Evaluating Non-Medical Prescription Opioid Demand Using Commodity Purchase Tasks: Test-Retest Reliability and Incremental Validity

Abstract

Rationale

Non-medical prescription opioid use and opioid use disorder (OUD) present a significant public health concern. Identifying behavioral mechanisms underlying OUD will assist in developing improved prevention and intervention approaches. Behavioral economic demand has been extensively evaluated as a measure of reinforcer valuation for alcohol and cigarettes, whereas prescription opioids have received comparatively little attention.

Objectives

Utilize a purchase task procedure to measure the incremental validity and test-retest reliability of opioid demand.

Methods

Individuals reporting past year non-medical prescription opioid use were recruited using the crowdsourcing platform Amazon Mechanical Turk (mTurk). Participants completed an opioid purchase task as well as measures of cannabis demand, delay discounting, and self-reported pain. A one-month follow up was used to evaluate test-retest reliability.

Results

More intense and inelastic opioid demand was associated with OUD and more intense cannabis demand was associated with cannabis use disorder. Multivariable models indicated that higher opioid intensity and steeper opioid delay discounting rates each significantly and uniquely predicted OUD. Increased opioid demand intensity, but not elasticity, was associated with higher self-reported pain, and no relationship was observed with perceived pain relief from opioids. Opioid demand showed acceptable-to-good test-retest reliability (e.g., intensity r_xx=.75; elasticity r_xx=.63). Temporal reliability was lower for cannabis demand (e.g., intensity r_xx=.53; elasticity r_xx=.58) and discounting rates (r_xx=.42-.61).

Conclusions

Opioid demand was incrementally valid and test-retest reliable as measured by purchase tasks. These findings support behavioral economic demand as a clinically useful measure of drug valuation that is sensitive to individual difference variables.

Introduction

The non-medical use of prescription opioids and opioid use disorder (OUD) present a significant and growing public health concern. Over 2 million people in the United States reported initiation of non-medical prescription opioid use in 2017 and over 11 million reported past year use (Center for Behavioral Health Statistics 2018). A steady rise in overdose fatalities attributable to prescription opioids has also occurred, with a four-fold increase since 1999 (Hedegaard et al. 2017). Improvements in monitoring systems and pill reformulations have shown some promise for deterring use, but in many at-risk populations (e.g., rural Appalachia) use-related burden remains high (e.g., Brown et al. 2018; Mack et al. 2017; Schranz et al. 2018; Van Handel et al. 2016). One research priority is to identify behavioral mechanisms underlying OUD to assist in developing improved prevention and intervention approaches.

The marriage of perspectives from behavioral economics and operant theory has resulted in numerous advances for psychological science, broadly (Hursh and Roma 2013), and addiction science, specifically (Bickel et al. 2017; MacKillop 2016). These theoretical accounts propose that systematic choice and decision-making processes described by behavioral economics may help to reveal the behavioral mechanisms contributing to the etiology and persistence of drug-taking behavior. A reinforcer pathology approach specifically posits that substance use disorder is characterized by high reinforcer valuation combined with an extreme bias for smaller, immediate reinforcers over delayed, larger reinforcers (Bickel et al. 2017). These two decision-making processes are thought to translate into diminished health outcomes by placing an overvaluation on harmful commodities (e.g., tobacco cigarettes) with a preference for these immediate reinforcers despite longer-term health consequences (e.g., lung cancer).

Behavioral economic demand (i.e., the relationship between commodity price and purchase) has received specific attention within behavioral economics as a measure of reinforcer valuation. Demand analysis presents several advantages over traditional measures of relative reinforcer value, including accounting for the multi-dimensional nature of reinforcement rather than treating reinforcement as a homogenous construct (Hursh and Silberberg 2008; Johnson and Bickel 2006). The study of demand has been facilitated, in part, by the development of the commodity purchase task (Jacobs and Bickel 1999; see reviews by Kaplan et al. 2018; MacKillop 2016). In this task, participants are asked to report hypothetical consumption of a commodity across varying prices to effectively and efficiently generate demand curves. These simulated procedures allow for data collection in the absence of drug administration thereby affording the opportunity to work with populations that cannot be evaluated using traditional drug self-administration methods (e.g., treatment-seeking individuals, individuals with compromised health). A growing body of literature has supported the clinical relevance of the purchase task procedure by utilizing demand to understand mechanisms by which interventions are clinically effective or as prognostic variables predicting reductions in substance use following intervention delivery (e.g., Bujarski et al. 2012; MacKillop and Murphy 2007; Murphy et al. 2015).

