Contextual correlates of alcohol use: Evidence from scattered perspectives

Our understanding of peoples’ behaviors in their immediate environment largely depends on the source of the collected data. In sociology, a distinction is traditionally made between the people involved in a situation (the insiders) and those who observe the former (the outsiders). The insiders are the actors–or ‘subjects’–in the situation and can provide knowledge permeated by the history and symbolic meaning of their current situation [6]. In contrast, the outsiders–or observers–are blinded to such specificities and rather consider the situation as an ‘object’ of observation [6]. Collecting information from both perspectives–subjective experience of the participants and objective observation–are therefore ideal to gain a comprehensive picture of a given drinking occasion [7, 8]. However, this has not been implemented in alcohol research and most studies only reflect one perspective at a time.

Self-reports are typically used to identify which characteristics of the context, perceived subjectively by the participants, are associated with alcohol use. Cross-sectional retrospective surveys [e.g., 9–11] and diary-based methods, such as ecological momentary assessment (EMA) [e.g., 12–15], concur with the observation that people are more likely to drink alcohol in specific locations and in the presence of large groups of friends. For example, Canadian students reported consuming more alcohol per occasion when they were at a party (as opposed to any other drinking occasions), in a bar or at home than in a restaurant, as well as with large groups of friends [10]. Similarly, young Swiss adults were found to drink higher amounts of alcohol when drinking occasions started in private places before going out [14], and in the presence of a higher number of friends [12, 13]. Additionally, evidence from qualitative interviews revealed that young people take control over their experience of drunkenness in private settings by staging the atmosphere, notably with selected music styles and lighting [16, 17]. However, a major drawback of self-reports is the risk of omissions and recall errors, as respondents are known to forget details of their experiences within a couple of days [18, 19].

In-situ observations and experiments are generally used to document characteristics of drinking settings and their association with alcohol use from an external observer viewpoint [20–22]. For instance, in-bar experiments manipulating levels of music loudness revealed that patrons’ drinking pace and amounts increased in louder environments [23, 24]. To explain this phenomenon, the authors argued that high sound levels create a high level of arousal among patrons, who enhance their behavioral response toward the stimulus [24, 25]. While music at moderate sound levels eases the socialization process in pubs and nightclubs [26], loud music likely impedes conversation and likely increases patrons’ drinking pace. Besides this, observational studies have shown that bar patrons’ intoxication levels are also associated with promotion of soft or energy drinks, poor washroom facilities or the presence of a dance floor, but not with environmental brightness or perceived music sound level [27]. A major advantage of in-situ sensing (e.g. with sound level meters) and observation is that these methods provide standardized measures that can be compared across locations and times. However, such methods cannot easily be used outside of publicly accessible locations and cannot follow individuals when they change locations.

Overall, this body of findings provides concurring evidence that particular characteristics of the social (e.g., the size of the drinking group) and physical context (e.g., the type of location and loudness level) can be associated with peoples’ drinking behaviors, both when documented by participants and observers. However, an important limitation of the existing literature, besides the fact that each study only accounted for the perception of either the participants or the observers, is the limited types of locations investigated, especially with in-situ observations. While most evidence has been collected in commercial venues (bars and nightclubs), little is known about the drinking context characteristics of other public spaces, such as parks, streets, and means of transportation, as well as in private settings.

Capturing contextual characteristics with 10-second video clips

The present study elaborates on the technological developments of smartphones to explore the feasibility of assessing the physical and social characteristics of various drinking settings from both actors’ and observers’ perspectives. As part of a larger study on young adults’ nightlife and drinking behaviors, we developed a smartphone application that is able to collect various behavioral and contextual data in real time by means of questionnaires, built-in sensors, pictures, and short video clips [28]. Among other things, participants described the perceived levels of brightness, loudness, and the number of people present every time they had an alcoholic or a non-alcoholic drink. Additionally, for every drink in a new location, they were requested to record a 10-second panoramic video clip of the surrounding context, to be later annotated by research assistants and analyzed via computer algorithms.

