Comparing Behavioral Economic and Internalizing Symptom Predictors of Drinking Trajectories over Three Years During Emerging Adulthood

Abstract

Behavioral economics highlights external influences on alcohol’s reinforcing value and suggests life transitions in the mid-to-late 20s may increase costs of drinking, contributing to developmentally mediated drinking reductions among emerging adults. However, higher value for alcohol and fewer substance-free alternatives are theorized to predict persistent risky drinking. The current study compared behavioral economic indices of alcohol value with internalizing symptoms (e.g., anxiety, depression), an established correlate of drinking, as predictors of trajectories of heavy drinking days (HDD) and alcohol problems among emerging adults. Participants (N=495, 61.8% female) ages 21.5 to 25 years at enrollment who reported recent drinking (HDD M=1.58) completed questionnaires every four months for three years about alcohol consumption and problems, internalizing symptoms, alcohol and substance-free rewards, and demographics. HDD and alcohol problems were characterized with latent growth curve models. Baseline predictors were regressed onto alcohol growth curves. Both HDD and alcohol problems demonstrated quadratic reductions over time. Adjusting for demographics, demand intensity (consumption at zero price) and demand elasticity (sensitivity to changes in cost) were significantly associated with the HDD slope. Demand elasticity, internalizing symptoms, and environmental reward were significantly associated with the alcohol problems slope; lower elasticity and environmental reward scores were associated with accelerated early declines followed by deceleration of change in alcohol problems near the study’s end. Results suggest that elevated alcohol reward value, limited access to substance-free rewards, and internalizing symptoms confer persistent drinking risk, and that demand elasticity is a robust unique predictor of these outcomes.

General Scientific Summary:

Against the backdrop of normative alcohol changes in emerging adulthood, this study found that persistent motivation for alcohol use, operationalized as a willingness to consume alcohol at escalating levels of cost, was associated with shallower reductions in heavy drinking days and alcohol problems during emerging adulthood. Additional predictors of shallower reductions in alcohol problems over three years included higher internalizing symptoms and reward deprivation.

Introduction

Emerging adults (ages 18–25) drink more alcohol than other age groups (Substance Abuse and Mental Health Services Administration [SAMHSA], 2022). Although alcohol consumption may confer social benefits during this developmental period (Dumas et al., 2014), heavy drinking (5/4 drinks per occasion for males/females) is associated with acute consequences (e.g., blackouts and interpersonal conflict) and chronic risks (e.g., mortality and Alcohol Use Disorder [AUD] risk; Prince et al., 2019). Alcohol consumption peaks in emerging adulthood and declines across the lifespan (Bohm et al., 2021). Whereas developmental life transitions in the late teens and early 20s often precipitate alcohol consumption and heavy drinking, developmental tasks in the mid-to-late 20s are less compatible with heavy drinking. Accordingly, young adults often “age out” or reduce drinking during life transitions in their later twenties (Lee et al., 2015), though trajectories of heavy drinking appear to differ by demographic variables such as sex, race, and education. White and male participants tend to report the steepest increases in heavy drinking throughout adolescence and the greatest peak in the mid-20s with drinking levels that remain higher into the mid-30s (Chen & Jacobson, 2012). Those with a college education consume higher rates of alcohol (Crosnoe et al., 2017) but typically experience fewer alcohol-related harms over time compared to marginalized people and those without a college degree (i.e., alcohol harm paradox; Shuai et al., 2022).

Stable or escalating heavy drinking during emerging adulthood is associated with future AUD (Prince et al., 2019). Identification of predictors of drinking trajectories during emerging adulthood can thus enhance risk models that inform prevention and early interventions related to alcohol use disorder. In this study, we compare two robust and consistent correlates of harmful drinking in emerging adulthood, namely internalizing symptoms and behavioral economic indices of value, as predictors of drinking trajectories over time.

Internalizing Symptoms as Predictors of Alcohol Change in Emerging Adulthood

Internalizing psychiatric conditions, such as depression and anxiety, commonly co-occur with heavy drinking (Anker et al., 2023) and may serve as early risk markers for sustained alcohol consumption. Among emerging adults, alcohol consumption often occurs in social contexts, and social drinking may become less common following developmental transitions toward stages with increased responsibilities (e.g., family, work). However, stressors associated with new responsibilities may increase risk of alcohol-related harm, particularly among those consuming alcohol to cope. Greater depressive symptoms are consistently associated with higher levels of alcohol harm (Dennhardt & Murphy, 2011; Pedrelli et al., 2016), and longitudinal studies demonstrate that depressive symptoms are associated with alcohol problems 12 months later (Acuff et al., 2018). AUD and anxiety disorders are reciprocally related over time (Kushner et al., 1999), supporting negative affect as a marker for alcohol-related harm across emerging adulthood.

Behavioral Economics

Behavioral economics may also inform models of aging out. From the perspective of behavioral economic models of harmful alcohol use (Acuff, MacKillop, et al., 2023; Bickel et al., 2014), alcohol use confers benefits, such as euphoria and social connection, in addition to costs, such as interference with important life activities (e.g., school, relationships, work), and choices are made to maximize benefits and minimize costs probabilistically. Behavioral allocation typically occurs in the context of alternative activities and commodities, and likely choice outcomes are dependent upon the immediate and delayed costs and benefits of both alcohol and available alternative options. Researchers have quantified three critical factors: alcohol value (alcohol demand); alternative, substance-free reinforcement; and the tendency to prefer immediate rewards to larger, later rewards (delay discounting). Given the proposed stability of behavior, these three factors may serve as early detection markers of reduced probability of “aging out” among emerging adults (Acuff, Belisario, et al., 2023; Bird et al., 2024; Patrick et al., 2022).

