Stressor-Related Drinking and Future Alcohol Problems among University Students

Abstract

Research using daily designs has shown that daily stressors (i.e., conflict, school/work demands) are associated with alcohol use, and that the strength of within-person links between stressors and alcohol use differs from person to person. However, to our knowledge no research has tested whether individual differences in stressor-related drinking – characterized by within-person associations between daily stressors and drinking – predict risk for future alcohol problems, a relationship suggested by theoretical models. The current study used an internet-based daily diary design among 744 university students to (a) examine the day-level relationship between stressors and alcohol use during the first three years of college and (b) test whether individual differences in the stressor-drinking relationship, captured by person-specific slopes generated from multilevel models, predicted alcohol problems as measured by the Alcohol Use Disorders Identification Test (AUDIT) in the fourth year of college. Results showed that students were more likely to drink on days with many versus fewer stressors, and on drinking days, students consumed more drinks with each additional stressor they experienced. Next, using individual multilevel modeling slopes as predictors, we found that students whose odds of drinking alcohol increased more sharply on high- versus low-stressor days (steeper slopes) had more severe AUDIT alcohol problems in fourth year than students whose drinking odds increased less sharply (flatter slopes). Findings highlight the role of daily stressors in college student drinking and suggest stressor-related drinking as a risk factor for future alcohol problems.

For many students, the college years are a normative and transient period of experimentation with heavy alcohol use (Baer, 2002; Schulenberg & Maggs, 2002; Staff et al., 2010). Recent estimates suggest that heavy alcohol use among college students is prevalent, with 32% of students engaging in current heavy drinking (Johnston, O'Malley, Bachman, Schulenberg, & Miech, 2016), defined as consuming 5 or more drinks in a row on a single occasion in the past two weeks.¹ Although statistically normative, heavy alcohol use among college students can be problematic, as it is associated with a wide range of problems, including academic difficulties, physical and sexual assaults, alcohol-related traffic deaths, and increased risk for alcohol use disorder (AUD; Hingson, Zha, & Weitzman, 2009; NIAAA, 2015).

The college years can also be a time of stress. Between 75% and 80% of college students report being “moderately stressed,” and 10%-12% report being “severely stressed” (Abouserie, 1994; Pierceall & Keim, 2007). The transition from adolescence to young adulthood is inherently a period of challenge in Western society, as young people strive to establish independent identities through the accomplishment of salient (friendship, academic) and emerging (romantic, occupational) developmental tasks (Arnett, 2000; Roisman, Masten, Coatsworth, & Tellegen, 2004). While in college, students must simultaneously balance academics and independent living with relationships (social and romantic) and family demands. Achieving balance among these competing priorities is likely to be experienced as stressful (Robotham & Julian, 2006).

For some students, alcohol use may become a way of responding to, or coping with, day-to-day stressors. Alcohol has been perceived as a stress reliever since antiquity (Sayette, 1999), and the use of alcohol for stress reduction has played a central role in theoretical models of alcohol use (e.g., Conger, 1956; Cooper, Frone, Russell, & Mudar, 1995; Sher, 1987). For example, the tension-reduction theory suggests that alcohol use will reduce stress in most situations, and that this stress relief will lead people to increase their drinking when stressful situations are encountered, which may lead to the development of alcohol problems (Conger, 1956). The stress-alcohol use relationship is conceptualized as bidirectional, with stress leading to increased drinking and increased drinking leading to additional stress (e.g., Anthenelli, 2012; Armeli, DeHart, Tennen, Todd, & Affleck, 2007; McEwen & Stellar, 1993), potentially creating a deviation amplifying cycle that may foster psychological dependence on alcohol. Motivational models of alcohol use suggest that people drink, in part, to regulate emotion–to either enhance positive affect (drinking to enhance; DTE) or to dampen stress or negative affect (drinking to cope; DTC; Cooper et al., 1995; Cox & Klinger, 1988). DTC motivation predicts future alcohol problems beyond alcohol consumption levels (Cooper et al., 1995; Merrill & Read, 2010; Simons, Gaher, Correia, Hansen, & Christopher, 2005); some argue that DTC leads to alcohol problems by supplanting healthier and more adaptive coping and emotion regulation strategies (Armeli, O’Hara, Ehrenberg, Sullivan, & Tennen, 2014; Cooper, Russell, & George, 1988).

Studies assessing DTC motivation have typically used static, retrospective measures, such as the drug and alcohol subscale of the COPE (Carver, Scheier, & Weintraub, 1989), the Motivations for Alcohol Use Scale (MAUS; Cooper, 1994), or the Reasons for Drinking Questionnaire (RFDQ; Farber, Khavari, & Douglass, 1980). These measures ask respondents to self-assess, retrospectively and generally, whether coping-related motivations underlie their drinking. Retrospective measures have documented the deleterious effects of coping-related drinking, but they have limited validity relative to ambulatory assessment strategies–including daily diary or daily process studies–which shorten the recall window from weeks/months to hours/days, thereby reducing cognitive distortions associated with longer-term retrospective reporting (Bolger, Davis, & Rafaeli, 2003; Bradburn, Rips, & Shevell, 1987; Shiffman, Stone, & Hufford, 2008).