To date, the majority of research on behavioral economic demand with the purchase task has focused on alcohol and cigarette use. A smaller body of work has examined non-medical prescription opioid and heroin use (Jacobs and Bickel 1999; Pickover et al. 2016). Those studies that do exist have focused on samples of more narrowly defined populations and on demand in the absence of other behavioral economic measures (e.g., delay discounting). For example, Jacobs and Bickel (1999) evaluated heroin demand in individuals recruited from an outpatient opioid clinic and found that hypothetical heroin demand was well described by quantitative models. Pickover and colleagues (2016) measured demand for non-medical prescription drugs among college students and found that demand was predictive of opioid use frequency and OUD. Both of these studies emphasized opioid demand, however, which precludes determining a unique (i.e., above and beyond other measures) and commodity-specific (i.e., selective to opioid valuation over general reinforcer or drug valuation) contribution of demand to an understanding of the behavioral mechanisms underlying non-medical prescription opioid use. The purchase task developed for prescription opioids was also validated in a college student population and was limited in scope with respect to the clinical variables examined other than opioid use disorder (Pickover et al. 2016).

The overall purpose of the current study was to replicate and extend prior work on the use and validity of the commodity purchase task for evaluating behavioral economic demand for prescription opioids (Jacobs and Bickel 1999; Pickover et al. 2016). This study was designed to expand upon those prior findings by providing validation in a general community sample, determining temporal reliability over a one-month time period, incorporating measurement of delayed reward discounting, and examining incremental validity as it relates to a variety of clinically relevant outcomes (e.g., substance use disorder, pain). To this end, a battery of commodity purchase and delay discounting tasks were included to evaluate the unique and stimulus-selective contribution of opioid demand to OUD and other clinical outcomes. It was hypothesized that prescription opioid demand would be related to OUD in an incremental and commodity-selective manner above and beyond other commodity’s demand and delay discounting. Pain measures were included to establish the association between opioid demand and self-reported pain. Chronic pain has been posited to play a key role in the etiology of the opioid crisis and represents a potentially important pathway by which non-medical prescription opioid use develops (Kolodny et al. 2015; Sehgal et al. 2012; Volkow and Collins 2017). It was therefore hypothesized that pain would positively associate with opioid demand. Finally, opioid demand was collected at two time points separated by one month to establish test-retest reliability and temporal stability. It was hypothesized that opioid demand would show acceptable test-retest reliability consistent with purchase tasks for other substances (Acuff and Murphy 2017; Few et al. 2012; Murphy et al. 2009).

Methods

Participants and Screening

Participants were recruited using the crowdsourcing site Amazon Mechanical Turk (mTurk). Crowdsourced sampling utilizes the Internet to sample individuals from varied geographic regions and with varied health histories. Prior research has demonstrated the validity of crowdsourcing in psychological and addiction science (see reviews by Chandler and Shapiro 2016; Strickland and Stoops 2019; see further description in the Discussion).

Inclusion criteria were 1) past year non-medical prescription opioid use, 2) 30 or more lifetime prescription opioid uses, and 3) age 18 or older. Inclusion criteria were verified using a screener. Access was limited to individuals with at least 50 completed mTurk tasks, a ≥95% approval rating, and United States residence (see similar qualifications in Cunningham et al. 2017; Strickland and Stoops 2015). Qualifying participants completed behavioral economic measures as well as other cognitive-behavioral measures to be reported elsewhere. Participants were asked to complete a follow up approximately one month after the initial survey, which contained the purchase task and delay discounting measures. The University of Kentucky Institution Review Board approved all procedures and participants reviewed an informed consent prior to participation.

Measures

Commodity Purchase Tasks

Behavioral economic demand for prescription opioids and cannabis was evaluated using commodity purchase tasks. Cannabis was selected for testing stimulus-selectivity given putative similarities in behavioral responses and neurobiological pathways associated with pain, which clinically has led to a proposed substitution of cannabis for prescription opioids in the medical management of chronic pain (see discussion in Choo et al. 2016; Hill 2015; Lucas 2012).

A standard instructional vignette was provided in which participants were instructed to consume all purchases in a single day, could not stockpile or get the commodity from another source, and had no commodity available from previous days (see Supplemental Materials). Understanding of these instructions was verified by a correct response to two task questions. Opioids were quantified as “the standard dose that you use when you use these pills” consistent with prior work (Pickover et al. 2016). Cannabis hits were quantified as 10 hits/joint with 1 joint equal to 0.9 g cannabis (~0.09 g/hit) (Aston et al. 2015; Strickland et al. 2017). The price range included 17 prices from $0.00 [free] to $20/unit, presented sequentially (full range: $0.00 [free], $0.25, $0.50, $1, $1.50, $2, $2.50, $3, $4, $5, $6, $7, $8, $9, $10, $15, $20). These prices were selected from other purchase task studies (e.g., Murphy et al. 2015) and were intended to provide a range suitable for both commodities. Pilot data indicated that the prescription opioid task was feasible for delivery via the mTurk platform and provide preliminary support for validity (see Supplemental Materials Pilot Experiment).