Alcohol Demand

Alcohol value, or demand, is commonly measured with alcohol purchase tasks that quantify the degree of alcohol-seeking behavior (i.e., alcohol purchasing) as a function of increasing financial costs. In a typical alcohol purchase task, participants report how many standard drinks they would consume across increasing monetary prices. Alcohol consumption characteristically decreases as prices increase in a typical purchase task, indicating that people are sensitive to cost increases. However, there are individual differences in sensitivity to these costs that are captured by demand indices. These can be extracted from the purchase task data and represent different elements of motivation to consume alcohol: intensity (consumption at zero price, indicating raw alcohol value when there are no costs), O_max (maximum monetary expenditure over the course of the task), and elasticity (an index of sensitivity of demand to changes in cost). Greater intensity and O_max, and lower elasticity (or more “inelastic” demand) are indicative of greater motivation to consume alcohol, or alcohol value.

Higher alcohol demand may serve as a risk factor for drinking over time during the normative aging-out period. Longitudinal studies suggest that alcohol demand changes dynamically in conjunction with changes in alcohol consumption (Gex et al., 2022) and that changes precede changes in alcohol problems (Acuff, Belisario, et al., 2023; Bird et al., 2024; Cheong et al., 2022; Murphy et al., 2015). One study (Guame et al., 2022) explored demand indices as a baseline predictor over a single extended temporal interval of four years using a sample of 4,594 young Swiss men. Demand intensity, O_max, and elasticity all predicted self-reported weekly drinking amount, monthly heavy drinking days (HDD), alcohol-related negative consequences, and DSM-5 AUD symptoms after controlling for the baseline drinking-related and demographic factors. These studies suggest that alcohol demand may predict drinking trajectories among emerging adults. However, the Gaume paper focused on Swiss men young men only, and no study has explored alcohol demand indices as time-invariant predictors of drinking trajectories assessed at multiple timepoints over an extended period among emerging adults.

Alternative Substance-Free Reinforcement

Substance-free reinforcement is often measured using reinforcement surveys assessing frequency and enjoyment of activities while using a substance and while sober (see Acuff et al., 2019 for a review). From these data, a ratio is formed quantifying reinforcement from substances relative to reinforcement from substance-free activities. Generally, studies support an inverse association, such that higher levels of substance-free reinforcement are associated with less alcohol consumption and related problems (Acuff, MacKillop, et al., 2023; Kuhlemeier et al., 2024). Those with high levels of proportionate substance-related reward, or with low levels of environmental reward, may be particularly susceptible to stable, rather than declining, levels of heavy alcohol consumption as there are fewer alternatives to compete with alcohol. Two studies investigating associations between alcohol problems and indices of alternative reinforcement have demonstrated that higher proportions of substance-related, relative to substance-free, reinforcement predicted greater alcohol problems four months later, whereas alcohol problems did not predict greater proportionate reinforcement (Acuff, Belisario, et al., 2023; Bird et al., 2024). Acuff et al. (2023) also found that greater HDD preceded increases in proportionate substance-related reinforcement, suggesting that heavy drinking may lead to overall shifts in proportionate time allocation and enjoyment related to substance-related vs. substance-free activities. However, it is unknown if proportionate substance-related reinforcement at a single assessment in emerging adulthood predicts trajectories of drinking over extended periods of time. Other studies using alternative operationalizations of reward, such as the reward probability index (Carvalho et al., 2011), suggest that emerging adults with low perceived levels of reward in their environment are at heightened risk for greater alcohol severity and may uniquely benefit from brief alcohol interventions that enhance access to substance-free rewards (Murphy et al., 2021).

Delay Discounting

Delay discounting is often conceptualized as a trait-level index of the degree to which a delayed reward decreases in value as its receipt moves further away in time (Madden et al., 2003). Delay discounting tasks present respondents with a series of choices between smaller rewards now or larger rewards later. Responses on these tasks are used to identify the hyperbolic discounting function, which highlights that rewards with closer delays are discounted less steeply, whereas the discounting rate increases as the reward receipt moves further away in time. Meta-analyses have demonstrated robust associations between delay discounting and alcohol-related risk cross-sectionally (Amlung et al., 2017). However, there is a relative dearth of longitudinal studies establishing associations between delay discounting and alcohol outcomes over time. Studies conducted among adolescents generally suggest that higher rates of delay discounting early in adolescence (age 14), and increases in delay discounting trajectories across adolescence, predict escalating trajectories of alcohol use in late adolescence and emerging adulthood (Felton et al., 2020). More research is needed to understand whether delay discounting is associated with trajectories of drinking reductions during emerging adulthood.

Current Study

Research has established associations between harmful alcohol use and symptoms of internalizing disorders, which commonly co-occur with harmful alcohol use and are thought to be one potential mechanism leading to greater alcohol use severity. Research has also begun to establish behavioral economic indices of alcohol demand, alternative substance-free reinforcement, and delay discounting with alcohol outcomes cross-sectionally and over short periods of time. The current study compared internalizing symptoms and behavioral economic indices of value as predictors of trajectories of heavy drinking days and alcohol problems. The first aim of these analyses was to characterize changes in HDD and alcohol-related problems among emerging adults in their early-to-mid 20s. We hypothesized that HDD and alcohol-related problems would continue to decrease over the 3-year period (beyond those established in previous analyses; Acuff et al., 2023). The second aim was to explore time-invariant predictors of changes in HDD and alcohol-related problems. We hypothesized that greater alcohol demand, delay discounting, proportionate substance-related reinforcement, environmental reward, and internalizing symptoms at baseline would be associated with greater HDD and alcohol-related problems, and a shallower decline over time. Support for behavioral economic predictors as baseline time-invariant predictors of drinking trajectories across emerging adulthood would suggest their utility as early intervention markers of harmful alcohol use among emerging adults.