Daily process studies allow researchers to measure the actual–as opposed to the remembered–correlation between reported stressful experiences and drinking episodes, the contingency that is at the heart of the DTC construct. Daily process measurement is important for DTC motivation because typical questionnaire approaches to assessing DTC rely on individuals to have correct memories of or insights into the underlying reasons for their behavior – that is, participants are expected to be able to accurately report the degree to which their drinking is influenced by stress and attempts to cope with such stress. Such reports of DTC may be biased by illusory correlation (Chapman, 1967) – a tendency to report events as correlated because individuals believe or remember them to be, regardless of whether or not they actually are (Todd et al., 2005). This could lead people to over- or underestimate the extent to which their drinking is influenced by stress, providing an inaccurate characterization of their drinking tendencies. Using repeated measurements of stress and alcohol consumption, researchers can estimate a daily stressor-drinking slope for each individual by comparing each person’s alcohol consumption on high- versus low-stressor days via multilevel modeling. This approach offers a number of advantages for assessing DTC because (a) reports are provided closer in time to when the events occurred (thus reducing recall bias), (b) drinking and stressor reports are collected in the flow of people’s day-to-day lives, thereby enhancing ecological validity of the stressor-drinking relationship (Bolger et al., 2003); and (c) the stress-drinking association is directly calculated from daily data, rather than relying on the accuracy of participants’ memories or insights into their drinking motivations. Although daily stressor-drinking slopes capture the root contingency of the DTC construct, they may also be influenced by a number of other factors, including beliefs about and effectiveness of alcohol as a stress reliever (Mohr et al., 2013; Sayette, 1999; Sher, 1987). Ultimately, the daily process approach to estimating stressor-related drinking in day-to-day life may be informative for theory and intervention because it allows researchers to capture the degree to which stressors and drinking actually co-occur in the flow of people’s day-to-day lives, and to use this information to inform our understanding of whether or not people with high stressor-drinking contingencies are at greater risk for alcohol problems.

Once estimated, daily stressor-drinking slopes can be used as predictors in models directly testing a central hypothesis of motivational models of alcohol use: that increasing alcohol consumption to cope with stress increases risk for future alcohol problems (Cooper et al., 1995). People who increase drinking when stressed may be at greater risk for developing alcohol problems due to the negative reinforcement of reducing stress and negative affect (Armeli et al., 2003; Sher, 1987). Similar to DTC motivations, frequent stress-related drinking may supplant more adaptive coping and emotion regulation strategies, lead to alcohol consumption as the person’s primary means of coping with stress (e.g., Armeli et al., 2014; Cooper et al., 1988), and ultimately lead to alcohol abuse episodes and increased risk for psychological dependence.

A number of daily process studies have examined within-person relationships between stress-related measures – including stressor load, perceived stress, negative mood or distress, negative interpersonal or work events – and alcohol consumption. Samples include community samples of adults (Armeli, Carney, Tennen, Affleck, & O'Neil, 2000; Armeli et al., 2003; Grzywacz & Almeida, 2008; Todd, Armeli, & Tennen, 2009; Todd, Armeli, Tennen, Carney, & Affleck, 2003) as well as college student samples (Armeli, Todd, Conner, & Tennen, 2008; DeHart, Tennen, Armeli, Todd, & Mohr, 2009; Howard, Patrick, & Maggs, 2015; Hussong, 2007; Hussong, Galloway, & Feagans, 2005; Hussong, Hicks, Levy, & Curran, 2001). Although these studies often find significant within-person relationships between stress-related measures and alcohol consumption, the direction of the relationships is inconsistent: some studies show negative associations (more stress associated with less drinking, e.g., Armeli, Carney, et al., 2000; Armeli, Todd, & Mohr, 2005; Carney, Armeli, Tennen, Affleck, & O'Neil, 2000), and others show positive associations (more stress associated with more drinking; Armeli, Tennen, Affleck, & Kranzler, 2000; Mohr et al., 2001; Park, Armeli, & Tennen, 2004). Daily stress has been shown to predict next-day drinking, and drinking has been shown to predict next–day stress levels (Armeli et al., 2007; Ayer, Harder, Rose, & Helzer, 2011; Levitt & Cooper, 2010).

Daily process studies demonstrate that the size and direction of stress-drinking slopes differ substantially between persons and are predicted by person-level factors, including biological sex, alcohol outcome expectancies, personality traits, and drinking motivations including DTC and DTE (see reviews by Armeli et al., 2005; Mohr, Armeli, Tennen, & Todd, 2010). Although many studies have tested individual difference factors that predict stress-drinking slopes, to our knowledge only one study has used daily stress-drinking slopes as predictors of future alcohol-related problems (Mohr et al., 2013), and none have used daily stress-drinking slopes as predictors of alcohol problems among college students.

The current study estimates students’ daily stress-drinking slopes during their first three years of college and uses these slopes to predict alcohol problems during the fourth year of college. We tested the following research questions: (1) Are students more likely to drink on high- versus low-stressor days? (2) On drinking days, do students drink more when they experience stressors? and (3) Does stressor-related drinking in daily life predict future alcohol problems? Our approach is novel in two ways. First, because drinking has previously been conceptualized as a two-step process (Armeli et al., 2005), we estimated daily relationships between stressors and alcohol use in two separate models testing the associations (1) between stressful events in daily life (daily stressors) and drinking events (any versus no drinking) across all study days and (2) between stressors and the number of drinks consumed on drinking days only. Although many studies have the requisite data, few have estimated stressor-drinking associations at both steps of the drinking process (see Huh, Kaysen, & Atkins, 2015 for an exception). Second, we used multilevel modeling to estimate stressor-drinking slopes for each student. These slopes captured the increase in each student’s odds of drinking on high- versus low-stressor days. Once estimated, these individualized slopes were used as predictors of alcohol problems in the fourth year of college. That is, we tested whether students who showed larger drinking increases on stressor versus non-stressor days (steeper stressor-drinking slopes) during the first three years of college were at greater risk for fourth-year alcohol problems than students with smaller (or no) drinking increases on stressor versus non-stressor days (flatter stressor-drinking slopes). This approach aligns with Mischel and Shoda’s (1995) behavioral signatures, which are conceptualized as idiosyncratic and stable if-then contingencies between situations and behavior that differ from person to person. To our knowledge, Mohr and colleagues (2013) were the first to link negative experience-drinking slopes to Mischel and Shoda’s (1995) behavioral signatures framework. The current study follows Mohr and colleagues’ lead by invoking the behavioral signatures framework to hypothesize that some students would show an increase in drinking when experiencing stressors (steeper positive stressor-drinking slopes), whereas others will show no change or even a decrease (flat or steeper negative stressor-drinking slopes, respectively). Although recent research has demonstrated the predictive validity of within-person slopes as predictors of various health outcomes, including alcohol problems among adults (Charles, Piazza, Mogle, Sliwinski, & Almeida, 2013; Mohr et al., 2013; Piazza, Charles, Sliwinski, Mogle, & Almeida, 2013), the current study is the first to our knowledge to apply this approach to assess stress-related drinking and future risk for alcohol problems in university students.

