of these methods prohibit advanced clinical application. Naturalistic substance cues can be complex and involve several contextual factors that are difficult, if not impossible, to replicate in standard laboratory cue‐based paradigms, thereby limiting the translatability of results to the real world (Bouton 2000). Thus, additional studies using eye‐tracking indices of attentional bias in naturalistic, yet translatable, settings are needed to enhance our understanding of the underlying mechanisms of alcohol attentional bias, its relation to subjective craving, and improve the utility of these paradigms for lab and clinical applications. Virtual reality (VR) is increasingly being studied as a technique to enhance the ecological validity of cue‐based paradigms. VR allows for the presentation of immersive and relevant three‐dimensional (3D) scenes with interactive cues, greatly enriching our ability to invoke craving for various substances in the laboratory (Acker and Mackillop 2013; Baumann and Sayette 2006; Bordnick et al. 2009; Culbertson et al. 2010; Hone‐Blanchet et al. 2014; Tamburin et al. 2021; Traylor et al. 2011); however, the magnitude of this effect often depends on how present one feels in the virtual world (i.e., VR presence; Ferrer‐Garcia et al. 2010; Hernández‐Serrano et al. 2021). Studies using VR cue‐exposure have supported its effectiveness in inducing subjective craving for a variety of substances including alcohol (Bordnick et al. 2008; Ghita and Gutierrez‐Maldonado 2018; Ryan et al. 2010) and provide initial evidence of an effect of VR cue exposure on alcohol attentional bias assessed following the VR paradigm (Ghita et al. 2021). Recent advances in VR implementation now allow for real‐time assessment of eye movement in a noninvasive manner, with cameras built into the VR headset to avoid obstructing the field‐of‐view. These technological advances provide additional possibilities for increasing the translatability of various psychological paradigms with concurrent direct assessment of psychophysiological metrics, such as eye‐gaze fixation time in the context of attentional bias toward naturalistic substance cues. The present study sought to: (a) test the construct validity of a novel Alcohol Cue VR Paradigm as an effective cue‐exposure paradigm capable of inducing subjective alcohol craving in individuals with varying drinking histories; (b) identify the relevant alcohol predictors of alcohol attentional bias derived from eye‐gaze fixation; (c) examine the moderating effect of VR presence (i.e., how present one felt within the virtual scenes) on identified alcohol predictors; and (d) compare attentional bias estimates from the Alcohol Cue VR Paradigm against a validated visual probe task to investigate the potential for the VR‐based attentional bias metric to improve upon traditional response time and two‐dimensional (2D) eye‐gaze metrics. Based on available literature and in line with our funded grant aims, our exploratory hypotheses were that (a) the active scenes of the Alcohol Cue VR Paradigm would induce grater subjective craving than the neutral scenes, across all participants; (b) alcohol use variables (identified via data‐driven exploratory approach) would predict attentional bias metrics; (c) VR presence would moderate the effects of identified alcohol predictors on attentional bias estimates; and (d) the VR‐based attentional bias metric would perform better (greater R 2) at predicting alcohol outcomes than the traditional 2D eye‐gaze metrics. The study sample consisted of adult community members with a range of alcohol use characteristics to facilitate generalization of findings. Materials and Methods Participant Recruitment and Screening Procedures Recruitment of participants was completed through flyers and online (e.g., Craigslist, Nextdoor) advertisements posted in the broader San Diego community. Interested individuals contacted the study team and completed a telephone‐screening interview to determine initial eligibility. Inclusion criteria for the study included: ages 21+; alcohol use on at least three occasions in the past 3 months. Efforts to oversample heavy alcohol use, defined as > 4 drinks males/> 3 drinks for females on > 14 days in the previous 3 months, were employed to ensure variability in drinking histories, resulting in approximately 33% of the sample meeting these criteria. Exclusion criteria included: medical or psychiatric history affecting brain development (i.e., history and/or treatment of neurologic disorders, severe head trauma with loss of consciousness > 2 min, or current severe DSM‐5 psychiatric disorders other than alcohol or tobacco use disorders); nonfluency in English; and visual/vestibular problems that may make study participation difficult (e.g., severe motion sickness, blindness, glasses). Eligible participants were invited for one in‐person study visit where study procedures were explained and participants provided informed consent consistent with the University of California, San Diego Institutional Review Board policies. Prior to the study session, participants were instructed to abstain from cannabis and alcohol use (confirmed by negative breathalyzer results) for at least 24 h. Alcohol Cue VR Paradigm Development The Alcohol Cue VR Paradigm is based on the Nicotine and Tobacco Product (NTP) Cue VR Paradigm developed previously by our group (Courtney