Area Under the Curve as a Novel Metric of Behavioral Economic Demand for Alcohol

Abstract

Behavioral economic purchase tasks can be readily used to assess demand for a number of addictive substances including alcohol, tobacco and illicit drugs. However, several methodological limitations associated with the techniques used to quantify demand may reduce the utility of demand measures. In the present study, we sought to introduce area under the curve (AUC), commonly used to quantify degree of delay discounting, as a novel index of demand. A sample of 207 heavy drinking college students completed a standard alcohol purchase task and provided information about typical weekly drinking patterns and alcohol-related problems. Level of alcohol demand was quantified using AUC – which reflects the entire amount of consumption across all drink prices - as well as the standard demand indices (e.g., intensity, breakpoint, O_max, P_max, and elasticity). Results indicated that AUC was significantly correlated with each of the other demand indices (rs = .42–.92), with particularly strong associations with O_max (r = .92). In regression models, AUC and intensity were significant predictors of weekly drinking quantity and AUC uniquely predicted alcohol-related problems, even after controlling for drinking level. In a parallel set of analyses, O_max also predicted drinking quantity and alcohol problems, although O_max was not a unique predictor of the latter. These results offer initial support for using AUC as an index of alcohol demand. Additional research is necessary to further validate this approach and to examine its utility in quantifying demand for other addictive substances such as tobacco and illicit drugs.

Behavioral economics integrates principles from psychology and microeconomics to investigate how individuals make transactions with the world around them. This field has shown particular promise in characterizing pathological decision-making processes in individuals who exhibit excessive consumption of addictive substances. From the standpoint of behavioral economics, the potential for substance abuse is theorized to be highest when substance-related rewards are overvalued relative to substance-free alternatives (Rachlin, 1997). This perspective has spurred a substantial body of laboratory and survey-based research that has examined how individuals with substance use disorders attribute value to various rewards and how this value, in turn, influences decisions to use addictive drugs (for a review, see Bickel, Johnson, Koffarnus, MacKillop, & Murphy, 2014).

One behavioral economic variable that is well suited for examining the relative value of substance-related rewards is demand, which refers to the quantitative relationship between consumption of a commodity and its cost, and putatively reflects the value of the commodity to the individual. Demand can be readily assessed via self-report purchase tasks that ask individuals to report how much of an addictive commodity they would consume across a range of prices. Hypothetical purchase task assessments have been developed for alcohol (Murphy & MacKillop, 2006), tobacco (Jacobs & Bickel, 1999; MacKillop et al., 2008), and illicit substances (Collins, Vincent, Yu, Liu, & Epstein, 2014; Goudie, Sumnall, Field, Clayton, & Cole, 2007). Incentivized purchase tasks using actual drug and money outcomes have also been implemented in laboratory studies (Amlung, Acker, Stojek, Murphy, & MacKillop, 2012; MacKillop et al., 2012, 2014). Importantly, prior studies have found close correspondence between demand on hypothetical and actual-reward purchase tasks (Amlung et al., 2012) and self-reported consumption on purchase tasks is highly correlated with actual drug consumption (Amlung et al., 2012; Collins et al., 2014).

Several methods are commonly used to quantify individuals’ level of demand. The first step typically involves plotting consumption and amount of money spent on the drug as a function of price to generate demand and expenditure curves, respectively (Figure 1). Four demand indices can be operationalized directly from these curves using an observed values approach (Murphy & MacKillop, 2006). Intensity is consumption when price is zero. Breakpoint is the price at which consumption is suppressed to zero. O_max is the maximum expenditure, and P_max is the price associated with O_max, as well as the price at which demand becomes elastic. An additional index, elasticity, is a measure of proportionate price sensitivity that is commonly derived using an exponential demand curve model (Hursh & Silberberg, 2008). These indices are theorized to be conceptually related to each other, but nonetheless to reflect distinct aspects of drug value. Factor analytic studies further suggest that individual indices appear to load on two higher-order factors (Bidwell, MacKillop, Murphy, Tidey, & Colby, 2012; MacKillop et al., 2009), reflecting volumetric motivation (peak consumption and expenditure) and price sensitivity.