Video clips were requested to capture data that could not be obtained by the smartphones built-in sensors and that represented an additional perspective other than the participants’ subjective perception of the situation. Our intention in requesting a 360-degree panorama of the situation was to obtain a comprehensive audio-visual representation, similar to the one that could be obtained from an external observer’s perspective [7, 8]. Additionally, given the videos were recorded at the same time and place as the consumed beverage, they provided data with a very high degree of ecological validity. Lastly, an important asset of the video clips is the ability to document contextual elements that participants may be unaware of, or do not pay attention to, and would therefore not report accurately a posteriori.

Exploratory analyses of the videos content revealed that typical levels of brightness varied across location, with bars and nightclubs being darker and louder than other public spaces and private places [5]. Participants’ ratings of brightness and loudness correlated poorly to moderately with those of annotators based on the videos. In contrast, annotators’ ratings were globally consistent with those measured by computer algorithms [5]. These findings suggest that the video content can effectively be used to assess contextual characteristics and that annotators and computer algorithms rated some ambiance attributes differently compared to the participants. However, limitations in analytic approach, such as the absence of a systematic annotation procedure (i.e., several annotators worked on distinct subsets of videos) and the absence of conversion of loudness and brightness levels to the human perception specificities (i.e., human perception of loudness and brightness is not linear; see Measures section below) called for an improved approach.

For the present study, five independent annotators annotated the entire collection of 843 video clips, and the brightness and loudness extraction algorithms were redeveloped to match human perception specificities. Additionally, we extended the analytic frame to the social context, namely the number of people appearing in the videos, as the presence and number of people was frequently highlighted as a core component of the context by previous research [10, 12, 13]. Although other momentary circumstances were shown to be associated with participants’ choice to drink alcohol or not (e.g., event-specific drinking motives and expectancies [29, 30], parental supervision [31, 32]), we selected characteristics that could be measured in a comparable way by participants, annotators, and computer algorithms. Given that algorithms are not as capable as humans to identify social bounds or cognitions, we restricted the focus on factual characteristics of the physical and social context.

Study aims

The Youth@Night study was conceived as a natural experiment in which participants documented drinking situations, in various settings, where they could freely choose to drink alcohol or not. Given the public health burden of heavy alcohol use among young people on weekend nights [33, 34], the main aim of this study was to identify which contextual characteristics were associated with participants’ choice to drink alcohol or not. Additionally, given the unique opportunity to compare three measures of the same contextual characteristics from different data sources, the study also aimed to explore the specificities of the data from each source and their ability to predict participants’ choice to drink alcohol or not in various contexts.

The first part of the analysis explored the distributions and bivariate associations of the raw ratings of brightness, loudness, and attendance provided by participants, annotators, and computer algorithms.

The second part investigated how levels of brightness, loudness, and attendance measured by each source vary depending on whether the participants were drinking an alcoholic or a non-alcoholic drink in seven different types of locations including bars, nightclubs, restaurants, public spaces, and homes.

The third part aimed to identify which of the three contextual characteristics were associated with participants’ choice to drink alcohol or not, above and beyond each other, using series of logistic regression models.

The last part of the analysis evaluated the ability of each set of contextual measures from each of the sources to predict participants’ choice to drink alcohol or not, by comparing the percentage of variance explained, the goodness of fit, and the inference accuracy of each model.

Study design

Participants were recruited in the streets of the two major nightlife hubs in Switzerland (Lausanne and Zurich) between 9pm and midnight on Friday and Saturday nights in the first three weekends of September 2014. In defined nightlife areas, research assistants approached every nth person crossing a ‘virtual line’ on the street [35]. Eligibility criteria were being aged between 16 and 25, owning an Android smartphone on which the Youth@Night app could be installed, having consumed alcohol at least once in the past month (legal drinking age is 16 for beer and wine in Switzerland), and having been out in the city at least twice in the past month. After explaining the aim and procedure of the study, recruiters recorded volunteers’ email address, and volunteers then automatically received an email containing a link to the study website and the online consent form. After signing the consent form and completing the baseline questionnaire, participants installed the Youth@Night application on their smartphone. This app was specifically developed to record various aspects of the participants’ Friday and Saturday nights, including the types of drinks consumed and the social and physical characteristics of locations attended over seven consecutive weekends, using questionnaires, pictures, video clips, and sensors [28]. The Lausanne and Zurich Cantonal Ethics Committees for Research on Human Beings (protocol 145/14) approved the study and did not request consent from parents of participants under the age of 18.