et al. 2024; Liu et al. 2022) and utilizes a Unity‐based VR environment. As outlined in Courtney et al. (2024), use of the HTC Vive Pro Eye VR headset (HTC, Taoyuan City, Taiwan) enabled VR capabilities and captured eye‐related data. HTC's SRanipal SDK (https://developer‐express.vive.com/resources/vive‐sense/eye‐and‐facial‐tracking‐sdk/) was used in conjunction with Tobii's (Tobii Technology, Stockholm, Sweden) Tobii XR SDK (https://vr.tobii.com/sdk/) to provide access to data from the eye tracking hardware. Specifically, Tobii XR SDK handled object selections, determining what participants were looking at with its Gaze‐to‐Object‐Mapping (G2OM) algorithm while the rest of the data were retrieved from the SRanipal SDK. Using two hand‐held Vive controllers, the participants were free to move around (via teleportation) and interact with various objects within the VR environment. Surveys [visual analog scale (VAS): 0–100] assessing depressed mood and anxiety were presented at the start of the paradigm (following the initial training and test scenes) and additional surveys assessing subjective craving (“How much are you craving alcohol right now?”) and scene relevance (“How relevant was that scene to your own life?”) were presented following each scene within the headset; survey responses were made by adjusting a slide bar using one of the controllers. Participants were told “During the task, we will be measuring what you pay attention to, and we will be asking you to rate your craving level between each scene. Just explore everything around you until the scene changes.” Three Active scenes (Living Room, Bar, Outdoor BBQ) and three Neutral scenes (Art Gallery, Waiting Room, Library) were included in the final paradigm (see Figure 1 for screenshots of the scenes). The Active scenes included alcohol‐related and control cues, whereas the Neutral scenes contained only control cues. The 3D models for the cue objects were acquired through multiple online 3D asset stores (e.g., Sketchfab, Turbosquid, CG Trader, Unity Asset Store). Alcohol cues included various common brands of beer and liquor bottles (e.g., Budweiser, Pabst, Heineken, Jager, Leffe, Glenfiddich, Jack Daniels, Johnnie Walker, Hennessy, Malibu, multiple wines and champagnes), martini, beer, and shot glasses, red Solo cups, cocktail shakers, beer kegs, as well as the presence of human models engaged in alcohol drinking behaviors. Appetitive and nonappetitive control cues (e.g., water bottles, glasses of water, Coca‐Cola and Pepsi soda cans and bottles, cellphones, pens/pencils, magazines, candies) varied depending on the scene context. Alcohol and control cues were generally comparable in approximate size and placement within the scenes. All cues were interactable such that the participants were able to pick‐up, throw, and collide the items with other items in the scene. Four scenes (two Active and two Neutral) included the presence of at least one animated human model, whereas one Active (Living Room) and one Neutral (Waiting Room) scene were devoid of human models. All scenes included background music and audio effects consistent with the scene and the participants' interaction. The overall rendering environment was standardized within the paradigm scenes, including lighting, exposure, and other scene‐level visual settings, such that all cues (alcohol and control) had the same values within scenes. Screen captures of the alcohol cue VR paradigm. Neutral VR scenes: (a) library, (b) art gallery, (c) empty waiting room; Active VR scenes: (d) backyard BBQ, (e) bar, (f) home. Alcohol Cue VR Paradigm Procedure The Alcohol Cue VR paradigm begins with three “test scenes” which are approximately 3 min in duration depending on participant comfort and abilities with the VR hardware, which includes a practice room, increasing familiarity with controls/interface, and calibration of blinks [detailed further in Liu et al. 2022, (Courtney et al. 2024)]. Following the completion of the initial test scenes, the two mood surveys are presented (described above), and then the six test scenes are presented for 5 min each, which are pseudo‐randomized within scene type such that the general scene order is maintained (Active, Neutral, Active, Neutral, Active, and Neutral). Thus, the entire paradigm totals approximately 30 min. Alcohol Cue VR Paradigm Data Collection As described previously (Courtney et al. 2024; Liu et al. 2022), two types of data were recorded within each scene, regular time series and event‐based data. Regular time series data was collected every 10 milliseconds (100 Hz), independent of the frame time. The following data were recorded periodically: (1) timestamp, (2) raw gaze intersection point, (3) position and forward direction of the participants' headset, and (4) pupil diameter and eye openness (calculated by SRanipal SDK). The following events and corresponding timestamps are recorded when they occur: (1) blinks, including number of blinks and the object of gaze at the time of the blink, (2) button presses on the controller, including time, button pressed, and object of interaction (if applicable), and (3) object of gaze when eye gaze switches to a new object. Alcohol Cue VR Paradigm Gaze Statistics Calculation As described previously (Liu et al. 2023), for small objects of interest, the G2OM algorithm provided by the Tobii XR SDK was used, which is a