Method

Data were from the University Life Study (ULS), a daily diary study of 744 first-time college students beginning in fall of their first year and continuing through fall of their fourth year. The sample was diverse (25.1% Hispanic/Latino, 27.4% non-Hispanic (NH) European-American, 23.3% NH Asian-American, 15.7% NH African-American, and 8.5% NH multiracial), nearly evenly split by gender (49.2% male), with an average age of 18.4 years at study entry (SD=.43, min=16.9, max=20.8). Details of the ULS are described elsewhere (Howard et al., 2015). Briefly, students were invited via email to complete short daily web-based surveys on stressors and drinking over 14 consecutive days once in each of seven semesters from fall of Year 1 to fall of Year 4 (≤98 days per student). During daily diary periods, students received emails with a link to a web-based survey early each morning; they could access surveys for two days. All questions referred to a full experienced day, such as “yesterday, that is, from the time you woke up until the time you went to sleep.” Completion and retention rates were high. Participants typically provided 77.8% (SD=29.4%)–or 76.2– of 98 possible diary reports (14 reports × 7 semesters). Retention at each measurement burst ranged from 96.2% in Fall of Year 1 to 79.4% in Fall of Year 4. Due to our interest in prospective links of stress-related drinking with the development of alcohol problems, we used diary reports of stressors and alcohol use from the first three years only, excluding diary data from fall of Year 4. All study procedures were approved by the university’s Institutional Review Board.

Measures

Daily stressors were measured using six items adapted from the Daily Interview of Stressful Experiences (DISE; Almeida, Wethington, & Kessler, 2002) for web-based administration (see also Galambos, Dalton, & Maggs, 2009). Questions referred to the previous day, and asked about arguments/disagreements (8.1% of all days from years 1–3, ignoring nesting of days within students), non-argument interpersonal tensions (7.3%), school or work stressors (5.1%), home stressors (2.7%), friend or family stressors (3.0%), and other general stressors (5.5%), with responses of Yes (stressor happened) or No (it did not). Across Years 1–3, students reported, on average, at least one stressor on 20.0% of their diary days (SD=18.0%). Most students experienced at least one stressor day; only 5.3% (n=39) reported no stressor days. Among students who reported stressors, measures of central tendency varied widely; the mean number of stressor days was 14.1 (SD=13.9, 21% days on average), with a median of 9 stressor days and a mode of 2 stressor days. Looking across all stressor days, ignoring nesting of stressor days within students, 62% of stressor days had only one stressor, 24% had two, 9% had three, and 5% had four or more. A stressor count variable was created for analyses to characterize the stressor load of each day (M=0.32, SD=0.76, min=0, max=6).

Alcohol use

Alcohol use. Students reported on the number of drinks consumed using a pull-down menu permitting responses from 0 to 25+ drinks. A drink was defined for students as “half an ounce of absolute alcohol, for example, 12-ounce can or bottle of beer or cooler, 5-ounce glass of wine, a drink containing one shot of liquor or spirits.” Across Years 1–3, students reported drinking on 12.8% of their diary days, on average (SD=13.0%). Most students (81.4%, n=598) reported at least one drinking day across Years 1–3. Those who reported drinking during Years 1–3 reported an average of 10.6 drinking days (SD=9.2; 15.7% of their daily reports on average), median of 8 drinking days, and mode of 2 drinking days. Across all students and all drinking days, drinking days tended to be heavy drinking days according to a variety of descriptive measures (M=6.0, Median=5.0, Mode=5.0) with somewhat truncated variability (SD=4.1, interquartile range [IQR]=8–3) in the number of drinks consumed. Because we were interested in modeling daily alcohol use as a two-step process, we split the alcohol count variable into two parts: (1) whether or not students engaged in any drinking on that day (drinking episode; Years 1–3, M=0.13, SD=0.33), and (2) the number of drinks students consumed on drinking days (drinking count on drinking days; Years 1–3, M=6.01, SD=4.08, min=1, max=26).

Alcohol problems were measured during fall of Year 4 using dependence and harmful alcohol use items from the Alcohol Use Disorders Identification Test (AUDIT; Babor, Higgins-Biddle, Saunders, & Monteiro, 2001). Dependence items asked about the frequency of not being able to stop drinking once started, having failed to do what was expected because of drinking, and morning drinking. These items were scored on a 5-point scale ranging from never (0) to daily or almost daily (4). Harmful alcohol use items asked about (a) frequency of guilt after drinking and blackouts on the same 5-point scale as the dependence items; and (b) occurrence of alcohol-related injuries and others expressing concern about the student’s drinking on a 3-point scale with response options 0=no, 2=yes, but not in the last year, and 4=yes, during the last year. An alcohol problems severity score was created by taking the sum of all 7 dependence and harmful use items (M=2.47, SD=332, Min=0, Max=19). In addition, we created an indicator of alcohol problem high-risk status by dividing the distribution of AUDIT severity scores into low risk (<8) versus high risk (≥8), following Babor and colleagues (2001). In total, 54 students (8.9%) showed high risk for alcohol problems during the fourth year of college whereas 551 students showed low risk (91.7%).

Alcohol-problem history was measured in the fall of Years 1–3 of college using the 23-item Rutgers Alcohol Problem Index (RAPI; White & Labouvie, 1989). Students were asked whether they had experienced problems related to alcohol such as “felt that you had a problem with alcohol,” “not able to do homework or study for a test,” and “kept drinking when you promised yourself not to.” During Year 1, students were asked about problems experienced in their lifetimes; in Years 2 and 3, students were asked about the past 12 months. Ratings were collected using a 4-point scale: none (0), 1-2 times (1), 3-5 times (2), and more than 5 times (3). Alcohol problem history indexes were created at Year 1 (M=0.24, SD=033, min=0, max=2.48), Year 2 (M=0.20, SD=0.29, min=0, max=1.70), and Year 3 (M=0.21, SD=034, min=0, max=2.61) by taking the mean of the 23 items at each year. These three index scores were then averaged across years to create an alcohol-problem history index (M=0.22, SD=0.27, min=0, max=1.91), which was standardized (M=0, SD=1) and used as a covariate in models predicting AUDIT alcohol problems in Year 4.