We previously examined bidirectional associations between behavioral economic indices and alcohol problems and heavy drinking using a subset of data used for the current study (Acuff, Belisario, et al., 2023). We found reductions in alcohol problems and HDD across assessments, significant between-person effects between behavioral economic indices (reinforcement ratio, demand elasticity, intensity, and O_max) and alcohol problems and HDD. Further, changes in reinforcement ratio were associated with decreases in alcohol problems across the full sample, demand intensity and O_max predicted change in alcohol problems for male participants, and change in intensity predicted change in alcohol problems for non-White participants. The analyses for the current study extended these findings because they (1) model the latent growth of heavy drinking and alcohol-related problems over multiple assessments for a longer period of time (3 years versus 1.5 in the previous analyses), and (2) evaluate baseline time-invariant predictors as early risk markers of growth trajectories.

There is also some evidence that the COVID-19 pandemic influenced emerging adult alcohol use (Acuff et al., 2022; Baptist Mohseni et al., 2022; Creswell et al., 2024); however, the impact of COVID-19 on the emerging adult aging out trajectory has not been studied. For the current study, data previously organized by data collection wave was reorganized into 4-month calendar intervals to better disaggregate aging out trajectories from COVID-19 effects, differentiating these data from those previously published.

Methods

Participants and Procedures

The sample is a subset of participants (n=495) from a larger study investigating trajectories of drinking behavior in emerging adulthood (N = 601). Initial enrollment eligibility criteria were: (i) 2+ binge drinking days in the past month, defined as >4 standard drinks/>3 standard drinks for males/females;¹ (ii) between 21and 25 years of age; (iii) fluency in English; and (iv) no past/current psychotic disorders. Enrollment occurred on a rolling basis, starting in September 2017, with follow-up assessments occurring every 4 months; all study data collection ended July 2022. Participants were initially recruited from a midsize city in the southern United Sates. Those eligible participated in a 2-hour session in a university research laboratory, where they completed an assessment battery. Follow-up sessions occurred every four months alternating between in-person and remote assessments. In March 2020, all assessments were shifted to remote data collection in response to the COVID-19 pandemic. This study was approved by the University of Memphis Institutional Review Board (project #4320).

To assess the impact of the COVID-19 pandemic, participant data were re-organized based on assessment date into 4-month intervals, resulting in distinct time points separated by the declaration of COVID-19 as a global pandemic by the World Health Organization on March 11th, 2020. This resulted in six pre-COVID and four intra-COVID waves (up to and including July 10th, 2021). The first timepoint included a longer period than 4 months due to the nature of the ongoing enrollment into the study.

Measures

Alcohol Outcomes

Heavy Drinking Days.

HDD were measured using the Daily Drinking Questionnaire (DDQ). Participants reported the typical number of standard drinks consumed on each calendar day of the week over the past month (range = 0–7). HDD were defined using the NIAAA definition of 5+/4+ drinks per occasions for males/females.

Alcohol-related Problems.

The 24-item Brief Young Adult Alcohol Consequences Questionnaire (B-YAACQ; Kahler et al., 2008) was used to count the number of alcohol-related problems participants experienced in the past month (e.g., “While drinking, I have said or done embarrassing things”). Items are summed to create a total score. The B-YAACQ is a reliable and validated measure for measuring alcohol problems in emerging adult populations (Kahler et al., 2008).

Time-Invariant Baseline Predictors

Demographic Predictors.

Participants reported their sex assigned at birth (male, female), race, age, perceived financial status, and education level. Participants were able to report their race in a “check all that apply format”. For analyses, race was dichotomized (Non-Hispanic White and minoritized racial identity). For financial status, following Najdzionek et al. (2023), participants responded to: “Without giving exact dollars, how would you describe your household’s financial situation right now?” Participants could answer on a scale from 1–4: “Not enough to pay some bills no matter how hard you try”, “Enough to pay bills, but have had to cut back”, “Enough to pay bills without cutting back but no ‘extras’”, “Enough money for ‘extras’”. Perceived income was dichotomized (1–2, 3–4) for the current analyses. We created a dichotomous variable for education level: Those without a four-year degree or not currently enrolled in a four-year program were categorized as “no four-year/non-college,” while those enrolled in or having completed a four-year degree were categorized as “four-year college degree.

Internalizing Symptoms.

Depression symptomatology was measured for the past month using the Patient Health Questionnaire (PHQ-9; Kroenke et al., 2001). The nine PHQ-9 items are rated by frequency on a scale from not at all (0) to nearly every day (3). Scores on the PHQ-9 have been shown to predict diagnosis of depression (Kroenke et al., 2001). Internal consistency for the PHQ-9 in the current sample was strong (α=0.90 [0.87 – 0.91], ω=0.90 [0.87–0.92]. The 7-item Generalized Anxiety Disorder Questionnaire (GAD-7) was used to measure anxiety symptomatology experienced in the past month (Spitzer et al., 2006). Participants report the frequency of symptoms of anxiety on a scale from not at all (0) to nearly every day (3). Scores on the GAD-7 have been shown to predict diagnosis of anxiety (Spitzer et al., 2006). Internal consistency for the GAD-7 in the current sample was strong (α=0.93 [0.92–0.94], ω = 0.93 [0.92–0.94]). The Perceived Stress Scale (PSS; Cohen, 1988) consists of 4 questions pertaining to thoughts and feelings in the past month, with higher scores suggesting greater stress. Internal consistency for the PSS in the current sample was strong (α=0.76 [0.66–0.83], ω=0.76 [0.68–0.83]).