machine‐learning‐based object selection algorithm that aims to improve small object and fast‐moving object tracking. To further improve object selection, we introduced an additional mechanism to “lock” the object selection when an object is manipulated such that whenever a participant actively picks up a virtual object, the object selection algorithm always selects the picked object until the participant releases the object. If the participant is not interacting with an object, the G2OM algorithm is employed, or if no small objects are within the field, naïve ray‐casting is employed. To calculate eye‐gaze statistics toward alcohol and control cue objects, four dictionaries corresponding to four different types of objects (alcohol, control, miscellaneous, background) were initialized prior to the start of participant involvement in the paradigm. These dictionaries were then used to store the cumulative gaze fixation/dwell time durations as values for individual objects belonging to each object and type. Following the completion of the paradigm, total fixation time was calculated which reflects the sum of values within each dictionary and mean fixation time which reflects the total fixation time divided by the total number of objects gazed at within each object category. Mean fixation times for alcohol and control cue analyses presented here were both derived from the Active scenes only, thus effectively controlling for differences in luminosity and sound‐related effects. Visual Probe Task A previously validated visual probe task (VPT) (Marks et al. 2015) with alcohol cues was also administered to participants following the VR paradigm. Participants were seated in front of the monitor and placed their index fingers on two arrow keys on a keyboard. They first completed a practice set of 10 trials, followed by the test set of 80 trials (40 cue trials and 40 filler trials). For each trial, two 13 cm × 18 cm images (one alcohol and one matched neutral image) were presented 3 cm apart on a computer screen for 1000 ms. Duration of eye‐gaze fixation on each of the images was collected using a Tobii Eye X eye tracker mounted under a computer monitor. Following the offset of the image pair, a visual probe (“X”) appeared in the location of one of the images (left or right side of the screen). An equal number of alcohol (20) and neutral (20) probes were presented. Participants responded to the location of the probe by pressing the left or right arrow on the keyboard. Total fixation time (ms) and response time (ms) to the correctly responded probe trials were recorded and averaged across alcohol and neutral images independently. Psychological Measures Participants underwent an extensive clinical interview and completed several self‐report questionnaires covering demographic, a structured clinical interview for DSM‐5 disorders, and substance use [90‐day Timeline Follow‐Back (TLFB) (Sobell and Sobell 1996), Customary Drinking and Drug Use Record (CDDR) (Brown et al. 1998)], and craving [Alcohol Craving Questionnaire (ACQ) (Singleton et al. 2000)] domains. Alcohol use variables extracted from the TLFB and CDDR and used for subsequent analyses included: lifetime AUD diagnosis (> 2 symptoms) and total symptom count, past‐year AUD diagnosis (> 2 symptoms) and total symptom count, past 90‐day alcohol use drink count, past 90‐day alcohol use days, alcohol recency, past‐month alcohol use days, past‐month alcohol drinks, past‐month binge episodes, past‐year alcohol use days, past‐year alcohol drinks, maximum drinks in one drink episode in past year, number of > 5 drink episodes in the past year, number of > 10 drink episodes in the past year, age of alcohol initiation, age of regular alcohol use onset, age of first binge episode, lifetime nicotine use (yes/no), past‐month nicotine use (yes/no), past‐month nicotine use episodes, lifetime cannabis use (yes/no), and past‐month cannabis use episodes (correlations between alcohol use variables are presented in Table S6). All study interview and self‐report data were collected and managed using REDCap electronic data capture tools hosted at the University of California, San Diego. Upon completion of the VR paradigm, additional assessments on VR‐related outcomes such as VR presence [i.e., the sense of being present in a computer‐generated environment; Igroup presence questionnaire (IPQ) (Schubert 2003)] and VR‐related simulator/motion sickness [Simulator Sickness Questionnaire (SSQ) (Kennedy et al. 1993)] were administered, followed by administration of the alcohol VPT. The IPQ includes items such as “In the computer generated world I had a sense of ‘being there’” and “How real did the virtual world seem to you?,” where positive scores on the scale items represent greater feelings of presence and realism in the virtual world (full instrument available online at https://www.igroup.org/pq/ipq/index.php). The IPQ total score was calculated using a simple averaging method to obtain a single average perceived presence score ranging from 0 to 100. Similarly, the SSQ was scored in concordance with procedures outlined to assess virtual reality specific sickness [VRSQ (Kim et al. 2018)], involving a simple averaging method to obtain a single average sickness score ranging from 0 to 100. Statistical Analysis While the study was not preregistered, the analyses directly tested the primary hypothesis‐generating aims