Extraction of audio and visual cues from the videos

After the fieldwork, we developed an online annotation task to extract visual and audio cues from the videos recorded by the participants [36]. Five independent annotators watched, in a random order, the entire set of videos and annotated the type of location, the loudness and brightness levels, the number of people visible, and other situational cues (e.g., ongoing activities, people’s reaction to being filmed). After a training session, annotators completed the annotation task at their own pace over two months using their computer.

Samples of participants and situations

In total, 3,092 people were approached in the two cities. Of those, 1,119 (36.2%) did not have an Android smartphone, 859 (27.8%) were not interested in participating in the study and 233 (7.5%) were outside the required age range of 16 to 25. Of the 881 who agreed to participate, 629 (71.4%) signed the online consent form, 367 completed the baseline questionnaire (41.7%) and 241 documented their nights using the smartphone app (27.4%) [35]. The sample of 241 participants (mean age = 19.0 [SD = 2.4]; 46.5% women) was slightly younger than the rest of the eligible pool of passers-by approached on the streets (mean age = 19.6 [SD = 2.8]; t = 2.76, p = .006) but similar in terms of gender ratio (41.2% women; χ² = 2.24, p = .134).

Participants documented 2,420 drinking situations, each with a picture of the drink and a description of the context, 1,394 with labels of the locations, and 843 with video clips of the context (see Measures section below for the full sequence of questionnaires). While preparing the data for the analysis, we assigned the type of location to 987 additional drinking situations based on the sequence of events during the night, the GPS coordinates of the location and contextual cues visible in the background of drink pictures, and excluded 18 situations because the related videos were entirely black and silent. The analyses were conducted on a total sample of 2,358 drinking situations documented with data on the drink and the location, and 825 with a video. These situations were reported by 210 participants who were slightly older than the rest of the participants who installed the app (mean age = 19.2, SD = 2.4, t = -2.1; p = .037) but similar in terms of gender ratio (47.1% women; χ² = 0.07, p = .791). Half of the 2,358 situations were reported by 42 participants (20%).

Measures

Analytic strategy

Descriptive statistics were first used to report the average levels of brightness, loudness and attendance from the participants’, annotators’ and algorithms’ perspective. Level of agreement between participants’, annotators’ and algorithms’ ratings of brightness, loudness, and attendance were illustrated using correspondence matrices and measured with Pearson’s correlations and paired-sample t-tests. Effect sizes of correlations and t-tests were adjusted for the effect of the drinking situations being nested within participants using Stata 14 [35]. Correlation coefficients under .40 were considered as ‘poor’, between .40 and .59 as ‘fair’, between .60 and .74 as ‘good’ and above .75 as ‘excellent’, following the guidelines by Cicchetti [39].

Secondly, descriptive statistics were used to calculate the average levels of brightness, loudness and attendance from participants’, annotators’ and algorithms’ perspectives for each type of location. The magnitude of the difference in levels of brightness, loudness, and attendance between situations with and without alcohol use were assessed using Cohen’s d estimates and confidence intervals. Effect sizes of 0.2 were considered as small, 0.5 as medium and 0.8 or more as large [52] and the effects were considered significant if the confidence interval did not contain ‘0’ [53].

Thirdly, a series of multilevel logistic regression models were estimated to investigate the mutually adjusted associations of contextual variables with the likelihood of drinking an alcoholic versus a non-alcoholic drink. Nine models were estimated to explore the location-specific associations of the context with alcohol, separately for each data source and the three main location types (i.e., commercial venues, public spaces, and private places; due to the small number of observations in some locations types, the models were estimated only at the coarser level) using the following formula: ; for j = 1,…, n participants, with participant j consisting of i = 1,…,n_j observations. Situation-level predictors also included the number of alcoholic drinks consumed earlier in the night because participants’ perception of their context might be altered by their level of inebriation. All models were estimated using the maximum likelihood robust (MLR) estimator to account for deviation from normal distribution using Stata 14 [35]. Reported effect sizes were odds ratios (OR) and 95%-confidence intervals.