Statistical Analyses

Models were estimated to answer each research question; details are outlined below. Students were included in analyses whether or not they reported drinking in the daily diaries.

Question 1: Are students more likely to drink on high- versus low-stressor days?

We estimated whether students’ odds of drinking increased with each additional daily stressor using logistic multilevel modeling in SAS PROC GLIMMIX. The model is presented in Equation 1.

$$\text{Log Odds}({\text{ALC}}_{\text{ijk}})={\mathrm{\beta }}_0+{\mathrm{\beta }}_1(\text{stressors}\_{\text{day}}_{\text{ijk}})+{\mathrm{\beta }}_2(\text{stressors}\_{\text{sem}}_{\text{jk}})+{\mathrm{\beta }}_3(\text{stressors}\_{\text{person}}_{\mathrm{k}})+{\mathrm{\beta }}_4({\text{wknd}}_{\text{ijk}})+{\mathrm{v}}_{0\text{jk}}+{\mathrm{u}}_{0\mathrm{k}}+{\mathrm{u}}_{1\mathrm{k}}(\text{stressors}\_{\text{day}}_{\text{ijk}})$$ (1)

Equation 1 accounts for the nesting of days (Level 1) within semesters (Level 2) using a random semester-level intercept (v_0jk), as well as the nesting of semesters within students (Level 3) using a random student-level intercept (u_0k). Of particular importance in this model are (a) the β₁(stressors_day_ijk) fixed slope, which describes the sample average day-level increase in the likelihood of drinking (measured in log odds) with each additional stressor; and (b) the u1k(stressors_day_ijk) random slope, which captures individual differences in the β₁ slope. To improve interpretability of the stressor-drinking daily association, we included each student’s mean stressor count for each semester – β₂(stressors_sem_jk) – as well as each student’s mean stressor count across all days and semesters – β₃(stressors_person_k) – as covariates; this helps ensure that relationship between stressors and drinking is purely a within-person, daily-level relationship (Hoffman & Stawski, 2009; Sliwinski, Almeida, Smyth, & Stawski, 2009). Stressors_sem_jk and Stressors_person_k variables were centered on their overall means prior to model entry. β₄(wknd_ijk) is a daily variable to control for “weekend” effects, reflecting increased drinking likelihood on students’ typical drinking days (Del Boca, Darkes, Greenbaum, & Goldman, 2004; Maggs, Williams, & Lee, 2011). In this variable, Thursdays, Fridays, and Saturdays were coded 1; all other days were coded 0. Finally, the model included a Level 1 dispersion parameter to capture over- or under-dispersion in the dichotomous drinking outcome (description in Bolger & Laurenceau, 2013).

Given that the daily stressor variable varies at the daily, semester, and person levels, one approach to separate this variability would be to center daily stressor values on each student’s semester- and person-level means. However, such centering would imply that we were interested in how stressor levels relative to each student’s person- or semester-level averages were associated with drinking behavior. Instead, we chose to leave the daily stressor variable raw and uncentered, while statistically adjusting for differences in average stressor load across semesters and students.² This approach has the benefit of leaving the Level 1 predictor, daily stressors, in its original metric, thereby allowing the zero point to be the same for every student (a day with no stressors), while still adjusting for the fact that some semesters are more stressful than others and some students experienced more stressors than others overall. Approached this way, the Level 1 association is interpreted as the within-person, day-level association; the Level 2 association is interpreted as the within-person semester-level association, adjusted for day-level stressor load; and the Level 3 association becomes the between-person association, adjusted for semester- and day-level stressor load (Hoffman & Stawski, 2009; Sliwinski et al., 2009).

Question 2: On drinking days, do students drink more when they experience more stressors?

We ran Poisson multilevel models (Equation 2) only on drinking days to test the within-person association between stressor count and drinking count.

$$\text{Log}(\text{ALC}\_{\text{Count}}_{\text{ijk}})={\mathrm{\beta }}_0+{\mathrm{\beta }}_1(\text{stressors}\_{\text{day}}_{\text{ijk}})+{\mathrm{\beta }}_2(\text{stressors}\_{\text{sem}}_{\text{jk}})+{\mathrm{\beta }}_3(\text{stressors}\_{\text{person}}_{\mathrm{k}})+{\mathrm{\beta }}_4({\text{wknd}}_{\text{ijk}})+{\mathrm{v}}_{0\text{jk}}+{\mathrm{u}}_{0\mathrm{k}}+{\mathrm{u}}_{1\mathrm{k}}(\text{stressors}\_{\text{day}}_{\text{ijk}})$$ (2)

Drinking count on drinking days (ALC_Count_ijk; ranging from 1 to 25+ drinks) was modeled using a loglinear link, a Poisson distribution, and a Level 1 dispersion parameter in SAS PROC GLIMMIX. Person-mean and person-semester-mean stressor counts (calculated across drinking days only, and centered on their overall means) were included as covariates at Levels 3 and 2, respectively, along with the weekend (versus weekday) variable as a Level 1 covariate.

Question 3: Does stressor-related drinking in daily life predict future alcohol problems?