specified in the funded grant proposal. Despite this, all analyses should be considered exploratory. Statistical analyses were conducted using R (4.4.2). Participants with incomplete data on key outcome variables (i.e., VR‐related outcomes) were listwise deleted from relevant analyses. Appropriate data transformations were applied to all relevant predictor variables, including Winsorizing values ≥ 3 standard deviations from the mean, and imputing missing values from predictor variables with median values across the sample (Leys et al. 2019). Across analyses, statistical significance was determined at α = 0.05. Pearson correlations were conducted on VR, VPT, and alcohol use variables to determine interrelatedness. Spearman's ranked correlations were estimated on the mean fixation time for alcohol cues across the three Active scenes to test internal reliability within the Alcohol Cue VR Paradigm. Repeated measures t‐tests were conducted to compare within‐paradigm metrics of craving and mean fixation times across cues. The following procedures were conducted to test the four aims of the project. (a) Subjective alcohol craving ratings immediately following each scene were averaged within‐participant across Active and Neutral scenes independently and differences were tested using a paired t‐test. (b) Given the number of potential alcohol and substance related predictor variables, stepwise hierarchical regression was used to identify the best predictors of mean gaze time to alcohol cues, control cues, and their combination (mean alcohol + control cue gaze fixation time) in two blocks of variables: (1) a priori relevant covariates of noninterest and (2) alcohol and other substance use‐related variables. The first block consisting of covariates included demographic (sex and age), VR presence (IPQ total score), and the outcome equivalent metric from the other condition (e.g., for mean fixation for alcohol cue models, mean fixation for control cues was inputted as a covariate)—except in the model predicting fixation time across both cue types (alcohol + control). Then, variable selection for alcohol‐related effects was completed using a forward stepwise linear regression search approach for optimal Akaike information criterion (AIC), while holding all covariates from block 1, to identify alcohol and substance related effects on VR paradigm outcomes (mean fixation times). Alcohol and substance use‐related variables included 23 total summary variables across the MINI, TLFB, and CDDR (18 alcohol‐related, 3 nicotine‐related, and 2 cannabis‐related). On balance, blocks 1 and 2 (i.e., model selection process) were repeated with neutral control variables (i.e., mean fixation time to control cues) as an outcome variable and then reproduced while controlling for the equivalent metric from the active alcohol conditions and other covariates. (c) Next, to investigate potential moderation of VR presence, interactions were estimated between IPQ total score and alcohol‐related variables included from the forward stepwise process. (d) Lastly, R 2 metrics were compared between VR and VPT models by rerunning models from block 2 while removing IPQ total score (due to its specificity to the VR) and the equivalent metric from neutral conditions (due to the neutral conditions varying across VR and VPT). R 2 Δ was calculated by subtracting R 2 from models with only age and sex included as fixed effects, and then the selected alcohol‐related variables to determine contributory variance from substance use‐specific variables. Further, R 2 values were statistically compared through evaluating the correlational contrasts between the fitted model values and the respective outcome variables (i.e., VR and VPT) (Diedenhofen and Musch 2015; Steiger 1980). Results Participant Characteristics Sociodemographic information and relevant sample descriptors are presented in Table 1. The sample (N = 105) consisted of approximately equal numbers of males and females (sex at birth; 53.3% male) and averaged 37.6 years old. Approximately half of the sample self‐identified as White and 36% as Hispanic ethnicity. TABLE 1Sample demographics (N = 105).Variable% Frequency or mean (SD)Age37.6 (13.4)Sex—% male53.3%Race—% White54.3%Ethnicity—% non‐Hispanic63.8%Education (college level)84.8%Previous VR experienceNever34.3%Once24.8%A few times29.5%Many times/regularly11.4%Alcohol use days (past 90 days)32.2 (24.9)Alcohol Craving Questionnaire (baseline)37.3 (14.0)VR presence (IPQ)Spatial presence67.1 (20.3)Involvement61.2 (23.7)Experienced realism47.6 (28.2)Total58.6 (9.9)VR‐related sickness (VRSQ)Nausea23.4 (27.7)Oculomotor23.0 (24.2)Disorientation33.9 (49.9)Total29.6 (33.2)Abbreviations: IPQ, Igroup presence questionnaire; SD, standard deviation; VR, virtual reality. Sample demographics (N = 105). Abbreviations: IPQ, Igroup presence questionnaire; SD, standard deviation; VR, virtual reality. Alcohol Cue VR Paradigm Simulator Sickness and Presence Average raw ratings within each VR‐related sickness VRSQ subscale fell below the “slight” rating level (Oculomotor mean = 0.49, SD = 0.51; Disorientation mean = 0.34, SD = 0.49; subscale possible range 0–3); equating to a mean VRSQ total score of 29.6 (see Table 1). Average raw ratings within each VR presence IPQ subscale ranged from 3.13 to 3.91 (SD range: 1.20–1.69) out of a possible range of 0–6, corresponding to ratings slightly above