Areal-time examination of context effects on alcohol cognitions

Rebecca L. Monk and Derek Heim

of Edge Hill University, UK

Author Note

Rebecca Louise Monk and Derek Heim, Department of Psychology, Edge Hill University, St. Helens Road, Ormskirk, Lancashire, L39 4QP, UK. Email: ;

Correspondence concerning this article should be addressed to Rebecca Monk, Department of Psychology, EdgeHillUniversity, St. Helens Road, Ormskirk, Lancashire, L39 4QP, UK. Email: . Tel: +44 (0)1695 65 0940

Word count: 36953279

Running Head:Context effects on alcohol-related expectancies

“Areal-time examination of context effects on alcohol cognitions”

Background: This research used context aware experiential sampling to investigate the effect of contexts on invivo alcohol-related outcome expectancies. Method: A time-stratified random sampling strategy was adopted in order to assess 72 students and young professionals at 5-daily intervals over the course of a week using a specifically designed smart-phone application. Thisapplication recorded respondents' present situational and social contexts, alcohol consumption and alcohol-related cognitions in real-time. Results: In-vivo social and environmental contexts and current alcohol consumption accounted for a significant proportion of variance in outcome expectancies. For instance, prompts which occurred whilst participants were situated in a pub, bar or club and in a social group of friends were associated with heightened outcome expectancies in comparison with other settings. Conclusion: Alcohol-related expectancies do not appear to be static but instead demonstrate variation across social and environmental contexts. Modern technology can be usefully employed to provide a more ecologically valid means of measuring such beliefs.

Key Words: Alcohol, Social cognition, Social cognition models, Context, Expectancies, Smartphone technology, Real-time sampling

Despite longstanding awareness that people's immediate environments mediate behaviour (Bourdieu, 1977; Nyaronga, Greenfield, & McDaniel, 2009; Lott, 1996; Rosnow & Rosenthal, 1989), mostpsychological theories of behaviour and cognitions are formulated upon data which are obtained without sufficient consideration of contextual influences(Biglan Hayes, 1996;Biglan, 2001; Hayes, 2004). When using social cognition models to explain alcohol consumption this negligence might constitute a critical oversight in view of long-documented contextual influences on alcohol behaviours (MacAndrew & Edgerton, 1969).

Research indicates that alcohol-related beliefs predict consumption and, resultantly, interventions have been designed to target these beliefs to reduce drinking (c.f. Jones et al., 2001). Specifically, outcome expectancies – people’s beliefs about the likely consequences of drinking have been found to impact both the quantity and frequency of alcohol consumption (c.f. Ham & Hope, 2003; Oei & Morawska, 2004; Reich, Below, & Goldman, 2010). Specifically, high positive outcome expectancies appear to be associated with recurrent drinking in greater quantities (c.f. for example Andersson et al., 2012), whilst higher negative expectancies seem to be associated with reduced consumption (c.f. for example Stacy, Widaman, & Marlatt, 1990). While it has also been noted for some time that outcome expectancies may vary across different contexts (Wall, Mckee, & Hinson, 2000), this body of research has tended to rely on single occasion testing and on retrospective self-reports obtained within laboratory settings or non-alcohol-related environments (e.g. lecture theatres) without adequate consideration of possible contextual influences (Monk & Heim 2013a; in press). Accordingly, studies have begun to address these limitations by utilising more ecologically aware testing environments such as simulated bars (e.g., Larsen, Engels,Wiers, Granic, Spijkerman,2012) or wine tasting events (e.g., Kuendig& Kuntsche, 2012), and recent findings suggest that social contexts and alcohol-related environments are associated with increases in positive expectancies (Monk & Heim, 2013b; 2013c). While pointing to the importance of social and environmental contexts in shaping alcohol-related beliefs, these studies have tended to test participants in environments which, to a greater or lesser extent, are removed from real world drinking contexts. The current study addresses this by using an experience sampling method.