Proportionate Substance-related Reinforcement.

Proportionate substance-related (relative to total) reinforcement was assessed using the Activity Level Questionnaire (ALQ; Acuff, Belisario, et al., 2023; Meshesha et al., 2020), a reinforcement survey assessing past-month frequency and enjoyment of 36 activities (e.g., “Reading”, “Having a conversation with another person”), both with and without alcohol or other drug involvement. Frequency and enjoyment ratings were multiplied to obtain cross-product scores reflecting reinforcement derived from both substance-related and substance-free activities (Murphy et al., 2005). Next, the substance-related reinforcement cross-product derived in the last step was divided by the total reinforcement score (substance-free reinforcement total + substance-related reinforcement total) to calculate proportionate substance-related reinforcement (referred to as the reinforcement ratio). Proportions range from 0 to 1, with higher values indicating a greater proportion of total enjoyable activity engagement related to activities that included alcohol or drug use. To ensure satisfactory data quality on this lengthy questionnaire, for participants giving identical responses on >80% of the non-alcohol related frequency items (29+ of the 36 items), their alcohol reinforcement ratio was recoded to missing.

Environmental Reward.

The Environmental Suppressors scale of the Reward Probability Index (RPI; Carvalho et al., 2011) is a self-report measure designed to assess accessibility of environmental rewards and approximate overall response-contingent positive reinforcement (RCPR; defined as an increase in the frequency or duration of a behavior in response to a reinforcer). Participants were asked to rate each item (e.g., “Changes have happened in my life that have made it hard to find enjoyment.”) on a 4-point scale from 1 (strongly disagree) to 4 (strongly agree), and responses are summed for a total score (range = 9 to 36), with higher scores indicating greater availability of rewarding opportunities that should decrease risk for drinking. We used the 8-item version of the environmental reward subscale (Voss et al., 2021). Reliability for this index was excellent (α=0.90; ω=0.90).

Alcohol Demand.

The Alcohol Purchase Task (APT; Murphy & MacKillop, 2006) evaluates self-reported alcohol consumption and monetary expenditures over a range of drink prices in a typical drinking situation to produce an overall measure of alcohol demand. Participants were asked to report the number of standard drinks they would consume across 30 prices ranging from $0 to $40 per drink, with the contingencies that drinks could not be stockpiled for later consumption and no drinking could occur before or after the hypothetical purchase task. Alcohol demand corresponds with real-world drinking (Motschman et al., 2022) and with changes in drinking over time (Gex et al., 2022), supporting the concurrent validity of these items.

Delay Discounting.

Delay discounting was measured using the Effective Delay-50 5-item titrating delay discounting task (Koffarnus & Bickel, 2014). Participants choose between two hypothetical amounts of money: a smaller, immediately available amount or a larger, delayed amount. For all trials, the larger delayed reward was $1000, and the smaller immediate reward was $500. After each choice, the time to delayed reward was titrated based on the participant choice. The final choice outcome corresponds to a specific ED₅₀ delay, which can be used to estimate the discount rate or k value (log transformed). Higher k values reflect a greater proportion of selecting the smaller, immediate reward (i.e., steeper discounting). Hypothetical money choices have been shown to be valid and reliable estimates of discounting rates that are comparable to those generated from tasks using real monetary options (Madden et al., 2003).

Data Analysis

For each variable within each timepoint, outliers greater than four standard deviations from the mean were winsorized to one unit greater than the next highest non-outlying value. Values for time-invariant predictors were extracted from each participant’s earliest pre-COVID assessment. A latent internalizing symptom factor was created using total scores on the PSS, GAD-7, and PHQ-9. Indicators contributing covariance greater than .40 were considered adequate. We extracted the latent factor scores for use as indicators for the remaining models to improve convergence.

Demand data were screened for inconsistency using three criteria: (1) all things equal, consumption should decrease, rather than increase, as prices increase across the full range of a task (trend=0.025); (2) consumption should decrease or stay the same from one price to another (bounce=0.10); and (3) consumption will not reappear at a higher price once zero consumption has been attained at a lower price (reversal from zero; Stein et al., 2015). The current study used PThelper R macro to clean the demand data (Belisario et al., 2022). A total of 16 unique participants were removed due to data quality violations (reversal from zero n=2; trend violation [delta Q<0.025] n=0; bounce violation n=14). Two observed indices are reported: O_max and intensity. We also derived alpha as an index of elasticity (or change in consumption as a function of increasing price) using the exponentiated demand equation (Koffarnus et al., 2015):

$$Q=Q_0*{10}^{k(e-\alpha P-1)}$$

Where Q = quantity consumed, P = price, and α = elasticity of demand. This equation anchors outcomes to the span of consumption using a constant (k) which reflects consumption in the full sample from zero price to the maximum price. This is important because the change in consumption is relative to the prices set in the task. In our study this value (k) was 2.6 calculated as the span of consumption in log units using the mean average consumption at the lowest and highest price points, plus 0.5 (Kaplan et al., 2019). These indices were chosen for parsimony, as they are the better established indices for predicting alcohol-related outcomes (Martínez-Loredo et al., 2020).