Data from commodity purchase tasks were analyzed using the exponentiated demand equation (Koffarnus et al. 2015):

$$Q=Q_0\mathrm{*}{10}^{k\mathrm{*}{(e}^{\left(-\alpha \mathrm{*}{Q}_0\mathrm{*}C\right)}-1)}$$

where Q = consumption; Q₀ = derived demand intensity; k = a constant related to consumption range (a priori set to 2); C = commodity price; and α = derived demand elasticity. The exponentiated demand equation allowed for use of all zero values without transformation (Koffarnus et al. 2015). Demand intensity refers to a theoretical consumption of a commodity at zero price. Demand elasticity refers to the sensitivity of consumption to changes in price. Primary analyses focused on intensity and elasticity given that prior factor analytic studies have demonstrated improved stimulus-selectivity when using derived over curve-observed measures (Strickland and Stoops 2017) and given that these two measures reflect a two-factor structure underlying purchase task data (Aston et al. 2017; Bidwell et al. 2012; Epstein et al. 2018; Mackillop et al. 2009). Secondary analyses considered curve-observed metrics including P_max (point of unit elasticity approximated as price of maximum consumption), O_max (maximum consumption), and breakpoint (first price at which consumption is suppressed to zero). Individuals who did not reach a breakpoint were coded as one unit higher ($21). Breakpoint data were also analyzed as a dichotomous variable (breakpoint reached versus no breakpoint) given a high percentage of individuals that did not reach a breakpoint. Demand values were log-transformed to achieve normality.

Delay Discounting Rates

Delay discounting rates for money, prescription opioids, and cannabis were determined using a 5-trial adjusting delay task (Koffarnus and Bickel 2014). Participants made five choices between an immediate, smaller reinforcer ($500, $500 of opioids, or $500 of cannabis now) and a delayed, larger reinforcer ($1000, $1000 of opioids, or $1000 of cannabis delayed) at delays that titrated up or down based on selections. This task was selected for its utility in an online setting (e.g., Stein et al. 2017; Strickland et al. 2017) and validation against traditional test forms (Cox and Dallery 2016; Koffarnus and Bickel 2014). Delay discounting rates were log-transformed for normality. Visual inspection of frequency distributions confirmed that transformed rates followed an approximately normal distribution.

Brief DSM-5 Substance Use Disorder Diagnostic Assessment

DSM-5 substance use disorder was evaluated using an adapted version of the Brief DSM-5 Diagnostic Assessment (Hagman 2017). This questionnaire evaluated each of the 11 DSM-5 criteria for alcohol, cannabis, and opioid use disorders. Prior research has demonstrated the internal consistency reliability and validity of this assessment for alcohol use disorder (Hagman 2017).

Brief Pain Inventory

A modified version of the brief pain inventory was used to evaluate pain (Mendoza et al. 2006). Participants were asked to consider all daily and non-daily pain and indicate: 1) average past week pain (0–10 scale), 2) average interference due to pain across common daily activities (0–10 scale), and 3) typical relief from pain when using prescription opioids (0–100 scale).

Data Analysis

One hundred and five participants met inclusion criteria. Six failed one or more attention or validity checks and were removed from data analysis. An additional 16 provided non-systematic opioid purchase task data according to standardized criteria (Stein et al. 2015). This resulted in a final primary sample of 83 participants. Sixty-five of these participants completed the time 2 assessment (61 provided systematic data on the follow up opioid purchase task). Measures involving cannabis use were only completed by individuals reporting past year cannabis use (n = 76; 91.6% of the sample).

Bivariate relationships involving behavioral economic measures and opioid use or demographic variables were first evaluated. Pearson correlations were used in most cases, however negative binomial regression was used for past month use days given the observation of zero-inflation. Significant outcomes were then followed up with multivariable models evaluating the incremental and unique association for opioid demand intensity and elasticity controlling for demographic variables (i.e., age, sex, education, and income), opioid use frequency (i.e., past month opioid use), cannabis demand, and delay discounting rates. Test-retest reliability was first determined using bivariate correlations (r_xx) and intraclass correlations (ICCs) comparing time 1 and time 2 values. Interpretation of reliabilities was guided by recommended rules of thumb of < .40 = poor, .40−.59 = acceptable, .60−.74 = good, .75+ = excellent (Cicchetti 1994). Temporal stability for demand, discounting, and DSM-5 substance use disorder values was then evaluated using dependent-samples t-tests or McNemar tests for paired nominal data. Additional equivalence tests by two one-sided tests (TOST) were conducted as a supplemental analysis to determine if differences over time were smaller than a meaningful effect size (Lakens 2017). These tests involved building a 90% confidence interval (i.e., 100% - 2(α)) around the difference score and comparing the lower and upper bounds to the equivalence region. Values that 1) overlapped zero in the 95% confidence interval (i.e., were not significantly different) and 2) had a 90% confidence interval that contained this equivalence region were considered statistically equivalent. The approach for selecting an equivalence region was design-driven, such that a region was selected to provide 80% power to detect equivalence. Power analyses indicated that this region was a small-to-medium effect size Cohen’s d_z of .375. All tests were conducted with a type I error rate of 5% (α = .05).