The increasing accessibility of advanced mobile devices (Katz & Aakus, 2002) has facilitated the regular, day-to-day assessment of individuals in naturally diverging contexts and has opened the field for Ecological Momentary Assessment (EMA) or Experience Sampling (Collins, Lapp, Emmons, & Isaac, 1990; Collins et al., 1998; Courvoisier, Eid, Lischetzke, & Schreiber, 2010; Killingsworth & Gilbert, 2010; Kuntsche & Robert, 2009). The present research usedsmartphone technology to enable participants to provide real-time in vivo reports with a particular focus on alcohol-related expectancies. In line with previous research (Monk & Heim, 2013a; 2013b; Wall et al., 2000; 2001; Wiers et al., 2003), it was predicted that there would be an increase in alcohol-related expectancies when assessment occurred within alcohol-related environments and in the presence of a social group (in comparison with assessments that take place in alcohol neutral environments and in solitary contexts).

Method

Design

A within participant design was utilised to investigate the effect of environmental and social contexts on participant real-time responses to alcohol expectancy questions.

Participants

72 participants comprising students (n=43) and young professionals (n=29) who were aged 18-34 years (M = 21.73, S.D = 3.64) were recruited for this study from universities and businesses in the UK (North West). The majority of the sample were White British (88.9%) and 69% of this sample were female. Baseline average AUDIT scores were 9.02 (2.07) in the student sample and 8.72 (1.28) in the business sample.

Measures

Demographic information and reports regarding personal alcohol consumption (AUDIT-C) were recorded at participants’ initial briefings. These were anonymously combined with participants’ individual responses using a unique numeric identifier. The smart-phone application ascertained participants’ environment(home, work/lecture, bar/pub/club, restaurant, sporting event, party or other) and social contexts (alone, with one friend, with twoor more fiends, with family, work colleagues or other), whether they were drinking or had had a drink (yes or no), and if so what they had been drinking (quantity). Furthermore, all participants answered a random selection of items taken from the 34-item Alcohol Outcomes Expectancy Questionnaire (Leigh & Stacy, 1993) which covers a range of outcomes, including social, sexual and emotional outcomes. However,pilot studies (n= 42) which trialled the administration of full and abridged versions of this questionnaire revealed that participants were less likely to respond when all items were included.Furthermore, if all of the 34 items had been available for random allocation, analyses would be limited as any variation observed between contexts may have been the result of variation in the expectancy measure presented (e.g. social vs. sexual expectancy items). Resultantly, it was only the six social items that were part of the question pool (three positive and three negative).In each response session, two positive and two negative expectancy items were randomly selected from the question pool and separate average scores for positive and negative expectancieswere subsequently calculated, giving a standardised maximum and minimum score of 1-6.

Equipment

A web based smart-phone application designed specifically for this research enabled participants to respond to questioning via the use of their own mobile phone – when prompted by automated SMS messages. The application was a website built using HTML and JavaScript (JavaScript's jQuery mobile library) and answers were tracked and stored using Google Analytics. The survey was designed to work on mobile phones and native mobile browsers and was web-standardscompliant. Each response session was individually tracked and involved a personally interactive user experience using tree based logic. For example, only those who responded that they consumed alcohol were asked about what they had consumed. Participants’ response mechanisms were also interactive, determined by the users’ smart-phone - for example, Iphone or Android users could indicate their response by pressing or ‘dragging’ the onscreen response items whilst those without touch screen technology responded in a fashion compatible with their phone (e.g., ‘scroll and click’). The questions were randomly selected from the database of questions using a computer-generated randomisation code. The application was designed to make the user interface as intuitive/user friendly as possible and,in accordance with recommendations (c.f. Palmblad & Tiplady, 2004), there no default answers set..

Procedure

Following ethical approval, participants were recruited and given a demonstration of the response mechanism on their personal mobile phone. In accordance with similar EMA procedures (Csikszentmihalyi & Larson, 1992; Wichers et al., 2007) and recommendations by Larson and Delespaul (1992), a time-stratified random sampling strategy was adopted (c.f. Moberly & Watkins, 2008). Pilot questionnaire data examining perceptions of online vs. real-time assessments (Response N = 108) indicated that respondents preferred SMS reminders and that five daily prompts were deemed the most acceptable number of daily participation requests. Therefore, the volunteers received five randomly allocated SMS participation prompts every day for one week.No two prompts could occur within 15 minutes (ibid) or outside 0800 - 2300 hours. Each day of participation was divided into five equal three hour periods and one prompt was randomly sent within each period (e.g., once between 0800 and 1100, once between 11 and 1400 and so on).The exact time a participant was prompted at was determined usinga random number generator - each 3 hour section was split into 15 minute blocks and the generator selected the time that the prompt would be sent, making response sessions unpredictable Upon receiving the prompts, participants activated the Application by clicking on a link provided in the SMS. The questions provided were randomly selected from the question database in order to prevent the order effects (Csikszentmihalyi & Larson, 1992).