Statistical analyses were conducted using Mplus Statistical Software Version 8.10 (Muthén & Muthén, 2017). Data analysis took place in three steps. First, we tested unconditional Latent Growth Curve Models (LGCMs) for HDD per week and alcohol problems to determine the best model. We tested: (1) an intercept-only model, which assumes no change in slope over time; (2) a linear model, which assumes a linear rate of change over time; (3) a quadratic model, which assumes the rate of change either slowed or accelerated over time; and (4) a piecewise model, which modeled one linear slope from the first timepoint to the timepoint immediately preceding the COVID survey and another linear slope from the timepoint preceding the COVID survey to the final assessment. The best fitting model was guided by the following model fit indices in all analyses, suggested by Hu and Bentler (1999) and Kline (2016): Comparative Fit Index (CFI; acceptable fit >0.90), Standardized Root Mean Square Residual (SRMR; acceptable fit <.08), and Root Mean Square Error of Approximation (RMSEA; acceptable fit <.08). Missing data were handled with Full Information Maximum Likelihood (FIML). When fit indices were relatively equal, model selection favored parsimony. Second, we examined time-invariant predictors of latent growth drinking trajectories, and used a stepped approach to test: (a) an omnibus demographic model that included sex assigned at birth, race, age, financial status, and education as time-invariant predictors of growth trajectories; (b) a model that added the internalizing symptom latent variable to the omnibus demographic model; and (c) separate models that added each behavioral economic variable to the omnibus demographic model. Third, any variable (demographic, internalizing symptoms, or behavioral economic) demonstrating a significant association with a linear or quadratic slope across all models was included in a final multivariate model. All data and syntax have been made publicly available at the Open Science Framework and can be accessed at https://osf.io/zv3nw/?view_only=d265b1f2fed24d888ca7e158124fb656. Model fit data and demographic comparison between included and excluded participants can be found in the Supplemental Materials. The study was not preregistered.

Results

Demographics and Preliminary Analyses

Table 1 reports sample characteristics. The majority of participants had data for all waves, with a mean (± SD) of 5.35 (±1.57) pre-COVID waves (median of 6), and a mean ± SD average of 3.80±1.37 number of intra-COVID waves (median of 4). At baseline, the final sample reported a mean±SD age of 22.6±1.0; 61.8% were assigned female at birth; 44.9% were White and 40.8% were Black. More than 1.5 HDD were reported per week on average, and an average of 6.37 (SD=4.50) past one-month alcohol-related problems were reported.

Table 1. Sample characteristics at enrollment (n = 495).

Variable	M (SD) / Median	%
Age	22.63 (1.03)
Sex
Female		306 (61.82%)
Male		189 (38.18%)
Race
Asian		17 (3.43%)
Black		202 (40.81%)
Multiracial		20 (4.04%)
Native American/Pacific Islander		3 (0.61%)
Non-Hispanic White		222 (44.85%)
Other		30 (6.06%)
Median Household Income	$30,000 - $44,999
Educational Outcome
<4 Year Degree		166 (33.54%)
4 Year Degree		329 (66.46%)
Drinks per Week	16.52 (14.61)
HDD per week	1.58 (1.58)
B-YAACQ Score	6.37 (4.50)

Note. HDD = Weekly heavy drinking days; B-YAACQ = Brief Young Adult Alcohol Consequences Questionnaire.

The internalizing symptom measurement model estimating a latent variable was fully saturated and demonstrated excellent fit. PHQ-9 total (Standardized Estimate [Est.]=.89, Standard Error [SE]=.03, p<.001), GAD-7 total (Est.=.84, SE=.03, p<.001), and PSS total (Est.=.72, SE=.03, p<.001) all loaded well onto the internalizing factor.

Modeling Latent Growth Curves for HDD and Alcohol Problems

Table 2 presents fit indices for unconditional LGCMs, and Figure 1 presents observed mean values and estimated slopes for alcohol outcomes. Model fit indices suggested that the quadratic model best fit the data for both alcohol outcomes. For HDD, latent variable means were positive for the intercept (Est. M_intercept=1.56), indicating that participants reported 1.56 HDD per week on average at the first assessment; negative for the linear slope (Est. M_linear=−1.46), indicating a downward trend over time, and positive for the quadratic slope (Est. M_quadratic=0.52), indicating greater reductions earlier in the follow-up assessment interval with deceleration of reductions as the end of the follow-ups neared. Examination of mean observed values confirmed interpretation of the derived latent variable means. There was a minor increase in HDD at the start of the COVID-19 pandemic; however, this did not appear (visually or in model fit) to disrupt the overall trajectory.

Table 2. Model Fit Indices for Unconditional Latent Growth Curve Models.