the “moderate”/“neutral” anchors on items pertaining to presence, spatial presence, involvement, and experienced realism. No difference in IPQ total score was observed between individuals with no prior VR experience (n = 40) and those with one or more prior VR experiences [n = 69; t(79) = −0.8, p = 0.43]. Alcohol Cue VR Paradigm Subjective Craving and Cue Eye‐Gaze Fixation Time Craving ratings immediately following Active Scenes (M = 28.0) were greater than those following Neutral Scenes (M = 12.8): t(104) = 8.2, p < 0.001. Greater alcohol use in the past 90‐days was positively correlated with condition‐level craving means [r(208) = 0.28, p < 0.001]; see Figure 2. Mean fixation time across all cues was 2687.20 ms, SD = 1030.99 ms. Mean fixation time was greater for control cues (M = 3337.28 ms, SD = 1013.72 ms) relative to alcohol cues (M = 2037.11 ms, SD = 1048.25 ms) across all participants [d = −1.3, t(103) = −13.2, p < 0.001]. Spearman's ranked correlations for mean fixation time for alcohol cues from the three Active scenes produced rho values ranging from 0.45 to 0.52 (ps < 0.001). Observed craving rating following Active (red) and Neutral (blue) scenes by past 3‐month alcohol use days. Alcohol‐Related Predictors of VR Attentional Bias Model 1: Mean Fixation Time Toward Alcohol Cues The addition of block 2 of the model predicting mean fixation time towards alcohol cues (controlling for sex, age, IPQ total score, and mean fixation time towards neutral cues) accounted for 42% of variance, and revealed the following features as significant contributors: younger age of regular alcohol use onset [t(94) = −2.3, p = 0.025, f = 0.23], less time since last alcohol drink [t(94) = −2.2, p = 0.03, f = 0.22], and fewer past‐month binge episodes [t(94) = −2.7, p = 0.01, f = 0.28]. Additional, but non‐significant, features identified by the forward stepwise model selection process included: greater maximum drinks in one drinking episode in the past‐year (p = 0.13, f = 0.16) and lifetime AUD diagnosis (i.e., endorsement of > 2 AUD symptoms) (p = 0.17, f = 0.14). See Table 2 and Figure 3. TABLE 2Forward stepwise model selection predicting mean fixation toward alcohol cues.PredictorsStep 1: A priori covariates of noninterestStep 2: Selected model from alcohol variable searchStepwise smodel selection (AIC)EstimatesdfStatistic p EstimatesdfStatistic p Cohen's f (Intercept)−602.099−0.860.39−151.094−0.180.860.02Sex202.6991.170.248.5940.050.960.00Age−7.699−1.120.27−7.294−1.060.290.11IPQ total18.8992.15 0.03 23.9942.700.008 0.28Mean fixation to control cues0.5995.97 < 0.001 0.5945.99 < 0.001 0.621412.0Age of regular alcohol use onset−40.094−2.270.025 0.231411.1Alcohol recency−30.194−2.160.03 0.221410.9Past‐month binge episodes−55.594−2.680.009 0.281409.4Maximum drinks in one drink episode in past‐year19.1941.510.130.161407.5AUD diagnosis (lifetime)375.0941.380.170.141407.4Observations104104 R 2/R 2 adjusted0.33/0.310.42/0.37 Note: Values in bold are statistically significant at the p < 0.05 level. Forward stepwise model selection predicting mean fixation toward alcohol cues. Note: Values in bold are statistically significant at the p < 0.05 level. Predicted regression line (with 95% CI) of alcohol recency plotted for mean fixation time to alcohol cues (p = 0.03) and control cues (p = 0.09) in Active scenes. Model 2: Mean Fixation Time Toward Control Cues The addition of block 2 of the model predicting mean fixation time toward neutral control cues (controlling for sex, age, IPQ total score, and alcohol mean fixation time) accounted for 38% of variance, and revealed the following features as significant contributors: fewer past‐year AUD symptoms [t(95) = −2.1, p = 0.04, f = 0.21] and more past‐month alcohol use days [t(95) = 2.11, p = 0.038, f = 0.22]. Additional, but non‐significant, features identified by the forward stepwise model selection process included: more time since last alcohol drink (p = 0.09, f = 0.18) and fewer past‐month cannabis use episodes (p = 0.15, f = 0.15). See Table S1 and Figure 3. Model 3: Mean Fixation Time Across Alcohol and Control Cues The addition of block 2 of the model predicting mean fixation time averaged across alcohol and control cues (controlling for sex, age, and IPQ total score) accounted for 17% of variance and revealed the following features as significant contributors: fewer past‐year AUD symptoms [t(99) = −2.58, p = 0.011, f = 0.26] and younger age of any alcohol use [t(99) = −2.51, p = 0.014, f = 0.25]. No additional alcohol or other substance use features were identified by the forward stepwise model selection process (see Table S2). Moderation of IPQ Total Score on Alcohol‐Related Predictors IPQ total scores did not significantly moderate alcohol‐related predictors in the mean alcohol or control cue fixation time models (p > 0.05). VR and VPT Eye‐Tracking Comparisons The correlations between the VR and VPT behavioral and eye gaze metrics were nonsignificant for alcohol‐related cues: correlation between VR mean alcohol fixation time and VPT mean alcohol fixation time [r = −0.09, t(97) = −0.91, p = 0.37]; VR mean alcohol fixation time and VPT mean alcohol response time [r = −0.09, t(97) = −0.91, p = 0.36]. The correlation between VPT mean alcohol fixation time and VPT mean alcohol response time was also non‐significant [r = 0.17, t(102) = 1.74, p = 0.08]. Table 3 displays R 2 and R 2 Δ for VR