Average completion time was recorded at 2 minutes 27 seconds and the overall study retention rate was 84.7%. Only relatively few participants completely stopped responding and dropped out (n = 8). Furthermore, respondents were removed from the sample(n = 3) where the response rate was below 40 percent, based on previous research which indicates that low response rates on substance-use-related assessments have low reliability (Shiffman, 2009).

Over the course of the week, there was the potential for participants to respond to 35 prompted sessions (5 per day for a week). There was no substantial increase in the number of missed response sessions as interaction with the application increased, suggesting that order effects were limited by the use of this technology. The average percentage of failed responses (sessions which were not completed following a prompt) was 20% per participant, with the 0800-1100 time-slot eliciting the highest number of late or failed responses. The average percentage of late responses (> 15 minutes post prompt) was 5% per participant and these late responses were excluded from subsequent analyses in order to ensure that the results could reasonably be asserted to be a representativeaccount of the specific time as opposed to a retrospective report (Delespaul, 1995).The study therefore had an average overall valid response rate of 75% per participant (26 out of a total possible 35 prompts responded to).

Analytic Strategy

Multilevel modelling (MLM) is amethod of statistical analyses which is capable of advanced portioning of variance (Tabachnick & Fidell, 2001). MLM was used as this technique can incorporate the natural complex (and related) nature of the data (Heck, Thomas, & Tabata, 2010) and look for explained and unexplained variance both between and within groups (see Goldstein, 2011). MLM is also able to deal with missing data which are to be expected in experiential sampling (Tabachnick & Fidell, 2001). In the present study variances in outcome expectancies (the dependent variable) were modelled at a number of levels: Prompts were nested within days which were nested within participants. However, given that data were not recorded at the day level (e.g. day, weather etc), it was decided that this level did not warrant inclusion within the statistical modelling. Indeed, the day of the week in which participants began the researchwas not consistent in this study (participants chose their most personally convenient starting point). This meant that no specific predictors required modelling at this level and the lack of information at this level may have unduly reduced the overall explanatory power of the model. A series of 2 level random intercept multilevel models (prompts within participants) were therefore fitted – one for each of the alcohol-related cognitions (positive and negative outcome expectancies).MLM therefore allowed analysis of variance at theprompt level (context factors) and the person level (individual differences).The resultant hierarchical random intercept multilevel model was fitted with predictor variables which were justified by correlational analyses (see Table 1). Preliminary analyses revealed no evidence of multicollinearity, residuals were normally distributedand scatterplots indicated that the assumption of linearity and homoscedasticity were met. The MLM was designed to portion variance in outcome expectancies and the predicted variance from the null and fitted models were compared in each case. Table 1 outlines the correlational analysesand the findings of these analyses were used to inform the subsequent MLMs. Any variable which significantly correlated with at least one of the dependent variables was included.

Results

Full random intercept MLMs were calculated, one with positive expectancies as the dependent variable and another for negative expectancies. Predictorvariables were imputed at both levels (as specified in Table 1): Prompt level variables (jsocial context, environmental context, alcohol consumption -yes or no, and number of drinks), and individual level predictors (ijage, gender, ethnicity, student/professional status and raw (as opposed to therapeutic categories) AUDIT scores were used for analyses. In all analyses, binary variables (Gender, 1 = female; Student/Professional status, 1 = student ; Ethnicity, 1 = white British; Alcohol Consumption, 1 = yes) were dummy coded (for a more easily interpretable outcome),and the two categorical predictors (environmental and social context) were dummy coded usingHome and Alone conditions as the respective reference categories(k-1), and the remaining variable were left as continuous variables (Positive expectancies, Negative expectancies, Age, AUDIT, Number of drinks)..