Figure 1. Representative Demand and Expenditure Curves.

Panel A depicts demand curves from two hypothetical individuals exhibiting high demand (filled markers) and low demand (unfilled markers). The light gray shaded region reflects AUC for the low demand individual while AUC for the high demand individual is the sum of the dark and light gray shaded regions. Panel B depicts the corresponding expenditure curves. Drink prices are presented along the x-axis, with logarithmic scale used for proportionality. Demand indices and AUC values are also indicated. Note, the logarithmic scale of the x-axis results in a visual distortion of AUC due to stretching of low price intervals.

These demand indices have demonstrated good psychometric properties, including good-to-excellent test-retest reliability (Few, Acker, Murphy, & MacKillop, 2012; Murphy, MacKillop, Skidmore, & Pederson, 2009). With regard to external validity, prior research has found significant associations between level of demand and quantity/frequency of drug use, dependence severity, and treatment response (e.g., Collins et al., 2014; MacKillop & Murphy, 2007; MacKillop et al., 2008, 2010; Murphy et al., 2009). Finally, state-oriented purchase tasks have been shown to enhance the assessment of acute motivation in the lab. Specifically, demand has been shown to be dynamically altered by affective states and external influences that contribute to substance misuse, including negative affect (e.g., Amlung & MacKillop, 2014), drug withdrawal (e.g., MacKillop et al., 2012), and drug-related cues (e.g., Acker & MacKillop, 2013; Amlung et al., 2012; MacKillop et al., 2012; MacKillop, O’Hagen, et al., 2010).

These studies have helped solidify demand as a useful means of assessing individual differences in drug motivation. However, a number of methodological limitations in the methods currently being used to quantify demand may limit its application more broadly. For instance, there is currently no single index that provides an overall summary of demand on purchase tasks. Although Hursh and Silberburg (2008) argue that elasticity is the central index that reflects the “essential value” of a drug, a limitation of the elasticity parameter is that it does not take into account volumetric consumption. In addition, while each of the demand indices provide somewhat unique information about motivation for a substance, the specific indices that are the most important has varied across prior studies in a variety of populations. In many cases, intensity was found to be the strongest predictor of alcohol-related outcomes (e.g., Amlung et al., 2012; MacKillop, Miranda, et al., 2010; MacKillop, O’Hagen, et al., 2010; Murphy et al., 2009; Murphy & MacKillop, 2006), yet in other studies, O_max, breakpoint, or elasticity were more sensitive (Acker, Amlung, Stojek, Murphy, & MacKillop, 2012; MacKillop & Murphy, 2007). Similar ambiguities also exist in studies examining demand for tobacco and other substances (Acker & MacKillop, 2013; Collins et al., 2014; MacKillop et al., 2008). From the standpoint of statistical power, using all of the indices increases the number of statistical comparisons being conducted which may inflate Type I error rate. As a result, researchers must use stringent error correction (e.g., Bonferroni correction) that constrains power. Finally, subjects’ consumption patterns on purchase task assessments are also not always well-characterized by the exponential demand curve equation, which often results in a portion of the sample (often as many as 15–20% of participants) being excluded due to poor model fit.

An alternative approach to quantifying demand that can be imported from the delay discounting literature is area under the curve (AUC). Area under the demand curve can be calculated via a similar method as AUC for discounting curves (see Myerson, Green, & Warusawitharana, 2001). Starting with a demand curve with individual consumption values plotted as a function of price, vertical lines can be drawn from each data point to the x-axis, subdividing the graph into a series of trapezoids. The area of each trapezoid is equal to (x² − x¹)[(y¹ + y² / 2], where x¹ and x² are successive prices, and y¹ and y² are the consumption values associated with these prices. The area under the demand curve is equal to the sum of the areas of these individual trapezoids. Assembling these trapezoids together, the aggregate AUC is most often reported as a proportion of total curve area, ranging from 0.0 to 1.0. In the case of demand, the total curve area can be operationalized as the AUC value when the maximum consumption value across all prices is inputted at each price. Each individual’s AUC value can then be divided by this total AUC to generate proportionate AUC, with larger AUC values reflecting greater demand (see Figure 1).