and VPT‐related models. The addition of alcohol‐related variables to regression models with only age and sex as covariates adds 0.10 variance to the VR model, compared to 0.02 and 0.06 for mean alcohol fixation time and mean alcohol response time, respectively, on the VPT. To statistically test differences between these R 2 values, a z‐test was conducted between the fitted model values and the predicted outcome. Fitted values from resulting linear regressions had a statistically larger correlation when compared against the VPT mean alcohol fixation time (z = 3.27, p = 0.001) and mean alcohol response time (z = 3.56, p < 0.001). TABLE 3Model comparison between VR fixation eye‐tracking metrics and VPT fixation and response metrics.Outcome R 2 (sex + age model) R 2 (+ alcohol‐related variables model) R 2 Δ Pearson correlation between outcome and fitted valuesComparison Z‐statistic p VR—mean alcohol fixation time0.070.170.100.43——VPT—mean alcohol fixation time0.010.060.05−0.043.270.001VPT—mean alcohol response time0.070.140.06−0.093.56< 0.001 Note: IPQ total score and equivalent neutral variable were removed from these analyses to maintain consistency across models. See Tables S3–S5 for full model output. Pearson correlations were calculated from the linear regression fitted values and the outcome variables; then, a statistical comparison between these correlations was determined through z‐statistic comparison (Steiger 1980). Model comparison between VR fixation eye‐tracking metrics and VPT fixation and response metrics. Note: IPQ total score and equivalent neutral variable were removed from these analyses to maintain consistency across models. See Tables S3–S5 for full model output. Pearson correlations were calculated from the linear regression fitted values and the outcome variables; then, a statistical comparison between these correlations was determined through z‐statistic comparison (Steiger 1980). Notably, VPT response time and VR alcohol fixation time have relatively similar R 2 (0.14 and 0.17, respectively); however, when the equivalent neutral control variables are inputted into these models, all alcohol variables in the VPT models lose significance (p > 0.05), whereas age of regular alcohol use onset (p = 0.048) and past‐month alcohol binging episodes (p = 0.015) in the VR model maintain their statistical significance (see Tables S3–S5). Discussion In this study, we tested the construct validity of the Alcohol Cue VR Paradigm by evaluating its ability to induce subjective alcohol craving and exploratorily assess eye‐gaze markers of attentional bias to naturalistic 3D cues. Criterion validity of the VR‐derived attentional bias metric was further investigated through alcohol‐related characteristic predictions of eye‐gaze markers in adults with a range of alcohol use patterns. Last, the attentional bias metric derived from the VR paradigm was exploratorily compared against metrics obtained from a validated visual probe task (i.e., 3D vs. 2D eye‐gaze metrics) to assess potential improvement of construct measurement. Attentional bias represents a behavioral marker of incentive salience (Ostafin and Feyel 2019) and has been advanced as a useful clinical phenotype of addiction across substances (Field et al. 2009). While attentional bias has been linked with alcohol outcomes across a spectrum of alcohol use patterns (Escudero et al. 2024; Garland et al. 2012; Manchery et al. 2017; Qureshi et al. 2019; Wilcockson et al. 2019), past investigations were constrained to 2D presentations which are limited by a lack of naturalistic cues (Jones et al. 2018; Shiffman et al. 2015) and frequently rely on indirect markers of bias (e.g., response time). This has led to inconsistencies in findings and questions regarding generalizability to real‐world settings (Pennington et al. 2019; Suffoletto et al. 2020; Topbas et al. 2025). VR cue exposure paradigms have enhanced researcher's ability to invoke craving across substances (Acker and Mackillop 2013; Baumann and Sayette 2006; Bordnick et al. 2008, 2009; Culbertson et al. 2010; Ghita and Gutierrez‐Maldonado 2018; Hone‐Blanchet et al. 2014; Ryan et al. 2010; Tamburin et al. 2021; Traylor et al. 2011) and many VR head‐mounted displays now include built‐in eye‐tracking which make objective assessment of attentional bias easily attainable. Yet this is the first study, to the authors' knowledge, to investigate direct measures of attentional bias in the context of alcohol‐related cues using VR. Participants' subjective ratings of craving following Active scenes were greater than those following Neutral scenes, supporting the construct validity of the Alcohol Cue VR paradigm as an effective cue‐exposure task. These findings are in line with prior research demonstrating increased craving following substance cue‐presentations in 3D VR environments (Acker and Mackillop 2013; Baumann and Sayette 2006; Bordnick et al. 2008, 2009; Culbertson et al. 2010; Ghita and Gutierrez‐Maldonado 2018; Hone‐Blanchet et al. 2014; Ryan et al. 2010; Traylor et al. 2011). Greater density of past 90‐day alcohol use was associated with higher craving ratings across both active and control scenes (see Figure 2), which aligns with prior research demonstrating greater craving is associated with increased alcohol use (Murphy et al. 2014; Vafaie and Kober 2022). When averaging across all participants, mean fixation time