Results

Demographics and Opioid Use

Table 1 contains demographic and opioid use information. A majority of participants were white and female with an average age of 34.0 years old. Two-thirds of participants endorsed statements indicative of DSM-5 criteria for opioid use disorder and a quarter reported a preference for a risky route of opioid administration (i.e., intranasal, smoked, or injection versus oral).

Table 1.

Demographics and Substance Use Variables

	Mean	SD	IQR
Age	34	8	29 to 37
Female	63.90%
White	89.20%
College	51.80%
Income (USD)	43000	28000	20k to 70k
Substance Use
Past Month Opioid Use Days	7.6	9.5	0 to 12
Risky Route of Administration	26.50%
Past Month Cannabis Use Days	15	12.6	1 to 30
Grams Cannabis/Week	7.5	9.5	2 to 10
Treatment History
Current Treatment-Seeking	8.40%
Treatment Plans in Next Year	19.30%
DSM-5 SUD
OUD	67.50%
OUD Criteria Endorsed	4.9	4	1 to 8
CUD	41.00%
CUD Criteria Endorsed	1.7	2.1	0 to 2
AUD	59.00%
AUD Criteria Endorsed	4	3.9	0 to 7
Pain
Average Pain (0–10)	4.6	2.4	3 to 6
Pain Effect on Daily Life (0–10)	4.1	2.8	1.6 to 6.3
Opioid Relief of Pain (0–100)	55.1	29.6	30 to 79
Opioid Demand
Intensity (Q0)	1.14	0.55	0.70 to 1.49
Elasticity (α)	−2.33	0.58	−2.72 to −1.95
Pmax	0.62	0.51	0.40 to 0.95
Omax	1.39	0.53	1.02 to 1.78
Breakpoint	0.9	0.45	0.70 to 1.32
No Breakpoint Reached	33.70%
Cannabis Demanda
Intensity (Q0)	1.57	0.57	1.16 to 2.05
Elasticity (α)	−2.14	0.52	−2.47 to −1.79
Pmax	0.19	0.56	−0.30 to 0.50
Omax	1.21	0.51	0.89 to 1.60
Breakpoint	0.6	0.55	0.18 to 1.18
No Breakpoint Reached	23.00%
Delay Discounting Rates (k)a
Money	−2.08	0.69	−2.48 to −1.63
Opioid	−1.65	0.79	−2.17 to −1.08
Cannabis	−1.69	0.98	−2.43 to −1.23

Note. SUD= substance use disorder; OUD = opioid use disorder; CUD = cannabis use disorder; AUD = alcohol use disorder; Risky Route of Administration = preferred use of intranasal, smoked, or injection administration.

^aDiscounting and demand variables were log-transformed prior to analysis and are presented in log values

Behavioral Economic Demand and Delay Discounting

Opioid and cannabis demand were well characterized by decreased consumption with increased price (Figure 1). Good model fits were also observed with the exponentiated demand equation for individual opioid (median R² = .86; IQR = .79 to .93) and cannabis (median R² = .93; IQR = .84 to .97) data. Money delay discounting rates were shallower than for opioids, t₈₂ = 6.33, p < .001, d_z = 0.69, or cannabis, t₇₅ = 4.16, p < .001, d_z = 0.48, which did not significantly differ from each other, t₇₅ = 0.64, p =.52, d_z = 0.07.

Fig. 1. Behavioral economic demand for prescription opioids and cannabis.

Participants completed commodity purchase tasks in which prescription opioids (top) or cannabis (bottom) were available. Price varied in United States dollars (USD). Plotted are mean (SEM) group data fit using the exponentiated model for individuals with (closed circles) and without (open circles) DSM-5 opioid use disorder (OUD).

Bivariate Associations

Table 2 contains bivariate associations and significance values. A preference for risky opioid routes was associated with less elastic opioid demand. Plans for reduction or treatment in the next year, but not current-treatment seeking, were also correlated with greater opioid demand intensity and lower opioid elasticity. OUD was significantly associated with more intense and less elastic opioid demand (see Figure 1 for group mean plots). Similar results were observed for cannabis demand intensity, which was associated with cannabis use quantity (i.e., grams/week) and cannabis use disorder. Neither opioid nor cannabis demand were associated with the alcohol use disorder. Steeper opioid discounting rates were also associated with OUD, but no relationship was observed with monetary discounting rates. Average pain and the typical impact of pain on everyday life were both associated with more intense opioid demand and steeper monetary and opioid discounting rates. In contrast, typical pain relief from opioids was not related to any demand or discounting variables.