Using AUC as an index of demand may be particularly useful for several reasons. From a practical standpoint, AUC is a relatively straightforward and can be easily calculated for an entire dataset using computer graphing programs, such as GraphPad Prism (GraphPad Software, Inc., La Jolla, CA). Area under the curve does not assume any particular shape of the demand curve or rely on any specific mathematical model to describe demand. Thus, AUC may be especially useful in cases where individual consumption deviates from the prototypical demand curve shape and is not well-characterized by the exponential models. AUC also encompasses the full conceptual notion of demand, in contrast to the other indices that, by definition, capture more focal aspects of the demand curve such as the intercept (intensity), peak (O_max), slope (elasticity), or terminus (breakpoint). As a more comprehensive index of the entire curve, AUC values will be influenced by unrestricted consumption at low prices (i.e., intensity), the rate of decrease in consumption with increased price (i.e., elasticity), and point at which consumption is suppressed to zero (i.e., breakpoint). Finally, demonstrating that AUC is a viable index of demand will further increase methodological convergence with the delay discounting literature in which AUC has been increasingly used. Specifically, the lack of dependence on adequate curve fits to generate parameter estimates has in part contributed to AUC becoming a preferred method for quantifying delay discounting. This convergence is important given that a priority in behavioral economics is to better understand the unique and combined contributions of demand and discounting to reinforcement pathology in addiction (Bickel et al., 2014).

The goal of the present study was to introduce AUC as a novel approach to quantifying behavioral economic demand for alcohol on an alcohol purchase task (APT). Furthermore, we also sought to examine of the relationship between AUC and indicators of alcohol use and misuse. We predicted that AUC would be correlated with other common demand indices, further supporting the internal validity AUC as an index of alcohol demand. Consistent with prior research, we predicted that greater alcohol demand—as indicated by higher AUC—would be significantly associated with typical drinking quantity and self-reported alcohol problems.

Method

Participants

Participants were 207 college students from a large urban Southern university who were recruited for one of two randomized clinical trials for brief alcohol interventions (Murphy, Dennhardt, Skidmore, Martens, & McDevitt-Murphy, 2010). Eligibility criteria included reporting one or more heavy drinking episodes (five/four or more drinks in one occasion for males/females) in the past month. Participants were identified via a larger screening survey and were not seeking alcohol treatment. The mean age of the sample was 19.50 (SD = 1.99) years and approximately half of the sample (53%) was female. Self-identified ethnicity was as follows: White/Caucasian (65%), Black/African American (26%), multi-racial (4%), Hispanic/Latino (3%), Asian (1%) and American Indian/Alaskan Native (<1%). Participants reported consuming an average of 16.45 (SD = 15.17) drinks/week and endorsed a total of 12.64 (SD = 8.55) alcohol-related problems in the past six months.

Procedure

All procedures were approved by the University of Memphis Institutional Review Board. Participants were recruited using surveys completed at the student health center or in classrooms. All measures included in the present study were completed during an individual laboratory testing session that occurred prior to any intervention. Participants provided written informed consent and received course credit or monetary payment for their participation.

Measures

Alcohol purchase task

Behavioral economic demand for alcohol was assessed with an alcohol purchase task (APT; Murphy & MacKillop, 2006). Participants reported the number of standard-sized drinks (e.g., 12 oz. beer, 5 oz. glass of wine, or a mixed drink containing 1.5 oz. of liquor) they would purchase and consume in a typical drinking situation if the drinks cost 17 different prices: Free, $0.25, $0.50, $1, $1.50, $2, $2.50, $3, $4, $5, $6, $7, $8, $9, $10, $15, and $20.

Alcohol use and alcohol-related problems

The Daily Drinking Questionnaire (DDQ: Collins, Parks, & Marlatt, 1985) was used to assess the total number of standard drinks consumed on each day during a typical drinking week in the past month. Daily totals were summed to generate an estimate of typical weekly drinking. Alcohol-related problems over the past six months were measured using the Young Adult Alcohol Consequences Questionnaire (YAACQ; Read, Kahler, Strong, & Colder, 2006). The YAACQ total score was used as an overall index of alcohol problems.