Step 1: Estimating student-specific daily-stressor drinking slopes

We generated daily stressor-drinking event slopes for individual students using logistic multilevel models; these slopes represent stressor-related drinking, or the day-level association between stressors and drinking events, for each student. Daily stressor-drinking event slopes were in log odds metric, describing the increase in each student’s log odds of drinking with each additional stressor on a given day. Daily stressor-drinking event slopes were calculated by (a) estimating the model in Equation 1, with daily stressor count as the predictor of drinking events; then (b) adding the fixed slope describing the average association between stressor count and drinking events among students (β₁(stress_sday_ijk) in Equation 1) to the model-estimated best linear unbiased predictor (BLUP; calculated from the u_1k(stress_sday_ijk) random effect) for each student. To avoid spurious results, daily stressor-drinking slopes were estimated in this way only for students who showed at least some intraindividual variance (iVar) in both stressors and drinking events across diary days (N students = 573, N days = 39,838). iVar was calculated for each person using SAS PROC MEANS with participant ID number on the CLASS statement and VAR on the OUTPUT statement to estimate a variance specific to student. This was done to identify students who reported the same value for all their stressor reports (n students = 137) and/or all their drinking reports (n students = 37). All students with drinking iVar=0 reported not drinking across all days; all students with stressor iVar=0 reported no stressors across all days. Students (n=123) with at least some variance (iVar > 0) in stressor exposure but no variance in drinking events (iVar = 0) were not included in the estimation of individual-specific slopes; instead, they were assigned a slope of 0 because they did not drink, even though they did experience differing stressor levels across days. Individual-specific slopes were set to missing for students with no variance in stressors, as no estimate could be calculated (n students=37). Two students provided only one day of stressor reports each (both non-stressor days); their iVars were missing as variance could not be calculated.

Step 2: Using daily stressor-drinking event slopes to predict alcohol problems

After estimating individual slopes, we tested whether differences in the daily stressor-drinking association predicted alcohol problems in Year 4. This was tested using the daily stressor-drinking event slope as a predictor of alcohol problem outcomes (as measured by the AUDIT) in regression models. These models controlled for students’ drinking event probabilities on non-stressor days, defined as each student’s probability of engaging in drinking on days in which no stressors occurred (M=0.12, SD=0.13, Min=0, Max=1) and average stressor exposure, defined as the average number of stressors reported by each student across all of their diary days (M=0.32, SD=0.34, Min=0, Max=2.82). The scores for daily stressor-drinking event slopes, drinking event probabilities on non-stressor days, and average stressor exposure levels were standardized (M=0, SD=1) prior to model entry in order to facilitate interpretation and comparison. Drinking and non-drinking students were included in these models to provide inferences relevant to risk of Year 4 alcohol problems in the entire sample, not just among those who reported drinking in their daily diaries. Negative binomial regression was used to predict the alcohol problem severity count; logistic regression was used to predict AUDIT high-risk status. These regression models were run in Mplus using full information maximum likelihood, which boosts statistical power and minimizes bias by retaining all participants with complete or partially complete data across assessment waves (Enders & Bandalos, 2001).

Results

Question 1: Are students more likely to drink on high- versus low-stressor days?

Results of models testing the daily relationship between stressors and drinking events are presented in Table 1; four findings are apparent. First, a significant within-person, day-level association between stressors and drinking events suggested that on average, students’ drinking odds significantly increased with each additional stressor experienced on a given day. The odds ratio (OR) of 1.08 (p=.002) suggested that drinking odds for a typical student increased by 8% with each additional stressor experienced. Second, a significant daily stressors random effect (Est.=0.09, SE=0.02, p<.001) provided evidence that the strength of the daily stressor-drinking event association varied randomly between students, suggesting that drinking events were more strongly associated with daily stressors for some students than others. Third, adjusting for day-level stressor load, there was no evidence for a within-person, semester-level association between stressors and drinking events (OR=0.99, p=.85), suggesting that students’ odds of a drinking event on any given day did not differ across high- versus low-stress semesters. Fourth, there was a significant student-level association between stressors and drinking events (OR=1.69, p=.003), suggesting that, adjusting for daily and semester-level stressor load, students experiencing more stressors on average had higher odds of a drinking event on any given day than students experiencing fewer stressors on average.

Table 1.

Logistic multilevel model showing daily relationship between daily stressors and drinking events (N students=735, N days=49,341)

Fixed Effects	b (SE)	OR	95% CI
Daily Stressors Slope	0.08 (0.02)**	1.08	1.03, 1.13
Semester-Mean Stressors Slope	−0.01 (0.06)	0.99	0.87, 1.12
Person-Mean Stressors Slope	0.53 (0.18)**	1.69	1.20, 2.39
Weekend Slope	0.42 (0.03)***	1.52	1.44, 1.59
Intercept	−2.47 (0.05)***	0.08	0.08, 0.09
Random Effects (variances, covariances)	Estimate	SE	95% CI
Person Level
Intercept Variance	1.74***	0.12	1.52, 2.01
Daily Stressors Slope Variance	0.09***	0.02	0.07, 0.14
Intercept, Daily Stressors Slope Covariance	−0.02	0.04	−0.11, 0.06
Semester Level
Intercept Variance	0.47***	0.03	0.42, 0.53
Daily Level
Dispersion	0.67a	--	0.66, 0.68

OR=odds ratio.

^***p<.001,

^**p<.01,

^*p<.05,

⁺p<.10

^aThe dispersion parameter is tested for significance by comparing its 95% confidence interval (CI) to 1, a value that indicates expected dispersion for a binomial outcome (Bolger & Laurenceau, 2013). Significant underdispersion was present.

Question 2: On drinking days, do students drink more when they experience more stressors?

Models testing the association between daily stressors and number of drinks on drinking days (see Table 2) revealed four main findings. First, at the day level, the number of drinks consumed significantly increased with each additional stressor experienced on a drinking day (IRR=1.04, p<.001). The IRR of 1.04 suggested that consumption increased by 4% with each additional stressor. Model-predicted counts, holding covariates constant at their means, suggested that the typical student consumed 4.8 drinks on drinking days with no stressors and 5.9 drinks on drinking days with 6 stressors (the maximum reportable). Second, the random effect for this association was near zero (Est.=0.001, p=.20), suggesting that students did not differ in their day-level association between stressor load and number of drinks on drinking days. Third, at the within-person, semester level, a significant relationship suggested that students drank fewer drinks on drinking days during high- versus low-stress semesters, adjusting for day-level stressor load (IRR=0.96, p=.033). Fourth, we found no evidence for a student-level association between stressors and drinking level on drinking days, suggesting that a student’s overall average number of drinks across drinking days did not vary based on their average number of stressors after adjusting for day- and semester-level stressor load (IRR=1.06, p=.23).