Table 2.

Bivariate Relationships for Derived Behavioral Economic Demand and Discounting Variables

	Opioid Demand		Cannabis Demand		Discounting Rate (k)
	Intensity	Elasticity	Intensity	Elasticity	Money	Opioids	Cannabis
Age	0.06	0.13	−0.13	0.17	−0.04	−0.01	−0.05
Female	0.07	0.06	−0.12	.26*	0.15	0	0.05
White	−0.02	−0.16	0.09	−.30**	0.16	0.12	0.12
College	−0.09	0.11	0.02	−0.04	−.26*	−.25*	−0.2
Income (USD)	−.40***	0.15	−0.14	−0.07	−0.13	0.01	−0.08
Substance Use
Past Month Opioid Use Daysa	1.44	0.62	1.18	0.47	1.13	1.23	1.11
Risky Route of Administration	0.1	−.38***	0.19	−0.21	−0.06	0.05	0.03
Past Month Cannabis Use Daysa	1.31	0.89	2.08**	0.69	0.95	1.07	1.05
Grams Cannabis/Week	0.19	−0.2	.47***	−.28*	0.08	0.14	0.05
Treatment History
Current Treatment-Seeking	0.16	−0.13	0.14	−0.12	0.05	0.04	0.13
Treatment Plans in Next Year	.29**	−.25*	.23*	−0.16	0.19	0.08	0.07
DSM-5 SUD
OUD	.52***	−.28*	0.15	−0.05	0.11	.30**	0.2
OUD Criteria Endorsed	.40***	−.31**	0.1	−0.06	0.13	.29**	.23*
CUD	−0.08	0.15	.25*	−0.01	−0.02	−0.06	−0.01
CUD Criteria Endorsed	−0.01	0.09	.27*	−0.08	0.03	0.01	−0.01
AUD	0.05	0	0.03	−0.04	−0.09	0	0.14
AUD Criteria Endorsed	0.11	−0.04	−0.01	−0.04	−0.12	−0.01	0.05
Pain
Average Pain (0–10)	.29**	−0.06	−0.06	0.07	.27*	.22*	0.16
Pain Effect on Daily Life (0–10)	.32**	−0.11	0.03	0.02	.30**	.25*	0.14
Opioid Relief of Pain (0–100)	−0.06	0.09	−0.01	0.16	0.18	0.08	0.09

Note. OUD = opioid use disorder; CUD = cannabis use disorder; AUD = alcohol use disorder; Risky Route = preferred use of intranasal, smoked, or injection administration; BP1 = raw breakpoint; BP2 = dichotomous breakpoint (coded 0) versus no breakpoint (coded 1).

^aThese variables were evaluated using negative binomial regression given the observation of zero-inflation. Values represent rate ratios.

^*p < .05

^**p < .01

^***p < .001 (bold = statistically significant)

Curve-derived P_max and breakpoint measures generally showed less robust and consistent associations with substance use and pain (Supplemental Table 3). Risky routes of administration were associated with higher opioid O_max and breakpoint values. Higher opioid O_max was also related to OUD as well as plans to reduce or seek treatment for opioid use in the next year. In contrast, lower opioid P_max values were related to OUD. Higher cannabis O_max was related to cannabis grams/week, but other cannabis curve-derived measures were not related to cannabis use. No curve-derived measures were related to self-reported pain.

Multivariable Models

Multivariable models including opioid demand intensity and elasticity with monetary and opioid discounting rates were conducted to test incremental and unique associations with OUD, next-year reduction/treatment intentions, risky opioid route preference, and average pain (i.e., variables with significant bivariate associations) controlling for demographic variables and opioid use frequency. Higher opioid intensity (OR = 31.30, p =.004) and higher opioid discounting rates (OR = 7.46, p = .018) were each significant and independent predictors of OUD in multivariable models. More inelastic opioid demand was significantly associated with risky opioid route preferences (OR = 0.07, p = .003) and greater opioid demand intensity was significantly associated with higher average pain levels (β = .31, p = .027) and reduction intentions (OR = 4.89, p = .04). Other behavioral economic variables were also not significant in multivariable models.

Test-Retest Reliability

Good test-retest reliabilities were observed for opioid demand (Q₀ r_xx = .75; α r_xx = .63) and acceptable reliabilities were observed for cannabis demand (Q₀ r_xx = .53; α r_xx = .58) (Table 3). Temporal reliabilities were also mostly acceptable and significant for discounting rates with lower and bordering poor reliability for money compared to commodity discounting (money r_xx = .42, opioid r_xx = .58, cannabis r_xx = .61). These values were temporally stable with no significant changes in demand intensity, demand elasticity, or discounting values from time 1 to time 2, p values > .05 (see Table 3 for effect sizes of change over time). ICC values were consistent with the results of the bivariate correlations (Table 3).