Demographics

Participant demographics were assessed using a self-report questionnaire that included age, sex, income, ethnicity, and other relevant demographic variables.

Data Analysis

Area under the curve

Responses on the APT were used to generate individual consumption and expenditure curves for each participant. GraphPad Prism 6 was used to calculate AUC values for the consumption curves. The total area was operationalized as the AUC value when the maximum consumption value across the entire sample was inputted at each price. For this sample, the maximum number of drinks purchased across all prices was 60. Therefore, total AUC was calculated by generating a demand curve with 60 at each of the 17 price intervals. Proportionate AUC values (ranging from 0.0 to 1.0) were then generated by dividing each participant’s raw AUC value by the total AUC. Higher AUC values reflect greater alcohol demand.

Observed and derived demand indices

Four observed demand indices were generated for each participant (Murphy & MacKillop, 2006). Intensity is defined as the number of drinks consumed when drinks were free. Breakpoint is the first price at which reported consumption is zero. O_max reflects the greatest expenditure value. P_max is the price associated with O_max and the point at which demand becomes elastic. Finally, elasticity of demand (α) was derived from the following exponential equation (Hursh & Silberberg, 2008):

$$\text{ln}\mathrm{Q}=\text{ln}{\mathrm{Q}}_0+\mathrm{k}({\mathrm{e}}^{-\mathrm{\alpha }\mathrm{P}}-1)$$ (1)

where Q = quantity consumed, Q₀ = derived intensity, k = the range of the dependent variable (standard drinks) in logarithmic units, P = price, and α = elasticity of demand. GraphPad Prism 6 was used to fit the data to Equation 1 using the program available through the Institute for Behavioral Resources website (ibrinc.org). The overall mean performance was first analyzed to find the best-fitting k parameter, which was determined to be 2.6 and was used across all demand curve fits. Zero values (which cannot be log transformed) were omitted prior to curve fitting.

Outliers and normality

All variables were initially screened for missing data, outliers, and distribution abnormalities (Tabachnick & Fidell, 2001). For each variable, values that were greater than or equal to 3.29 standard deviations above the mean were winsorized to be one unit greater than the greatest non-outlier value. Variables that were skewed or kurtotic were transformed prior to analysis when appropriate.

Primary analysis

Associations between conventional demand curve metrics and drinking measures were originally reported in Skidmore et al. (Skidmore, Murphy, & Martens, 2014). The current report focuses on the relative utility of the AUC approach to quantifying demand. Bivariate correlations (Pearson’s r) were conducted to assess the associations among the indices of alcohol demand (AUC, Intensity, Breakpoint, O_max, P_max, and Elasticity) and the alcohol use variables (drinks/week and YAACQ total score). Next, hierarchical regressions were conducted to assess the utility of the alcohol demand metrics to predict drinking quantity and alcohol problems. The first step in these controlled for sex and income. AUC was then added in the second step. The final step included the remaining demand indices that were significantly correlated with drinks/week or YAACQ total score to examine whether these indices contribute to alcohol use above and beyond AUC. A conventional significance level of p < .05 was used in all analyses.

Results

Preliminary analyses

One participant did not provide any responses on the APT and was excluded from all analyses. Another participant did not provide a response for the free drink price and this participant was excluded from all analyses involving the intensity index. A small number of outliers (< 1% of all data points) were identified and were winsorized prior to analysis. Square root transformations were used to normalize five variables that were significantly skewed or kurtotic (breakpoint, O_max, P_max, elasticity, and YAACQ total). Logarithmic transformations were applied to three additional variables (AUC, intensity, and drinks/week). Alcohol demand on the APT was prototypical, with reported consumption decreasing at increasing prices, and expenditure conforming to an inverted U-shaped curve (Figure 2). Equation 1 provided a very good fit to the aggregate sample mean data (mean R² = .87) but only an adequate fit to the individual participant data (mean R² = .59). Elasticity could not be calculated for five participants due to reported consumption that did not conform to the demand curves (e.g., missing data, no reported consumption, or rectangular demand). The observed demand indices and AUC for these participants were still included in statistical analyses.