to control cues was approximately 1 s longer than alcohol cues, suggesting the absence of an overall alcohol attentional bias effect during the VR paradigm. However, several alcohol‐related variables were identified by model selection to be significantly associated with increased alcohol‐specific attentional bias (controlling for control cue fixation time), including earlier age of regular alcohol use onset, more recent time since last alcohol drink, and fewer past‐month alcohol binge episodes (Table 2). Additional, nonsignificant, alcohol‐related variables identified as important for the model included greater maximum drinks in one drinking episode in the past‐year and endorsement of lifetime alcohol use disorder diagnosis. Although nonsignificant as independent predictors, these variables were retained by the model selection process due to their contributions to overall model fit indices and, when combined with the significant alcohol‐related predictors, accounted for approximately 10% of the variance of alcohol attentional bias over and above the known covariates of sex, age, and VR presence. Notably, five additional, nonalcohol substance use variables (i.e., cannabis and nicotine use measures) did not survive alcohol attentional bias model inclusion showcasing the specificity of the Alcohol Cue VR Paradigm to alcohol‐specific characteristics. In contrast, greater attentional bias toward control cues, controlling for alcohol attentional bias (model 2), was significantly associated with predominately less severe alcohol and cannabis use characteristics (e.g., lower past‐year AUD symptom count, longer duration since last drink, less past‐month cannabis use), suggesting those with less severe alcohol and cannabis use histories are less drawn toward alcohol cues in the scenes. Consistent with previous research identifying attentional bias differences as a function of increased alcohol use severity (Bollen et al. 2022; Escudero et al. 2024), greater current AUD severity and older age of any alcohol use onset predicted less overall attention toward all cues in the present study (model 3). These results may be indicative of broader sustained attention/executive functioning weaknesses among those with heavier alcohol use patterns and more problematic alcohol use (Powell et al. 2024; Stephan et al. 2017). Thus, the traditional contrast scoring method for alcohol attentional bias estimation (alcohol minus control) may not be the best approach to capture clinically relevant alcohol‐specific relationships in a heterogenous sample of individuals with varying alcohol use histories; instead, controlling for baseline (control) values in the statistical models appears to capture specific alcohol‐related effects while accounting for differences in baseline attentional processes. In contrast to Escudero et al. (2024) who reported reduced attentional bias following prolonged abstinence in AUD, the current analyses identified both remote (e.g., age of onset of regular use) and recent (e.g., days since last drink) alcohol use variables as important predictors of bias. This discrepancy suggests that there may be more nuance to the stability of the attentional bias construct, and the stepwise regression approach yields further hypotheses to follow up on in future analyses. Moreover, findings of these exploratory analyses warrant more longitudinal research to confirm if and how alcohol attentional bias changes over time in individuals with varying alcohol use histories. An important consideration for VR research is the extent to which agent perception of presence/immersion in the virtual world affects the outcome of the study (Melo et al. 2023). It was hypothesized that participants' rating of VR presence in the Alcohol Cue VR Paradigm would moderate the predictive ability of certain alcohol‐related characteristics on eye‐gaze metrics [i.e., higher VR presence and higher alcohol use characteristics would theoretically associate with increased attentional bias (Ferrer‐Garcia et al. 2010; Hernández‐Serrano et al. 2021)]. Surprisingly, VR presence did not moderate alcohol‐related variables for eye‐gaze fixation time to alcohol and control cues. This may reflect the robustness of the attentional bias construct and is consistent with the numerous investigations with positive results that employed 2D (low immersion) stimuli. Consistent with this theory, clinically relevant alcohol‐related characteristics, collectively, were associated with eye‐gaze metrics from the alcohol VPT. While not substantial (i.e., large increases in R 2 Δ), alcohol‐related predictors explained a significantly larger proportion of the variance in the VR attentional bias metrics relative to the VPT derived response and eye‐gaze fixation metrics (Table 3). This suggests a potential for small increases in attentional bias effect sizes when substance cue stimuli are presented as 3D objects within interactive environments versus 2D pictorial objects (Girondini et al. 2025). Importantly, VR‐based alcohol attentional bias appears to be more specific to alcohol cues as compared to 2D‐derived metrics, as evidenced by the survival of significant alcohol predictors in the VR models while controlling for the corresponding neutral control condition (Tables S3–S5). This distinction between VR and VPT models on alcohol outcomes may account for the low correlation between measures and