Table 2.

Poisson multilevel model showing daily relationship between stressor exposure and drinking level on drinking days (N students=597, N drinking days=6,337)

Fixed Effects (intercepts, slopes)	b (SE)	IRR	95% CI
Daily Stressors Slope	0.04 (0.01)***	1.04	1.02, 1.06
Semester-Mean Stressors Slope	−0.04 (0.02)*	0.96	0.92, 1.00
Person-Mean Stressors Slope	0.06 (0.05)	1.06	0.97, 1.16
Weekend Slope	0.29 (0.02)***	1.34	1.30, 1.39
Intercept	1.56 (0.02)***	4.76	4.57, 4.97
Random Effects (variances, covariances)	Estimate	SE	95% CI
Person Level
Intercept Variance	0.18***	0.02	0.16, 0.22
Daily Stressors Slope Variance	0.001	0.002	0.00, 0.24
Intercept, Daily Stressors Slope Covariance	−0.005	0.004	−0.01, 0.00
Semester Level
Intercept Variance	0.03***	0.004	0.02, 0.04
Daily Level
Dispersion	1.37a	--	1.31, 1.42

IRR=Incident rate ratio

^***p<.001,

^**p<.01,

^*p<.05,

⁺p<.10

^aThe dispersion parameter is tested for significance by comparing its 95% confidence interval (CI) to 1, a value that indicates expected dispersion for a Poisson outcome. Significant overdispersion was present.

Question 3: Does stressor-related drinking in daily life predict future alcohol problems?

Predicting future risk for alcohol problems from daily stressor-drinking event slopes

To test whether students with stronger daily stressor-drinking event relationships (steeper stressor-drinking event slopes) were at greater risk for future alcohol problems than students with weaker daily stressor-drinking event relationships (flatter stressor-drinking slopes), we computed zero-order correlations between daily stressor-drinking slopes event and alcohol problems (see Table 3). Significant positive correlations existed between the daily stressor-drinking event slope and alcohol problem history (r=0.16), Year 4 alcohol problem severity (r=0.13) and Year 4 high-risk cutoff (r=0.13), suggesting that people with a sharper increase in drinking odds on high-versus low-stressor days had greater alcohol problem histories and more alcohol problems in Year 4.

Table 3.

Zero-Order Correlations between daily stressor-related variables in years 1–3 and year-4 AUDIT alcohol problem outcomes

Variable	1	2	3	4	5	6
1. Daily Stressor-Drinking Event Slope (Years 1–3)	--
2. Drinking Event Probability on Non-Stressor Days (Years 1–3)	−0.07+	--
3. Average Stressor Exposure (Years 1–3)	−0.07+	0.19***	--
4. Alcohol Problem History (Years 1–3)	0.16***	0.35***	0.19***	--
5. AUDIT Alcohol Problem Severity (Year 4)	0.13**	0.35***	0.07+	0.48***	--
6. AUDIT High-Risk Cutoff (Year 4)	0.13**	0.14***	0.04	0.33***	0.78***	--

^***p<.001,

^**p <.01,

⁺p<.10

To test whether stressor-related drinking in daily life independently predicted Year 4 alcohol problems, we ran multiple regression models with daily stressor-drinking event slopes as predictors of Year 4 alcohol problem severity and high-risk status (see Table 4). Models adjusted for drinking event probability on non-stressor days, average stressor exposure, and alcohol problem history. Table 4 reveals three findings relevant to alcohol problem severity. First, Model 1 shows that students’ daily stressor-drinking event slopes and drinking event probabilities on non-stressor days were significant predictors of alcohol problem severity in Year 4. This suggested that (a) students with a sharper increase in drinking event likelihood on high- versus low-stressor days had greater alcohol problem severity in Year 4 than students whose drinking event likelihood increased less sharply, and (b) students with a higher probability of drinking events on non-stressor days had more severe alcohol problems than students with a lower probability. Second, Model 2 showed that these associations remained significant after controlling for alcohol-problem history. Third, average stressor exposure was not significantly associated with Year 4 alcohol problem severity, suggesting that students’ typical number of daily stressful experiences was not related to their risk for future alcohol problems.

Table 4.

Regression models predicting Year-4 AUDIT outcomes from daily stressor and drinking-related daily parameters

	AUDIT Alcohol Problem Severity (Dependence & Harmful Drinking)		AUDIT High-Risk Status (8+)
	Model 1	Model 2	Model 1	Model 2
	IRR [95% CI]	IRR [95% CI]	OR [95% CI]	OR [95% CI]
Daily Stressor-Drinking Slope	1.29***[1.16, 1.44]	1.19**[1.07, 1.32]	1.52**[1.18, 1.96]	1.29+ [0.98, 1.69]
Drinking Event Probability on Non-Stressor Days	1.71***[1.51, 1.95]	1.42***[1.25, 1.61]	1.57**[1.22, 2.03]	1.23 [0.92, 1.63]
Average Stressor Exposure	0.99 [0.87, 1.12]	0.95 [0.85, 1.07]	1.01 [0.76, 1.33]	0.94 [0.69, 1.28]
Alcohol Problem History (Years 1–3)	--	1.71***[1.48, 1.98]	--	2.29***[1.70, 3.08]

Note: Model 1 tests regression models with daily stressor and drinking-related parameters as predictors of alcohol problem outcomes without controlling for alcohol problem history in years 1–3. Model 2 adds year 1–3 alcohol problem history as a covariate to regression models.

^***p<.001,

^**p<.01,

^*p<.05,

⁺p<.10.

IRR= incident rate ratio, OR=odds ratio. Significant results are shown in bold.