	χ2 (df), p-value	RMSEA (90% CI)	CFI	SRMR	AIC	BIC
Heavy Drinking Days
Intercept only	262.43 (53), <.001	0.09 (0.08, 0.10)	0.74	0.13	12422.25	12472.71
Linear	78.29 (50), <.001	0.03 (0.02, 0.05)	0.96	0.07	12064.11	12127.18
Quadratic	49.47 (46), 0.34	0.01 (0.00, 0.03)	1	0.04	12016.3	12096.19
Piecewise	49.84 (46), 0.32	0.01 (0.00, 0.03)	1	0.04	12015.5	12095.38
Alcohol Problems
Intercept only	357.93 (53), <.001	0.11 (0.10, 0.12)	0.68	0.15	19087.66	19138.12
Linear	105.22 (50), <.001	0.05 (0.4, 0.06)	0.94	0.08	18608.38	18671.45
Quadratic	76.17 (46), .003	0.04 (0.02, 0.05)	0.97	0.06	18558.68	18638.56
Piecewise	78.44 (46), .002	0.04 (0.02, 0.05)	0.97	0.06	18564.06	18643.94

Note. Bolding indicates best across models for each model fit index. Df = degrees of freedom; RMSEA = Root Mean Square Error of Approximation; CFI = Comparative Fit Index; SRMR = Standardized Root Mean Squared Residual; AIC =Akaike Information Criterion; BIC = Bayesian Information Criterion.

^aThe initial model was nonpositive definite due to high correlations amongst the growth factors. As a result, the variance of the quadratic slope was fixed to zero.

Figure 1. Trajectories of Heavy Drinking Days and Alcohol Problems from July 2018 through July 2021.

Note. Trajectories of (A) weekly heavy drinking days and (B) alcohol problems from July 2018 through July of 2021. For both panels, dots represent raw data, whereas the line represents the estimated slope.

For alcohol-related problems, latent variable means were positive for the intercept (Est. M_intercept=6.56); negative for the linear slope (Est. M_linear=−6.06), indicating a downward trend over time; and positive for the quadratic slope (Est. M_quadratic=2.55), indicating greater reductions earlier in the assessment period with deceleration of reductions as the end of the study period neared. The mean observed values for alcohol problems similarly confirmed interpretation of the derived latent variable means. There was no evidence that the COVID-19 pandemic disrupted the trajectory in alcohol problems over time.

Time-Invariant Predictors of HDD and Alcohol Problems

Results for all time-invariant predictor models are reported in Table 3.

Table 3. Time Invariant Predictors of Trajectories of Heavy Drinking Days per Week and Past Month Alcohol-Related Problems.

	Heavy Drinking Days per Week			Past-month Alcohol-Related Problems
	Intercept	Linear Slope	Quadratic Slope	Intercept	Linear Slope	Quadratic Slope
	Est. (S.E.)	Est. (S.E.)	Est. (S.E.)	Est. (S.E.)	Est. (S.E.)	Est. (S.E.)
Separate Models
Model 1: Demographic Omnibus
Sex (Ref.=Males)	.004 (.06)	−.01 (.09)	.01 (.09)	−.09 (.06)	.06 (.09)	−.05 (.10)
Race (Ref.=White)	.08 (.06)	−.06 (.08)	.10 (.08)	.08 (.06)	−.02 (.08)	.03 (.09)
Age	.13 (.06)	−.14 (.09)	.08 (.08)	.02 (.05)	−.03 (.08)	−.003 (.09)
Perceived Income Status	−.21 (.06)***	.10 (.09)	−.03 (.09)	−.15 (.06)**	.14 (.08)	−.15 (.09)
Education	−.27 (.06)***	.11(.10)	−.04 (.10)	−.09 (.06)	.12 (.09)	−.10 (.10)
Model 2A: Model 1 + Internalizing
Internalizing Latent Variable	.02 (.02)	−.04 (.06)	.04 (.06)	.48 (.05)***	−.22 (.09)*	.12 (.10)
Model 2B-2G: Model 1 + Separate Inclusion of Behavioral Economic Predictors
Environmental Reward Probability Index	−.08 (.06)	.13 (.09)	−.13 (.09)	−.46 (.05)***	.27 (.08)**	−.19 (.09)*
Reinforcement ratio	.33 (.07)***	−.12 (.10)	.04 (.10)	.33 (.06)***	.03 (.10)	−.16 (.10)
Intensity	.54 (.06)***	−.22 (.09)*	.10 (.09)	.25 (.04)***	−.11 (.13)	.02 (.13)
Omax	.47 (.06)***	−.17 (.10)	.01 (.11)	.28 (.06)***	−.16 (.09)	.10 (.11)
Elasticity	−.38 (.08)***	.26 (.10)**	−.15 (.09)	−.26 (.06)***	.30 (.09)**	−.29 (.11)**
Delayed Discounting	.15 (.06)**	−.14 (.10)	.06 (.09)	.07 (.06)	−.05 (.08)	−.02 (.08)
Model 3: Model 1 + Significant Model 2 Predictors
Internalizing Latent Variable	-	-	-	.33 (.07)***	−.09 (.12)	−01 (.13)
RPI - Environmental Reward	-	-	-	−.21 (.07)**	.18 (.11)	−.18 (.12)
Intensity	.48 (.06)***	−.18 (.09)*	.08 (.09)	-	-	-
Elasticity	−.26 (.06)***	.19 (.09)*	−.11 (.09)	−.25 (.06)***	.30 (.09)**	−.29 (.10)**

Note. RPI = Reward Probability Index; Est. = Estimate; S.E. = Standard Error.

HDD.

There was a significant negative association between both financial status and education with the intercept for HDD, suggesting that greater financial status and education levels were associated with fewer weekly HDD. There were no other significant demographic predictors.

The internalizing latent factor was not significantly associated with any HDD indicator. Separate behavioral economic models were also tested. Controlling for all demographic indicators, higher reinforcement-ratio, intensity, O_max, discounting and lower elasticity was associated with higher HDD intercepts. Higher demand intensity was associated with a smaller negative slope, indicating less reductions in HDD over time. Finally, there was a positive association between demand elasticity and the negative linear HDD slope, suggesting that inelastic demand was associated with less reductions in HDD over time. No other associations were significant.