Figure 2. Demand and Expenditure Curves on the Alcohol Purchase Task.

Panels A and B depict demand and expenditure curves, respectively, from the alcohol purchase task assessment. Data points reflect the sample mean value at each price with error bars depicting standard error of the mean.

Primary analyses

Table 1 presents the results of the bivariate correlations between each of the demand metrics, drinks / week, and YAACQ total problems. Significant correlations were observed between the alcohol use variables and the majority of the alcohol demand indices, with the exception of P_max and breakpoint. Specifically, greater weekly alcohol use was associated with higher AUC, higher intensity, greater O_max, and decreased elasticity. The pattern of associations with YAACQ total score was similar, with an additional significant positive association with breakpoint. Importantly, the magnitude of the association between AUC and alcohol use was comparable to the highest magnitude associations for the other alcohol demand indices. AUC values were also significantly correlated with each of the other demand indices, albeit at varying magnitudes, with a particularly high association between AUC and O_max (r = .92), suggesting collinearity. As such, O_max was omitted from the primary regression analyses that included AUC.

Table 1.

Bivariate Correlations among Alcohol Demand and Alcohol Use

	1.	2.	3.	4.	5.	6.	7.
1. AUC	—
2. Intensity	.67**	—
3. Breakpoint	.69**	.23**	—
4. Omax	.92**	.61**	.64**	—
5. Pmax	.42**	−.06	.65**	.53**	—
6. Elasticity	−.62**	−.48**	−.47**	−.70**	−.43**	—
7. Drinking Quantity	.51**	.57**	.10	.51**	−.05	−.40**	—
8. YAACQ	.43**	.38**	.16*	.46**	.13	−.33**	.49**

Note: n = 206;

^*p<.05;

^**p<.01;

Drinking quantity reflects typical drinks/week in last month;

AUC = area under curve; YAACQ = Young Adult Alcohol Consequences Scale total score

Hierarchical linear regressions were conducted to determine the specific associations between alcohol demand and drinking quantity and alcohol problems. The results of the regression predicting weekly drinking are presented in Table 2. In step one, sex was a significant predictor of drinking quantity while income was not significant. In step two, AUC was a significant predictor of weekly drinking, accounting for an additional 20% of the variance in alcohol use (R² = .36). The final step, in which intensity and elasticity were added as predictors, accounted for an additional 5% of the variance in drinking quantity (R² = .42). AUC and intensity, but not elasticity, were significant predictors of greater weekly alcohol consumption.

Table 2.

Hierarchical Linear Regression Predicting Typical Weekly Drinking

Variable	B	SE B	β	R2	ΔR2
Step 1				.16***
Sex	0.30	0.05	.39***
Income	0.00	0.00	−.10
Step 2				.36***	.20***
Sex	0.25	0.04	.33***
Income	0.00	0.00	−.06
AUC	20.97	2.67	.45***
Step 3				.42***	.05*
Sex	0.19	0.05	.24***
Income	0.00	0.00	−.03
AUC	9.24	3.92	.20*
Intensity	0.52	0.14	.30***
Elasticity	−13.86	9.90	−.10

Note:

^*p<.05;

^**p<.01;

^***p<.001;

n = 200; AUC = area under curve; Sex: Female = 0, Male = 1 Dependent variable = drinks/week

The results of the regression predicting YAACQ total problems are presented in Table 3. Neither sex nor income was a significant predictor of alcohol problems in the first step of the model. Adding AUC to the model explained an additional 18% of the variance in alcohol problems (R² = .20). In the final step, adding intensity, breakpoint, elasticity explained 4% more of the variance in alcohol problems (R² = .25). However, only AUC was a significant predictor of alcohol problems in the full model. An additional regression analysis was conducted to examine the incremental utility of AUC in predicting problems above drinking quantity. In this analysis, AUC was a significant predictor of YAACQ total score, β = .22, t(201) = 3.15, p = .002, after accounting for weekly drinking, which also significantly predicted YAACQ total, β = .43, t(201) = 5.86, p < .001.