Table 4 also reveals two findings relevant to AUDIT high-risk status. First, students’ daily stressor-drinking event slopes and drinking event probabilities on non-stressor days showed significant and positive associations with high alcohol problem risk, whereas students’ average levels of stressor exposure were not significantly associated (Model 1). These results suggest that (a) students showing a steeper increase in drinking odds on high-versus low-stressor days and (b) students who had higher versus lower drinking event probabilities on non-stressor days had greater odds of high alcohol problem risk in Year 4. Second, after controlling for alcohol-problem history, drinking event probabilities on non-stressor days no longer significantly predicted high alcohol problem risk (p=.16), whereas the predictive association for students’ daily stressor-drinking event slopes remained but was reduced to marginal significance (p=.065).

Additional analysis using the RAPI

We also tested whether daily stressor-drinking slopes predicted alcohol problems after swapping in the RAPI at year 4 for the AUDIT in regression models. Models showed that daily stressor-drinking slopes significantly predicted RAPI past-year alcohol problems at year 4 (b=0.04, SE=0.01, p=.002, β=0.13) after adjusting for stressor exposure (b=0.02, SE=0.01, p=.099, β=0.07) and drinking probability on non-stressor days (b=0.11, SE=0.01, p<.001, β=0.32). However, after adjusting for alcohol problem history using RAPI years 1–3, daily stressor-drinking slopes no longer significantly predicted RAPI alcohol problems at year 4 (b=0.02, SE=0.01, p=0.16, β=0.05).

Discussion

In a longitudinal, web-based, daily diary study of university students, we tested three research questions: (1) Are students more likely to drink on high- versus low-stressor days? (2) On drinking days, do students drink more when they experience more stressors? and (3) Does stressor-related drinking in daily life predict future alcohol problems? Daily stressors were associated with increases in odds of a drinking event on a given day and in the number of drinks consumed. Our day-level finding of a link between stressors and drinking is in line with past research showing positive associations between stress and alcohol consumption (Armeli, Tennen, et al., 2000; Mohr et al., 2001; Park et al., 2004), although prior findings have been inconsistent. The day-to-day relationship between stressors and drinking events differed significantly between students, whereas the relationship between stressors and number of drinks on drinking days did not differ between students. Differences between students in daily stressor-drinking event slopes predicted alcohol problems in fourth year. That is, students with sharper increases in odds of a drinking event on high- versus low-stressor days had more alcohol problems in the fourth year of college than students whose drinking event odds increased less sharply or did not increase. Daily stressor-drinking event slopes remained significantly predictive of alcohol problem severity (and marginally predictive of AUDIT high-risk status) after controlling for students’ drinking probabilities on non-stressor days, average levels of stressor exposure, and alcohol problem history. However, effect sizes of significant associations were small (Olivier, May, & Bell, in press), suggesting that these daily stressor- and drinking-related parameters may be small pieces of the puzzle when it comes to predicting daily drinking and the development of alcohol problems.

These findings are novel because they are the first to show that an increased tendency to drink on stressor- versus non-stressor days may presage the development of alcohol problems among college students, consistent with theoretical models (e.g., Conger, 1956; Cooper et al., 1995; Sher, 1987). Moreover, because the majority of daily reports (91%) occurred when students were under age 21, and because the majority of students (91%) were age 21 or older at the Year 4 assessment, our results add to the body of evidence relating underage drinking to alcohol problems at legal drinking age (e.g., Brown et al., 2008). Using individualized stressor-drinking slopes as predictors of future alcohol problems is innovative. An emerging body of research has used within-person slopes to predict future outcomes (Charles et al., 2013; L. H. Cohen, Gunthert, Butler, O'Neill, & Tolpin, 2005; Piazza et al., 2013), including one study that used this approach to predict alcohol problems in community adults (Mohr et al., 2013), but our study is the first to apply this approach to stressor-related drinking and alcohol problems among college students.

In students’ daily lives, we found that daily stressors were associated with increased drinking at both stages of the drinking process. That is, compared to themselves, students were more likely to engage in a drinking event on high- than low-stressor days, and they drank more on drinking days with a higher versus lower number of daily stressors. These findings are important because they suggest that exposure to daily stressors may influence both students’ likelihood of drinking and the number of drinks they consume on days they drink. Additionally, because these relationships captured changes in each student’s drinking compared to him- or herself across days, our results provide a stringent test of stressor-drinking links by holding constant stable individual differences through within-person comparisons (e.g., gender or family history of alcohol use; Allison, 2005; Bolger & Laurenceau, 2013).

Although the relationship between stressors and drinking events was stronger for some students than others, we found no difference between students in the relationship between stressors and drinking count on drinking days. This may have been partly because drinking days in this sample tended to be heavy drinking days. For example, students consumed a median of 5 drinks on drinking days with an IQR of 3–8 drinks, thus creating a bit of a ceiling effect on the slopes capturing the daily stressor-drinking count relationship. Other research shows similar patterns among young adults who engaged in binge drinking (i.e., median of 6 drinks per episode (e.g., Esser, Kanny, Brewer, & Naimi, 2012; Naimi, Nelson, & Brewer, 2010). This suggests that individual differences in how stressors affect students may lie primarily in the decision to drink, not in the number of drinks per episode. Future research is needed to verify these observations and examine the extent to which they also apply to non-college adult drinkers.

Stressor-related drinking, as indexed by daily stressor-drinking multilevel modeling slopes, predicted alcohol problem severity (assessed as a criterion count) and marginally predicted alcohol problem risk (AUDIT score ≥ 8) in the fourth year of college. Importantly, the daily stressor-drinking event slope predicted alcohol problems beyond drinking probabilities on non-stressor days, suggesting that these findings are not explained by individual differences in students’ typical (non-stress) drinking event likelihood on a given day. These findings align with previous research showing that coping motivations predict alcohol problems beyond consumption levels (Cooper et al., 1995; Merrill & Read, 2010; Simons et al., 2005). The lack of a significant association between daily stressor-drinking slopes and the AUDIT high-risk status indicator may have resulted from limited power, as only 54 students exceeded the threshold. We also ran additional regression models testing whether daily stressor-drinking slopes predicted alcohol problems at year 4 using the RAPI in place of the AUDIT. Although daily stressor-drinking slopes significantly predicted RAPI-measured alcohol problems at year 4 after adjusting for drinking probabilities on non-stressor days and stressor exposure, the association did not hold when alcohol problem history was added to the model. Thus, caution is warranted when interpreting our findings and replication of our significant results in independent samples is necessary before any causal inferences can be drawn.