Finally, the ability of each set of contextual measures to predict participants’ choice to drink alcohol or not were compared based on the percentage of variance explained, goodness of fit, and inference accuracy of the nine logistic regression models. These indices were computed using the following approaches:

1. The percentage of variance explained was estimated using the McKelvey & Zavoina’s pseudo R-squared using the “fit_meologit_2lev” package for Stata [54]. In Table 3, we report the estimate based on the fixed effects only, which is a good estimate of how the model might fare in a different sample (in opposition to the estimate based on the fixed and random effects, which is a good estimate of how well the current model is calibrated to the current data).
2. The goodness of fit was estimated using a variation of the Hosmer-Lemeshow test for multilevel logistic models in Stata [55]. We report the Hosmer-Lemeshow Chi-squared and p-value based on the fixed effects only (see rationale above), and a p-value below .05 indicates that the model is not a good fit.
3. Inference accuracy was estimated using a two-class inference task, with "drinking alcohol" versus "drinking no alcohol" as a target classes, using python with scikitlearn and keras [56]. We trained and tested six models types: random forest classifier, naive Bayes, gradient boosting, XGboost, AdaBoost, and support vector machines (in this decreasing order of overall inference accuracies), and repeated each inference task for 10 iterations. Because of the data imbalance between the number of points for each class (i.e., situations with or without alcohol), we oversampled the minority class (i.e., the class with fewer data points) [57] in order to obtain balanced datasets in each iteration, and then, evaluated the models using 10-fold cross-validation [58]. In this paper, we report the results for random forest classifiers (RFs) which provided the highest accuracy values using n-tree values between 200 and 500 (results of the other model types to be obtained from authors upon request).

Levels of brightness, loudness, and attendance per data source

As seen in Table 1 and Fig 1, annotators provided the lowest ratings for all three characteristics, while the loudness algorithm rated the environment as being louder than the two human sources. Annotators rarely rated brightness and loudness levels using the maximum level of ‘4’, resulting in narrower and lower average ratings than those of the participants (Fig 1). Similarly, the brightness algorithm never attributed the maximum score for brightness levels. Regarding attendance, participants and the computer vision algorithm reported the presence of six people on average, while annotators reported an average of five. For each characteristic, ratings from the three sources were all positively correlated (Fig 1). Yet, the level of agreement between participants and both annotators and algorithms was lower (r = .21 to .57; all p < .001) than between annotators and algorithms (r = .62 to .82; all p < .001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Correspondence matrix, Pearson’s correlations and mean difference tests of participants’, annotators’ and computer algorithms’ ratings of brightness, loudness and attendance in the 825 drinking situations.

https://doi.org/10.1371/journal.pone.0250443.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Average ratings of brightness, loudness and number of people per data source, and bivariate correlations with drinking behaviors and participants’ demographics.

https://doi.org/10.1371/journal.pone.0250443.t001

For all sources, loudness was positively correlated with attendance (r = .34 to .63; all p < .001; Table 1). In addition, based on annotators’ and algorithms’ ratings, brightness and loudness were negatively correlated (r = -.27 and -.31; all p < .001). Concurrent alcohol use was positively correlated with loudness and attendance for all sources (r = .16 to .47; all p < .001), and negatively correlated with brightness for annotators (r = -.23; p < .001) and algorithms (r = -.27; p < .001). Overall, patterns of ratings appear very similar across all three sources, and especially between annotators and algorithms.

Variations across location types and alcohol use

As shown in Table 2, the average levels of brightness, loudness, and attendance varied across types of locations (see superscript letters indicating mean differences in each column). For example, all three sources rated restaurants, modes of transport, and events among the brightest environments, while public parks and nightclubs were rated among the darkest. Nightclubs, events, and bars were rated among the loudest environments, and private places as the most quiet. Lastly, nightclubs, events, bars, and restaurants were rated as the most attended places, whereas private places were the least attended. Interestingly, in dark environments the algorithm detected a lower number of people than the participants, highlighting the difficulty the algorithm had in detecting shapes in dark videos.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Number of situations, and levels of brightness, loudness and attendance, per data source, location type, and type of drink consumed (alcoholic vs. non-alcoholic).