Table 3.

Hierarchical Linear Regression Predicting Alcohol Problems

Variable	B	SE B	β	R2	ΔR2
Step 1				.02
Sex	0.33	0.17	.14
Income	0.00	0.00	−.06
Step 2				.20***	.18***
Sex	0.19	0.15	.08
Income	0.00	0.00	−.02
AUC	60.57	9.25	.43***
Step 3				.24***	.04*
Sex	0.00	0.16	.00
Income	0.00	0.00	.00
AUC	53.25	16.66	.37**
Intensity	0.89	0.55	.17
Breakpoint	−0.23	0.16	−.15
Elasticity	−35.56	35.40	−.08

Note:

^*p<.05;

^**p<.01;

^***p<.001;

n = 200; AUC = area under curve; Sex: Female = 0, Male = 1 Dependent variable = YAACQ total score

Given the potential collinearity between AUC and O_max, we conducted a parallel set of regression analyses substituting O_max for AUC in the second step of the models. The results of these analyses were largely similar to the AUC findings. Specifically, intensity (β = .31, t(194) = 4.14, p < .001) and O_max (β = .27, t(194) = 3.10, p = .002) were significant predictors of drinking quantity in the full model. Both intensity (β = .20, t(194) = 2.30, p = .02) and O_max (β = .43, t(194) = 3.76, p < .001) significantly predicted YAACQ scores, in contrast to the AUC analysis, where AUC alone was a significant predictor of alcohol problems. Thus, our findings suggest that both AUC and O_max appear to be robust associates of alcohol problems among young adults, but that AUC may be more efficient.

Discussion

Purchase task assessments have been shown to be a useful tool for quantifying individual differences in behavioral economic demand and motivation for alcohol and other addictive substances (Amlung et al., 2012; Collins et al., 2014; Goudie et al., 2007; Jacobs & Bickel, 1999; MacKillop et al., 2008; MacKillop, Miranda, et al., 2010; MacKillop, O’Hagen, et al., 2010; Murphy & MacKillop, 2006). However, methodological limitations in the techniques used to quantify demand on these tasks—including the need to generate up to five separate demand indices and relying on mathematical models that may not adequately fit the observed data—may diminish their use in some contexts. Accordingly, the goal of this study was to propose a new analytic approach to quantifying alcohol demand using AUC and to offer preliminary evidence of the utility of this novel demand metric in predicting alcohol misuse in a sample of heavy drinkers.

The present results offer initial proof-of-concept for the AUC approach. In terms of internal validity, AUC was significantly correlated with each of the other commonly used demand indices, suggesting that this variable captures many relevant facets of the demand curve. More importantly, AUC predicted both quantity of alcohol use and alcohol-related problems in regression models, and this association remained significant even after the other indices were included. In the case of drinking quantity, both AUC and intensity were significant predictors in the combined model, with intensity exhibiting a larger magnitude association. This is consistent with prior studies that have found that intensity of demand is the strongest predictor of weekly drinking quantity and frequency of heavy drinking episodes (e.g., MacKillop, O’Hagen, et al., 2010; Murphy et al., 2009; Murphy & MacKillop, 2006). However, the present results suggest that other aspects of the demand curve that are captured by the broader AUC variable are also associated with drinking behavior. For drinking problems, however, AUC captured all of the relevant variance, and also predicted problems after first accounting for overall drinking quantity. This suggests that AUC may be most useful when the variable of interest is alcohol-related negative consequences. Taken together, these results provide initial support for potential of AUC as an index for quantifying alcohol demand. Importantly, as this is the first study to examine AUC in this context, these results should not be considered definitive. This is especially true considering that our results seem to indicate that other indices may perform as well as or even slightly better than AUC in predicting alcohol-related outcomes. Nonetheless, AUC is a promising new approach for quantifying demand and future research is clearly needed to further clarify its utility for behavioral economics research in addiction.