suggests that 3D VR cues may be especially useful for assessing more subtle changes in alcohol attentional bias over time. Alternatively, it is possible the two tasks may be measuring slightly distinct cognitive processes that account for differences in explained variance in alcohol use outcomes. Notably, the potential benefits of 3D VR over 2D image presentation extend beyond increased immersion. VR allows for the ease of presentation of multiple cues, providing a form of cue personalization without the need to change stimuli between participants, and facilitates a more directly translatable experience to complex real‐world settings in which alcohol is typically consumed. Additionally, the free‐viewing nature of the Alcohol Cue VR Paradigm also facilitates an unconstrained assessment of gaze times which may be better able to capture difficulties with cue disengagement. Traditional attentional bias tasks such as the visual probe and Stroop tasks often have short durations of stimulus onset asynchrony (SOA), leading to potential trial carryover effects. Even with tasks designed to explicitly measure attention disengagement, the SOA is often only 500–2000 ms (Field et al. 2009); however, the current data suggest that when free‐viewing a complex scene with many cues present, individuals spend on average 2687 ms engaged and attending to a single cue irrespective of alcohol use history and cue type. It will be important for follow‐up studies to continue to investigate the role of VR‐related immersion, as well as the translatability of eye‐gaze metrics derived from free‐viewing complex scenes, on clinically relevant substance use outcomes. There are several limitations to be considered in the context of this study and analysis. First, this investigation is cross‐sectional in nature, therefore causality regarding the relationships between attentional bias and alcohol‐related characteristics cannot be ascertained. Similarly, given the single assessment session, test–retest reliability of the Alcohol Cue VR Paradigm attentional bias estimates cannot be derived from this data; however, internal reliability estimates suggest the paradigm has moderately strong reliability across scenes. Future longitudinal studies with multiple VR and prospective alcohol use outcome assessments are planned to address these gaps. In line with this, an important future direction is to investigate whether attentional bias to alcohol‐related VR cues can be modified through treatment (e.g., psychosocial or pharmacological) and serve as a clinical treatment endpoint. Second, exploratory stepwise regressions were employed here which carry limitations (Whittingham et al. 2006) but can be beneficial for hypothesis generation to confirm in independent samples. Future studies could utilize the present findings for subsequent model selection approaches. Third, recruitment criteria for the study consisted of excluding those with current and severe DSM‐5 psychiatric disorders and visual/vestibular problems, which limits generalizability to populations with comorbid AUD and other severe psychopathology and those with the inability to complete VR‐related paradigms (e.g., blindness, medical conditions impacting vestibular systems, etc.). It will be important for future studies to continue to address overall generalizability of the Alcohol Cue VR Paradigm across a diversity of samples. Finally, the Alcohol Cue VR Paradigm consists of scenes with interactive cues where participants can freely participate with their environment with no specific task‐demands; therefore, these findings may not directly align with other cue‐exposure or attentional bias tasks, including the VPT. Explicit task demands could easily be integrated within the VR environment to directly assess related cognitive constructs of interest (e.g., response inhibition, approach/avoidance). In summary, the present study supports the advancement of the Alcohol Cue VR Paradigm as a valid alcohol cue exposure and attentional bias task, with multiple exploratory relationships observed between eye‐gaze fixation towards alcohol cues and clinically relevant alcohol use characteristics. These results, combined with the extant literature, suggest that this type of immersive VR attentional bias assessment may provide small alcohol‐specific increases in the sensitivity to explain alcohol use characteristics compared to traditional 2D presentations. Further analyses on the nuanced relationships between alcohol attentional bias and level of alcohol use severity are planned. Funding This work was supported by the National Institutes of Health (Grant Nos R21 AA030284, T32 AA013525, F32 DA064409, UL1 TR001442) and California Tobacco‐Related Disease Research Program (TRDRP) grant number T30IP0962. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: Forward stepwise model selection for mean fixation to control cues. Table S2: Forward stepwise model selection for mean fixation across all cue objects (alcohol and control). Table S3: Stepwise model building of VR alcohol mean fixation for VPT comparison. Table S4: Stepwise model building of VPT mean fixation time for alcohol cue model testing for VR comparison. Table S5: Stepwise model building of VPT mean response time for alcohol cue model testing for VR comparison. Table S6: Pearson correlational matrix for all 23 variables used for forward model selection process.