https://doi.org/10.1371/journal.pone.0250443.t002

Table 2 also shows the average levels of brightness, loudness, and attendance per type of location depending on whether participants documented an alcoholic or a non-alcoholic drink. Nearly all drinks reported in nightclubs (94.2%), pubs (93.3%), and public streets and parks (86.9%) contained alcohol. Inversely, only half of the drinks reported in private places (52.9%) and restaurants (50%) contained alcohol. Overall, similar variations in contextual characteristic levels between situations with and without alcohol use could be observed for all sources (e.g., all sources rated nightclubs as being darker, louder, and less attended when alcohol was consumed) but most effects were either of small (absolute value of Cohen’s d under 0.2) to medium magnitude (under 0.5) or were not significant (Cohen’s d 95%-confidence interval contains ‘0’). Nevertheless, a couple of noteworthy effects can be observed. First, all sources rated private places as much louder (+1 on the 5-point scale; d = 0.80 to 1.21) and much more crowded (about 3 times more people; d = 0.67 to 0.87) when alcohol was consumed. Additionally, private places were rated as darker by annotators (d = -0.29) and the algorithm (d = -.0.40) when alcohol was consumed. Regarding bars/pubs, annotators (d = 0.70) and the algorithm (d = 0.87) rated the location as being brighter, and participants as being louder (d = 0.50) when alcohol was consumed. Lastly, annotators rated events as being louder (d = 0.87) and more attended (d = 1.07) when alcohol was consumed. Unsurprisingly, the high attendance at festivals, concerts, or sporting events resulted in a particularly high number of people identified by the algorithm.

Mutlivariate associations of the context with alcohol use

Table 3 presents the extent to which variations in brightness, loudness, and attendance, perceived by either participants, annotators, or computer algorithms are associated with the consumption of an alcoholic drink (versus a non-alcoholic drink, as the reference category) in the three major types of nightlife settings. In commercial venues, results show that participants were more likely to drink alcohol when the context was reported as being less bright by all three sources (OR_participants = 0.67 [CI: 0.47–0.94]; OR_annotators = 0.53 [CI: 0.33–0.85]; OR_algorithm = 0.58 [CI: 0.37–0.88]) and louder by participants (OR = 1.40 [CI: 1.02–1.92]) and annotators (OR = 2.45 [CI: 1.25–4.80]). In public spaces, no clear association of alcohol use with brightness, loudness and attendance levels was found for all three sources. Finally, the likelihood of drinking alcohol in private places was associated with all three investigated contextual characteristics. In private, alcohol use was more likely with increased loudness according to all three sources (OR_participants = 1.74 [CI: 1.31–2.32]; OR_annotators = 3.22 [CI: 2.06–5.03]; OR_algorithm = 2.62 [CI: 1.83–3.76]), and with higher attendance according to the participants (OR = 1.10 [CI: 1.03–1.18]) and the algorithm (OR = 1.19 [CI: 1.07–1.32]), and with reduced brightness according to the algorithm (OR = 0.64 [CI: 0.44–0.94]). No effect of the number of prior drinks consumed was found.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Multivariate associations of contextual characteristics with the likelihood of consuming alcoholic drinks in different settings, separately per data source, and models performance indices.

https://doi.org/10.1371/journal.pone.0250443.t003

Models fit and inference accuracy

Table 3 also presents the percentage of variance explained, the goodness of fit, and the inference accuracy for each of the nine logistic regression models. For all models, the p-value of the Hosmer-Lemeshow test was non-significant (p >.05) showing no evidence of poor fit. Regarding the percentage of variance explained, contextual variables explained more variance in private places (R-squared between 0.34 and 0.46) than in the two other location types regardless of the data source type. Yet, no data source seemed to outperform the others overall since annotators’ reports explained slightly more variance in commercial venues than the two other sources, while it was the case for participants’ ratings in public spaces and for algorithms’ ratings in private places. Lastly, results of the cross-validation task showed that the inference accuracies of the models based on the annotators’ ratings (80% to 84%) and the algorithms’ ratings (76% to 86%) were higher than those of the models based on the participants’ ratings (69% to 71%).