In the final omnibus model, intensity and elasticity were included with all demographic covariates. Consistent with hypotheses, demand intensity was significantly positive associated with the HHD intercept and negatively associated with the linear slope, whereas demand elasticity was negatively associated with the HHD intercept and positively associated with the linear slope. No other associations were significant (Figure 2).

Figure 2. Trajectories of Alcohol Problems Split by Elasticity, and Heavy Drinking Days Split by Intensity, from July 2018 through July 2021.

Note. Dots represent raw data; lines represent estimated slopes. (A) Trajectory of alcohol problems from July 2018 through July of 2021 split by elasticity at the first assessment. Among those in the top 75% of elasticity values (less demand for alcohol), alcohol problems were lower at the first assessment, and the slope was more linear. Among those in the bottom 25% of elasticity values (greater demand for alcohol), alcohol problems were greater at the first assessment, and alcohol problems demonstrated greater quadracity. (B) Trajectory of weekly heavy drinking days (HHD) from July 2018 through July of 2021. Those with the lowest demand intensity demonstrated the shallowest decline in weekly HHD, whereas those with higher levels of demand intensity demonstrated slightly steeper reductions over the study period; those with high demand experienced, on average, one additional HHD per week relative to the low demand intensity group. Weekly HDD were slightly quadratic in the low moderate intensity group.

Alcohol-related Problems.

There was a significant negative association between financial status and the intercept for alcohol problems, suggesting that higher financial status was associated with fewer alcohol problems. There were no other significant associations.

The internalizing latent factor was significantly associated with the alcohol-related problems intercept and linear slope. Higher scores on the internalizing latent factor were associated with higher alcohol problems intercepts and shallower reductions over time. Separate behavioral economic models were also tested. Controlling for all demographic indicators, higher reinforcement-ratio, intensity, and O_max, and lower environmental reward and demand elasticity, were associated with higher alcohol problem intercepts. Higher environmental reward was associated with a larger negative slope, indicating greater reductions in alcohol problems over time. There was also a positive association between demand elasticity and the negative alcohol problems linear slope, suggesting that inelastic demand was associated with less reductions in alcohol problems over time. Finally, both elasticity and environmental reward were significantly associated with the quadratic slope. These results suggest that higher elasticity and environmental reward were associated with stable, consistent reductions in alcohol problems over time, whereas lower elasticity and environmental reward scores were associated with accelerated early declines followed by deceleration and flattening of changes in alcohol-related problems near the end of the assessment period. No other associations were significant.

In the final omnibus model, internalizing symptoms, environmental reward, and elasticity were included with all demographic covariates. The same effects were present when significant predictors were included in the same model. Associations between elasticity and the linear and quadratic slopes also remained. However, the associations between environmental reward and the linear and quadratic slopes for alcohol problems were no longer significant. No other associations were significant (Figure 2).

Discussion

In the current secondary data analysis, we hypothesized that HDD and alcohol-related problems would decrease over time, and that greater alcohol demand, delay discounting, proportionate substance-related reinforcement, and internalizing symptoms would be associated with greater harmful alcohol use at baseline and shallower declines over the course of the longitudinal study. Consistent with our first hypothesis and broadly consistent with the aging out literature (Lee et al., 2015), our results demonstrated overall reductions in both HDD and alcohol-related problems across the study period, with steeper reductions occurring early. Our results did not support a sustained impact of the COVID-19 pandemic on drinking trajectories. While there was a slight increase in HDD during the initial 4-month COVID-19 shut-down period, this deviation disappeared in subsequent assessments.

Alcohol problems demonstrated a consistent, monotonic and quadratic decline over the assessment period with no apparent impact of the COVID-19 shutdown on mean levels of alcohol problems in the current sample. Notably, the standard deviation for alcohol problems during the initial 4-month COVID-19 shut-down period was much greater than during later assessments, reflecting greater variability relative to other assessments. However, this was not replicated in later periods, suggesting, at most, a transient COVID-19 effect not captured by the current assessment window. This is consistent with most research that emerged in the immediate aftermath of the COVID-19 pandemic suggesting that drinking increased for some and decreased for others, in part depending upon an individual’s environment and psychosocial stressors (Acuff et al., 2022).

Previous research has demonstrated variability in slopes for HDD and alcohol problems across important demographic characteristics (Chen & Jacobson, 2012). The current study suggests that higher financial and education status were associated with the lower baseline HDD, whereas higher financial status only was associated with a lower baseline levels of alcohol problems among emerging adults. These effect sizes were small in magnitude. No other association with any linear or quadratic slope was significant, indicating that, within our sample, demographic predictors were not associated with differential declining slopes across these demographic groups. Previous research indicated differential slopes over large developmental periods from late adolescence to middle adulthood (Chen & Jacobson, 2012). Given an intercept effect for financial and educational status, there may have been differences in slopes that emerged during the accelerating slope, or during the peak, that occurred developmentally prior to the start of our sampling window and account for differences in the intercepts.

The time invariant predictor analyses partially supported hypotheses in the second study aim. All behavioral economic variables were associated with the HDD baseline, and most were associated with the alcohol problems intercept (except delay discounting). These findings reflect the broad literature on cross-sectional associations between behavior economic and alcohol risk outcomes. Demand intensity and elasticity at baseline also predicted changes in HDD, and elasticity also predicted changes in alcohol-related problems, over three years. These results persisted in the omnibus model. Our results extended previous research that found, among college students reporting heavy drinking, changes in demand intensity paralleled changes in both heavy drinking and alcohol-related negative consequences over the course of 16-months. The magnitude of the effect was small to medium. These findings suggest some long-term predictive stability even across a developmental period defined by changes in drinking patterns.