Table 3.

Temporal Reliability and Stability of Study Variables

	rxx	ICC	Cohen’s dz
Opioid Demand
Intensity (Q0)	0.75	0.74	0.16
Elasticity (α)	0.63	0.63	−0.07
Pmax	0.42	0.41	−0.01
Omax	0.7	0.69	0.11
Breakpoint	0.52	0.52	−0.08
Cannabis Demand
Intensity (Q0)	0.53	0.53	0.24
Elasticity (α)	0.58	0.58	0.2
Pmax	0.44	0.43	−0.28
Omax	0.61	0.61	−0.13
Breakpoint	0.57	0.56	−0.41
Delay Discounting Rates (k)
Money	0.42	0.41	−0.14
Opioid	0.58	0.56	0.03
Cannabis	0.61	0.61	0.08
DSM-5 SUD
OUD Criteria Endorsed	0.76	0.76	0.2
CUD Criteria Endorsed	0.63	0.63	−0.07
AUD Criteria Endorsed	0.77	0.77	0.03

Note. OUD = opioid use disorder; CUD = cannabis use disorder; AUD = alcohol use disorder; ICC = intraclass correlation.

Reliabilities for curve-observed variables P_max and Breakpoint were generally lower than derived measures for both commodities (r_xx < .57). O_max, in contrast, showed acceptable-to-good reliability for opioid demand (r_xx = .70) and cannabis demand (r_xx = .61). Cannabis P_max, p = .049 and breakpoint, p = .004, were also significantly greater at Time 2. Other curve-derived measures did not significantly change over time, p values > .05 (see Table 3 for effect sizes of change over time).

Temporal reliability was also good for scores on the Brief DSM-5 Substance Use Disorder Diagnostic Assessment for OUD, r_xx = . 76, as well as for cannabis and alcohol use disorder, r_xx = .63 and .77, respectively. Substance use disorder classifications were stable over the one-month period as indicated by non-significant McNemar tests for paired nominal data.

Similar results as the reliability and stability assessments were observed for supplemental equivalence testing (Supplemental Figure 2). Specifically, these tests provided support for the statistical equivalence of the opioid demand measures and delay discounting. DSM-5 cannabis and alcohol use disorder criteria endorsed were also equivalent, but OUD criteria fell outside the upper boundary. Cannabis demand variables, with the exception of O_max, did not show statistical equivalence.

Discussion

The overall purpose of this study was to evaluate the utility of and further validate the purchase task procedure for describing non-medical prescription opioid use. Overall, participants reporting diagnostically relevant opioid use showed more intense and inelastic opioid demand. That elevated opioid demand was related to OUD is consistent with prior work conducted in college student samples (Pickover et al. 2016) and indicates that this relationship is replicable and generalizes to a general adult population. This association was also selective to opioids in that opioid demand variables were associated with OUD and not cannabis or alcohol use disorders. A similar selectivity was observed for cannabis demand wherein cannabis consumption and cannabis use disorder, but not other substance use variables, were associated with cannabis demand intensity. These findings contribute to a growing body of literature demonstrating the stimulus selectivity of the purchase task procedure for indexing valuation that is specific to the substance of interest (Chase et al. 2013; Strickland and Stoops 2017). More broadly, these findings indicate that more intense and inelastic opioid demand could be a behavioral mechanism underlying the progression to diagnostically relevant use among individuals reporting non-medical prescription opioid use. Although current treatment seeking was not associated with demand, intentions to quit or seek treatment in the future were related to higher opioid demand intensity and more inelastic opioid demand. A possible explanation for these associations is a recognition or insight into more problematic levels of use that motivates an intention, albeit not an immediate action, to reduce use. Future longitudinal work will be important for establishing the causal relationship between variations in opioid demand and the development of and progression to OUD as well as corresponding clinical outcomes.

Multivariable models indicated that behavioral economic demand provided unique and incremental information about OUD above and beyond delay discounting rates and frequency of opioid use. These models specified that higher opioid demand intensity and steeper opioid delay discounting rates each significantly and uniquely predicted the presence of OUD. This finding that demand accounted for unique information about OUD provides evidence for distinct behavioral mechanisms that could underlie clinically relevant non-medical opioid use. This evidence is consistent with previous work demonstrating the relationship of demand (Pickover et al. 2016) and discounting (e.g., Kirby and Petry 2004; Kirby et al. 1999) with heroin and prescription opioid use when measured alone. Unexpectedly, monetary discounting rates did not show similar associations with opioid use in bivariate or multivariable models. These findings could be explained by the use of a 5-choice procedure rather than longer adjusting amount tasks typically used in delay discounting work. This discrepancy does highlight, however, the importance of the commodity discounted and is in accordance with prior work demonstrating improved prediction of substance use and other health behaviors when using commodity-relevant discounting rates (e.g., Johnson and Bruner 2012; Rasmussen et al. 2010; Strickland et al. 2017; Tsukayama and Duckworth 2010).