Which data source to favor

Given the unique opportunity to compare three measures of the same contextual characteristics from different data sources, one study aim was to explore the specificities of the data collected from each source and their ability to predict participants’ choice to drink alcohol or not, in general and in different types of locations. Alongside the development of smartphones and sensors to unobtrusively collect data on people’s real-life behaviors [4, 5], this study explored whether the participants’ reports of the context can be replaced by similar measures from other sources.

Differences in raw ratings (e.g., the algorithm attributed higher loudness levels than the other sources) show that neither algorithms (despite calibrating brightness and loudness levels to match human perception abilities), nor external annotators can effectively substitute for the actors’ in-situ experience in absolute terms. For instance, unlike participants, annotators and the algorithm almost never selected the highest score for brightness. Additionally, the algorithm almost exclusively rated nightclubs, streets, and parks as dark or very dark, while several participants reported them as being relatively bright. This discrepancy can partly be explained by the different ways smartphone cameras and human eyes function: camera sensors adapt by increasing contrast in very bright environments (so that videos appear moderately bright) [59] and the human eye adapts to dark conditions [60]. Similarly, the finding that the algorithm tended to rate the context as louder than humans (participants and annotators) might be explained by the fact that the audio sensors capture and account for very high energy sounds that cannot be perceived by humans. Differences between participants’ and observers’ raw ratings may also result from the conditions in which participants recorded the videos. Participants may have provided biased representations of the situation by, for example, standing close to loudspeakers, talking while recording in a silent place, or recording videos in another place than the one described in the in-situ questionnaire (e.g., chill-out room of a nightclub). They might also have not complied with the instructions on how to record the video, e.g., by failing to record a 360-degree panorama or focusing more on bright or dark zones. Occasional lack of compliance to instructions appears likely, given the high consistency between annotators’ and algorithms’ ratings.

However, above and beyond these differences in raw ratings, the patterns of bivariate associations across contextual characteristics and the ratings order per location type (e.g., loudness: nightclubs > pubs > travelling > private places) were similar for all three sources. These results suggest that, in relative terms, videos captured relevant variations of contextual characteristics in a similar way as was experienced by the participants. The three data sources can thus be considered as being partly interchangeable for correlational and regression analyses.

Regarding external validity, results of the inference task, which estimated the ability of any random subset of observations to replicate the findings of the rest of the sample, revealed higher inference accuracies for the models based on the annotators’ and on the algorithms’ ratings than those based on the participants’ ratings. This result highlights the utility of annotations and of algorithmic analyses to obtain measures of the context that are as standardized as possible, notably by having multiple rounds of annotations with each annotator working on the entire set of videos. Their ratings are therefore less fluctuant, or more predictable, than those of participants. Yet, this does not mean that annotators’ and the algorithms’ ratings would necessarily perform better at predicting alcohol use. In fact, results showed that annotators’ and the algorithms’ ratings explained more variance in private places and in commercial venues, but that participants’ ratings explained more variance in outdoor locations. In the present case, in line with the finding that contextual characteristics were weakly associated with alcohol use in public spaces, these results suggest that the choice of drinking alcohol in such locations might rather relate to other factors (e.g., night-level drinking intentions [61], off-licensed alcohol prices [62]) that could be better reported subjectively by the participants than sensed by observers.

These results have important implications for future research on context-dependent behaviors. Unless researchers are primarily interested in raw ratings and despite the fact that annotators or algorithms perform slightly worse in some locations, future studies are recommended to collect contextual data via sensors or media files whenever possible. This would reduce participant burden [63], and save participants’ time for reporting impressions, cognitions or other momentary circumstances that cannot be documented in another way. Additionally, the high correlations between annotators’ and algorithms’ ratings, as well as the high inference accuracy scores obtained for algorithms’ ratings, suggest that part of the research cost, burden, and privacy issues may be alleviated by using algorithmic analyses for extracting basic information from videos rather than annotators.