Greater environmental reward was also associated with a greater linear decline over the follow-up period, extending previous work linking lower environmental reward with risk for alcohol problem severity and AUD (Murphy et al., 2021). Importantly, this variable was no longer significant in the omnibus models, suggesting that the covariance is better accounted for by demand elasticity. Similarly, higher perceived income and education, both of which may facilitate access to substance-free rewards, were associated with less heavy drinking at baseline. Perceived income was also inversely associated with alcohol-related problems. These results are consistent with models of reward deprivation that suggest that access to important environmental reinforcers is protective against substance use (Acuff, MacKillop, et al., 2023; Shuai et al., 2022).

In several cases, our hypotheses were not directly supported, and several methodological and theoretical points may explain these results. While the lack of baseline association between discounting and alcohol problems is surprising, previous research has suggested that effect sizes for discounting tend to be larger in (1) more clinically severe samples, and (2) when longer form measures are used (Amlung et al., 2017; MacKillop et al., 2011). The lack of a longitudinal association between discounting, O_max, and substance-related reinforcement with HDD and alcohol problem slopes may also be explained by the developmental nature of emerging adulthood. Emerging adulthood is marked by shifting financial stability, and both O_max and discounting are directly related to monetary reward and are associated with socioeconomic status (Owens et al., 2015; Reimers et al., 2009). Conceivably, shifts during emerging adulthood may wash out any potential associations. Previous research demonstrated that proportionate alcohol-related reinforcement and O_max are predictive of drinking over time (Cheong et al., 2022; Gaume et al., 2022; Kuhlemeier et al., 2024; Murphy et al., 2005, 2015). Differences in alcohol change stages across samples (e.g., unmotivated heavy drinking college students prior to aging out and middle-aged adults attempting to reduce drinking or seek treatment) may partially explain differences in findings. Studies have shown that proportionate alcohol-related reinforcement does change with changes in drinking (Acuff et al., 2023), suggesting that a single timepoint predicting future alcohol consumption during a period known to be volatile may be less predictive than assessments conducted repeatedly over time. Further, environmental reward was associated with the alcohol problem slope, providing some support for this hypothesis. It is also conceivable that discounting, O_max, and substance-related reinforcement are not predictive of trajectories of drinking. More research is needed to eliminate alternative explanations to confirm a consistent null effect.

Partially supporting our hypotheses in our second aim, our findings indicated that internalizing symptoms were not associated with HDD intercept or slopes but demonstrated a small-to-medium magnitude association with those of alcohol-related problems. These results are broadly consistent with literature suggesting that depression and anxiety are associated with negative consequences of drinking, though not with levels of alcohol consumption (Pedrelli et al., 2016; Roberts et al., 2010; Schry & White, 2013). Our research extends this work by suggesting that internalizing symptoms may be an indicator of risk for sustained alcohol problem trajectories. People with high levels of depression or anxiety may become more reliant on alcohol to fulfill a function, such as providing relief, facilitating social interactions, or calming racing thoughts. Even if peers are using alcohol at similar levels for social or enhancement reasons, people with higher internalizing symptoms may experience greater problems due to the manner or setting in which alcohol is consumed. It is unclear whether high internalizing symptoms are a causal risk factor or perhaps a marker of a third variable. When included with behavioral economic predictors, however, internalizing symptoms were only associated with the intercept, suggesting that the variance is better accounted for by demand elasticity.

Limitations & Future Directions

Analyses used robust complex models in a moderately large sample to determine whether time-invariant behavioral economic predictors are risk factors for trajectories of heavy drinking and alcohol-related problems during a period when many young people reduce drinking, but the conclusions are nonetheless constrained by several considerations. First, the time scale for some effects may be longer or shorter than the four-month intervals captured by the current assessment structure. Future research should consider studying specific but variable timescales for the predictive utility of risk factors. Second, participants were recruited from a single city in the US and, although the sample was diverse with respect to race, sex assigned at birth, and education level, it was not nationally representative. Third, repeated assessments, while necessary in longitudinal research, may result in assessment reactivity. Fourth, externalizing psychiatric syndromes have also demonstrated associations with alcohol use and should be explored in future research as predictors of emerging adult drinking trajectories. Fifth, results may be different if we examined time-varying effects of predictors in parallel to changes in drinking, and future research may consider examining bidirectional effects over time.

Conclusion

The current analyses demonstrated reductions in both HDD and alcohol-related problems over the study period, with no clear impact of the COVID-19 pandemic. In separate models, greater alcohol demand intensity and inelasticity was associated with shallower linear declines in HDD over the assessment period, whereas higher internalizing symptoms, and lower demand elasticity and environmental reward, was associated with shallower linear declines in past month alcohol-related problems. Lower demand elasticity and environmental reward at the baseline was also associated with accelerated early declines followed by deceleration and flattening of changes in alcohol-related problems. After controlling for demographic and theoretically relevant covariates, demand elasticity and intensity were associated with HDD, and demand elasticity was associated with the trajectory of alcohol problems. These results support behavioral economic demand elasticity and intensity as predictors of longitudinal drinking trajectories and suggest that the reinforcing effects of alcohol are strong predictors of “aging out” of harmful alcohol use among emerging adults.