Reinforcer pathology theory predicts similar behavioral economic to substance use associations that involve decision-making characterized by high reinforcer valuation (i.e., behavioral economic demand) and an extreme preference for immediate reinforcers (i.e., delay discounting) (Bickel et al. 2017). These decision-making processes are thought to reflect a shift in the balance towards more reward and present-focused processes from controlled and future-focused processes that co-occur with negative health behaviors such as substance use. The current findings suggest a similar shift in OUD of more intense and inelastic opioid valuation combined with a preference for immediate opioid reinforcers. These uniquely predictive relationships also support incremental validity insofar as relevant information about OUD was offered above and beyond frequency of opioid use and valuation for alternative commodities.

Adjusted models revealed that opioid demand elasticity was a significant predictor of a preference for risky routes of administration and was unique among behavioral economic variables in this regard. Intranasal, smoked, and intravenous routes of opioid administration are associated with increased health risks, such as STI transmission and overdose (Conrad et al. 2015; Strathdee and Beyrer 2015). The transition from oral to non-oral routes of administration also represents an important risk factor for the initiation of heroin and other illicit substance use (Carlson et al. 2016; Compton et al. 2016; Young and Havens 2012). Continued use in the face of these putative health consequences is consistent with the association reported here in which a preference for non-oral routes was related to more inelastic opioid demand. Such a relationship suggests that these preferences may be mechanistically related to a decreased sensitivity to the costs and consequences of substance use as reflected by less sensitive changes in use to increases in unit price (i.e., more inelastic demand).

Average self-reported pain and level of interference in daily function were associated with increased opioid demand intensity even after controlling for other relevant demographic and behavioral economic variables. To our knowledge, this is the first study to describe an association of drug demand, broadly, and opioid demand, specifically, with pain. This relationship between pain and the relative intensity of non-medical prescription opioid use is consistent with the notion that self-medication of un- or under-managed chronic pain could contribute to problematic opioid use and is contrasted with the observation that perceived pain relief from prescription opioids was not related to opioid demand. The discrepancy between these two outcomes could signify a decoupling between the level of opioid intake and strength of opioid relief due to processes such as pharmacological tolerance or an ineffective targeting by opioids of underlying causes of chronic pain (e.g., Arner and Meyerson 1988; Ashburn and Staats 1999). This pain-demand relationship provides a putative behavioral mechanism linking the use of opioids in chronic pain and the emergence of OUD. Future longitudinal studies evaluating the incidence of OUD in chronic pain settings will be important for further exploring this potential clinical application.

Good support for the reliability of opioid demand intensity (r_xx = .75) and elasticity (r_xx = .63) was observed over one month of testing. These reliabilities are similar to those for alcohol demand when measured over a one-month period in college students (intensity r_xx = .67, elasticity r_xx = .71; Acuff and Murphy 2017). Similar support for temporal reliability was observed with the curve-observed measure O_max (r_xx = .70), but was lower for the curve-observed measures P_max (r_xx = .43) and breakpoint (r_xx = .52). Reliabilities for cannabis demand were also lower (r_xx = .43 to .61) and in some cases approached an unacceptable range. Supplemental equivalence tests were generally consistent with these results in that opioid demand showed statistical equivalence over time whereas cannabis demand measures other than O_max did not. It is possible that cannabis demand showed more variability over time given that this substance was not the primary substance used to define the sample and may be more subject to environmental/state-like changes in drug valuation. Nevertheless, this is the first study to evaluate the temporal reliability of purchase tasks for substances other than alcohol or cigarettes. The temporally stability of opioid demand supports a continued use in repeated measure or longitudinal designs of laboratory and clinic research.

Clinical classifications based on the Brief DSM-5 Substance Use Disorder Diagnostic Assessment were also temporally reliable and stable. This finding is important for at least two reasons. First, the test-retest reliability of this brief assessment has not been previously established. Prior research has demonstrated strong internal consistency reliability and construct validity for the assessment when evaluating alcohol use disorder in college students (Hagman 2017). The current study extends this research by showing that this measure can be easily adapted for other substance use disorders and that these classifications show good stability over at least a month period. Second, the stability of these clinically classifications supports the validity of self-reported substance use behaviors in a crowdsourced sample. This outcome is particularly important given the inability to biologically verify substance use over the mTurk platform. Offsetting this limitation are previous studies indicating that crowdsourced samples do not engage in higher rates of problematic responding, such as socially desirable bias, and that these samples report feeling more comfortable sharing sensitive materials, such as substance use, over an online platform than in person (Kim and Hodgins 2017; Necka et al. 2016; Strickland and Stoops 2018).