INSERT TABLE 1 HERE

How much variance in positive and negative outcome expectancies is found and can be subsequently explained at the individual level (variance between participants) and the group level (prompt level, variance between prompts/within participants)?

Empty models (also known as the variance component models - models without imputed predictor variables) indicated that there was a significant proportion of variance (ICC = 95.55%) to be explained at the prompt (μ0j = 3.68, p < .001) and the individual level (ICC = 4.41%, μ0ij = .17, p < .01). The same was also true of negative expectancies, with 46.36% (μ0j = .61, p < .001) and 19.74% (μ0ij = .15, p < .01) of unexplained variability being identified at the prompt and the individual levels respectively. 2* log likelihood statistics (using chi square) and ICC calculations revealed that the full positive expectancy model resulted in a significant reduction of unexplained variance (χ² (30, n = 61) = 978.06, p < .001), explaining 36.7%.and 35.3% of the identified variability in positive expectancies at the prompt and individual levels respectively. The negative expectancy model also produced significant reduction in the amount of unexplained variance (χ² = (9, n = 61) = 575.88, p < .001), with 22.95% and 15.38% of variance in negative expectancies being explained at the prompt and individual levels respectively.

Which predictors are significant predictors of variance in expectancies?

No single individual level predictor was significant within the MLM model of negative expectancies. However, for positive expectancies, the only individual level predictor that was significant was student/professional status (β0ij = -.23, p < .01), such that being a young professional was associated with reduced positive expectancies to a significant degree, whilst being a university student was associated with an increase in positive expectancies. At the prompt level, having consumed alcohol within the last hour of prompting was a significant predictor of both increased positive (β0j = -.82, p < .001) and negative expectancies (β0j = -.51, p < .001) whilst number of drinks was not a significant predictor of positive expectancies but it did negatively predict variance in negative expectancies (β0j = -.09, p < .001). This suggests that any level of alcohol consumption may increase both positive and negative expectancies. Nonetheless,, whilst the number of drinks did not appear to alter positive beliefs (they remained heightened during consumption), negative beliefs began to decrease as alcohol consumption increased.

Both prompt level categorical predictor variables (social and environmental context) were also significant predictors of positive and negative outcome expectancies. Specifically, responses whilst situated within alcohol-related contexts including bars (β0j = -.52, p < .05), parties (β0j = -.91, p < .01) and sporting events (β0j = - .79, p < .05) were associated with increased positive expectancies. Similarly, negative expectancies were significantly predicted by being in a bar/pub/club (β0j = -.25, p < .01), although sporting and party venues did not account for significant variance. Being at a friend or family member’s house was also a significant predictor of increased positive (β0j = -1.10, p < .001) and negative expectancies (β0j = -.67, p < .001). Being at work was also a significant predictor of positive (β0j = .61, p < .01) and negative expectancies (β0j = -.28, p < .05). Here, being outside of work was associated with an increase in positive expectancies, and a decrease in negative expectancies. Being at home during responses was the reference category for both expectancy types and this context also appeared to be associated with decreased positive and negative expectancies..

The social context sub-categories also varied to a statistically significant degree. Prompts that occurred whilst participants were with 1 friend (β0j = -1.78, p < .001: β0j = -.74, p < .001), 2 or more friends (β0j = -1.75, p < .001: β0j = -.84, p < .001) or family members (β0j = -1.10, p < .001: β0j = -.79, p < .001) were significant predictors associated with increases in positive and negative expectancies respectively. However, being with work colleagues predicted significant decreases in positive expectancies (β0j = .72, p < .05) and increases in negative expectancies (β0j = -.43, p < .001). Being alone during responses was the reference category for both expectancies categories, meaning that this context also appears to be associated with decreased expectancies. The ‘other’ response for social context was also a significant predictor of positive expectancies (β0j = 2.44, p < .01) but the large standard error here (.92) suggests a high degree of variability in participants’ responses in this category, perhaps due to the diversity of contexts captured by this response. Any attempt to interpret this finding without any further contextual information would therefore be unwise.