We found that daily stressor-drinking event slopes predicted future alcohol problems as measured by the AUDIT, whereas mean stressor exposure did not, suggesting that amount of stressor exposure may be less important than students’ reactions to such stressors when predicting alcohol problems. Other research has shown that emotional reactivity to daily stressors is a stronger predictor than exposure to stressors of many health indicators, including chronic physical conditions (Piazza et al., 2013), depression (Charles et al., 2013), inflammation (Sin, Graham-Engeland, Ong, & Almeida, 2015), and mortality (Mroczek et al., 2015). Our design prevents us from distinguishing whether drinking occurred before or after stressor exposure, however, so interpretation of this result as informative for stress reactivity must remain speculative.

Stressors and drinking were also associated at the student and semester levels. At the student level, the average number of stressors was independently associated with higher odds of a drinking event above semester and daily stressor levels, but not with a greater number of drinks on drinking days. At the semester level, after adjusting for daily stressor level, average number of stressors was not associated with that semester’s likelihood of drinking, but among drinking days only, semester-level stress was associated with a lower average number of drinks that semester after adjusting for daily stressor level. Although findings at the student level generally paralleled those seen at the day level, findings at the semester level were opposite to the day-level findings and may reflect decreased levels of drinking during times of challenge or other constraints (e.g., high academic demands; Greene & Maggs, 2015). Importantly, inconsistencies in the stressor-drinking relationship across student, semester, and day levels of analysis highlight the likely possibility that relationships across these levels are driven by different causal processes (Hoffman & Stawski, 2009) that unfold across separate timescales (Ram & Gerstorf, 2009).

This study has limitations. First, as noted above, the within-person stressor-drinking links we identified are same-day relationships. Previous studies show a likely bidirectional relationship between stressor load and alcohol use (Armeli et al., 2007; Ayer et al., 2011; Levitt & Cooper, 2010). However, our study design is unable to sort out temporality. Techniques such as ecological momentary assessment (EMA; Shiffman, 2009; Wray, Merrill, & Monti, 2015), which can assess stressors and alcohol use multiple times in a single day, are needed. Second, we measured the AUDIT only during the fourth year of college. Thus, we could not control for prior AUDIT levels in our predictive models, which would have been ideal for establishing evidence of a prospective association. Instead, the RAPI, a conceptually similar measure of alcohol problems, served as our control for prior alcohol problem history. Our daily stressor-drinking event slope association with AUDIT alcohol problems remained even after controlling for the highly significant association between RAPI alcohol problem history and AUDIT, providing some assurance of validity. Although we did include both the AUDIT and the RAPI at year 4 and could model either as outcomes, we chose to model the AUDIT as the primary outcome because its validity as a measure of alcohol problems among college students has been well-established, whereas less is currently known about the validity of the RAPI in college populations (see reviews by Devos-Comby & Lange, 2008; Martens, Arterberry, Cadigan, & Smith, 2012), although some evidence supports the use of a selected item set among college students (Neal, Corbin, & Fromme, 2006). Importantly, the non-equivalence of measures means that caution is advised in drawing causal inferences from these predictive results, especially given that additional analyses using the RAPI as the alcohol problem outcome did not converge with those using the AUDIT. Third, our findings rely on student self-report of both stressors and alcohol use, which may inflate associations due to shared method variance. Moreover, because the particular events that constitute stressors vary from person to person, these ratings are inherently subjective, and we cannot be sure that the objective severity of experienced stressors is comparable across individuals (Almeida, Stawski, & Cichy, 2011). Fourth, our findings may not generalize to younger, older, or non-university populations. Indeed, although we found that daily-stressor drinking slopes predicted alcohol problems, Mohr and colleagues (2013) found that daily negative mood-drinking slopes did not predict future alcohol problems in a sample of community adults. Fifth, alcohol problems were assessed in the fall semester of Year 4 of college; thus, we are unable to predict how stressor-related drinking patterns across the college years may impact alcohol problems after graduation and further into adulthood. In fact, many young adults mature out of problematic alcohol use in their 20s (Maggs & Schulenberg, 2005), but this developmental pattern is not universal—some young adults continue heavy use and escalate into lifelong AUDs (see Schulenberg, O'Malley, Bachman, Wadsworth, & Johnston, 1996). Stressor-related drinking may help identify those who continue to follow problem trajectories. Long-term longitudinal research could effectively address this question.

Alcohol use among college students is a prevalent and costly problem, and stressors endemic to the college transition may play a role in alcohol problems among students. In this daily diary study, we found that students were more likely to drink on days with many stressors than days with fewer, and the number of drinks they consumed on drinking days was greater with each additional stressor experienced. Using multilevel modeling slopes to capture individual differences in stressor-related drinking during the first three years of college, we found that students whose drinking likelihood increased more sharply on high- versus low-stressor days were at greater risk for alcohol problems (as measured by the AUDIT) in the fourth year of college. Our results provide evidence of stressor-drinking slopes as predictors of alcohol problems in college students, adding to existing research showing the utility of within-person slopes and other measures of intraindividual variability as predictors of important health outcomes (e.g., Charles et al., 2013; L. H. Cohen et al., 2005; Mohr et al., 2013; Piazza et al., 2013). Moreover, our results are informative for theory and research on stressor-related drinking because our daily process measurement strategy allowed stressor-drinking slopes to be defined via experiences in people’s day-to-day lives, thus reducing recall bias and enhancing ecological validity relative to laboratory studies and retrospective reports (Bolger et al., 2003; Shiffman et al., 2008). Our findings suggest stressor-related drinking as a potential risk factor for future alcohol problems and suggest the need for continued research into the effectiveness of stress management training as a prevention strategy for alcohol problems among university students.