These results have several potential implications. First, AUC is relatively straightforward to calculate and may be a more accessible means of quantifying demand. Second, AUC provides a single index of demand that reduces the required number of statistical tests compared to when traditional demand indices are used. Third, AUC is a model-free, data-driven index of demand that is not tied to any particular demand curve shape. This is important because individual demand curves may not always conform to an exponential function (e.g., Eq. 1). Although several of the other observed demand indices are similarly not constrained by a particular mathematical model, these indices often only capture one aspect of the demand curve (e.g., unrestricted consumption as in the case of intensity). Area under the curve, on the other hand, has the advantage of quantifying demand based directly on the consumption values across all prices.

More broadly, the utility of AUC as an index of demand converges with the larger behavioral economics literature that supports using AUC as an index of delay discounting (e.g., Myserson et al., 2001). From a theoretical standpoint, both demand and delay discounting have been argued to jointly contribute to the reinforcement pathology that characterizes substance use disorders (e.g., Bickel et al., 2014). These disorders are increasingly being viewed as a consequence of persistent high valuation of drugs of abuse (demand) and impulsive discounting of delayed rewards. Thus, one potential benefit of an AUC approach to quantifying demand is that it provides a common metric that can be used in future studies examining combined influence of demand and discounting on substance misuse.

The very strong association between O_max and AUC also warrants discussion. In correlation analyses, these two indices were correlated at r = .92, which was of considerably higher magnitude than the correlations between AUC and the other indices (rs = .42–.69). O_max also performed similarly to AUC in the regression models. Despite their high association, there was still ~15% unique variance between O_max and AUC suggesting that although these variables are highly related, they do not appear to reflect identical aspects of demand. Furthermore, only AUC uniquely predicted alcohol-related problems in the combined models suggesting that AUC may be a more efficient predictor of alcohol problems in young adult drinkers. Nonetheless, these results do raise the possibility that AUC is simply capturing maximum alcohol expenditure but with a different label. A priority for future studies will be to disentangle the unique aspects of demand that are captured by these two indices. Given the fact that AUC and O_max are conceptually distinct, it is possible that their near statistical redundancy is an artifact of the particular APT measure and sample included in this study. Whether AUC or O_max is more closely associated with alcohol misuse is an important empirical question for the future.

In terms of other future research directions, a main priority is to replicate and extend the present findings in other samples of drinkers with varying levels of alcohol involvement (e.g., social drinkers, individuals with alcohol use disorders). Additional studies are also needed to further examine the psychometric properties of AUC as an index of demand. For instance, although the other demand indices have been shown to have good-to-excellent test-retest reliability (Few et al., 2012; Murphy et al., 2009), the reliability of the AUC variable remains an empirical question. In addition, the purchase task assessment used in the present study was designed to assess trait-level demand. Previous studies have utilized purchase tasks that are specifically designed to measure state-based alcohol demand using real alcohol and money rewards (e.g., Amlung et al., 2012; MacKillop et al., 2014). Future studies are needed to address the utility of AUC in quantifying demand from these state-based measures. Finally, in order to examine the AUC approach more broadly, this method should be employed in studies examining demand for other addictive substances (e.g., Collins et al., 2014; Goudie et al., 2007; Jacobs & Bickel, 1999; MacKillop et al., 2008).

The present findings should be considered in the context of the study’s limitations. First, the APT used estimated alcohol consumption, not actual consumption, which may be biased in a number of ways. Although previous research has indicated that self-reported consumption values are highly similar across hypothetical and actual-reward purchase tasks (Amlung et al., 2012), future research should validate the AUC approach using real-reward APT assessments. Second, participants in the study were all young adult college students from a single geographic area, which may limit the generalizability of these findings to other samples of drinkers.

In summary, the present study provides initial support for using AUC to quantify level of behavioral economic demand for alcohol in heavy drinkers. This approach has a number of potential advantages over traditional methods for quantifying demand, particularly reducing the risk of Type I error. Our results also offer preliminary evidence that AUC can be used without sacrificing substantial variance in examining associations between level of demand and markers of alcohol misuse. Future research is needed to further validate this approach and extend its application to other substances beyond alcohol.