Contextual correlates of alcohol use

A second aim of the study was to investigate how variations in brightness, loudness, and attendance identified by the participants, annotators, and algorithms relate to whether alcohol is consumed. The results notably corroborate previous evidence showing that increased odds of drinking alcohol are associated with larger numbers of people present based on participants’ reports [12, 61, 64], as well as with higher levels of loudness in pubs and nightclubs from an external observer’s perspective [21, 24]. Although the study design does not allow us to determine whether participants choose to attend darker venues for drinking or whether changes in the venues context influenced the choice to order an alcoholic drink, the consistent associations between characteristics of the physical context and alcohol use might have implications for public health. In the same way that alcohol use was experimentally proven to increase with music loudness levels [23, 24], this study suggests that manipulating the brightness level might influence alcohol use. Therefore, similarly to policies regulating maximum loudness levels in nightclubs and events to prevent hearing loss [65], minimum brightness levels might be implemented and evaluated to determine whether those constitute an effective structural prevention measure to reduce alcohol intoxication and related harms.

The present study also extends the existing literature by providing detailed and systematic results on contextual characteristics outside of commercial venues. Among all location types investigated, the interplay between contextual characteristics and drinking was particularly evident in private places. Qualitative research provided convergent indications that young people can influence their experience of drunkenness by staging the atmosphere (e.g., music types and volume, lighting) in private places [16, 17, 66]. Yet, such evidence had not been replicated in quantitative research because private places only recently became accessible for large-scale in-situ observations and measurements thanks to the growth of smartphone ownership. The findings of this study that private places were louder (and darker according to algorithms) when alcohol was consumed, confirms previous evidence that people prepare their homes when drinking on a weekend night, e.g., by moving furniture, manipulating lighting, and changing the music type and volume. In fact, unlike commercial venues, private settings can be configured by their users by manipulating the attributes of the place depending on the number of attendees and the planned activities (music selection, leaving enough space for socializing, etc.) [67, 68].

Limitations and strengths

Several limitations of the current analyses should be acknowledged. Firstly, results do not provide evidence on whether changes in the investigated contextual characteristics directly increase the likelihood of drinking, or are the consequence of it, and should not be interpreted as causal relationships. The present results need replication in controlled conditions to better understand, at the individual level, if people choose to enter venues with specific contextual characteristics with the intention to drink alcohol, and, at the venue level, which contextual characteristics can promote or prevent alcohol use [24]. Secondly, a large number of analyzes were carried out in order to compare the levels of the three contextual characteristics across three sources, in multiple location types, and in two types of situations (with and without alcohol). The purpose of presenting all these results was to provide a complete overview of the differences across sources for future studies. This, however, bears the risk of an increased type 1 error. The present results should thus be considered as mainly exploratory and require replication. Thirdly, annotators worked in uncontrolled conditions, probably with self-set screen brightness settings and audio rendering devices. This may have caused some variations in conditions between annotation sessions within and between annotators. Fourthly, the study focused on contextual characteristics that could reliably be annotated by external observers and identified by computer algorithms from a short video clip. This approach de facto excluded many aspects that have also been shown to influence drinking behaviors at the event level [3, 15], such as motives and expectancies [29, 30], drinking of peers [69], activities done while drinking (e.g. chatting, watching TV, playing games) [70], or the gender composition of the drinking group [71], but that could not be identified by computer algorithms.

While questionnaire-based ecological momentary assessment studies generally request participants document their behavior within several minutes or hours [15, 72, 73], a major asset of recording contexts in videos is to force participants to provide an instant snapshot of the momentary circumstances. Thus, this study has the advantage of collecting behavioral and contextual data at the very event level, namely, the exact same time and space as the event of interest, enhancing, therefore, the ecological and internal validity of the findings. Yet, qualitative feedback after the seventh week of the app-fieldwork revealed that recording panoramic videos clips was not an ordinary action for young adults on weekend nights [74]. While recording selfies might be common on nights out, some participants may find intentionally filming the location and the people present intrusive and burdensome [28]. This might explain why only 210 out of the 241 study participants documented their drinking events with pictures and videos. To keep response burden as low as possible, we requested participants, for example, take videos only when they changed location rather than for every drink consumed. Future research should also consider the balance between data quantity and participant burden [28].