Secondary demand analyses focused on curve-observed variables. Relationships involving O_max were similar in significance and magnitude to elasticity reflecting similar factor loadings for these two measures within a two-factor demand structure. In contrast, associations for the curve-observed values of P_max and breakpoint were generally less robust and consistent for opioid and cannabis demand. For example, opioid P_max showed a relationship in the opposite direction than expected with OUD (i.e., lower P_max with OUD). P_max and breakpoint values also showed lower temporal reliability, which could explain the inconsistent associations with substance use behaviors. It is possible that relying on a single point along the demand curve for these curve-observed values results in increased measurement error compared to derived measures, which incorporate information along the entire curve in their estimation. Future work specifically addressing potential differences between curve-observed and derived measure as well as summaries of existing work (e.g., meta-analyses) will be important for clarifying these discrepancies.

This study should be considered in the context of its limitations. First, the use of crowdsourcing methods does present potential concerns related to sampling bias and generalization. A substantive body of literature has documented the ways in which crowdsourced sampling may differ from nationally representative sources. These studies have found that individuals recruited from crowdsourcing platforms tend to be younger, more educated, and less likely to be employed, married, or a racial minority compared to nationally representative sources (Berinsky et al. 2012; Huff and Tingley 2015; Paolacci and Chandler 2014). Higher rates of alcohol and illicit substance use has also been observed in crowdsourced samples (Shapiro et al. 2013; Strickland and Stoops in press) (but see Caulkins et al. 2015 for information on the under-estimation of substance use in nationally representative sources). Other research, however, has provided good evidence for the validity of data collected via crowdsourced methods by demonstrating a correspondence between outcomes observed in laboratory, clinic, and online settings (see reviews in Chandler and Shapiro 2016; Strickland and Stoops in press). The current study similarly replicated findings reported elsewhere both related to and independent of opioid use. Relationships between opioid demand and OUD, for example, were consistent with prior research conducted in college samples. Discounting rates were steeper for opioids and cannabis compared to money, which also replicates a canonical finding that consumable goods are more steeply discounted than money (e.g., Baker et al. 2003; Bickel et al. 2011; Charlton and Fantino 2008; Johnson et al. 2007). Although limitations associated with the convenience nature of crowdsourced sampling should be considered, ultimately the combination of research from laboratory, clinical, and crowdsourced sources should benefit the rigor, reproducibility, and scope of research conducted in addiction science.

Second, several aspects of the procedure could be manipulated or examined in future work. Prescription opioids were defined broadly in the purchase task procedure as “the standard dose that you use when you use these pills”. This approach has been used successfully previously and likely helps provide for a more general task accounting for the variations in prescription opioids. However, alternative procedures, such as defining specific medication types and/or equivalent doses should be explored (for similar problems in defining quantities with an e-cigarette purchase task see Cassidy et al. 2017). A substantive number of participants also did not reach a breakpoint. Future applications should explore larger price ranges or evaluate adaptive approaches that present values until a breakpoint is reached to ensure that full demand curves are collected for all participants.

Third, some unexpected, inconclusive, or exploratory findings could be evaluated in future work. It is unclear why opioid demand intensity was negatively related to income. One potential explanation is a “third variable” relationship wherein income was associated with another variable that explained this positive association (e.g., more severe opioid use in lower income participants). Follow up partial correlations controlling for OUD indicated that this demand-income association remained significant and of a medium effect size (r = −.32). It is also possible that participants with higher income had greater access to alternatives to opioids for indications like pain management. Even though such alternatives were supposed to be considered inaccessible in the hypothetical vignette, it is possible that general experience and knowledge of these substitutes could have altered decision-making. We also did not specifically design this study to evaluate differences by patterns of co-drug use. Indirect evaluations examining differences between individuals with alcohol, cannabis, and opioid substance use disorder (22.9% of the sample) did not reveal significant differences in opioid demand. Future work would benefit from a priori designs intended to evaluate questions related to comorbid substance use and associations with demand and discounting variables (e.g., Morris et al. 2018; Yurasek et al. 2013). Finally, the study was not designed to conduct the tests of statistical equivalence that were used to supplement the primary focus on tests of temporal reliability. Future work would benefit from larger sample studies to determine if the positive results regarding statistical equivalence of opioid demand replicate.

The primary finding of this study was that the commodity purchase task provided an incrementally valid and temporally reliable measure of opioid demand. These findings are consistent other research indexing valuation for alcohol and cigarettes using the purchase task procedure. Coupled with the present data, this body of work demonstrates that the purchase task procedure provides a clinically useful measure of drug valuation that is sensitive to individual difference variables relevant to drug-taking behavior. These studies also provide clear evidence for the utility of demand in providing relevant information about the behavioral mechanisms underlying the relative reinforcing effects of drugs that can be used to inform prevention and treatment efforts targeting substance use disorders.